AI innovation concentration and the governance challenge

Lippold , Douglas
Wo king Pape
AI inno a ion concen a ion and he go e nance challenge
CIGI Pape s, No. 292
P o ided in Coope a ion wi h:
Cen e o In e na ional Go e nance Inno a ion (CIGI), Wa e loo, On a io
Sugges ed Ci a ion: Lippold , Douglas (2024) : AI inno a ion concen a ion and he go e nance
challenge, CIGI Pape s, No. 292, Cen e o In e na ional Go e nance Inno a ion (CIGI), Wa e loo
(On a io)
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/299989
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by/4.0/
CIGI Pape s No. 292 — Ap il 2024
AI Inno a ion Concen a ion
and he Go e nance
Challenge
Douglas Lippold
CIGI Pape s No. 292 — Ap il 2024
AI Inno a ion Concen a ion
and he Go e nance
Challenge
Douglas Lippold
Abou CIGI
The Cen e o In e na ional Go e nance Inno a ion (CIGI) is an independen ,
non-pa isan hink ank whose pee - e iewed esea ch and us ed analysis
in luence policy make s o inno a e. Ou global ne wo k o mul idisciplina y
esea che s and s a egic pa ne ships p o ide policy solu ions o he digi al
e a wi h one goal: o imp o e people’s li es e e ywhe e. Headqua e ed
in Wa e loo, Canada, CIGI has ecei ed suppo om he Go e nmen o
Canada, he Go e nmen o On a io and ounde Jim Balsillie.
À p opos du CIGI
Le Cen e pou l’inno a ion dans la gou e nance in e na ionale (CIGI) es un
g oupe de é lexion indépendan e non pa isan don les eche ches é aluées
pa des pai s e les analyses iables inci en les décideu s à inno e . G âce
à son éseau mondial de che cheu s plu idisciplinai es e de pa ena ia s
s a égiques, le CIGI o e des solu ions poli iques adap ées à l’è e numé ique
dans le seul bu d’amélio e la ie des gens du monde en ie . Le CIGI, don le
siège se ou e à Wa e loo, au Canada, béné icie du sou ien du gou e nemen
du Canada, du gou e nemen de l’On a io e de son onda eu , Jim Balsillie.
C edi s
Managing Di ec o o Digi al Economy (un il Feb ua y 2024) Robe Fay
Di ec o , P og am Managemen Dianna English
P og am Manage Jenny Thiel
Publica ions Edi o Lynn Schellenbe g
Senio Publica ions Edi o Jenni e Goyde
G aphic Designe Sami Chouhda y
Copy igh © 2024 by he Cen e o In e na ional Go e nance Inno a ion
The opinions exp essed in his publica ion a e hose o he au ho and do no
necessa ily e lec he iews o he Cen e o In e na ional Go e nance Inno a ion
o i s Boa d o Di ec o s.
Fo publica ions enqui ies, please con ac [email p o ec ed].
The ex o his wo k is licensed unde CC BY 4.0. To iew a copy o his licence,
isi h p://c ea i ecommons.o g/licenses/by/4.0/.
Fo euse o dis ibu ion, please include his copy igh no ice. This wo k may
con ain con en (including bu no limi ed o g aphics, cha s and pho og aphs)
used o ep oduced unde licence o wi h pe mission om hi d pa ies.
Pe mission o ep oduce his con en mus be ob ained om hi d pa ies di ec ly.
Fo publica ion enqui ies, please con ac [email p o ec ed].
Cen e o In e na ional Go e nance Inno a ion and CIGI a e egis e ed
adema ks.
67 E b S ee Wes
Wa e loo, ON, Canada N2L 6C2
www.cigionline.o g

Table o Con en s
i Abou he Au ho
ii Ac onyms and Abb e ia ions
1 Execu i e Summa y
2 In oduc ion
3 Li e a u e: Rapid Technological De elopmen , Unse led Schola ship
9 Leading AI-In ensi e Fi ms: A Few S ylized Fac s
19 Regula o y De elopmen s
30 Conclusions
35 Appendix 1: Sc eens Employed in Selec ion o
Top AI Fi ms among R&D Leade s
37 Appendix 2: Obse a ions on AI Regula o y
Regimes in Selec ed Sample Coun ies
45 Wo ks Ci ed
i CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
Abou he Au ho
Douglas (Doug) Lippold is a CIGI senio ellow
and an in e na ional ade economis based in
Cla emon , Cali o nia. He se ed as chie ade
economis a HSBC Global Resea ch om 2014 o
2020. P e iously, he se ed in a ious oles as a
senio economis a he O ganisa ion o Economic
Co-ope a ion and De elopmen in Pa is, F ance,
du ing a enu e o 22 yea s. Doug’s ea ly ca ee
included se en yea s as an in e na ional economis
wi h he US Depa men o Labo .
He has published ex ensi ely on ade opics
as well as on ela ed aspec s o economic
de elopmen , labou ma ke adjus men ,
inno a ion and in ellec ual p ope y. Doug holds
a Ph.D. in economics om he Ins i u d’é udes
poli iques de Pa is (Sciences Po), an M.A. in
in e na ional s udies om he Uni e si y o Den e
and a B.A. in in e na ional s udies om Washing on
College in Ma yland. He was a Fulb igh Schola a
he Uni e si y o Cologne, Ge many, and a Peace
Co ps olun ee in Bu kina Faso. Du ing 2015–2019,
he ep esen ed HSBC as he depu y delega e on he
Business 20 T ade and In es men Task Fo ce. He is
cu en ly a con ibu ing au ho on a Think20 ade
ask o ce eam.
iiAI Inno a ion Concen a ion and he Go e nance Challenge
Ac onyms and
Abb e ia ions
AI a i icial in elligence
AIDA A i icial In elligence and Da a Ac
AIIA A i icial In elligence
Indus y Alliance
APEC Asia-Paci ic Economic
Coope a ion egional o um
B2B business- o-business
BIS Bu eau o Indus y
and Secu i y (US)
CAC Cybe space
Adminis a ion o China
capex capi al expendi u e
CPTPP Comp ehensi e and
P og essi e Ag eemen o
T ans-Paci ic Pa ne ship
CUSMA Canada-Uni ed S a es-
Mexico Ag eemen
DPA Digi al Policy Ale
DSIT Depa men o Science,
Inno a ion and Technology (UK)
ESG en i onmen al, social
and (co po a e) go e nance
FATF Financial Ac ion Task Fo ce
FTC Fede al T ade Commission (US)
G7 G oup o Se en
G20 G oup o Twen y
GDPR Gene al Da a
P o ec ion Regula ion
ICB Indus y Classi ica ion Benchma k
IFAC-BIAC In e na ional Fede a ion o
Accoun an s and Business
a OECD
IP in ellec ual p ope y
IPRs in ellec ual p ope y igh s
LLMs la ge language models
NIST Na ional Ins i u e o
S anda ds and Technology (US)
OECD O ganisa ion o Economic
Co-ope a ion and De elopmen
p.p. pe cen age poin
R&D esea ch and de elopmen
SAG-AFTRA Sc een Ac o s Guild-
Ame ican Fede a ion o
Tele ision and Radio A is s
UNESCO Uni ed Na ions Educa ional,
Scien i ic and Cul u al
O ganiza ion
USTR O ice o he Uni ed S a es
T ade Rep esen a i e
VC en u e capi al
WIPO Wo ld In ellec ual
P ope y O ganiza ion
WTO Wo ld T ade O ganiza ion
7AI Inno a ion Concen a ion and he Go e nance Challenge
sha ing economy.16 The e a e many possibili ies.
Rega dless which one, IP is likely be pa o he
discussion. Ge ing he IPRs p o ec ion uned
o add ess challenges a ound a new echnology
may equi e a pe iod o collec ing expe ience o
iden i y he key issues, impedimen s and gaps in
he sys em. Use and adjudica ion o IPRs based on
exis ing laws may p o ide essen ial in o ma ion o
subsequen ly legisla ing adjus men s in he sys em.
Wi h AI, IPR issues a e al eady coming o he o e.
Gene a i e AI is aising copy igh issues bo h wi h
espec o aining da a employed (which may be
subjec o copy igh p o ec ion) and wi h espec o
asse ing copy igh o p oduc s de eloped wi hou
su icien human inpu . A ecen assessmen by he
US Cong essional Resea ch Se ice ad ised ha
i may be p ema u e o legisla e on such ma e s
un il he cou s ha e a chance o adjudica e and
iden i y any sho alls in he legal amewo k
(Zi poli 2023). (Nex up o dispu e may be he
IPR-adjacen issue o con ol o e one’s own image,
oice and ac ions in AI-gene a ed con en .)17
Simila issues may a ise wi h espec o pa en s,
whe e he no ion o human-di ec ed c ea ion may
need legal cla i ica ion. In ade sec e s, challenges
a e eme ging in ela ion o disclosu e o algo i hms,
es ic ions on labou mobili y and app op ia ion
o da a, among o he issues, pa icula ly in ligh o
he high alue o R&D in he AI sec o . Regula o y
p ocesses such as cus oms con ols o sou ce code
o egis a ion o models o app o als ( o example,
al eady equi ed in China and pending in he Uni ed
S a es and he Eu opean Union; see Appendix2)
may p esen u he challenges o p o ec ing
ade sec e s. I is unclea whe he he cu en IPR
16 Fo example, copy igh p o ec ion migh be used by a so wa e o igina o
o main ain ce ain igh s o e hei open-sou ce so wa e c ea ion while
making he code eadily and eely a ailable o use by subsequen
inno a o s o build upon o e en modi y, depending on he e ms o an
open-sou ce licence.
17 The complexi y o gene a i e AI conce ns was illus a ed in he Uni ed
S a es by he con ac nego ia ions be ween he Sc een Ac o s Guild–
Ame ican Fede a ion o Tele ision and Radio A is s (SAG-AFTRA) and
he Alliance o Mo ion Pic u e and Tele ision P oduce s. The deal, a i ied
on Decembe 5, 2023, de ines AI pe o mance concep s such as an
“independen ly c ea ed digi al eplica...used o po ay he pe o me in
scenes hey did no ac ually shoo ” and gene a i e AI “used o simula e
a pe o me ’s oice, acial exp essions, and mo emen s o c ea e en i ely
new con en ” (SAG-AFTRA 2023a, 2). Each use o AI is subjec o he
consen o he subjec , clea delinea ion o he AI con en o be c ea ed,
and app op ia e compensa ion. The con ac includes a p inciple o
p imacy o human pe o mance in he ealm o ac ing, he holding
o egula employe -union mee ings on he e ol ing use o AI, and
moni o ing o implemen a ion (SAG-AFTRA 2023b).
amewo k will p o e adequa e o no el challenges
will equi e changes in legisla ion and egula ion.
Ge ing he IPRs sys em uning igh can make
a di e ence o he di usion o a echnology.
Wo k a he OECD and elsewhe e has p o ided
e idence ha adequa e IPR p o ec ion can be
conduci e o ade and in es men in echnology-
in ensi e p oduc s and o echnology ans e
ia licensing o IP (Ca azos Cepeda and Lippold
2010; B ans e e , Foley and Saggi 2010; Lippold
and Schul z 2014). These e ec s appea o be
s onges wi h espec o pa en s, bu posi i e
associa ions ha e also been ound in some cases
o copy igh , adema ks and ade sec e s.
Whe e holde s o IP ha e means o p o ec ing
hei asse s, hey a e mo e inclined o ans e
con ol o una ilia ed pa ies (Pa k and Lippold
2005). And likewise, when o eign i ms looking
o in es di ec ly in a na ion a e able o p o ec
hei IPRs, hey a e mo e likely o ans e ecen
gene a ion echnology along wi h he knowhow o
success ully exploi i (Pa k and Lippold 2014). I
is likely ha simila endencies will mani es wi h
a leas some elemen s o AI sec o echnology.
A signi ican dis inc ion wi h espec o IP and
AI conce ns he di e ence in me hods be ween
he open-sou ce LLM de elopmen communi y,
which employs an open app oach wi h libe al
licensing e ms, and he de elope s building
p op ie a y LLMs, asse ing hei IPRs and
holding back some in o ma ion ha con eys
comme cial ad an age. The o me a e pe haps
able o le e age hei echnical e o s due o
inpu s om ou side con ibu o s, leading in some
cases o ad ances in model de elopmen . The
la e may be incen i ized in hei e o s by be e
access o in es o capi al and ad ice in some
cases (al hough, no all will a ail hemsel es o
hese op ions), as well as he p ospec o inancial
gains in he e en o comme cial success.
The open-sou ce e sus p op ie a y di ide echoes
h ough he AI ecosys em (CB Insigh s 2023b),
al hough he e is signi ican c oss- e iliza ion
among de elope s using ei he app oach (Kaye
2022). Open-sou ce ools such as PyTo ch,
Tenso Flow and PaddlePaddle ha e been de eloped
and eleased by majo AI de elope s (Me a,18 Google
and Baidu, espec i ely). Such ools ha e bene i ed,
18 No e ha in Sep embe 2022 Me a ans e ed con ol o PyTo ch o he
Linux Founda ion (Kaye 2022).

8CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
among o he s, smalle independen applica ion
de elope s, including hose collabo a ing
in e na ionally and he eby p omo ing di usion
o inno a ion. Hos ing se ices such as Gi Hub
and Hugging Face ha e enabled de elope s o
sha e code and da a. AI as a se ice p o ide
may deli e p op ie a y cloud-based solu ions
o expedi e AI de elopmen and deploymen
o use s, while in some cases d awing on open-
sou ced inpu s ( o example, see W en 2023).
Recen op-pe o ming LLMs ha e ended o be buil
using p op ie a y app oaches. Howe e , he e is an
ongoing exchange wi h he open-sou ce communi y,
and by some measu es he pe o mance o open-
sou ce LLMs is ca ching up (Chen e al. 2023;
B emme and Suleyman 2023, 30). Gian i ms such
as Me a (wi h Llama 2) and Mic oso (wi h Phi),
and academic ins i u ions such as S an o d (wi h
Alpaca) ha e eleased p e- ained, open-sou ce
models, bene i ing indi ec ly om new insigh s
as ou side esea che s and de elope s con ibu e
o ad ance he echnology. This le e ages alen
om he open-sou ce communi y and can add new
i ali y o inno a ion p ocesses. I may also pe mi
small businesses o pa icipa e in he AI economy
by ailo ing and ma ke ing hei own elemen s
buil on an open-sou ce AI ounda ion. Gi en he
subs an ial o iginal LLM de elopmen cos s, an
open-sou ce app oach p o ides some small i ms
wi h a use ul poin o en y o he AI ma ke .
The accele a ed pace o de elopmen om
bo h app oaches has led o AI sa e y conce ns,
pa icula ly wi h espec o on ie model
de elopmen ha may unleash new o poo ly
unde s ood capabili ies (Depa men o Science,
Inno a ion and Technology [DSIT] 2023). Open-
sou ce de elopmen s aise pa icula conce ns.
Downs eam e olu ion o open models may no
be adequa ely acked; exis ing gua d ails and
buil -in mi iga ion measu es may no be su icien
o ensu e AI sa e y (Fa hi and MacKinnon 2023;
Raskin and Ha is 2023). As Ian B emme and
Mus a a Suleyman (2023, 41) pu i in a ecen
a icle, “The de ol ed na u e o AI de elopmen
and co e cha ac e is ics o he echnology, such as
open-sou ce p oli e a ion, inc ease he likelihood
ha i will be weaponized by cybe c iminals,
s a e-sponso ed ac o s, and lone wol es.”
As a consequence o such conce ns, he Uni ed
Kingdom con ened a sa e y summi among
29 go e nmen s in No embe 2023. One esul
was he Ble chley Pa k Decla a ion whe eby
pa icipan s ag eed o edouble hei coope a ion
on AI sa e y ( o example, h ough iden i ica ion
o sa e y isks and building isk-based sa e y
policies) and o engage s akeholde s o ake
esponsibili y when unde aking ad anced model
de elopmen wo k ( o example, h ough ex e nal
p e- elease sa e y es ing) (GOV.UK 2023).19
Go e nance
Ma di Wi zel and Ni aj Bha ga a (2023) p o ide a
concise o e iew o he s a e o play wi h espec
o AI go e nance.20 They b ie ly conside key
elemen s o go e nance such as na ional s a egies,
e hical AI amewo ks, cu en and p oposed
legal and egula o y ins umen s (no ably in he
Eu opean Union, he Uni ed S a es and Canada),
AI s anda ds (highligh ing he In e na ional
S anda ds O ganiza ion, ISO 42001 s anda d o
AI managemen sys ems and he NIST AI Risk
Managemen F amewo k). The au ho s’ e iew
makes clea ha hese a ious elemen s each ha e
a ole and hei de elopmen emains a wo k in
p og ess. Wi h espec o balancing among he
a ious elemen s o go e nance, Tom Wheele
(2023, 20) no es, “All mode n egula ions walk a
igh ope be ween p o ec ing he public in e es
and p omo ing inno a ion and in es men . In
he AI e a, a e sing he egula o y igh ope
means accep ing ha di e en AI applica ions
pose di e en isks and iden i ying a plan ha
pai s he egula ion wi h he isk while a oiding
inno a ion-choking egula o y mic omanagemen .”
B emme and Suleyman (2023) make he case
o ea ly ac ion on he es ablishmen o a global
go e nance mechanism o AI. They highligh he
na u e o AI as a gene al pu pose echnology ha is
apidly e ol ing and capable o my iad applica ions
and, po en ially, sel -imp o emen . A some poin ,
AI will exceed human men al pe o mance in a
gene al manne . This has signi ican po en ial
o good, such as h ough i s abili y o accele a e
inno a ion p ocesses, i s po en ial o enhance
19 In ela ion o in e na ional coope a ion on AI sa e y, he Ble chley
Pa k summi was a imely ollow-up o he ecen G oup o Se en (G7)
Hi oshima mee ing. The G7 Leade s’ Communiqué, May 20, 2023,
pa ag aph 1, s a ed: “We a e de e mined o wo k oge he and wi h
o he s o: …ad ance in e na ional discussions on inclusi e a i icial
in elligence (AI) go e nance and in e ope abili y o achie e ou common
ision and goal o us wo hy AI, in line wi h ou sha ed democ a ic
alues.” See Whi e House (2023a).
20 The au ho s (2023) ackle AI go e nance om a unique en i onmen al,
social and (co po a e) go e nance (ESG) pe spec i e, whe eby AI could
be in eg a ed in o ESG epo ing equi emen s.
9AI Inno a ion Concen a ion and he Go e nance Challenge
p oduc i i y, and i s an icipa ed capaci y o assis in
esponding o global c ises such as heal h o clima e
eme gencies. Howe e , AI will cause disloca ion in
labou ma ke s and i has he po en ial o misuse
in a eas such as disin o ma ion, su eillance o e en
he c ea ion o au onomous weapons, among o he
possibili ies. I may also exace ba e inequali ies.
Leading AI na ions such as China and he Uni ed
S a es will compe e o AI leade ship, bu mos
coun ies will access AI in a seconda y manne .
Gi en he low ma ginal cos o eplica ion o
AI models, he unknowable na u e o u u e AI
de elopmen s and he isks o misuse o hese
echnologies, B emme and Suleyman (ibid.)
a gue o es ablishmen o a h ee-pa global
AI go e nance egime. Fi s , hey sugges
es ablishmen o a global scien i ic body, simila
o he UN In e go e nmen al Panel on Clima e
Change, o ad ise go e nmen s on eme ging AI
challenges. Second, hey p opose an in e na ional
body o manage ensions among majo AI powe s
( o example, China and he Uni ed S a es), o
es ablish gua d ails and moni o ing, and o p e en
p oli e a ion o dange ous ad anced AI sys ems.
Thi d, he au ho s p opose a echnoc a ic body
o AI isk — a Geo echnology S abili y Boa d21—
o engage na ional egula o s, in e na ional
s anda ds bodies and p i a e sec o mul ina ional
s akeholde s in p e en ing o esponding o AI
c ises, and go e ning open-sou ce AI. Ac oss
hese h ee bodies, egula o y mechanisms migh
be conside ed such as “know you cus ome ”
anspa ency s anda ds, licensing equi emen s,
sa e y es ing p o ocols and p oduc egis a ion
and app o al p ocesses, among o he s.
Acco ding o B emme and Suleyman (ibid., 36)
he AI go e nance egime should “iden i y and
mi iga e isks o global s abili y wi hou choking
o AI inno a ion and he oppo uni ies ha low
om i .” They call his “ echnop uden ialism,”
an app oach ha would be guided by common
p inciples: an app op ia e deg ee o p ecau ion
( i s , do no ha m); egula o y agili y in an icipa ion
o echnological changes; inclusi i y (engaging
AI s akeholde s including go e nmen s, p i a e
sec o AI i ms and ci il socie y expe s such as
scien is s, e hicis s, ade unions, among o he s);
impe meabili y, meaning blanke co e age globally;
21 Ci ing he example o he Financial S abili y Boa d, Robe Fay and
Rohin on Medho a (2021) make a simila p oposal o a Digi al S abili y
Boa d ha would ope a e wi h espec o egula ion o he in e na ional
digi al economy mo e b oadly.
and a ge ing ac ions depending on he isks in
speci ic con ex s ( a he han one-size- i s-all).
Ho e al. (2023) call o an Ad anced AI Go e nance
Agency o se in e na ional s anda ds, suppo
implemen a ion o s anda ds and moni o
compliance. They sugges es ablishmen
o ins i u ional capaci y o se e al u he
unc ionali ies, ei he ia s and-alone ins i u ions
o in some combina ion: an AI sa e y p ojec
o conduc o suppo esea ch; a scien i ically
o ien ed commission on on ie AI o iden i y
oppo uni ies and isks, and con ibu e o policy
consensus a ound hese; and a public-p i a e
on ie AI collabo a i e o de elop his echnology,
manage dual-use applica ions, educa e, and
suppo di usion and access. In de eloping
hese unc ionali ies ca e will be equi ed o
a oid encumb ances o he ype ha hinde
he ins i u ion’s agili y and esponsi eness.
Leading AI-In ensi e
Fi ms: A Few S ylized
Fac s
A ew compa a i e, s ylized ac s may help o
p o ide a sense o he di e en ia ing cha ac e is ics
o AI-in ensi e, inno a i e i ms. In he sea ch
o such i ms, he s udy employs wo sepa a e
i m-le el analyses d awing on wo independen
en e p ise da a se s ha co e a b oad ange
o indus ies and geog aphies. These co e wo
dis inc segmen s om he popula ion o inno a i e
en e p ises. The i s co e s medium and la ge
publicly lis ed, R&D-expendi u e leading i ms.
The second co e s unico n i ms — success ul
s a -ups ha ha e achie ed ma ke alua ions
o app oxima ely US$1 billion o mo e. Each o
hese wo pools o i ms is sc eened o iden i y
i ms ha a e likely o be among he leade s in AI
echnology de elopmen . The sc eens a e based
on published assessmen s om independen
indus y expe s and each i m’s membe ship in
ele an indus y bodies. The esul s a e illus a i e
a he han s ic ly s a is ically ep esen a i e.
The analysis he e ocuses on i ms ha ha e
al eady achie ed scale in he AI ield. This
ma e s o some dimensions o AI inno a ion
10 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
including, o example, he cos ly de elopmen
and aining o o iginal LLMs. While he e is
conside able po en ial o smalle i ms o
inno a e and build AI applica ion and se ice-
p o ision businesses, scale may con e ad an ages
in some undamen al a eas o de elopmen
( o example, ounda ion model de elopmen ).
Conside a ion o such a selec ion o la ge i ms
p o ides an oppo uni y o examine a eas o
concen a ion, which in u n may ha e implica ions
o di usion o some ypes o AI inno a ion.
AI Leade s among he Top Fi ms
o R&D Expendi u e Globally
The i s assessmen begins wi h conside a ion o
he op 2,500 i ms o R&D expendi u e globally.22
The analysis is based on da a om he “EU
Indus ial R&D In es men Sco eboa d,” which
is p oduced annually unde he auspices o he
Eu opean Commission (G assano e al. 2022). In
addi ion o R&D expendi u e, his EU publica ion
p esen s a ange o pe o mance indica o s o
each i m such as ne sales, employmen , capi al
expendi u e and p o i abili y. Each i m’s coun y
o egis y and indus y sec o a e also no ed.23
Ac oss all ields, hese i ms in es ed collec i ely
a o al o US$1.3 illion in R&D du ing 2021.24 This
ep esen s abou 86 pe cen o he es ima ed o al
global business expendi u e on R&D ha yea ,25
and oughly hal o he o al global inno a ion
expendi u e om all sec o s, including business,
go e nmen , academia and non-p o i ins i u ions.26
22 R&D co e s basic esea ch, applied esea ch and expe imen al
de elopmen .
23 See Table 1 no es o mo e in o ma ion on he indus y classi ica ion
employed by he EU s udy eam.
24 See Nicola G assano e al. (2022); an online o e iew is a ailable a
h ps://i i.j c.ec.eu opa.eu/sco eboa d/2022-eu-indus ial- d-in es men -
sco eboa d#block-ecl- heme-page- i le.
25 Ibid.
26 The Wo ld Bank p o ides da a on R&D expendi u e as a pe cen age o
GDP and also p o ides GDP alue (cu en US$). In 2020, his amoun ed
o 2.63 pe cen o global GDP (la es yea a ailable, Uni ed Na ions
Educa ional, Scien i ic and Cul u al O ganiza ion [UNESCO] da a as
epo ed by he Wo ld Bank). Global GDP was US$96.88 illion in 2021
(Wo ld Bank da a). R&D includes capi al and cu en expendi u es in he
ou main sec o s: business en e p ise, go e nmen , highe educa ion and
p i a e non-p o i . The alue o global R&D om all sou ces o alled abou
US$2.55 illion. This da a is a ailable om he Wo ld Bank a
h ps://da a.wo ldbank.o g/indica o /GB.XPD.RSDV.GD.ZS?name_
desc= alse (R&D as a pe cen age o GDP) and a
h ps://da a.wo ldbank.o g/indica o /NY.GDP.MKTP.CD?name_
desc= alse (GDP, alue).
In sc eening o leading AI-in ensi e i ms among
he op 2,500 R&D expendi u e leade s, he
analysis ini ially employs h ee sc eens: a lis ing
o leading AI inno a i e i ms om a widely ead
echnology and business publica ion (En e p ise
Weekly, eWeek, “Top 100+ A i icial In elligence
Companies 2023” lis ing, as o May 29, 202327); i ms
included in Schwab’s “AI Thema ic Resea ch” lis ing
as o Augus 18, 202328; and publicly lis ed i ms in
HSBC’s Global Resea ch equi y co e age o AI i ms
in China.29 AI in ensi y in a i m is aken he e o
mean ha he i m’s AI echnology is expec ed o
con ibu e signi ican ly o he i m’s own economic
pe o mance, and ha he i m’s AI inno a ion o
applica ion o AI inno a ion may ha e economically
impo an implica ions o o he i ms in e ms o
hei own pe o mance ( o example, whe eby some
may bene i due o inbound AI echnology ans e
o imi a ion, and o he s such as di ec compe i o s
may ind hemsel es a a disad an age). The h ee
sc eens we e selec ed based on wo p incipal
conside a ions: o ap in o he i m-le el expe ise o
a b oad selec ion o p o essional indus y analys s
acking i ms in a as -e ol ing sec o o ge a sense
o hei economic po en ial; and o ensu e b oad
geog aphic co e age o leading AI i ms ac oss
No h Ame ica, Eu ope and Asia (including China).
Pa ial con i ma ion o he sc eening selec ion
esul s was conduc ed by c oss checking agains
AI indus y g oup membe ship lis s o he US-
ancho ed Pa ne ship on AI as o Augus 202330 and
he China-ancho ed A i icial In elligence Indus y
27 The cu en eWeek AI company lis ing is a ailable a www.eweek.com/
a i icial-in elligence/ai-companies/. In he May 2023 edi ion employed
o he p esen esea ch, based on he expe opinion o he edi o ial
eam, a lis o i ms conside ed o be leading in each segmen o hei
axonomy o AI sec o i ms was assembled. The axonomy was as
ollows: con e sa ional AI, cybe secu i y, educa ion, en e p ise majo s,
inancial se ices, gene a i e AI, gian s, heal h-ca e AI, pionee s, e ail
and AI, obo ic p ocess au oma ion, and isiona ies. Pe he eWeek
axonomy, “ isiona ies” a e AI s a -ups ha a e “close o he edge,
in en ing he gene a i e AI landscape in eal ime”; “pionee s” a e
o wa d-looking AI companies in en ing and suppo ing he sys ems ha
a e p opelling AI de elopmen .
28 Schwab esea ch da a is a ailable o clien s a www.schwab.com/clien -
home.
29 HSBC Global Resea ch e e ences include F ank He (2023) o
AI so wa e and se ices, and Cha lene Liu e al. (2023) o AI
in as uc u e.
30 Acco ding o he “Abou us” page on i s websi e, he Pa ne ship on AI
is a non-p o i pa ne ship o academic, ci il socie y, indus y and media
o ganiza ions c ea ing solu ions so ha AI ad ances posi i e ou comes o
people and socie y. The membe ship lis is a ailable ia he Pa ne ship
on AI a h ps://pa ne shiponai.o g/wo k/.
11AI Inno a ion Concen a ion and he Go e nance Challenge
Alliance as o Ma ch 2020.31 Bo h indus y g oupings
ha e in e na ional membe ship. In six ins ances,
i ms om among he op 2,500 R&D inno a o s
we e selec ed o inclusion in he illus a i e lis
o AI leade s based on hei membe ship in one o
hese AI indus y g oups and hei signi ican AI
de elopmen s as discussed elsewhe e in he ade
p ess: Apple and Samsung due o hei wo k on
AI chips and consume - acing AI inno a ion, as
well as hei scale; Synposis due o inno a ion in
AI chip design so wa e; Seaga e due o AI s o age
inno a ion; China Telecom due o AI pla o m
and LLM inno a ion; and Nokia o AI inno a ion
in elecoms. (Fu he de ails o he sc eening a e
p esen ed in Appendix 1, Table A.1, including a
no e on i ms omi ed due o da a limi a ions in
he EU lis ing o he op 2,500 R&D inno a o s.)
AI Leade s: Findings and Ca ea s
The sc eening p ocess iden i ies 57 leading AI
inno a o i ms, 56 o which a e lis ed on majo
s ock exchanges (see Table 1 and Appendix 1,
Table A.132). One s iking esul is he geog aphic
concen a ion o hese i ms (see Table2). Fully
wo- hi ds o hese i ms a e based in he Uni ed
S a es. Among he o he geog aphies, China
accoun ed o less han one in se en o he
lis ings and he Eu opean Union o less han one
in 10. Collec i ely in 2021, hese i ms eco ded
R&D expendi u e o US$264 billion, wi h some
wo- hi ds o ha expendi u e by US- egis e ed
i ms. Fi ms in China accoun ed o jus unde
one- i h o he o al R&D expendi u e and he EU
o jus one-six een h. Toge he , he 57 leading
AI-in ensi e i ms in he panel accoun ed o
oughly one- en h o global in es men in R&D
om all sou ces. While he da a does no pe mi
an assessmen o he sha es o panel i ms’ R&D
wi h an explici ocus on AI, i is none heless
31 The A i icial In elligence Indus y Alliance (AIIA) websi e is a ailable
a h p://aiiao g.cn/index.php?m=alliance&c=index&a=s uc u e.
The e sion o he membe ship lis employed in he p esen s udy was
de eloped by he Cen e o Secu i y and Eme ging Technology (Luong
and A nold 2021). The au ho s desc ibe he alliance as ollows: “Th ough
he AIIA, he Chinese go e nmen aims o os e collabo a ion among
local go e nmen s, academic ins i u ions, and companies. In some cases,
he Chinese s a e uses he AIIA o ‘pick winne s,’ choosing among
a o ed companies in he AI indus y o ecei e go e nmen subsidies”
(ibid., 1).
32 Fo de ailed i m-le el pe o mance indica o s o hese i ms see
G assano e al. (2022), and he unde lying da a se a ailable om he
EU websi e a h ps://i i.j c.ec.eu opa.eu/sco eboa d/2022-eu-indus ial-
d-in es men -sco eboa d# ield_ epo sco eboa d.
clea ha his panel o leading AI-in ensi e i ms
deli e s esea ch a a globally signi ican scale.
Wi h espec o indus y (see Table 2), he
concen a ions in he Indus y Classi ica ion
Benchma k (ICB) sec o s “Technology Ha dwa e &
Equipmen ” and “So wa e & Compu e Se ices”
a e in line wi h p io expec a ions. Howe e , he
anking includes qui e a ew la ge and di e si ied
co po a ions ( o example, Alibaba, Alphabe ,
Apple, Baidu, Me a, Mic oso , Samsung and
Sony) wi h no able p oduc o e ings beyond hei
lis ed indus y, including o he ields whe e AI
applica ions migh be employed (such as media
o leisu e goods). Alongside global mul i- illion
USdolla behemo hs such as Apple and Mic oso ,
a numbe o smalle i ms wi h a pa icula ocus
on AI a e included in he ankings (see Figu e1).
One in h ee had a ma ke capi aliza ion o less
han US$25 billion (G assano e al. 2022). Thus,
while la ge i ms domina e he leading AI-
in ensi e g ouping (p o iding an upwa d shi
o he a e ages), he e is a signi ican subse
o smalle i ms ha ha e go en ac ion.
Table 3 p o ides compa isons be ween he leading
AI-in ensi e i ms, he emainde o he so wa e
and compu e se ices i ms (excluding he leading
AI-in ensi e i ms33), and he op 2,500 R&D i m
sample as a whole. Conside ing he median sco es
o each g oup o he yea 2021, he assessmen
e eals ha he median leading AI-in ensi e
i m is subs an ially la ge han he median i m
among he op 2,500 in absolu e e ms wi h
espec o ne sales (3.8x), employmen (3.1x), R&D
expendi u e (8.4x), capi al expendi u e (capex)
(3.3x), and ope a ing p o i s (6.0x).34 This is due in
pa o he p esence o e y la ge and di e si ied
co po a ions in he AI ankings. De elopmen
o o iginal AI models can be cos ly in e ms o
compu ing equi emen s and i m scale may
p o ide an ad an age in his ega d. La ge i ms
may also ha e a b oade ange o in-house use cases
o uel AI demand. This is no o say ha smalle
AI in ensi e i ms a e excluded om he ma ke . A
numbe o smalle i ms a e p o ing compe i i e
in adap a ion o exis ing LLMs o pa icula uses
and in de eloping AI-d i en applica ions. In some
33 As no ed in Table 2, mos , bu no all, o he leading AI-in ensi e i ms a e
in he ICB sec ion “So wa e & Compu e Se ices.”
34 In his no a ion, “x” signi ies “ imes.” Fo example, ne sales o he
median leading AI-in ensi e i m a e 3.8 imes hose o he median op
2,500 R&D i m.
12 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
cases, hese smalle i ms may compensa e o lack
o scale by d awing upon open-sou ce inpu s.
A u he con i ma ion o he impo ance o scale
o some segmen s o he AI sec o can be ound
in Table 3 by conside ing he gap be ween he
median leading AI-in ensi e i m and he median
so wa e and compu e se ices i m (excluding
leading AI-in ensi es). This gap is e en la ge han
o he op 2,500 R&D i m sample as a whole.
The median so wa e and compu e se ices i m
(excluding he AI in ensi es) has lowe ne sales,
lowe employmen , less capi al expendi u e and
much lowe ope a ing p o i s han he median
o he op 2,500 as a whole. Only wi h espec o
R&D expendi u e does he median so wa e and
compu e se ice i m (excluding AI in ensi es)
ou pe o m he median i m among he op 2,500
sample as a whole. Indeed, he median R&D
in ensi y (R&D expendi u e ela i e o ne sales)
o bo h he leading AI-in ensi e i ms and he
emainde o he “So wa e & Compu e Se ices”
sec o is much g ea e han o he op 2,500
sample as a whole ( espec i ely 2.6x and 3.0x).
Ca ea s
The da a employed in his assessmen may be
subjec o ce ain biases. The e a e wo impo an
ca ea s: Fi s , he EU eam depends on he public
a ailabili y o R&D expendi u e da a. Only i ms
ha e eal his in o ma ion a e co e ed. This esul s
in co e age p ima ily o publicly lis ed companies,
hough e en among publicly lis ed i ms he
disclosu es a y. Co e age is be e o i ms in he
ad anced economies, which end o ha e be e
co po a e disclosu e o in o ma ion on R&D. Thus,
some po en ially impo an AI inno a o s may be
le o he Eu opean Union’s anking. Acco ding
o he chosen AI-sc eening c i e ia, among la ge
de elope s and inno a i e use s o AI, some
signi ican omissions include Amazon, Capi al
One, GE Heal hCa e, JD.com, McDonald’s, Lowe’s
and Wip o. Second, some leading AI inno a o s
a e ac ually small in e ms o he scale o R&D
expendi u e and hus may no mee he EU op
2,500 c i e ia, e en i hey a e publicly lis ed and
disclose R&D expendi u e in o ma ion. Among
smalle publicly lis ed i ms iden i ied in he sc eens
as leading in e ms o AI echnology bu missing
om he EU ankings one inds: Da k ace PLC,
Inno a i e Eyewea Inc., Ko e G oup Holdings,
Nano Dimension LTD, SoundHound AI Inc., and
Xome y Inc. As a consequence o hese ca ea s,
his i m-le el assessmen may be conside ed as
illus a i e bu no necessa ily ully ep esen a i e.
AI Leade s among he Unico ns
The e is an ac i e insu gency in he AI-in ensi e
i m segmen . A signi ican po ion o he inno a ion
in AI has been uelled by s a -ups challenging —
Table 1: Leading AI-In ensi e Fi ms among Global R&D Expendi u e Leade s, Publicly Lis ed, 2021
Company Coun y Indus y (ICB Sec o ) Company Coun y Indus y (ICB Sec o )
Accen u e I eland Suppo Se ices In ui Uni ed S a es So wa e & Compu e
Se ices
Adobe Uni ed S a es So wa e & Compu e
Se ices
Med onic Public
Limi ed
I eland Heal h Ca e Equipmen
& Se ices
Alibaba G oup
Holding
China So wa e & Compu e
Se ices
Me a Uni ed S a es So wa e & Compu e
Se ices
Alphabe Uni ed S a es So wa e & Compu e
Se ices
Mic oso Uni ed S a es So wa e & Compu e
Se ices
Al ai Enginee ing Uni ed S a es So wa e & Compu e
Se ices
Ne Ease China So wa e & Compu e
Se ices
Al e yx Uni ed S a es So wa e & Compu e
Se ices
NICE Is ael So wa e & Compu e
Se ices
Amba ella Uni ed S a es Technology Ha dwa e
& Equipmen
Nokia Finland Technology Ha dwa e &
Equipmen
Apple Uni ed S a es Technology Ha dwa e
& Equipmen
N idia Uni ed S a es Technology Ha dwa e &
Equipmen

13AI Inno a ion Concen a ion and he Go e nance Challenge
Company Coun y Indus y (ICB Sec o ) Company Coun y Indus y (ICB Sec o )
Baidu China So wa e & Compu e
Se ices
O acle Uni ed S a es So wa e & Compu e
Se ices
B oadcom Uni ed S a es Technology Ha dwa e
& Equipmen
Palo Al o Ne wo ks Uni ed S a es So wa e & Compu e
Se ices
Bu e ly Ne wo k Uni ed S a es Heal h Ca e Equipmen
& Se ices
Pegasys ems Uni ed S a es So wa e & Compu e
Se ices
C3.ai Uni ed S a es So wa e & Compu e
Se ices
Rockwell
Au oma ion
Uni ed S a es Indus ial Enginee ing
Camb icon
Technologies
China Technology Ha dwa e
& Equipmen
Sales o ce Uni ed S a es So wa e & Compu e
Se ices
Ce ence Uni ed S a es So wa e & Compu e
Se ices
Samsung Elec onics Sou h Ko ea Elec onic & Elec ical
Equipmen
Ce a Uni ed S a es Mobile
Telecommunica ions
SAP Ge many So wa e & Compu e
Se ices
Check Poin
So wa e
Technologies
Is ael So wa e & Compu e
Se ices
Seaga e Technology I eland Technology Ha dwa e &
Equipmen
China Telecom China Technology Ha dwa e
& Equipmen
Se iceNow Uni ed S a es So wa e & Compu e
Se ices
Cohe en Uni ed S a es Elec onic & Elec ical
Equipmen
Shu e s ock Uni ed S a es Media
C owdS ike
Holdings
Uni ed S a es So wa e & Compu e
Se ices
Snow lake Uni ed S a es So wa e & Compu e
Se ices
Dell Technologies Uni ed S a es Technology Ha dwa e
& Equipmen
Sony Japan Leisu e Goods
Duolingo Uni ed S a es So wa e & Compu e
Se ices
SS&C Technologies Uni ed S a es So wa e & Compu e
Se ices
Fo ine Uni ed S a es So wa e & Compu e
Se ices
S yke Uni ed S a es Heal h Ca e Equipmen
& Se ices
Hewle Packa d
En e p ise
Uni ed S a es So wa e & Compu e
Se ices
Synap ics Uni ed S a es Technology Ha dwa e &
Equipmen
Huawei In es men
& Holding
China Technology Ha dwa e
& Equipmen
Synopsys Uni ed S a es So wa e & Compu e
Se ices
IBM Uni ed S a es So wa e & Compu e
Se ices
Tencen China So wa e & Compu e
Se ices
iFly ek China So wa e & Compu e
Se ices
Thomson Reu e s Canada Suppo Se ices
In o ma ica Uni ed S a es So wa e & Compu e
Se ices
UiPa h Uni ed S a es So wa e & Compu e
Se ices
In osys India So wa e & Compu e
Se ices
Zscale Uni ed S a es So wa e & Compu e
Se ices
In el Uni ed S a es Technology Ha dwa e
& Equipmen
Sou ce: G assano e al. (2022); au ho ’s abula ions. The unde lying da a se is a ailable om he EU websi e a
h ps://i i.j c.ec.eu opa.eu/sco eboa d/2022-eu-indus ial- d-in es men -sco eboa d# ield_ epo sco eboa d.
No es: Huawei In es men & Holding is employee-owned. Also, he ICB axonomy is employed he e. The i m FTSE Russell manages he
axonomy. Fo de ails, see www.lseg.com/en/ se- ussell/indus y-classi ica ion-benchma k-icb and G assano e al. (2022, 11).
Table 1 (con inued)
14 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
Table 2: Coun s o Top Publicly Lis ed AI Inno a o Fi ms, by Coun y o Regis y and Sec o ,
2021
Coun y Indus y (ICB Sec o )
Canada 1 Elec onic & Elec ical Equipmen 2
China 8 Heal h Ca e Equipmen & Se ices 3
Finland 1 Indus ial Enginee ing 1
Ge many 1 Leisu e Goods 1
India 1 Media 1
I eland 3 Mobile Telecommunica ions 1
Is ael 2 So wa e & Compu e Se ices 34
Japan 1 Suppo Se ices 2
Sou h Ko ea 1 Technology Ha dwa e & Equipmen 12
Uni ed S a es 38
To al AI inno a o i ms 57 To al AI inno a o i ms 57
O which egis e ed in EU
membe coun ies
5
Sou ce and No es: Re e o Table 1 sou ce and no es.
and hen in some cases e en ually collabo a ing
wi h— he la ge incumben s. Among he mos
well-known, o example, a e OpenAI (de elope
o he Cha GPT model, now a ilia ed wi h
Mic oso ) and By eDance (de elope o TikTok
and a go e nmen -app o ed LLM called Yunque;
By eDance is a p i a e i m wi h a public lis ing
pending). The mos success ul AI-in ensi e s a -ups,
such as hese wo, can g ow o become unico ns
many imes o e (be alued a many US$ billions).
In his second i m-le el assessmen , he analysis
shi s o conside he global da abase o 1,220
unico ns de eloped by CB Insigh s, a lis co e ing
i ms om a b oad ange o sec o s. By c oss-
e e encing his da abase agains he os e o
AI leade s published by eWeek (Magui e 2023),
he analysis de elops a wo king lis o op
AI-in ensi e unico ns (see Table 4). Based on
e e ences in he li e a u e, he analysis added
wo u he s a -ups. The i s is Epic Games, a
US-based unico n a ilia ed wi h Tencen . This i m
main ains inno a i e anspa ency p o isions
o con en managemen . Gene a i e AI con en
is welcomed om endo s on he i m’s Un eal
Engine Ma ke place, bu mus be labelled as such.
Vendo s may also exclude hei con en om use
by o he s in AI aining (Epic Games 2023). The
second addi ional s a -up is Zhipu AI, an inno a i e
s a -up selling access o i s AI models as a se ice.
Zhipu AI also has a LLM ha ecei ed go e nmen
app o al in China in Augus 2023 (Liu e al. 2023).
Unico ns: Findings and Ca ea s
One s iking cha ac e is ic o his lis o 22 unico ns
is he geog aphic concen a ion. All bu ou o
hese unico ns a e US-based i ms, he o he s being
based in China ( wo) and he Uni ed Kingdom ( wo).
Looking a bi close , he ex en o he geog aphic
concen a ion is e ealed. Only a ew ci ies wi hin
hese h ee coun ies ac ually hos hese unico ns
(see Figu e 2). This iew ound 10 unico ns eme ging
in he San F ancisco-Moun ain View-San Jose
co ido o Cali o nia. Ano he i e a e based in
New Yo k Ci y, wi h Bos on, Massachuse s, Ca y,
No h Ca olina, and Aus in, Texas, accoun ing
o he emaining US en ies on he lis . London
accoun s o wo and Beijing o wo. In e ms
o sec o , nea ly wo- hi ds (14) o he unico ns
a e ocused on en e p ise echnology. Th ee a e
conce ned wi h consume and e ail applica ions,
h ee wi h media and en e ainmen , and wo
wi h heal h ca e and li e sciences. Mos o he
cu en AI leade s among he unico ns a e ocused
on p oduc s o businesses (business- o-business
[B2B]), e en hough some o he B2B p oduc s will
be employed in consume - acing applica ions. A
ew a e deli e ing di ec - o-consume p oduc s
such as TikTok (By eDance), Fo ni e (Epic
Games) o Cha GPT and DALL-E (OpenAI).
15AI Inno a ion Concen a ion and he Go e nance Challenge
Coun s Median Sco es
Numbe
o i ms
Numbe
o
coun ies
Numbe
o egions
Numbe
o indus y
(ICB) sec o
names
Median
wo ld
ank
R&D 2021
(US$M)
R&D
one-yea
g ow h
(%)
Ne sales
(US$M)
Ne sales
one-yea
g ow h (%)
R&D in ensi y
(%)
AI-in ensi e
i ms 57 10 5 9 200 1,168.1 18.3 10,061.7 18.0 16.9
So wa e and
compu e
se ices i ms
(excluding AI-
in ensi e)
302 23 5 1 1143 156.7 22.7 782.5 21.8 19.3
Top 2,500 R&D
i ms ( ull
sample)
2,500 41 5 38 1251 139.4 15.7 2,627.2 16.9 6.4
Median Sco es
Capex
(US$M)
Capex
one-yea
g ow h
(%)
Capex
in ensi y
(%)
Ope a ing
p o i s
(US$M)
Ope a ing
p o i s
one-yea
g ow h
(%)
P o i abili y
(%) Employees
Employees,
one-yea
g ow h
(%)
Ma ke
capi aliza ion
(US$M)
Ma ke
capi aliza ion
one-yea
g ow h (%)
AI-in ensi e
i ms 341.4 5.6 3.1 1,297.6 26.8 14.0 23,437 11.3 53,759.0 36.7
So wa e and
compu e
se ices i ms
(excluding AI-
in ensi e)
18.2 11.3 2.0 4.6 12.0 1.2 3,610 15.0 5,682.4 17.2
Top 2,500 R&D
i ms ( ull
sample)
104.3 13.2 4.1 216.1 21.1 8.0 7,522 5.3 5,200.4 26.6
Sou ce: G assano e al. (2022); unde lying da a se is a ailable om he EU websi e a h ps://i i.j c.ec.eu opa.eu/sco eboa d/2022-eu-indus ial-
d-in es men -sco eboa d# ield_ epo sco eboa d; au ho ’s abula ions. US$ igu es a e con e ed om eu os using he US Fede al Rese e
annual exchange a e (G.5A), a ailable a www. ede al ese e.go / eleases/g5a/cu en /.
No e: As de ined in G assano e al. (2022), he egional axonomy includes China, he Eu opean Union, Japan, he Uni ed S a es, and he es o
he wo ld.
Table 3: Coun s and Median Sco es o he G oup o Leading AI-In ensi e Fi ms e sus O he
G oupings, 2021
16 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
Figu e 1: Top Publicly Lis ed AI-In ensi e Fi ms wi h Respec o R&D Expendi u e, Ranked by
Ma ke Capi aliza ion
0 500,000
Wo ld ank o R&D expendi u e
Ma ke capi aliza ion (US$M, 2021)
1,000,000 2,000,000 2,500,000 3,000,0001,500,000
Apple
Mic oso
Me a
Alphabe
Tencen
NVIDIA
Alibaba G oup Holding
Samsumg Elec onics
Adobe
Sales o ce
O acle
Accen u e
In el
B oadcom
Med onic Public Limi ed
SAP
In ui
Se iceNow
IBM
Sony
S yke
In osys
Snow lake
Ne Ease
C owdS ike Holdings
Thomson Reu e s
Fo ine
Synopsys
Palo Al o Ne wo ks
Baidu
Zscale
Rockwell Au oma ion
Nokia
Dell Technologies
UiPa h
Seaga e Technology
Hewle Packa d En e p ise
SS&C Technologies
NICE
Check Poin So wa e Technologies
iFly ek
Pegasys ems
Synap ics
Cohe en
C3.ai
China Telecom
Al e yx
Shu e s ock
Ce ence
Amba ella
Al ai Enginee ing
Bu e ly Ne wo k
Ce a
Sou ce: G assano e al. (2022); unde lying da a se is a ailable om he EU websi e a h ps://i i.j c.ec.eu opa.eu/
sco eboa d/2022-eu-indus ial- d-in es men -sco eboa d# ield_ epo sco eboa d; au ho ’s abula ions.
No e: The EU da abase does no include ma ke capi aliza ion da a o Camb icon Technologies, Duolingo, Huawei
In es men & Holdings o In o ma ica.
23AI Inno a ion Concen a ion and he Go e nance Challenge
Table 5, Pa 2: Coun s o Policy Ini ia i es as o 2021 in Rela ion o AI in an Illus a i e G oup o
O he Coun ies
Sample
Coun ies
Non-sample Coun ies
Policy Ins umen Type Median B azil Indonesia Singapo e Sou h A ica Vie nam
AI coo dina ion and/o moni o ing bodies 2.0 1 0 0 0 0
AI compu ing and esea ch in as uc u e 2.5 0 0 0 0 1
AI skills and educa ion 1.0 1 0 0 0 0
AI use in he public sec o 1.0 13 0 6 0 0
Cen es o excellence g an s 0.5 1 0 0 0 0
Da a access and sha ing 1.0 0 0 1 0 0
Eme ging AI- ela ed egula ion 5.0 2 0 6 0 1
Equi y inancing 0.0 0 0 0 0 0
Fellowships and pos g adua e loans and schola ships 0.0 0 0 0 0 0
G an s o business R&D and inno a ion 0.5 0 0 1 0 0
Indi ec inancial suppo 0.0 0 0 0 0 0
Ins i u ional unding o public esea ch 0.0 1 0 1 0 0
Knowledge ans e s and business ad iso y se ices 1.0 1 0 2 0 0
Labou ma ke policies 0.0 0 0 0 0 0
Labou mobili y egula ion and incen i es 0.0 0 0 0 0 0
Na ional s a egies, agendas and plans 7.0 5 1 9 0 2
Ne wo king and collabo a i e pla o ms 4.0 0 0 2 1 0
P ocu emen p og ams o AI R&D and inno a ion 1.0 0 0 0 0 1
P ojec g an s o public esea ch 2.5 1 0 1 0 1
Public awa eness campaigns and ci ic pa icipa ion
ac i i ies
1.0 0 0 2 0 1
Public consul a ions o s akeholde s o expe s 5.0 2 1 1 1 1
Regula o y o e sigh and e hical ad ice bodies 1.5 0 0 1 1 0
S anda ds and ce i ica ion o echnology de elopmen
and adop ion
0.0 1 0 4 0 0
Science and inno a ion challenges, p izes and awa ds 0.0 0 0 0 0 0
To al coun 43.5 29 2 37 3 8
Memo I ems
Policy ini ia i es in he OECD.AI policy da abase ha
men ion “inno a ion” in he objec i e desc ip ion
9.0 19 2 11 1 2
Sha e o ini ia i es agged o ha ing a men ion o
“inno a ion” in he objec i e desc ip ion (pe cen age)
20.4 65.5 100.0 29.7 33.3 25.0
Policy ini ia i es in he OECD.AI policy da abase
add essing anspa ency issues (e.g., in policy
o ma ion and egula o y p ocesses o in he ope a ion
o algo i hms)
11.5 11 2 10 0 0
Sha e o ini ia i es agged in ela ion o anspa ency
(pe cen age)
32.4 37.9 100.0 27.0 0.0 0.0
Sou ces and No es: See Table 5, Pa 1, sou ces and no es.

24 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
impedimen s o di usion o AI inno a ion may
also be add essed as pa o hese e o s.40
DPA: Machine Lea ning and AI Policy Ac i i y
In o de o ge a be e sense o he ecen policy
dynamics a ound he accele a ion in AI, he analysis
u ns o a mo e cu en esou ce o acking AI
policy ac i i y. This is DPA, an ini ia i e o he S .
Gallen Endowmen o P ospe i y h ough T ade.
I co e s policy de elopmen s in G20 coun ies,
EU membe s a es and Swi ze land. The sample
employed o he p esen analysis is o ac ions
ele an o machine lea ning41 and AI. The
ac ions acked include laws, o de s, s anda ds
and guidelines egula ing he de elopmen and
use o AI sys ems, as well as ela ed egula o y
ac ions and policy de elopmen wo k. Ac i i ies
may ocus on one o mo e policies and a single
policy may be he objec o one o mo e ac i i ies
o e ime. The da abase is upda ed in an ongoing
manne wi h he suppo o cus omized so wa e.
The s uc u e o he DPA da abase is conduci e
o conside a ion o policy ac i i y o e ime o
ecen yea s s a ing in 2020. Table 6 p o ides an
illus a ion o he accele a ion in policy ac i i y
be ween wo pe iods, 2020–2021 and 2022–
No embe 2023. F om he ini ial pe iod, all o he
sample coun ies hos ing leading AI-in ensi e
i ms we e al eady engaged in machine lea ning
and AI policy ac i i ies. The Uni ed S a es led he
g oup wi h 15 inciden s o policy ac i i y, ollowed
by he Eu opean Union (10), he Uni ed Kingdom
(eigh ) and China (se en). A ew in e na ional
coun y g oupings also ook some s eps in he
i s wo yea s co e ed he e (2020–2021). In he
second pe iod, om 2022 o No embe 2023, he
olume o ac i i y accele a ed by 4.2 imes o e all.
The leading AI-in ensi e i m hos coun ies all
s epped up policy ac i i y, including he Uni ed
S a es ( o 99 ac i i ies), China ( o 26), he Eu opean
Union ( o 20) and he Uni ed Kingdom ( o 20). The
numbe o ac ions ac oss in e na ional bo de s
also amped up, om i e in he ini ial wo yea s
o 37 in he pe iod om 2022 o No embe 2023.
40 Fo mo e on his pe spec i e o egula o y e o m, see he OECD e icien
egula ion p inciples o ma ke openness (Tsai e al., 2011, 18, box 1)
and he de ailed discussion o he applica ion o hese p inciples in he
case o Is ael (ibid., 18–39).
41 Machine lea ning in ol es use o a model ained and e ined on la ge
da a se s o classi y in o ma ion, make p edic ions and gene a e con en .
I enables compu e s o lea n wi hou being explici ly p og ammed.
Each o hese in e na ional ac ions engaged
one o mo e coun ies om he sample.
The subjec ma e co e ed inc eased in scope and
in ensi y (see Table 7). Du ing 2020–2021, 12 subjec
a eas we e ackled. In 2022–2023, his ose o 30.
In bo h pe iods, “Algo i hm design and echnical
s anda ds” exhibi ed he mos ac i i y. “Da a
p o ec ion egula ion” ac i i y ose d ama ically
om he i s pe iod o he second. Also du ing
he second pe iod, “Cybe secu i y egula ion,”
“Con en mode a ion egula ion,” “In ellec ual
p ope y (copy igh and pa en s)” and “Consume
p o ec ion egula ion” all ose o ill ou he op
anks o policy a eas. Mo eo e , he inc eased
olume o policy ac i i y led o subs an i e
esul s. Ac i i ies o policies adop ed ose om
40 du ing he i s biennium, o 114 du ing he
second. A u he 35 ac ions conce ned policies in
o ce du ing 2022–2023. In addi ion, he pipeline o
policy ac i i y poin s o a con inued high olume o
ac i i y going o wa d. Du ing 2022–2023, some 45
ins umen s we e a some poin in he consul a ion
p ocess and 80 we e unde delibe a ion. Among
he leading issues o ac i i ies conce ning policy
ins umen s unde delibe a ion, “Algo i hm design
and echnical s anda ds” emains in i s place and
“Da a p o ec ion egula ion” places a s ong second.
Findings om Analysis o DPA T acking
The DPA acking highligh s he esponsi eness o
policy make s o he accele a ion in AI inno a ion,
wi h he equency o policy ini ia i es inc easing
as well as he coun y co e age. Some o he
subjec a eas equen ly a ge ed, such as echnical
s anda ds and IP igh s, can ha e an in luence on
di usion o echnology ac oss bo de s. Alignmen
in egula ion in hese a eas can acili a e ade in
p oduc s and se ices by pe mi ing de elope s
o aim o a mo e uni o m p oduc wi h less
ailo ing equi ed, po en ially also educing he
bu den o es ing and ce i ica ion whe e equi ed.
The inc eased co e age o p o ec ion o IP igh s
and da a p o ec ion may also enable de elope s
o sell AI p oduc s in o an expanded ange o
ma ke s ( o example, i he e o ms enable
de elope s o be e p o ec hei p oduc s om
abuses o i he e o ms educe unce ain y wi h
espec o liabili y). Also, conce ning egula o y
alignmen , i is encou aging o see he g owing
numbe o coun y g oupings aking join policy
ini ia i es, o en including he majo economies.
25AI Inno a ion Concen a ion and he Go e nance Challenge
Table 6: Dynamics o Recen Machine Lea ning and AI Policy De elopmen s, Janua y 2020–
No embe 2023 (Coun s)
2020–2021 2022–2023
Coun y o G oup o Coun ies Policy
Ac i i y
Coun y o G oup o Coun ies Policy
Ac i i y
— A gen ina 5
Aus alia 3 Aus alia 6
— Aus ia 1
B azil 2 B azil 7
Canada 2 Canada 9
China 7 China 26
— Denma k 1
Eu opean Union 10 Eu opean Union 20
— F ance 6
Ge many 3 Ge many 9
— G eece 1
Hong Kong 1 —
— Hunga y 1
India 2 India 3
Indonesia 1 Indonesia 1
I aly 1 I aly 4
— Japan 8
— Ne he lands 3
— New Zealand 2
— Pe u 2
— Poland 1
Russia 1 Russia 1
—Saudi A abia 1
— Singapo e 4
—Sou h A ica 1
Sou h Ko ea 4 Sou h Ko ea 8
Spain 1 Spain 5
Swi ze land 1 Swi ze land 2
Taiwan 1 —
Tü kiye 2 —
— Uk aine 1
—Uni ed A ab Emi a es 1
Uni ed Kingdom 8 Uni ed Kingdom 20
Uni ed S a es 15 Uni ed S a es 99
Aus alia, Eu opean Union, India, Japan, Mexico,
New Zealand, Singapo e, Sou h Ko ea, Uni ed
Kingdom, Uni ed S a es
1 —
Canada, Uni ed Kingdom, Uni ed S a es 1 —
Chile, New Zealand, Singapo e, Sou h Ko ea 1 —
UNESCO membe ship 2 —
26 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
2020–2021 2022–2023
Coun y o G oup o Coun ies Policy
Ac i i y
Coun y o G oup o Coun ies Policy
Ac i i y
— Albania, A gen ina, B azil, Chile, China, Colombia,
Ecuado , Egyp , Es onia, E hiopia, Finland, F ance,
Ge many, Guyana, India, Is ael, I aly, Japan, Kenya,
Mexico, Ne he lands, New Zealand, Pakis an, Pe u,
Po ugal, Russia, Saudi A abia, Senegal, Singapo e,
Sou h A ica, Sou h Ko ea, Spain, Su iname,
Swi ze land, T inidad & Tobago, Uni ed A ab
Emi a es, Uni ed Kingdom, Uni ed S a es, Zimbabwe
1
—Ibe o-Ame ican Ne wo k: Ando a, A gen ina, B azil,
Cape Ve de, Chile, Colombia, Cos a Rica, Dominican
Republic, Ecuado , El Sal ado , Gua emala, Hondu as,
Mexico, Nica agua, Panama, Pa aguay, Pe u, Po ugal,
Sao Tome & P incipe, Spain, U uguay
1
— A gen ina, Aus alia, Canada, Colombia, Hong Kong,
Mexico, Mo occo, New Zealand, No way, Swi ze land,
Uni ed Kingdom
1
— Aus alia, Aus ia, Belgium, B azil, Bulga ia, Canada,
Chile, China, C oa ia, Cyp us, Czechia, Denma k,
Es onia, Finland, F ance, Ge many, G eece, Hunga y,
India, Indonesia, I eland, Is ael, I aly, Japan, Kenya,
La ia, Li huania, Luxembou g, Mal a, Ne he lands,
Nige ia, Philippines, Poland, Po ugal, Romania,
Rwanda, Saudi A abia, Singapo e, Slo akia, Slo enia,
Sou h Ko ea, Spain, Sweden, Swi ze land, Tü kiye,
Uk aine, Uni ed A ab Emi a es, Uni ed Kingdom,
Uni ed S a es
1
— OECD: Aus alia, Aus ia, Belgium, Canada, Chile,
Colombia, Cos a Rica, Czechia, Denma k, Es onia,
Finland, F ance, Ge many, G eece, Hunga y, Iceland,
I eland, Is ael, I aly, Japan, La ia, Li huania,
Luxembou g, Mexico, Ne he lands, New Zealand,
No way, Poland, Po ugal, Slo akia, Slo enia, Sou h
Ko ea, Spain, Sweden, Swi ze land, Tü kiye, Uni ed
Kingdom, Uni ed S a es
2
— Aus alia, Canada, Ge many, Ne he lands, New
Zealand, Uni ed Kingdom, Uni ed S a es
1
— Aus alia, Uni ed Kingdom 1
— G7: Canada, F ance, Ge many, I aly, Japan, Uni ed
Kingdom, Uni ed S a es
4
—Canada, Japan 2
—China, Saudi A abia 1
— Eu opean Union, India 1
— Eu opean Union, Sou h Ko ea 1
— Eu opean Union, Uni ed S a es 3
—Ge many, Japan 1
— India, Is ael 1
— India, Uni ed Kingdom 1
— India, Uni ed S a es 1
—Indonesia, Japan 1
— Japan, Uni ed Kingdom 2
— Ge many, Sou h Ko ea 1
— Singapo e, Sou h Ko ea 1
— Sou h Ko ea, Uni ed S a es 1
Table 6 (con inued)
27AI Inno a ion Concen a ion and he Go e nance Challenge
Table 6 (con inued)
Table 7: Na u e o Recen Machine Lea ning and AI Policy De elopmen s, Janua y 2020–
No embe 2023 (Coun s)
Panel A: Policy Ins umen s
2020–2021 2022–2023
Policy Ins umen s Policy
Ac i i y
Policy Ins umen s Policy
Ac i i y
Algo i hm design and echnical s anda ds 23 Algo i hm design and echnical s anda ds 99
— Au ho iza ion o goods o se ices 8
— Bo de measu e, n.e.s. 1
— Business egis a ion equi emen 10
— Compe i ion au ho i y go e nance 1
— Consume p o ec ion egula ion 11
— Con en mode a ion egula ion 12
— Con en emune a ion egula ion 1
Con en - ela ed egula ion and IP, n.e.s. 1Con en - ela ed egula ion and IP, n.e.s. 4
— Cybe secu i y egula ion 14
— Da a p o ec ion au ho i y go e nance 7
Da a p o ec ion egula ion 9 Da a p o ec ion egula ion 58
Di ec axes, including digi al se ice axes 1 —
Expo licensing equi emen s 2 Expo licensing equi emen s 5
Financial g an 5 Financial g an 6
Ins umen unspeci ied 13 Ins umen unspeci ied 5
— In ellec ual p ope y: copy igh 6
In ellec ual p ope y: pa en s 4 In ellec ual p ope y: pa en s 6
Me ge con ol egula ion 1 Me ge con ol egula ion 1
MNE: Financial incen i e 1 —
Ope a ing condi ion, n.e.s. 9 Ope a ing condi ion, n.e.s. 4
— Ope a ional licence equi emen 1
— P oduc ion subsidy 1
— Public p ocu emen access 4
2020–2021 2022–2023
Coun y o G oup o Coun ies Policy
Ac i i y
Coun y o G oup o Coun ies Policy
Ac i i y
— Singapo e, Uni ed Kingdom 1
— Singapo e, Uni ed S a es 2
— Uk aine, Uni ed Kingdom 1
— Uni ed Kingdom, Uni ed S a es 3
To al 70 To al 296
Sou ces: DPA (2023); au ho ’s abula ions using he unde lying da a.
No es: “—” indica es ha he e is no compa able ac i i y en y o he cell du ing he ime pe iod co e ed by he column. Also no e ha
he DPA Ac i i y T acke p o ides in o ma ion on ecen de elopmen s in legisla u es, judicia ies and he execu i e b anches o he G20,
EU membe s a es and Swi ze land. Among he AI leade coun ies mos a e co e ed. The DPA acking does no ha e sepa a e lis ings o
I eland o Finland, bu hey a e pa ly co e ed ia he EU lis ings. Is ael is only co e ed o i s in e na ional acco ds, no o i s domes ic
policy ac i i y.
28 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
Panel A: Policy Ins umen s
2020–2021 2022–2023
Policy Ins umen s Policy
Ac i i y
Policy Ins umen s Policy
Ac i i y
— Public p ocu emen egula ion, n.e.s. 1
— Quali y o se ice equi emen s 7
— Regis a ion and licensing, n.e.s. 1
Se ice access es ic ions 1 Se ice access es ic ions 1
— S a e aid, n.e.s. 7
— Unila e al conduc egula ion 3
— Wo ke moni o ing egula ion 9
— Wo ke s a us classi ica ion egula ion 2
To al 70 To al 296
Panel B: S a us o Policy
2020–2021 2022–2023
S a us Types Policy
Ac i i y
S a us Types Policy
Ac i i y
Adop ed 40 Adop ed 114
Concluded 1 Concluded 3
In consul a ion 1 In consul a ion 7
In o ce 11 In o ce 35
— In g ace pe iod 1
P ocessing consul a ion 6 P ocessing consul a ion 38
Rejec ed 1 Rejec ed 12
Unde delibe a ion 10 Unde delibe a ion 80
— Unde in es iga ion 6
To al 70 To al 296
Sou ces and No es: See Table 6. In addi ion, MNE = mul ina ional en e p ise; n.e.s. = no elsewhe e speci ied.
Table 7 (con inued)
Leading AI-En e p ise Hos
Coun y Regula o y Speci ics
As he global go e nance amewo k o AI is s ill
unde de elopmen , s eps aken by he majo
economies (China, he Eu opean Union and
he Uni ed S a es; see Appendix 2) may ha e a
signi ican in luence. This in luence de i es no only
om he ma ke powe o policy and egula o y
ac ions aken in hei own domes ic ma ke s, bu
also om hei in luence in in e na ional bodies
( o example, he OECD, he Asia-Paci ic Economic
Coope a ion egional o um (APEC),42 he G7 and
he G20),43 h ough hei indus y associa ions
( o example, he AI Indus y Alliance in China
o he Pa ne ship on AI and he Da a and T us
42 Fo in o ma ion abou APEC, see www.apec.o g/Abou -Us/Abou -APEC/.
43 A he in e na ional le el, o example, hese na ions we e engaged in
he UNESCO “Recommenda ions on he E hics o AI in 2021” (UNESCO
2022) and he Ble chley Decla a ion on AI Sa e y ag eed by mos G20
coun ies and o he s in 2023 (GOV.UK 2023). Fu he mo e, in a newly
es ablished EU-US ech pa ne ship ( he T ade and Technology Council),
he Eu opean Union and he Uni ed S a es a e seeking o de elop a
mu ual unde s anding on he p inciples unde lining us wo hy and
esponsible AI (Madiega 2023). They bo h ha e endo sed he OECD’s
Recommenda ion o he Council on AI, which lays ou basic p inciples
(OECD 2019).

29AI Inno a ion Concen a ion and he Go e nance Challenge
Alliance in he Uni ed S a es), and h ough hei
ade measu es ( o example, US expo con ols
a ec ing hi d pa ies). China has been a i s
mo e in AI egula ion. Canadian,44 EU and US
egula o s also ha e signi ican elemen s in place,
o en d awing on p e-exis ing law and egula ions
applicable o he digi al economy mo e gene ally
(Aa onson 2024). Bu , hey a e also mo ing
o wa d igo ously wi h AI-speci ic measu es.
Gi en he po en ial o c oss-bo de di usion o AI
echnology, i is clea ha inclusi e in e na ional
coo dina ion is needed o add ess conce ns
(B emme and Suleyman 2023). This would apply,
o example, wi h espec o ma e s anging
om aceabili y and sa e y o assignmen o IPRs
and p o ec ions ( o example, o algo i hms).
Some aspec s will be sec o speci ic ( o
example, pa ien sa e y and p i acy in heal h
ca e), whe eas o he s will be c osscu ing ( o
example, p o isions o egula o s’ access and
o e sigh o algo i hms). A he same ime, in
ac ing o add ess such conce ns, he in e na ional
communi y may educe impedimen s o esponsible
di usion o AI echnology. Such a esul could
p omo e mo e inclusi e economic gains om
his echnology in e na ionally (DSIT 2023).
China’s i s -mo e s a us in AI egula ion has
gi en he au ho i ies aluable hands-on expe ience
o in o m u u e egula ion. I appea s a ange
o mo i a ions a e in ol ed: a goal o ensu ing
adequa e con ols on in o ma ion con en in line
wi h go e nmen policy o poli ical and social
s abili y; a desi e o anspa ency and adequa e
p o ec ions; and an objec i e o ha ing a policy
en i onmen conduci e o Chinese leade ship in AI
echnologically and comme cially, including wi h
espec o e hical no ms and s anda ds (Sheehan
2023). China’s h ee-pa egula o y amewo k
(Cybe space Adminis a ion o China [CAC] 2021,
2022, 2023) co e s algo i hmic ecommenda ions,
deep syn hesis in e ne in o ma ion se ices
( o example, i ual eali y, deep lea ning and
gene a i e sequencing algo i hms) and gene a i e
AI se ices. The amewo k p o ides o egis a ion
o LLMs, wi h he i s ba ch o 12 models epo edly
being au ho ized in Augus 2023 (Liu e al. 2023).
The EU au ho i ies ha e sough o balance suppo
o AI sec o de elopmen wi h conce ns abou AI
sa e y and condi ions o a compe i i e ma ke .
44 The case o Canada is co e ed in Appendix 2.
A subs an ial landma k piece o legisla ion, he
A i icial In elligence Ac , is in ad anced s ages o
de elopmen and en y in o o ce is an icipa ed
by mid-yea 2024 ( o example, see Bambe g
e al. 2024; Den ons 2023; Heikkilä 2023).45 This
legisla ion includes a ie ed AI isk managemen
amewo k ha would ban unaccep able isks ( o
example, models ha employ ha m ul manipula i e
subliminal echniques), s ic ly egula e high- isk
AI sys ems ( o example, biome ic iden i ica ion
and ca ego iza ion o na u al pe sons), impose
anspa ency and ope a ional obliga ions on
limi ed isk AI sys ems ( o example, syn he ic
ou pu om gene a i e AI sys ems will need o be
labelled as such), and pe mi ee use o low and
minimal isk AI sys ems ( o example, spam il e s
o ideo games), al hough po en ially subjec ing
hem o a u u e code o conduc . Gene al pu pose
AI models ( ounda ion models, LLMs) will en ail
addi ional compliance equi emen s o main ain
and p o ide cu en echnical documen a ion,
sys em anspa ency, copy igh compliance
and in o ma ion on aining da a. The AI Ac
also es ablishes an EU AI o ice in he Eu opean
Commission o supe ise and en o ce ce ain
aspec s o compliance in coope a ion wi h a new
EU AI boa d o membe coun y ep esen a i es,
a new scien i ic ad iso y panel o independen
expe s and a new ne wo k o na ional compe en
au ho i ies in each membe s a e. Penal ies o
i m-le el iola ions o compliance equi emen s
could be subs an ial, anging om 1.5 pe cen o
se en pe cen o annual u no e (subjec o ce ain
minimum and maximum h esholds). The AI Ac
will i in o a sys em o exis ing measu es go e ning
he digi al economy in he Eu opean Union
including he Gene al Da a P o ec ion Regula ion,
he Cybe secu i y Ac , he Digi al Se ices Ac
and he Digi al Ma ke s Ac , among o he s.
Likewise in he Uni ed S a es, he go e nmen has
elied on exis ing legal au ho i ies o discipline
AI s akeholde s. As US Fede al T ade Commission
Chai Lina M. Khan (2023, 1) no ed, “The e is no AI
exemp ion o he laws on he books.” So, a c ime
such as aud may emain a c ime e en i he
aud akes place using a new ype o echnology.
Wi h espec o AI-speci ic egula ion, he Uni ed
S a es had ini ially aken a lenien app oach
be o e ad ancing o e he pas wo yea s in a
s epwise ashion, allowing inno a ion in he
p i a e sec o o h i e. I se ou basic p inciples
45 See also Eu opean Pa liamen (2024).
30 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
o AI go e nance in Oc obe 2022, consul ed wi h
s akeholde s du ing he i s hal o 2023, ag eed
a olun a y egime wi h 15 leading i ms du ing
summe 2023, hen ollowed wi h a comp ehensi e
Execu i e O de d awing on exis ing legisla i e
au ho i ies in Oc obe 2023. This measu e imposes
no i ica ion equi emen s on de elope s o
dual-use ounda ion models.46 I also manda es
ede al agencies o p oceed wi h esponsible and
s a egic exploi a ion o AI sys ems, as well as
adminis a i e s eps o suppo AI s anda ds, ools
and es s, human capi al de elopmen , esea ch and
in e na ional coope a ion, among o he elemen s.
O he economies engaged in leading-edge wo k
on AI ace he challenge o es ablishing hei own
domes ic AI go e nance amewo ks and de ending
na ional in e es s, while also s i ing o meaning ul
coope a ion wi h ele an in e na ional ins i u ions
and he h ee majo economies (China, he Eu opean
Union and he Uni ed S a es). Canada p o ides a
use ul example (see Appendix 2). Domes ically, i is
wo king o comple e legisla ion p o iding a basis o
a p incipled, isk-based amewo k o egula ion o
AI. The A i icial In elligence and Da a Ac (AIDA)47
would p o ide o u he de elopmen o AI
go e nance in he na ion, while allowing lexibili y
o main ain cohe ence wi h key pa ne economies
in e na ionally, as easible. In e na ionally, Canada
is ela i ely well posi ioned wi h espec o some
ele an digi al economy ma e s co e ed ia
p o isions o i s ee ade ag eemen s wi h he
Eu opean Union ( he Comp ehensi e Economic and
T ade Ag eemen , and ela ed acco ds including
he Gene al Da a P o ec ion Regula ion [GDPR]
adequacy), wi h he Uni ed S a es and Mexico
(CUSMA) and 10 Paci ic Basin pa ne s (CPTPP).
Fo example, hese acco ds can be suppo i e
o AI de elopmen wi h espec o access o
da a. Canada is also suppo ing collabo a ion
ela ed o he digi al economy (including AI) wi h
46 The e m “dual-use ounda ion model” means an AI model ha is ained
on b oad da a; gene ally uses sel -supe ision; con ains a leas ens o
billions o pa ame e s; is applicable ac oss a wide ange o con ex s;
and ha exhibi s, o could be easily modi ied o exhibi , high le els
o pe o mance a asks ha pose a se ious isk o secu i y, na ional
economic secu i y, na ional public heal h o sa e y, o any combina ion
o hose ma e s (Biden 2023). Ellio Jones (2023) p o ides u he
elabo a ion on his ype o model.
47 The House o Commons concluded i s second eading o AIDA
on Ap il 24, 2023. As o Janua y 31, 2024, i is cu en ly unde
conside a ion be o e he S anding Commi ee on Indus y and
Technology. See LEGISin o, “C-27, 44 h Pa liamen ,” a ailable a
www.pa l.ca/legisin o/en/bill/44-1/C-27.
membe s o he APEC o um, he WTO and he
OECD, and o he in e na ional ins i u ions.
Conclusions
AI Inno a ion Fuels
P essu e on Regula o s
O e he pas decade, he pace o echnological
ad ance in he AI sec o has accele a ed wi h
espec o so wa e a chi ec u e and algo i hms,
aining da a assembly and mobiliza ion o powe ul
compu ing ha dwa e, among o he a eas. The
eme gence o eadily accessible gene a i e AI has
opened he doo o b oad use o he echnology by
he wide public. LLMs, in pa icula , ha e ma ked
some subs an ial and highly isible p og ess.
O e he pas wo yea s, he elease o ex -d i en
LLMs such as Cha GPT and Ba d su p ised some
obse e s wi h he le els o unc ionali y p o ided.
While such AI ad ances a e eal enough, he e is
also an elemen o eupho ia among some use s,
pe haps al eady an icipa ing he nex s eps in
AI de elopmen (Mu gia and Tho nhill 2023).
Indeed, despi e AI ha ing i s ea ly oo s da ing back
o he 1950s, and despi e he ecen wa e o p og ess
in AI de elopmen , i is s ill ea ly days in he AI
e olu ion.48 In eg a ion o AI in o many business
p ocesses is only now scaling up. Fo example,
inc eased applica ion o AI in indus ial au oma ion
and mul i asking obo ics has he po en ial o
unlock signi ican addi ional p oduc i i y gains.
The e emain subs an ial un ul illed AI objec i es
on he ho izon. Fo example, he long- ime AI
communi y goal o deli e ing a i icial gene al
in elligence — ha is, he human-equi alen
capaci y o pe cep ion, easoning, in e ence and
ac ion in an AI sys em — emains o be achie ed.
Wi h all ha has been accomplished in he sec o
and he p ospec o mo e powe ul sys ems unde
de elopmen , egula o s a e indeed eeling some
u gency in he need o imp o ed AI go e nance.
48 Fo example, al hough use o embedded AI is widesp ead in US
businesses, he US Census Bu eau ecen ly ound ha as o end o
Oc obe 2023 only 3.8 pe cen o US businesses epo ed using AI o
ac ually p oduce goods and se ices, hough he e is wide a ia ion
ac oss sec o s (B eaux and Dinle soz 2023).
31AI Inno a ion Concen a ion and he Go e nance Challenge
Concen a ion and Di usion
o AI Inno a ion
A he ou se o his pape , a cen al conce n was
highligh ed o examine he concen a ion o AI
inno a ion and o conside ac o s in he egula o y
en i onmen ha may suppo o hinde i s
esponsible di usion. The illus a i e assessmen o
leading AI-in ensi e i ms (see he sec ion “Leading
AI-In ensi e Fi ms: A Few S ylized Fac s” abo e)
con i med he ex en o concen a ion o a leas
wo signi ican popula ions o AI-in ensi e i ms:
hose AI-in ensi e i ms in he uppe echelon o
global businesses o R&D expendi u e (selec ed
om among he op 2,500 i ms globally o R&D
expendi u e in all ields); and hose AI-in ensi e
i ms wi h success as a s a -up in g owing hei
alua ion beyond US$1 billion (unico ns). In
bo h cases, he assessmen ound a ai ly high
concen a ion geog aphically in AI inno a o s.
The i ms in hese wo popula ions ep esen jus
11 na ions, h ee o which a e membe s o he
Eu opean Union. The Uni ed S a es is home o
a majo i y o he AI i ms in hese samples. The
concen a ion o i ms in jus 11 hos coun ies
may in pa be he esul o agglome a ion e ec s,
whe eby de elopmen in c owded echnology
cen es may o e some ad an ages o ad ancemen
o AI de elopmen and di usion. Fo example,
hick labou ma ke s may de elop and o e la ge
pools o sec o - ele an alen . And, imp o ed
communica ion a ound inno a ion could eme ge
due o he p oximi y o s akeholde s, he eby
con e ing u he in o ma ion ad an ages.49
In e na ionally, he a ailabili y o AI inno a ion
depends in pa on openness in channels o
echnology ans e . This may ake place ia such
means as he sale o licensing o p oduc s and
se ices, o eign di ec in es men , join en u es
and ex e nal s a aining ini ia i es, among o he
possibili ies.50 The e a e also oppo uni ies o
emo e access o AI inno a ion h ough online
hubs and so wa e-as-a-se ice p o ide s. F om
he li e a u e e iew (see he sec ion “Li e a u e:
Rapid Technological De elopmen , Unse led
Schola ship” abo e), some ac o s associa ed wi h
di usion o echnology can be iden i ied. Fo
49 Fo example, see he discussion on agglome a ion in Paul K ugman
(1995).
50 Fo example, Pa k and Lippold (2014) p o ide an examina ion o hese
echnology ans e issues om an IP pe spec i e.
example, a ailabili y o adequa e p o ec ion o IP
p o ec ion may play a ole in acili a ing ans e s
( o example, enabling iden i ica ion o igh s and
obliga ions wi h espec o ans e ed in ellec ual
asse s). Es ablishmen o an app op ia e, aligned
in e na ional egula o y egime may also help
i ms o manage isks a ising om echnology
ans e by cla i ying esponsibili ies wi h espec o
sa e y, consume p o ec ion, secu i y, and p omo e
eliabili y and us wo hiness in AI, among o he
issues. Indeed, egula o y misalignmen en ails
cos s ha may inhibi echnology ans e (see
Box 2). In he absence o igh e in e na ional
coo dina ion, he e is a isk ha inconsis encies
may eme ge in he egula o y egime in e na ionally
and po en ially e en domes ically ( o example,
be ween ju isdic ions o minis ies).
To da e, much o he AI egula o y ac i i y has
been a he na ional le el as highligh ed in he
egula o y e iew sec ion abo e. The e iew
poin s o ac ions unde way in a a ie y o a eas
po en ially ele an o di usion o AI inno a ion.
Fo example, al eady as o 2021, se en o he
e iewed economies — i e sample economies
(Canada, China, he Eu opean Union, I eland
and he Uni ed S a es) plus wo non-sample
coun ies (B azil and Singapo e) — had policy
ini ia i es conce ning s anda ds and ce i ica ion
o echnology de elopmen and adop ion. While
in e na ional coope a ion has suppo ed some
pee lea ning and some con e gence ( o example,
among OECD membe s), some di e gences ha e
appea ed. Fo example, con en mode a ion
app oaches in he Uni ed S a es o he Eu opean
Union a e inconsis en wi h he con ols on
con en in China wi h espec o poli ical and
social s abili y. Among he i s 12 LLMs egis e ed
unde Chinese egula o y equi emen s in 2023, all
o igina ed wi h Chinese i ms and ins i u ions. On
he o he hand, ce ain o he non-sample coun ies
conside ed in he e iew such as Indonesia, Sou h
A ica o Vie nam appea o be lagging in hei AI
egula o y de elopmen in e ms o he numbe
o ac ions unde way. Such gaps in egula o y
egimes could inc ease unce ain y a ound he
u u e condi ions o AI businesses and po en ially
ha e an inhibi ing e ec on echnology ans e .
The na ional p o isions ha e been complemen ed
in some cases by AI- ele an in e na ional acco ds
ha may help o ensu e some con e gence in
egimes. Fo example, his can be he case in digi al
economy p o isions in some ade ag eemen s.
32 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
These can a ec ma e s such as limi ing ecou se
o da a localiza ion, p ohibi ing imposi ion o
cus oms du ies on da a ansmission and digi al
p oduc s, ensu ing ee c oss-bo de ans e o
pe sonal da a, equi ing p o ec ion o consume s’
pe sonal da a and equi ing cybe secu i y measu es
( o ins ance, his is he case wi h he CPTPP and
CUSMA) (Suominen 2021). The EU AI Ac goes u he
o s anda dize he handling o AI isks ac oss he EU
membe ship, es ablishing a egional AI egula o y
egime and egional AI ins i u ions o guide and
o e see he implemen a ion o he legisla ion in
coope a ion wi h na ional au ho i ies. Howe e ,
egional ade acco ds and AI go e nance egimes
do no p eclude in e - egional misalignmen . Fo
example, he EU AI Ac builds in ex a e i o iali y in
ha i applies o o eign supplie s o he EU ma ke .
Bu , his does no p eclude he possible eme gence
o egula o y inconsis ency wi h o he egimes
ha hose same supplie s may ace. A he same
ime, e en less binding commi men s such as join
communiqués issued by he G20 o G7 may p omo e
ce ain ypes o aligned AI policy ac ion among
membe s and ia askings o o he in e na ional
bodies. Follow-up may ake place h ough epo ing
p ocesses, so powe means such as subsequen
pee e iew and o he me hods ( o example, e en
mo al suasion among pa icipa ing leade s).
Success in imp o ing egula o y coope a ion
(see Box 3) can yield angible economic esul s.
The OECD (2021, 19) e iewed e idence on
he pe o mance o in e na ional egula o y
coope a ion in a ious ields and ound ha such
e o s deli e ed imp o emen s in h ee a eas in
pa icula : egula o y e ec i eness (especially o
challenges ha a e se in e na ional bounda ies);
economic e iciency ( o example, limi ing undue
ade ic ion); and adminis a i e e iciency ( o
example, domes ic egula o s may bene i om
in e na ional in elligence and insigh s). In ligh o
his pas expe ience including aspec s o he digi al
economy, i may well be ha imp o ed coope a ion
in he a ea o AI would yield simila ypes o esul s.
Wha Is o Be Done?
Recommenda ions
In iew o he pace o echnological change
in he AI sec o , his pape has examined he
concen a ion o AI inno a ion and conside ed
ac o s in he egula o y en i onmen ha may
suppo o hinde i s esponsible di usion.
D awing on he combined indings om he h ee
analy ical sec ions, he pape ecommends:
→Designa ion o an in e na ional ins i u ion o
lead on AI egula o y coope a ion: Replacing
he exis ing agmen ed app oach, na ional
go e nmen s — pe haps ia a join decision o
he G20 — may wish o es ablish a designa ed
AI body a he global le el o moni o egula o y
Box 2: Regula o y Misalignmen Can Be Cos ly
The OECD (2021, 22) p esen s a axonomy o ade cos s o p oduce s and ade s ope a ing
in e na ionally in cases o egula o y di e gence. These include in o ma ion cos s ela ed o
ob aining and p ocessing in o ma ion on egula o y equi emen s (whe eby mo e opaque and
complex sys ems en ail highe cos s); speci ica ion cos s o adjus p oduc s and se ices o
di e en equi emen s (wi h po en ial educ ions in economies o scale, highe labou and o he
inpu cos s); con o mi y assessmen cos s (demons a ing compliance wi h di e en equi emen s
may equi e addi ional cos ly es s, ce i ica ion, inspec ions, audi s); and o he cos s, including o
he go e nmen s adminis a ing he egime.
The bu dens o misalignmen do no all equally. Businesses in small economies could be a a
pa icula disad an age wi h espec o accessing models, so wa e applica ions and quali y
da a. Small businesses in hose economies could be disp opo iona ely a ec ed, as hey o en
lack capaci y o ack de ails o egula o y de elopmen s in o eign ma ke s. In addi ion o
misalignmen , in some economies, he u bulence o changing egula ion may c ea e unce ain y
ha unde mines incen i es o in es . Fo la ge economies such as China, he Eu opean Union
and he Uni ed S a es, he e may be su icien scale in he domes ic ma ke o con inue o uel AI
de elopmen e en in a agmen ed global AI economy.
39AI Inno a ion Concen a ion and he Go e nance Challenge
i sel p o ides o penal ies o in he mos se ious
cases wi hd awal o use o gene a i e AI se ices.
A eam om HSBC Global Resea ch ecen ly
conside ed he impac o Chinese egula ion on
he de elopmen o gene a i e AI in China (Liu
e al. 2023). The au ho s no e ha he egis a ion
sys em o algo i hms was ope a ional and ha a
i s ba ch o 12 LLMs had epo edly been app o ed
by egula o s in Augus 2023. All 12 o igina ed wi h
Chinese de elope s, including Alibaba’s Tongyi
Qianwen, Baidu’s E nie Bo and Tencen ’s Hunyuan.
The HSBC analys s no e ha hese i ms will ha e a
numbe o op ions o p o i able comme cializa ion
o hei AI capabili ies, such as AI compu ing
powe (capi alizing on la ge g aphics p ocessing
uni in en o ies o supply compu ing se ices o
clien s); AI gene a ed con en ; and model-as-a-
se ice ( making models a ailable o cus omizing
models o cus ome s). While Ame ican and o he
o eign i ms ha e a lead in such comme cializa ion
globally, Chinese i ms such as Baidu, Alibaba
and Tencen ( he so-called BAT) a e mo ing o
ca ch-up and will ha e a egula o y ad an age
in hei domes ic ma ke . HSBC es ima es ha
he gene a i e AI ma ke in China could each a
scale o US$10 billion by 2026 (Liu e al. 2023).
The Eu opean Union
The EU au ho i ies a e posi i ely inclined owa d
AI inno a ion and use, subjec o compliance wi h
exis ing egula ion and pending upda es mean
o p o ec use s and ensu e a con es able ma ke .
The EU ulebook is se o a signi ican upg ade.
On Decembe 8, 2023, he Eu opean Commission,
Council and Pa liamen ag eed on he e ms o a
landma k comp ehensi e law on AI, he EU A i icial
In elligence Ac . The legal ex is now being inalized
and will be submi ed o he Eu opean Council
and Pa liamen o o mal adop ion in he coming
mon hs (Council o he Eu opean Union 2023).
Once adop ed, he o icial ex o he EU AI Ac will
be published. I en e s in o o ce 20 days la e . The
applica ion o he p o isions will oll ou in phases
o e he nex h ee yea s. The p ohibi ion on models
posing unaccep able isks will apply six mon hs
a e he law’s en y in o o ce. Fi ms de eloping
gene al pu pose AI models will need o comply
wi hin 12 mon hs. The Eu opean Commission will
ha e 18 mon hs o p o ide p ac ical guidelines o
he classi ica ion o high- isk AI sys ems. All AI Ac
ules will be applicable wi hin 24 mon hs in mos
cases, al hough ope a o s o ce ain special pu pose
AI sys ems ( o example, c i ical in as uc u e;
mig a ion, asylum and bo de con ol managemen )
will ha e an addi ional yea o comply.
The EU AI Ac go e nance amewo k uses a
echnology-neu al, isk-based app oach a ge ing
a ange o use cases ( o example, see, Den ons
2023; Ho mann 2023). The new legisla ion
will complemen o he AI- ele an elemen s
al eady in place as pa o he Eu opean Union’s
app oach o go e nance in he digi al sphe e.
And, i has an elemen o ex a e i o iali y:
i will apply o p o ide s se ing he EU
ma ke , ega dless o whe e hey a e based.
The EU AI Ac de ines AI sys ems b oadly, d awing
on wo k om he OECD (Bambe g e al. 2024).
As o he ime o w i ing, he p oposed ex o
he EU AI Ac de ined he scope o he legisla ion
as being AI sys ems ha a e machine-based
and “designed o ope a e wi h a ying le els o
au onomy and ha may exhibi adap i eness
a e deploymen and ha , o explici o implici
objec i es, in e s, om he inpu i ecei es,
how o gene a e ou pu s such as p edic ions,
con en , ecommenda ions, o decisions ha can
in luence physical o i ual en i onmen s.”58 The
ac u he elabo a es ha i is no mean o co e
adi ional so wa e p og ams based on ules
designed by na u al pe sons o execu e speci ic
ope a ions and lacking he capaci y o in e .59
The pending AI Ac employs a ou - ie ed
isk s uc u e (Madiega 2023; Ho mann
2023; Den ons 2023; Heikkilä 2023; Bambe g
e al. 2024), wi h special supplemen al
p o isions o gene al pu pose AI models.
→ The AI Ac would explici ly ban AI p ac ices ha
pose an unaccep able isk o people’s sa e y,
li elihoods and igh s. Acco ding o he pending
legisla ion, his would include AI sys ems ha
deploy:
– ha m ul cogni i e beha iou al manipula ion
echniques;
– emo ional ecogni ion in he wo kplace o
educa ional ins i u ions (excep o medical
and sa e y easons);
58 This ex d aws on language om he ma ked-up legisla i e ex o he EU
AI Ac , a . 3(1), which was accessed ia Bambe g e al. (2024).
59 This ex d aws on language om he ma ked-up legisla i e ex o he EU
AI Ac , eci al 6, and accessed ia Bambe g e al. (2024).

40 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
– algo i hms o exploi speci ic ulne able
g oups, o example, hose wi h physical o
men al disabili y) and he eby cause ha m;
– algo i hms used by public au ho i ies, o on
hei behal , o social-sco ing pu poses and
he eby causing de imen al o un a ou able
ea men ;
– un a ge ed sc aping o acial images om he
in e ne o closed-ci cui ele ision oo age
o c ea e acial ecogni ion da abases;
– eal- ime, emo e biome ic iden i ica ion
in publicly accessible spaces o law
en o cemen (albei wi h excep ions o
ce ain g ie ous and u gen ma e s such as
h ea o e o is a ack);
– biome ic ca ego iza ion o in e sensi i e
da a such as sexual o ien a ion o eligious
belie s; and
– p edic i e policing o indi iduals based
solely on p o iling o pe sonali y ai s.
→ The AI Ac would igh ly egula e high-
isk AI sys ems ha could ad e sely a ec
sa e y o undamen al igh s. P o ide s and
deploye s o such sys ems will ace de ailed
and comp ehensi e obliga ions. Among o he
equi emen s, such sys ems would be subjec o
go e nance and echnical s ipula ions co e ing
such a eas as anspa ency, isk managemen ,
accoun abili y, da a go e nance, accu acy,
obus ness and cybe secu i y. Such high- isk
AI sys ems will need o be subjec o human
o e sigh and con ol. Da a o aining and use
in hese AI sys ems will need o be high quali y,
ele an and unbiased. In ligh o he isks, he
Eu opean Union will equi e egis a ion o
such AI sys ems in a special da abase. High- isk
sys ems will be u he ca ego ized in o wo
main g oups, Annex II and Annex III sys ems
( he la e will ha e an ex a yea o each
compliance, 36 mon hs in o al):
– Annex II sys ems a e used as a sa e y
componen o a p oduc ha is i sel isk-
p one, co e ed by EU law and subjec o a
equi ed con o mi y assessmen .
– Annex III sys ems se e a ious speci ic
pu poses: non-banned biome ics; c i ical
in as uc u e; educa ional and oca ional
aining including sys ems ela ed o
p og am access, e alua ion and moni o ing
o es ing; human esou ce managemen
in employmen ; access and use o essen ial
public and p i a e se ices; law en o cemen ;
mig a ions, asylum and bo de con ol
managemen ; adminis a ion o jus ice and
democ a ic p ocesses.
→ Limi ed- isk AI sys ems such as cha bo s
and gene a i e AI sys ems would be mo e
ligh ly egula ed bu s ill be subjec o a se o
anspa ency obliga ions. Use s will need o be
in o med o hei in e ac ions wi h such an AI
sys em and he limi a ions o such sys ems. AI-
gene a ed con en would need o be lagged.
→ Low- and minimal- isk AI sys ems ( o example,
spam il e s o ideo games) could be de eloped
and used eely in he Eu opean Union, subjec
o o he applicable laws. Howe e , he EU AI
Ac en isages c ea ion o codes o conduc o
p o ide s o limi ed- and low- isk sys ems o
guide hem owa d olun a y compliance wi h
highe AI s anda ds.
Gene al-pu pose AI models ( ounda ion models,
LLMs) will be subjec o u he obliga ions o e
hei li e cycle (Bambe g e al. 2024). These include
main aining cu en echnical documen a ion
o he model (including aining and es ing
p ocess, wi h e alua ion esul s); anspa ency on
p ope ies o he model o downs eam sys em
p o ide s; de elopmen o a policy o compliance
wi h copy igh law; and a de ailed summa y
o he aining da a employed in he model’s
de elopmen . In cases o sys emic isk, gene al-
pu pose AI model de elope s mus implemen
isk assessmen and mi iga ion measu es, as well
as inciden esponse and epo ing p ocedu es.
In addi ion, he EU AI Ac es ablishes an
ins i u ional amewo k o suppo and en o ce he
ope a ion o he legisla ion. Each membe s a e
mus designa e a leas one na ional compe en
au ho i y o supe ise he AI Ac implemen a ion
and ack AI ma ke de elopmen s. A new Eu opean
A i icial In elligence Boa d, wi h one na ional
ep esen a i e pe membe , will o e see his
policy a ea a he EU le el. A new EU AI O ice
wi hin he Eu opean Commission is asked wi h
o e sigh o he gene al pu pose AI models egime,
suppo ing na ional au ho i ies, and ce ain aspec s
o en o cemen . A scien i ic panel o independen
expe s will suppo he EU ins i u ions in his wo k.
41AI Inno a ion Concen a ion and he Go e nance Challenge
Wi h espec o en o cemen , he legisla ion
speci ies subs an ial penal ies ha depend
on he se e i y o he in ingemen . These
a y be ween 1.5 pe cen and 7.0 pe cen o
annual u no e o a i m, subjec o ce ain
minimum and maximum h esholds.
As o o he AI- ele an EU legisla ion ha
would ope a e in conjunc ion wi h he AI Ac ,
a ew illus a i e examples a e ci ed below:
→ Gi en he expansi e exploi a ion o da a h ough
AI echnology, he EU GDPR will be qui e
ele an . The GDPR empowe s da a subjec s o
exe some con ol o e he use o hei pe sonal
da a ia consen equi emen s and equi es
some p i acy p o ec ions, pa icula ly in ela ion
o p o iling and decision making. As wi h any EU
business, AI p o ide s mus adhe e o p inciples
such as da a minimiza ion, pu pose limi a ion
and s o age limi a ion when collec ing and
p ocessing pe sonal da a. On he o he hand,
he GDPR does include allowances o da a use
o s a is ical and scien i ic esea ch pu poses. A
Eu opean Pa liamen s udy ound ha he GDPR
“can be in e p e ed and applied in such a way
ha i does no hinde bene icial applica ion o
AI o pe sonal da a, and ha i does no place EU
companies a a disad an age in compa ison wi h
non-Eu opean compe i o s” ( hough a a ie y
o speci ic issues s ill equi e cla i ica ion)
(Sa o and Lagioia 2020, 79–80). GDPR adequacy
de e mina ions p o ide a deg ee o in e na ional
access and po abili y wi h espec o EU da a o
ce ain i ms based in a lis o egis e ed pa ne
economies. As o Janua y 15, 2024, 15 economies
a e co e ed by hese decisions including leading
AI hos coun ies such as Canada, Is ael, Japan,
Sou h Ko ea, he Uni ed Kingdom and he Uni ed
S a es.60
→ The EU Cybe secu i y Ac es ablished a
cybe secu i y amewo k o p oduc s and
se ices, ope a ing unde he EU Agency o
Cybe secu i y, aiming o ha monize EU-wide
cybe secu i y ce i ica ion o in o ma ion
echnology p oduc s, se ices and p ocesses. As
wi h o he digi al sys ems, AI sys ems would be
co e ed.
→ The Digi al Se ices Ac (2022) imposes ie ed
obliga ions o online ma ke places and sea ch
60 See h ps://commission.eu opa.eu/law/law- opic/da a-p o ec ion/
in e na ional-dimension-da a-p o ec ion/adequacy-decisions_en.
engines, as well as p o ec ions o mino s.
Obliga ions include, among o he s, anspa ency
in a eas such as algo i hmic decision making and
ecommenda ions, and o e y la ge p o ide s,
due diligence wi h espec o annual assessmen
o sys emic isks and egula o y access o
da abases, algo i hms and p emises (Beck and
Wo m 2023).
→ I an AI ounda ional model pla o m61 we e o
a ain su icien scale, o example, he Digi al
Ma ke Ac (2022) p o isions migh be in oked
wi h consequences o AI sys em p o ide s.
Requi emen s could include hi d-pa y
in e ope abili y, limi s on he use o pe sonal
da a and cons ain s on p io i izing ankings
o own p oduc s o se ices, among o he
possibili ies ha aim o es o e ai ness and
con es abili y in he ma ke (Yasa e al. 2023).
→ The Digi al Go e nance Ac (2022), which
aims o acili a e access o publicly held da a
while main aining p i acy and con iden iali y
p o ec ions.62
→ The Un ai Comme cial P ac ices Di ec i e
(2005), as upda ed by subsequen guidance.63
The di ec i e and guidance go e n business-
o-consume ansac ions and ep esen a ions
made by endo s and o he s akeholde s ( o
example, p o ide s o compa ison ools). The
guidance add esses (wi h explici e e ences
o AI) such issues as acking and a ge ing
echnologies and algo i hmic pe sonaliza ion,
among o he isks.
The Uni ed S a es
As in o he coun ies, AI in he Uni ed S a es is
he subjec o an eme ging egula o y amewo k.
This is no o say ha i has been un egula ed.
A he ede al le el, he US go e nmen included
some measu es o de ine AI o add ess speci ic
conce ns such as AI applica ions in de ence o
anspo a ion sys ems. A de ence app op ia ions
61 AI ounda ional models a e la ge-scale and in ended o be used as a
pla o m o which de elope s could add addi ional unc ionali y.
62 See h ps://digi al-s a egy.ec.eu opa.eu/en/policies/da a-go e nance-
ac -explained.
63 EC, Guidance on he in e p e a ion and applica ion o Di ec i e
2005/29/EC o he Eu opean Pa liamen and o he Council conce ning
un ai business- o-consume comme cial p ac ices in he in e nal ma ke
(see “Digi al sec o ” a §4.2), [2021] OJ C 256/1; online:
h ps://eu -lex.eu opa.eu/legal-con en /EN/TXT/
PDF/?u i=CELEX:52021XC1229(05).
42 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
ac in 2020 legisla ed he c ea ion o he Na ional
A i icial In elligence Ini ia i e O ice in he Whi e
House o o e see he US na ional AI s a egy. (A
numbe o US s a es ha e also legisla ed measu es
pe aining o AI, as no ed in DPA [2023]).
Ope a ions in he AI sec o ha e also been subjec
o p io exis ing egula o y p o isions applicable
mo e gene ally in he economy. Exis ing domes ic
law and egula ion con inued o apply, e en in
he AI sec o . Fo example, a ecen Cong essional
Resea ch Se ice e iew o copy igh p o ec ion in
he sec o ound ha exis ing copy igh law may
p o e adequa e and ha addi ional expe ience wi h
cou challenges would be equi ed o highligh
any speci ic changes in he legal amewo k ha
migh be equi ed wi h espec o gene a i e AI
(Zi poli 2023). As he US Fede al T ade Commission
Chai Khan (2023) pu i in ano he con ex , “The e
is no AI exemp ion o he laws on he books.”
Ye , he US app oach o AI egula ion has been
g adually changing in ecen yea s, s a ing wi h
ha dwa e. In he compe i ion be ween China and
he Uni ed S a es o AI leade ship, he Uni ed S a es
has ocused on pe cei ed secu i y isks, mo ing o
mo e igh ly limi access o ad anced AI- ela ed
echnology o Ame ican o igin. By means o an
Execu i e O de da ed Sep embe 15, 2022 (Biden
2022), P esiden Biden exp essly manda ed ha he
Commi ee on Fo eign In es men in he Uni ed
S a es in conduc ing in es men e iews should
conside supply chain esilience and secu i y wi h
espec o AI and ela ed a eas o mic o-elec onics,
among o he a eas. Then, on Oc obe 7, 2022 (Bu eau
o Indus y and Secu i y [BIS] 2022), he Uni ed
S a es imposed con ols on expo s o China o
ce ain ad anced compu ing chips and equipmen
used o manu ac u e semiconduc o s, including
ce ain o eign-p oduced i ems om hi d coun ies
inco po a ing such ad anced US echnologies. The
public no ice (ibid., 1) jus i ied he es ic ions on
na ional secu i y and human igh s g ounds and
no ed he po en ial mili a y applica ions o he
co e ed i ems, as well as uses ela ed o in elligence
and secu i y se ices. These s eps wen hand-
in-hand wi h he CHIPS and Science Ac passed
ea lie in 2022 by he US Cong ess, which deli e ed
emendous suppo o he US semiconduc o
sec o (US$52.7 billion o manu ac u ing incen i es,
R&D, wo k o ce de elopmen and mo e) and
explici ly e e enced i s suppo o US leade ship in
he AI sec o as an objec i e (Whi e House 2022b).
Also in Oc obe 2022, he Whi e House eleased i s
“Bluep in o an AI Bill o Righ s” (Whi e House
2022a). The documen se ou i e ounda ional
p inciples o use in de elopmen o a amewo k
o AI go e nance. These p inciples a e mean o
guide he design, use and deploymen o au oma ed
sys ems. They co e p o ec ion om unsa e and
ine ec i e sys ems, p o ec ion om algo i hmic
disc imina ion, p o ec ion om abusi e da a
p ac ices (while ensu ing use s ha e agency o e
da a uses), use no ice and explana ion o AI use and
impac s, and he abili y o op ou and ha e access
o a human ( o example, o emedy p oblems).
In o de o ge a be e iew o de elopmen s and
conce ns wi h espec o AI echnology, membe s o
he US adminis a ion held s uc u ed discussions
wi h AI s akeholde s du ing he i s hal o 2023.
D awing on his inpu and c oss- e e encing he
“Bluep in o an AI Bill o Righ s” and o he ecen
adminis a i e ac ions conce ning AI, he Whi e
House p oceeded o secu e olun a y commi men s
om se en leading AI i ms o ensu e sa e y,
secu i y and us in de eloping AI (Whi e House
2023b). Amazon, An h opic, Google, In lec ion,
Me a, Mic oso and OpenAI signed on as o
July21, 2023. In Sep embe 2023, eigh u he i ms
signed on: Adobe, Cohe e, IBM, N idia, Palan i ,
Sales o ce, Scale AI and S abili y (Mi chell 2023).
The key ea u es o he eigh US olun a y
commi men s (Whi e House 2023b) a e:
→ in e nal and independen ex e nal secu i y
es ing o hei AI sys ems be o e hei elease;
→ sha ing in o ma ion ac oss he indus y and
wi h go e nmen s, ci il socie y and academia on
managing AI isks ( o example, bes p ac ices
o sa e y);
→ in es ing in cybe secu i y and inside h ea
sa egua ds o p o ec p op ie a y and un eleased
model weigh s;
→ acili a ing ongoing hi d-pa y disco e y and
epo ing o ulne abili ies in hei AI sys ems;
→ de eloping obus echnical mechanisms o
ensu e ha use s know when con en is AI
gene a ed, such as a wa e ma king sys em;
→ publicly epo ing hei AI sys ems’ capabili ies,
limi a ions and a eas o app op ia e and
inapp op ia e use;
43AI Inno a ion Concen a ion and he Go e nance Challenge
→ p io i izing esea ch on he socie al isks ha AI
sys ems can pose, including on a oiding ha m ul
bias and disc imina ion, and p o ec ing p i acy;
and
→ ag eeing o de elop and deploy ad anced AI
sys ems o help add ess socie y’s g ea es
challenges ( o example, cance p e en ion and
mi iga ing clima e change).
The Uni ed S a es announced a u he igh ening
o expo con ols on Oc obe 17, 2023 (BIS 2023).
Res ic ions we e expanded o co e a g ea e
ange o ad anced compu ing semiconduc o s
ha could po en ially “enable he de elopmen
and p oduc ion o echnologies such as a i icial
in elligence (AI) used in mili a y applica ions.”
The public no ice s a es (ibid., 2), “These con ols
we e s a egically c a ed o add ess, among o he
conce ns, he PRC’s [People’s Republic o China’s]
e o s o ob ain semiconduc o manu ac u ing
equipmen essen ial o p oducing ad anced
in eg a ed ci cui s needed o he nex gene a ion
o ad anced weapon sys ems, as well as high-end
ad anced compu ing semiconduc o s necessa y
o enable he de elopmen and p oduc ion o
echnologies such as a i icial in elligence (AI)
used in mili a y applica ions.” The es ic ions
conce n he 22 coun ies o which he Uni ed S a es
main ains an a ms emba go, including China.
As no ed by HSBC analys F ank He (2023, 1), “The
a ec ed p oduc co e age is wide han ma ke
expec a ions, which may cause a ma e ial impac
on China’s AI in as uc u e supply chain as we
es ima e ha he chips ha a e newly added o
he con ol lis cu en ly suppo o e 90% o
AI aining wo kloads in China.” The immedia e
dis up ion will be limi ed as Chinese impo e s
an icipa ed po en ial dis up ion and had buil up
in en o ies o equisi e chips. Also, i is no able
ha Chinese domes ic supplie s such as Huawei
and Hygon ha e imp o ed hei chip o e ings
and can supply iable subs i u es o some chips.
While He (ibid.) ecognizes ha he US expo
con ols will incen i ize u he de elopmen in he
domes ic sec o , he a gues, howe e , ha a mul i-
yea echnology gap may pe sis in sys em-le el
enginee ing and sees ha as a likely cons ain on
he AI sec o in China. A eas o conce n include, o
example, chip- o-chip in e connec ion, so wa e
ecosys ems and ad anced node chip ab ica ions.
The US go e nmen ook a subs an ial s ep o
s eng hen i s AI go e nance wi h he “Execu i e
O de (EO) on Sa e, Secu e, and T us wo hy
A i icial In elligence” issued by P esiden Biden
on Oc obe 30, 2023 (Biden 2023; Whi e House
2023c). The aim was o ad ance a coo dina ed,
ede al go e nmen -wide app oach o posi ion
he Uni ed S a es o esponsible leade ship in AI
de elopmen while add essing AI isks o sa e y
and secu i y. The amewo k seeks o capi alize
and p omo e bene i s om de elopmen o AI,
while also aking ca e o mi iga e ha ms. In iew
o pa isan blockage in he US Cong ess, he
Execu i e O de exploi s au ho i ies a ailable o
he p esiden unde exis ing law. S ill, his ac ion is
ela i ely expansi e conside ing he cons ain s.
The Execu i e O de di ec s ha execu i e agencies
use hei cu en powe s o de elop s anda ds,
ools and es s o ensu e ha AI sys ems a e
sa e, secu e and us wo hy. The app oach is isk
based, wi h special epo ing measu es included
co e ing de elope s o powe ul ounda ion
models. I akes s eps o add ess isks conce ning
p i acy, wo kplace and o he disc imina ion, and
labou ma ke disloca ions. The ini ia i e includes
measu es o suppo esponsible AI use in heal h
ca e and educa ion. The Execu i e O de aims o
boos AI- ela ed human capi al by suppo ing AI
esea ch, acili a ing ele an skilled immig a ion
and expansion o ede al s a ing o AI p o essionals.
Adminis a i e measu es would acili a e amping
up use o AI in ede al agencies. Openness would
be cul i a ed ia suppo o small business,
in e na ional collabo a ion and deploymen o
AI in mee ing challenges global challenges. The
documen emphasizes he goal o de eloping and
implemen ing in e na ional s anda ds o AI.
On No embe 21, 2023, he US Fede al T ade
Commission (FTC) (2023) ook a s ep mo ing
u he beyond he ea ly olun a y app oaches
o AI go e nance. Unde he FTC Ac and o he
laws, he FTC is conce ned wi h aud, decep ion,
in ingemen s on p i acy, un ai comme cial
p ac ices and compe i ion issues, including
ins ances in ol ing AI. Consequen ly, he FTC
au ho ized he use o a compulso y p ocess64 in
in es iga ions ela ed o p oduc s and se ices
p oduced o used in connec ion wi h AI. Recipien s
64 The e m “compulso y p ocess” e e s o a ious ypes o in o ma ion
o documen eques s — including bu no limi ed o subpoenas, ci il
in es iga i e demands, and o de s o special epo s — whe e he
ecipien o he eques may be compelled o comply wi h he eques by
o de o cou .
44 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
o eques s o in o ma ion o documen s may be
legally compelled o comply by a cou o de .
As o No embe 2023, he US Adminis a ion’s
O ice o Managemen and Budge has published
o public consul a ion a d a memo andum wi h
u he conc e e s eps o p omo e AI go e nance,
inno a ion and isk managemen in agencies
o he ede al go e nmen (Young 2023). Wi h
he excep ion o na ional secu i y agencies, he
memo andum would equi e each agency o
designa e a senio -le el chie AI o ice , de elop
an AI compliance plan co esponding o he
memo andum, in en o y he agency’s AI sys ems
and isks (emphasizing sa e y and igh s impac s),
iden i y po en ial esponsible uses o AI and emo e
ba ie s o deli e ing hose sys ems. Non-complian
AI sys ems a e o be e mina ed by Augus 1, 2024.
The memo andum iden i ies speci ic issues o
conside a ion wi h espec o each manda e, wi h
miles ones, s anda ds, and ela i ely igh imelines
o comple ion (measu ed in days and mon hs o
ini ial ac ions and hen o ongoing equi emen s,
depending on he ac ion, annually o biennially).
OECD Recommenda ion o he
Council on A i icial In elligence
The OECD p inciples on AI wi hin he
OECD’s “Recommenda ion o he Council
on A i icial In elligence”65 s a e ha :
→ AI should bene i people and he plane
by d i ing inclusi e g ow h, sus ainable
de elopmen and well-being.
→ AI sys ems should be designed in a way
ha espec s he ule o law, human igh s,
democ a ic alues and di e si y, and hey
should include app op ia e sa egua ds — o
example, enabling human in e en ion whe e
necessa y— o ensu e a ai and jus socie y.
→ The e should be anspa ency and esponsible
disclosu e a ound AI sys ems o ensu e ha
people unde s and when hey a e engaging wi h
hem and can challenge ou comes.
65 See OECD (2019). Also no e ha as o No embe 3, 2023, he OECD
Legal Ins umen s online da abase epo s ha adhe ence o he
ecommenda ion now co e s all 38 OECD membe coun ies, plus eigh
o he s: A gen ina, B azil, Egyp , Mal a, Pe u, Romania, Singapo e and
Uk aine.
→ AI sys ems mus unc ion in a obus , secu e
and sa e way h oughou hei li e imes, and
po en ial isks should be con inually assessed
and managed.
→ O ganiza ions and indi iduals de eloping,
deploying o ope a ing AI sys ems should be
held accoun able o hei p ope unc ioning in
line wi h he abo e p inciples.
The OECD (2019) ecommends ha go e nmen s:
→ acili a e public and p i a e in es men in R&D
o spu inno a ion in us wo hy AI;
→ os e accessible AI ecosys ems wi h
digi al in as uc u e and echnologies, and
mechanisms o sha e da a and knowledge;
→ c ea e a policy en i onmen ha will open he
way o deploymen o us wo hy AI sys em;
→ equip people wi h he skills o AI and suppo
wo ke s o ensu e a ai ansi ion; and
→ coope a e ac oss bo de s and sec o s o sha e
in o ma ion, de elop s anda ds and wo k
owa ds esponsible s ewa dship o AI.

45AI Inno a ion Concen a ion and he Go e nance Challenge
Wo ks Ci ed
Aa onson, Susan A iel. 2024. Da a Disquie : Conce ns
abou he Go e nance o Da a o Gene a i e AI.
CIGI Pape No. 290. Wa e loo, ON: CIGI.
www.cigionline.o g/publica ions/da a-disquie -conce ns-
abou - he-go e nance-o -da a- o -gene a i e-ai/.
Aa onson, Susan A iel and Pa ick Leblond. 2018. “Ano he
Digi al Di ide: The Rise o Da a Realms and i s Implica ions
o he WTO.” Jou nal o In e na ional Economic Law
21(2): 245–72. h ps://doi.o g/10.1093/jiel/jgy019.
Bambe g, Da id, Ma in P luege , Nicole Sau in, S e an Schuppe ,
Jaspe Siems, Dan Whi ehead, Edua do Us a an and
Leopold on Ge lach. 2024. “The EU AI Ac : an impac
analysis (pa 1).” Engage, Feb ua y 1. Hogan Lo ells.
www.engage.hoganlo ells.com/knowledgese ices/
news/ he-eu-ai-ac -an-impac -analysis-pa -1/.
Beck, Benjamin and Ul ich Wo m. 2023. “EU Digi al Se ices Ac ’s
E ec s on Algo i hmic T anspa ency and Accoun abili y.” Maye
B own, Ma ch 27. www.maye b own.com/en/pe spec i es-
e en s/publica ions/2023/03/eu-digi al-se ices-ac s-
e ec s-on-algo i hmic- anspa ency-and-accoun abili y.
Bende , Emily M., Timni Geb u, Angelina McMillan-Majo
and Shma ga e Shmi chell. 2021. “On he Dange s o
S ochas ic Pa o s: Can Language Models Be Too Big?” In
FAcc 2021: P oceedings o he 2021 ACM Con e ence
on Fai ness, Accoun abili y, and T anspa ency, 610–23.
New Yo k, NY: Associa ion o Compu ing Machine y.
h ps://doi.o g/10.1145/3442188.3445922.
Biden, Joseph R. 2022. “Ensu ing Robus Conside a ion o
E ol ing Na ional Secu i y Risks by he Commi ee on
Fo eign In es men in he Uni ed S a es.” Execu i e O de
14083, Sep embe 15. P esiden ial Documen . Fede al
Regis e (87) 181, Sep embe 20. www.go in o.go /
con en /pkg/FR-2022-09-20/pd /2022-20450.pd .
———. 2023. “Execu i e O de on he Sa e, Secu e, and T us wo hy
De elopmen and Use o A i icial In elligence.” The Whi e
House B ie ing Room, Oc obe 30. www.whi ehouse.go /
b ie ing- oom/p esiden ial-ac ions/2023/10/30/
execu i e-o de -on- he-sa e-secu e-and- us wo hy-
de elopmen -and-use-o -a i icial-in elligence/.
BIS. 2022. “Comme ce Implemen s New Expo Con ols on
Ad anced Compu ing and Semiconduc o Manu ac u ing I ems
o he People’s Republic o China (PRC).” BIS, US Depa men
o Comme ce. Oc obe 7. www.bis.doc.go /index.php/
documen s/abou -bis/news oom/p ess- eleases/3158-
2022-10-07-bis-p ess- elease-ad anced-compu ing-
and-semiconduc o -manu ac u ing-con ols- inal/ ile.
———. 2023. “Comme ce S eng hens Res ic ions on
Ad anced Compu ing Semiconduc o s, Semiconduc o
Manu ac u ing Equipmen , and Supe compu ing I ems o
Coun ies o Conce n.” BIS, US Depa men o Comme ce.
Oc obe 17. www.bis.doc.go /index.php/documen s/
abou -bis/news oom/p ess- eleases/3355-2023-10-
17-bis-p ess- elease-acs-and-sme- ules- inal-js/ ile.
B ans e e , Lee, C. F i z Foley and Kamal Saggi. 2010. “Has he Shi
o S onge In ellec ual P ope y Righ s P omo ed Technology
T ans e , FDI, and Indus ial De elopmen ?” The WIPO Jou nal
2 (1): 93–98. h ps:// ind.wipo.in / eco d/28665.
B eaux, Co y and Emin Dinle soz. 2023. “How Many U.S. Businesses
Use A i icial In elligence?” US Census Bu eau. No embe 28.
www.census.go /lib a y/s o ies/2023/11/businesses-
use-ai.h ml?u m_campaign=20231128msacos1ccs o s&u m_
medium=email&u m_sou ce=go deli e y.
B emme , Ian and Mus a a Suleyman. 2023. “The AI Powe Pa adox:
Can S a es Lea n o Go e n A i icial In elligence — Be o e I ’s
Too La e?” Fo eign A ai s 102 (5): 26–43.
www. o eigna ai s.com/wo ld/a i icial-in elligence-
powe -pa adox.
Buildd. 2022. “By eDance — The wo ld’s mos aluable s a up!”
Ap il 13. Bengalu u, India: Buildd.co. h ps://buildd.co/
p oduc /by edance-wo lds-mos - aluable-s a up.
CAC. 2021. “P o isions on he Managemen o Algo i hmic
Recommenda ions in In e ne In o ma ion Se ices.”
CAC, Minis y o Indus y and In o ma ion Technology,
Minis y o Public Sa e y, S a e Adminis a ion o Ma ke
Regula ion. T ansla ion by China Law T ansla e, 2022.
www.chinalaw ansla e.com/en/algo i hms/.
———. 2022. “P o isions on he Adminis a ion o Deep
Syn hesis In e ne In o ma ion Se ices.” CAC, Minis y
o Indus y and In o ma ion Technology, Minis y o
Public Sa e y. T ansla ion by China Law T ansla e, 2022.
www.chinalaw ansla e.com/en/deep-syn hesis/.
———. 2023. “Measu es o he Managemen o Gene a i e A i icial
In elligence Se ices (D a o Commen ).” Cybe space
Adminis a ion o China (CAC), Minis y o Science and
Technology, and he Minis y o Public Secu i y. T ansla ion by
Sea on Huang, Helen Tone , Zac Haluza, Rogie C eeme s
and G aham Webs e , a DigiChina, S an o d Uni e si y,
2023. h ps://digichina.s an o d.edu/wo k/ ansla ion-
measu es- o - he-managemen -o -gene a i e-a i icial-
in elligence-se ices-d a - o -commen -ap il-2023/.
Ca azos Cepeda, Rica do H. and Douglas C. Lippold .
2010. “The S eng hening o IPR P o ec ion: Policy
Complemen s.” The WIPO Jou nal 2 (1): 99–122.
h ps:// ind.wipo.in / eco d/28665.
46 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
CB Insigh s. 2023a. “Global Unico n Club.” Online
da abase. Oc obe 19. www.cbinsigh s.com/ esea ch/
epo /unico n-s a ups- alua ions-headcoun -
in es o s/?u m_sou ce=CB+Insigh s+Newsle e &u m_
campaign=b c39b1e66-newsle e _gene al_
hu s_2023_10_26&u m_medium=email&u m_
e m=0_9dc0513989-b c39b1e66-90194953.
———. 2023b. “The open-sou ce AI de elopmen
ma ke map.” Resea ch b ie , CB Insigh s Resea ch.
Oc obe 13. www.cbinsigh s.com/ esea ch/
open-sou ce-ai-de elopmen -ma ke -map/.
Chen, Hailin, Fangkai Jiao, Xingxuan Li, Chengwei Qin, Ma hieu
Ra au , Ruochen Zhao, Caiming Xiong and Sha iq Jo y.
2023. “Cha GPT’s One-yea Anni e sa y: A e Open-Sou ce
La ge Language Models Ca ching up?” a Xi , No embe 29.
h ps://doi.o g/10.48550/a Xi .2311.16989.
Ch is ensen, Clay on M. 2016. The Inno a o ’s Dilemma:
When New Technologies Cause G ea Fi ms o Fail.
Bos on, MA: Ha a d Business Re iew P ess.
Chui, Michael, B yce Hall, Helen Mayhew, Alex Singla and Alex
Sukha e sky. 2022. “The s a e o AI in 2022 — and a hal
decade in e iew.” McKinsey & Company. Decembe 6.
www.mckinsey.com/capabili ies/quan umblack/ou -insigh s/
he-s a e-o -ai-in-2022-and-a-hal -decade-in- e iew#/.
Ciu iak, Dan and Anna A yushina. 2023. “Canada Needs an
A i icial In elligence Agency.” Opinion, Cen e o In e na ional
Go e nance Inno a ion, Oc obe 5. www.cigionline.o g/
a icles/canada-needs-an-a i icial-in elligence-agency/.
Council o he Eu opean Union. 2023. “A i icial in elligence ac :
Council and Pa liamen s ike a deal on he i s ules
o AI in he wo ld.” P ess elease, Decembe 9.
www.consilium.eu opa.eu/en/p ess/p ess-
eleases/2023/12/09/a i icial-in elligence-ac -council-and-
pa liamen -s ike-a-deal-on- he- i s -wo ldwide- ules- o -ai/.
Den ons. 2023. “The New EU AI Ac – he 10 key hings you
need o know now.” Decembe 14. www.den ons.com/
en/insigh s/a icles/2023/decembe /14/ he-new-
eu-ai-ac - he-10-key- hings-you-need- o-know-now.
DPA. 2023. “Ac i i y T acke : ML and AI De elopmen .”
DPA websi e. Viewed on No embe 18, 2023.
h ps://digi alpolicyale .o g/ac i i y- acke ?o se =0&l
imi =10&ac i i y=9&pe iod=2019-12-31,2023-11-17.
DSIT. 2023. “Eme ging P ocesses o F on ie AI Sa e y.”
Con e ence Backg ound Pape , AI Sa e y Summi . DSIT,
Uni ed Kingdom, Oc obe . h ps://asse s.publishing.
se ice.go .uk/media/653aabbd80884d000d 71bdc/
eme ging-p ocesses- on ie -ai-sa e y.pd .
Ede e , Flo ian and B uno Pelleg ino. 2023. “The G ea
S a up Sellou and he Rise o Oligopoly.”
AEA Pape s and P oceedings 113 (May): 274–78.
h ps://pape s.ss n.com/sol3/pape s.
c m?abs ac _id=4322834.
Epic Games. 2023. “In oducing he NoAI and C ea edWi hAI ags.”
Cus ome Help, Feb ua y 20.
h ps://ma ke placehelp.epicgames.com/s/
a icle/In oducing- he-NoAI-and-
C ea edWi hAI- ags?language=en_US.
Eu opean Pa liamen . 2024. “A i icial In elligence Ac : MEPs
adop landma k law.” P ess elease, Ma ch 13.
www.eu opa l.eu opa.eu/news/en/
p ess- oom/20240308IPR19015/a i icial-
in elligence-ac -meps-adop -landma k-law.
Fa hi, Eli and Pe e MacKinnon. 2023. “AI Gua d ails A e
U gen ly Needed, bu Tha ’s Jus a S a .” Opinion,
Cen e o In e na ional Go e nance Inno a ion,
Oc obe 11. www.cigionline.o g/a icles/ai-gua d ails-
a e-u gen ly-needed-bu - ha s-jus -a-s a /.
Fay, Robe and Rohin on Medho a. 2021. “A Global Go e nance
F amewo k o Digi al Technologies.” T20 I aly 2021, Task
Fo ce 4, Policy B ie 5. www. 20i aly.o g/2021/08/26/a-
global-go e nance- amewo k- o -digi al- echnologies/.
Fe guson, Ch is ophe and Julie He. 2023. “China’s New Rules
Fo Gene a i e AI: An Eme ging Regula o y F amewo k.”
Fasken Technology, Media and Telecommunica ions Bulle in.
Augus 29. www. asken.com/en/knowledge/2023/08/
chinas-new- ules- o -gene a i e-ai#au ho s.
FTC. 2023. “FTC Au ho izes Compulso y P ocess o AI- ela ed
P oduc s and Se ices.” P ess elease, No embe 21.
www. c.go /news-e en s/news/p ess- eleases/2023/11/
c-au ho izes-compulso y-p ocess-ai- ela ed-p oduc s-se ices.
Fujian. 2021. “Ou line o he 14 h Fi e-Yea Plan (2021-2025) o
Na ional Economic and Social De elopmen and Vision 2035
o he People’s Republic o China.” The People’s Go e nmen o
Fujian P o ince. Sep embe 8. www. ujian.go .cn/
english/news/202108/ 20210809_5665713.h m.
GOV.UK. 2023. “The Ble chley Decla a ion by Coun ies A ending
he AI Sa e y Summi . No embe 1–2.” Policy Pape .
No embe 1. www.go .uk/go e nmen /publica ions/
ai-sa e y-summi -2023- he-ble chley-decla a ion/
he-ble chley-decla a ion-by-coun ies-a ending-
he-ai-sa e y-summi -1-2-no embe -2023.
G assano, Nicola, Héc o He nandez Gue a a, Pé e Fako,
Elisabe h Nindl, Aliki Geo gakaki, Ela Ince, Lo enzo
Napoli ano, F ancesco Ren occhini and Alexande
47AI Inno a ion Concen a ion and he Go e nance Challenge
Tuebke. 2022. “The 2022 EU Indus ial R&D In es men
Sco eboa d.” Luxembou g: Publica ions O ice o
he Eu opean Union. h ps://i i.j c.ec.eu opa.eu/
sco eboa d/2022-eu-indus ial- d-in es men -sco eboa d.
Ha zius, Jan, Dominic Wilson, Joseph B iggs, Vickie Chang, De esh
Kodnani and Gio anni Pie domenico. 2023. “Mac o Ou look
2024: The Ha d Pa is O e .” (Redac ed e sion.) Goldman
Sachs. No embe 8. www.goldmansachs.com/
in elligence/pages/gs- esea ch/mac o-ou look-
2024- he-ha d-pa -is-o e / epo .pd .
He, F ank. 2023. “China Technology — AI In as uc u e:
Looking o Al e na i es.” Equi y Resea ch Repo .
HSBC Global Resea ch. Oc obe 19.
Heikkilä, Melissa. 2023. “Fi e hings you need o know abou he
EU’s new AI Ac .” MIT Technology Re iew, Decembe 11.
www. echnology e iew.com/2023/12/11/1084942/
i e- hings-you-need- o-know-abou - he-eus-new-ai-ac /.
Ho, Lewis, Joslyn Ba nha , Robe T age , Yoshua Bengio,
Miles B undage, Allison Ca negie, Rumman Chowdhu y
e al. 2023. “In e na ional Ins i u ions o Ad anced AI.”
a Xi , July 11. h ps://a xi .o g/abs/2307.04699.
Ho mann, Mia, 2023. “The EU AI Ac : A P ime .” CSET Blog,
Sep embe 26. h ps://cse .geo ge own.edu/
a icle/ he-eu-ai-ac -a-p ime / .
IFAC-BIAC. 2018. “Regula o y Di e gence: Cos s, Risks, Impac s.”
www.i ac.o g/_ lysys em/azu e-p i a e/publica ions/
iles/IFAC-OECD-Regula o y-Di e gence.pd .
Inno a ion, Science and Economic De elopmen . 2023.
“The A i icial In elligence and Da a Ac (AIDA) –
Companion documen .” Go e nmen o Canada. Ma ch 13.
h ps://ised-isde.canada.ca/si e/inno a ion-be e -canada/
en/a i icial-in elligence-and-da a-ac -aida-
companion-documen .
Jones, Ellio . 2023. “Explaine : Wha is a ounda ion model?” Ada
Lo elace Ins i u e. July 17. www.adalo elaceins i u e.o g/
esou ce/ ounda ion-models-explaine .
Kaye, Ka e. 2022. “Will na ionalism end he golden age o
global AI collabo a ion?” P o ocol, No embe 1.
h ps://web.a chi e.o g/web/20221101100423/www.
p o ocol.com/en e p ise/china-us-ai-open-sou ce.
Khan, Lina M. 2023. “S a emen o Chai Lina M. Khan Rega ding he
Join In e agency S a emen on AI.” FTC, Ap il 23.
www. c.go /sys em/ iles/ c_go /pd / ema ks-o -chai -
lina-m-khan- e-join -in e agency-s a emen -on-ai.pd .
Ki iyama, Nobuo. 2012. “T ade and Inno a ion: Syn hesis Repo .”
OECD T ade Policy Pape s No. 135. Pa is, F ance: OECD
Publishing. h p://dx.doi.o g/10.1787/5k9gwp b xn-en.
K ugman, Paul. 1995. De elopmen , Geog aphy and
Economic Theo y. Camb idge, MA: The MIT P ess.
Leblond, Pa ick. 2021. “Uploading CPTPP and USMCA P o isions
o he WTO’s Digi al T ade Nego ia ions Poses Challenges
o Na ional Da a Regula ion: Example om Canada.”
In Big Da a and Global T ade Law, edi ed by Mi a Bu i,
301–15. Camb idge, UK: Camb idge Uni e si y P ess.
h p://dx.doi.o g/10.1017/9781108919234.019.
Lee, Kai-Fu. 2018. AI Supe powe s: China, Silicon Valley and he
New Wo ld O de . Bos on, MA: Hough on Mi lin Ha cou .
Lippold , Douglas and Ma k. F. Schul z. 2014. “Unco e ing
T ade Sec e s — An Empi ical Assessmen o Economic
Implica ions o P o ec ion o Undisclosed Da a.” OECD
T ade Policy Pape s No. 167. Pa is, F ance: OECD Publishing.
h p://dx.doi.o g/10.1787/5jxzl5w3j3s6-en.
Liu, Cha lene, Ri chie Sun, Peishan Wang and Ma k McDonald.
2023. “China Gene a i e AI — Pa One: A Guide o
Mone isa ion.” HSBE Global Resea ch. Oc obe 17.
Luong, Ngo and Zacha y A nold. 2021. “China’s A i icial
In elligence Indus y Alliance: Unde s anding China’s AI
S a egy Th ough Indus y Alliances.” Cen e o Secu i y and
Eme ging Technology. Da a B ie . May.
h ps://cse .geo ge own.edu/wp-con en /uploads/CSET-
Chinas-A i icial-In elligence-Indus y-Alliance-1.pd .
Madiega, Tambiama. 2023. “A i icial in elligence ac .” B ie ing.
Eu opean Pa liamen , June. www.eu opa l.eu opa.eu/
RegDa a/e udes/BRIE/2021/698792/
EPRS_BRI(2021)698792_EN.pd .
Magui e, James. 2023. “100+ Top A i icial In elligence
(AI) Companies 2023.” eWeek, May 29.
Maslej, Nes o , Lo edana Fa o ini, E ik B ynjol sson, John
E chemendy, Ka ina Lige , Te ah Lyons, James Manyika e
al. 2023. A i icial In elligence Index Repo 2023. S an o d,
CA: AI Index S ee ing Commi ee, Ins i u e o Human-
Cen e ed AI. h ps://aiindex.s an o d.edu/wp-con en /
uploads/2023/04/HAI_AI-Index-Repo _2023.pd .
McCa hy, John, Ma in Minsky, Na haniel Roches e and Claude
Shannon. 1955. “A P oposal o he Da mou h Summe
Resea ch P ojec on A i icial In elligence.” Da mou h College.
Augus 31. www- o mal.s an o d.edu/jmc/his o y/
da mou h.pd .
48 CIGI Pape s No. 292 — Ap il 2024 • Douglas Lippold
Mish a, Sau abh, Robe Koopman, Giudi a De P a o, Anand Rao,
Is ael Oso io-Roda e, Julie Kim, Nikola Spa a o a, Kei h
S ie and And ea Zacca ia. 2023. “AI specializa ion o
pa hways o economic di e si ica ion.” Scien i ic Repo s13(1).
h ps://doi.o g/10.1038/s41598-023-45723-x.
Mi chell, Billy. 2023. “Eigh mo e ech companies join Whi e House
commi men o manage AI isks.” Fedscoop, Sep embe 12.
h ps:// edscoop.com/eigh -mo e- ech-companies-
join-whi e-house-commi men - o-manage-ai- isks/.
Moo e, Oli ia. 2023. “How A e Consume s Using Gene a i e AI?”
And eessen Ho owi z, Sep embe 13. h ps://a16z.com/
how-a e-consume s-using-gene a i e-ai/.
Mu gia, Madhumi a and John Tho nill. 2023. “Supe in elligen
AI — can cha bo s hink?” No embe 27, in Tech Tonic,
p oduced by he Financial Times, podcas , 30:23. www.
.com/con en /36a 32c - a29-466d-bc -8a8a898829d2.
NIST. 2019. “U.S. Leade ship in AI: A Plan o Fede al Engagemen
in De eloping Technical S anda ds and Rela ed Tools.”
www.nis .go /sys em/ iles/documen s/2019/08/10/
ai_s anda ds_ edengagemen _plan_9aug2019.pd .
———. 2023. “A i icial In elligence Risk Managemen
F amewo k (AIRMF1.0).” Janua y.
h ps://doi.o g/10.6028/NIST.AI.100-1.
OECD. 2019. “Recommenda ion o he Council on A i icial
In elligence.” OECD Legal Ins umen s (OECD/LEGAL/0449).
May 21. h ps://legalins umen s.oecd.o g/
en/ins umen s/OECD-LEGAL-0449.
———. 2021. “In e na ional Regula o y Co-ope a ion.” OECD
Bes P ac ice P inciples o Regula o y Policy se ies. Pa is,
F ance: OECD. h ps://doi.o g/10.1787/5b28b589-en.
———. 2024. “Explana o y Memo andum on he Upda ed OECD
De ini ion o an AI Sys em.” OECD A i icial In elligence
Pape s No. 8. www.oecd-ilib a y.o g/science-and-
echnology/explana o y-memo andum-on- he-upda ed-
oecd-de ini ion-o -an-ai-sys em_623da898-en.
OECD.AI. 2021. “Da abase o na ional AI Policies,
powe ed by EC/OECD (2021).” OECD A i icial
In elligence Policy Obse a o y. Pa is, F ance: OECD.
h ps://oecd.ai/en/dashboa ds/o e iew.
O ice o Managemen and Budge . 2019. “Iden i ying P io i y Access
o Quali y Imp o emen s o Fede al Da a and Models o
A i icial In elligence Resea ch and De elopmen (R&D), and
Tes ing; Reques o In o ma ion.” Fede al Regis e Documen
No. 2019-14618. July 10. www. ede al egis e .go /
documen s/2019/07/10/2019-14618/
iden i ying-p io i y-access-o -quali y-imp o emen s-
o - ede al-da a-and-models- o -a i icial.
Pa k, Wal e G. and Douglas Lippold . 2005. “In e na ional
Licensing and he S eng hening o In ellec ual P ope y Righ s
in De eloping Coun ies du ing he 1990s.” OECD Economic
S udies 40: 7–48. h ps://doi.o g/10.1787/
eco_s udies- 2005-a 2-en.
———. 2014. “Channels o echnology ans e and in ellec ual
p ope y igh s in de eloping coun ies.” In In ellec ual P ope y
o Economic De elopmen , edi ed by Sanghoon Ahn, B onwyn
H. Hall and Keun Lee, 33–89. Chel enham, UK: Edwa d Elga .
Raskin, Asa and T is an Ha is. 2023. “The P omise and Pe il
o Open Sou ce AI wi h Elizabe h Sege and Je ey
Ladish,” No embe 21, in You Undi ided A en ion,
p oduced by Cen e o Humane Technology, podcas ,
38:44. h ps://you -undi ided-a en ion.simplecas .com/
episodes/ he-p omise-and-pe il-o -open-sou ce-ai-
wi h-elizabe h-sege -and-je ey-ladish-K 4pNW3s.
Robe s, Huw, Josh Cowls, Jessica Mo ley, Ma ia osa ia Taddeo,
Vincen Wang and Luciano Flo idi. 2021. “The Chinese
app oach o a i icial in elligence: an analysis o policy,
e hics, and egula ion.” AI and Socie y 36: 59–77.
h ps://doi.o g/10.1007/s00146-020-00992-2.
SAG-AFTRA. 2023a. “TV/Thea ical 2023: Regula ing
A i icial In elligence.” Los Angeles, CA: SAG AFTRA.
www.saga a.o g/ iles/sa_documen s/AI TVTH.pd .
———. 2023b. “TV/Thea ical Con ac s 2023: Summa y o
Ten a i e Ag eemen .” No embe 16. Los Angeles, CA:
SAG-AFTRA. www.saga a.o g/ iles/sa_documen s/
TV-Thea ical_23_Summa y_Ag eemen _Final.pd .
Sa o , Gio anni and F ancesca Lagioia. 2020. “The impac
o he Gene al Da a P o ec ion Regula ion (GDPR) on
a i icial in elligence.” Eu opean Pa liamen a y Resea ch
Se ice, June. www.eu opa l.eu opa.eu/RegDa a/e udes/
STUD/2020/641530/EPRS_STU(2020)641530_EN.pd .
Sheehan, Ma . 2023. “China’s AI Regula ions and How They Ge
Made.” Ca negie Endowmen o In e na ional Peace Wo king
Pape . July. h ps://ca negieendowmen .o g/2023/07/10/
china-s-ai- egula ions-and-how- hey-ge -made-pub-90117.
Suominen, Ka i. 2021. “The CPTPP’s Impac s on Digi al T ade and he
Pa h Fo wa d.” Cen e o S a egic and In e na ional S udies.
Oc obe . h ps://csis-websi e-p od.s3.amazonaws.com/
s3 s-public/publica ion/211027_Suominen_CPTPP_FullRepo .
pd ?Ve sionId=6TBKVom7Oc kpyAb1NDJmR0m0wwWpIyx.
Tsai, Cha les, Eloi Laou ou, Ma ias Gansland and Douglas Lippold .
2011. “Enhancing Ma ke Openness, In ellec ual P ope y
Righ s, and Compliance Th ough Regula o y Re o m in Is ael.”
Pa is, F ance: OECD. www.oecd.o g/is ael/48262991.pd .