MindWa e: A Socio echnical Cogni i e Middlewa e o
Coun e ing Coo dina ed Inau hen ic Beha io
P i h i Vasi eddy
College o Enginee ing, No heas e n Uni e si y, Bos on, MA
Abs ac
The p oli e a ion o Coo dina ed Inau hen ic Beha io (CIB)—speci ically he deploymen o
malicious bo - a ms and bo -he ds—p esen s a unique challenge o mode n in o ma ion ecosys-
ems. These en i ies do no me ely sp ead alsehoods; hey exploi he cogni i e heu is ics o
human use s o manu ac u e consensus.
T adi ional de enses agains CIB ha e his o ically elied on an algo i hm-cen e ed pa adigm,
p io i izing au oma ed de ec ion sys ems ha ope a e as "black boxes" o classi y ac o s as
au hen ic/ inau hen ic. Howe e , his echnological de e minism aces an inhe en "a ms ace"
dilemma: as de ec ion algo i hms imp o e, ad e sa ial ac o s e ol e mo e sophis ica ed me h-
ods o e ade hem, c ea ing a pe pe ual cycle o concealmen and de ec ion ha o en lea es
he human use — he ul ima e a ge o in luence ope a ions— ulne able and disempowe ed.
This epo p oposes "MindWa e," a cogni i e middlewa e designed o in e up he social
heu is ics exploi ed by bo ne wo ks. MindWa e le e ages Seam ul Design o s a egically e-
eal he "seams"— he ough edges, inconsis encies, and in as uc u al gaps—o digi al in e -
ac ions. I ope a ionalizes Social T anspa ency h ough a "4W F amewo k" (Who, Wha , When,
Why) o ende he in isible p o enance o in o ma ion isible. By ans o ming he implici so-
cial signals o CIB (e.g., empo al coo dina ion, ne wo k clus e ing) in o explici isual cues,
MindWa e ac s as a cogni i e middlewa e laye . I dis up s he cogni i e heu is ics — speci i-
cally bandwagon, au ho i y, simila i y, and social p esence e ec s— ha bo ne wo ks exploi o
bypass c i ical hinking. This epo de ails he heo e ical unde pinnings, sys em a chi ec u e,
and e alua ion ene s o MindWa e, o e ing a comp ehensi e bluep in o a esilience-based
app oach o in o ma ion in eg i y.
The Socio echnical Shi
To unde s and why we need a sys em like MindWa e, we mus i s unde s and he ajec o y o
Explainable AI (XAI) o e he las ew yea s. We a e wi nessing a depa u e om he idea ha
he "black box" o AI is he only hing ha ma e s, mo ing owa d a ealiza ion ha he ’human’
looking a he box is jus as impo an .
Ehsan e al. demons a ed ha simply seeing a poli ician’s wee isn’ enough; we need a
"clea box" o ace he cogni i e lineage o ha hough [5]. I es ablished ha i we can model
he ’belie s’ and ’s a egies’ behind a piece o con en , we can unde s and i . Bu unde s anding
is no a passi e ac . The backg ound o he pe son who ecei es he explana ion changes
e e y hing [4] . A da a scien is looks a a con idence sco e and sees a p obabili y; a layman
looks a he same numbe and sees an absolu e u h. This disc epancy c ea es a dange ous
1
gap—a c ea o -consume gap—whe e he e y ools designed o build us can acciden ally
igge delusion.
This dange is u he a icula ed wi h EP’s (Explainabili y Pi alls), whe e e en well-in en ioned
explana ions can back i e[2]. Use s o en place unwa an ed ai h in numbe s and algo i hmic
ou pu s simply because hey look au ho i a i e. In a s udy conduc ed on AI/non-AI g oups, he
non-AI g oup associa ed he incomp ehensibili y o a nume ical- easoning obo wi h ’highe ,
mo e in elligen exp ession’. Thei heu is ic was complex/c yp ic = sma [4].
This is he "zone o dange " whe e bo - a ms ope a e; hey exploi ou endency o us
olume, epe i ion, and pe cei ed consensus.
**To coun e his, we canno simply build "be e " algo i hms in a acuum**. We need
ahuman-cen e ed explainable AI, which ad oca es o a ’ e lec i e socio echnical app oach’.
The philosophical spine o ou na a i e a i s e y basic is ha we mus look a he ’social
apes y’ in which hese sys ems si , no jus op imize o ’accu acy’[1].
"Social T anspa ency" (ST) is a game-changing concep o XAI sys ems. The idea behind
ST is no o make o he s ac i i ies isible, bu o enable pe meabili y and obse abili y o e
o he s in e ac ions o unde s and he impac on explainabili y. In human decision-making, we
a ely ely on echnical specs alone. We ely on wha o he s hink, wha he "c ew" knows, and
he social his o y o an objec [3].
The concep o ST is behind he mechanism we p opose in MindWa e o make in isible so-
cial con ex isible. Mos AI sys ems a e enginee ed o be ’seamless’; howe e , i is essen ial
o s a egically disclose hei ulne abili ies when necessa y. A seamless map would hide
he ne wo k, ying o au o-connec (unknown unknowns). A seam ul map shows you exac ly
whe e he Wi-Fi is weak, so you can mo e o a be e spo (known unknowns). By e ealing
hese "seams"— he misma ches and he ough edges—we a e acili a ing he de elopmen o
’a be e AI’.[5]
Ad e sa ial ac o s a e unbu dened by e hical cons ain s. They apidly e e se-enginee
black boxes. They engage in "model ex ac ion a acks" by p obing he sys em’s bound-
a ies— weaking pos ing equencies, di e si ying con en sou ces, and aging accoun s o all
jus below de ec ion h esholds. The ’ igh ’ bo ( om he pe spec i e o a male olen ac o )
looks human, ac s human, and bypasses il e s seamlessly. A seamless design ha o ces
us on an in isible secu i y laye , has no cogni i e de enses emaining. The "seamlessness"
o he use expe ience he e becomes a ulne abili y, as i smoo hs o e he e y ic ion poin s
(e.g., he millisecond synch oniza ion o 1,000 likes) ha would o he wise be ay he inau hen-
ici y o he in e ac ion.
This is an a ms ace, and i is unwinnable i ough solely on he echnological plane.
The a acke always possesses he i s -mo e ad an age, inno a ing new o ms o coo dina-
ion (e.g., using Gene a i e AI o p oduce unique ex o e e y bo ) while he de ende plays
ca ch-up. To b eak his cycle, we mus mo e he locus o de ense om he se e -side algo-
i hm o he clien -side use cogni ion.
MindWa e is he culmina ion o his lineage. This sys em is inspi ed by he cogni i e mod-
eling app oach ound in Unpack ha Twee (Ehsan e al.). I espec s he who in he equa ion,
ope a es h ough a e lec i e socio echnical lens, and u ilizes seam ul e ela ions o achie e
social anspa ency.
2
MindWa e, he p oposal
In oduc ion: The Heu is ic Ba le ield
We p opose MindWa e no me ely as a con en il e , bu as a cogni i e middlewa e designed o
educe he impac o malicious bo - a ms and bo -he ds. To unde s and MindWa e, we mus
i s acknowledge ha CIB (Coo dina ed Inau hen ic Beha io ) does no hack code; i
hacks human sociology. Bo -he ds succeed because hey o e whelm ou na u al social
heu is ics.
When a use sees a pos wi h housands o likes, he ’Bandwagon Endo semen Heu is ic’
kicks in (e e yone likes his, so i mus be ue). When a bo p o ile mimics a e i ied expe ,
he ’Au ho i y Heu is ic’ is igge ed. When a swa m o accoun s uses local slang and amilia
image y, he ’Simila i y Heu is ic’ lowe s ou de enses. And simply by gene a ing massi e
olume, hey exploi he ’Social P esence Heu is ic’, making a inge idea eel like a dominan
public sen imen [3].
Coun e ing Cogni i e Heu is ics h ough Seam ul In e en ions:
The Heu is ic: The Bandwagon E ec
The De ense: MindWa e in oduces a "Consensus Seam." Ins ead o showing a aw coun
(e.g., "10,000 Likes"), MindWa e decomposes he me ic based on coo dina ion signals.
Visual Cue: The like bu on e ec i ely " ac u es." I migh show ha o he 10k likes, 9k
appea ed wi hin a 5-minu e window ( empo al clus e ing) o o igina ed om accoun s c ea ed
in he las 24 hou s (iden i y clus e ing).
Mechanism: This e eals he "seam" be ween o ganic popula i y (which g ows o e ime) and
ino ganic ampli ica ion (which appea s as ins an aneous spikes). By exposing he mechanism
o he popula i y, he bandwagon heu is ic is dis up ed; he use sees no a c owd o people,
bu a machine o sc ip s.
The Heu is ic: The Simila i y Heu is ic
The De ense: MindWa e uses "Iden i y Ne wo k Seams." I analyzes he linguis ic and ne wo k
neighbo hood o he accoun o check o syn he ic simila i y.
Visual Cue: I an accoun claims o be a "local mo he " bu i s ne wo k connec ions a e al-
mos exclusi ely o eign IP blocks/disconnec ed bo clus e s, MindWa e o e lays a "Ne wo k
Misma ch" isualiza ion — a seam showing he use is "ou o place" in he g aph.
Mechanism: This e eals he gap be ween he pe o mance o iden i y and he s uc u al eal-
i y o he ne wo k. I helps use s see ha he "simila " ac o is ac ually a "s ange " in he social
g aph.
The Heu is ic: Au ho i y Bias
The De ense: MindWa e implemen s "P o enance Seams." I isualizes he accoun ’s his o y
no as a s a ic bio, bu as a dynamic imeline.
Visual Cue: I an accoun suddenly shi s i s opical ocus (e.g., om "K-Pop" o "Geopoli ics")
o changes i s handle/loca ion equency, MindWa e highligh s his discon inui y as a isible
ea in he p o ile’s imeline.
Mechanism: This exposes he misma ch be ween he claimed au ho i y (a geopoli ical expe )
and he beha io al eali y (a epu posed en e ainmen an accoun ). I o ces he use o
e alua e he accoun based on i s ajec o y, no jus i s cu en s a e (i.e., he concep o
Expec a ion s. Reali y)[6].
The Heu is ic: Social P esence
The De ense: MindWa e deploys "Au oma ion Seams." I highligh s pa e ns ha a e physio-
logically impossible o s a is ically unlikely o humans.
3
Visual Cue: I an accoun eplies ins an ly o hund eds o h eads 24/7 wi hou sleep, Mind-
Wa e isualizes his "inhuman endu ance" as a imeline sa u a ion map—a solid block o ac i -
i y ha no human could sus ain.
Mechanism: This b eaks he illusion o social p esence by e ealing he indus ial scale o he
ope a ion. I shi s he use ’s pe cep ion om "cha ing wi h a i eless ac i is " o "in e ac ing
wi h a sc ip ."
MindWa e ac s as a lens ha si s be ween he con en s eam and he mode a o (o end-
use ). I u ilizes Social T anspa ency (ST) o make he nuance o in e ac ion isible, e ec i ely
sho -ci cui ing hese heu is ics.
The High-Le el A chi ec u e
MindWa e AI Co e (Golem)
Social Feed (Bo /Use Mix)
Backend (Wha )
In e ac ion T aces
(Accep /Rejec )
Backend (Who)
Iden i y & Usage
(Ve i ied/Bo )
F on end (Why)
Commen s
& Reasoning
Backend (When)
Con ex ual Timing
(E en Sync)
Technological AI
Socio echnical
Valida ion
Decision AI
C ew Knowledge
O ganiza ional AI
Me a-Knowledge
(o o g p ac ices)
Mode a o Agen | Seam ul Dashboa d (Insigh s) In e ace
Rele ance / Au hen ici y / Flags
Figu e 1: A chi ec u e, MindWa e
The 4W F amewo k: Cap u ing Nuance
Answe ing he * ’who’ did ’wha ’ wi h he sys em, ’when’ and ’why’ hey did wha hey did * can
g ea ly cap u e he nuances ou side he bounds o a machine, d i ing he algo i hm be e [3] .
In Mindwa e, CIB is iden i ied ia Social T anspa ency based on he 4W’s and he unde ly-
ing ’Golem’ (Co e-AI). The goal he e is o design a sys em ha d i es he algo i hm be e by
eeding i ich, socially si ua ed da a.
1. The F on end: WHY (The Commen ing Fea u e)
The "Why" is he mos quali a i e aspec o MindWa e, loca ed on he on end. I is based on
he commen ing ea u e. In a s anda d sys em, a " lag" is bina y (bad/good). In MindWa e, he
lag comes wi h an addi ional "Why" ha cap u es he a ionale. The "Why" looks o a ace
4
o easoning. Why was his lagged? Why was his sha ed? This cap u es he in en and he
"why" behind he in e ac ion, p o iding he na a i e con ex ha aw da a migh lack.
2. The Backend: WHAT (In e ac ion T aces)
The "Wha " esides in he backend and ocuses on he ’ aces o o he s in e ac ions’. This is
whe e we analyze he collec i e ou come o he con en . i.e.,
- Did o he us ed agen s Accep o Rejec his con en ?
- Was he con en deemed ele an o non- ele an ?
By analyzing hese aces, MindWa e can de ec i a piece o con en is being a i icially in la ed
by a bo -he d (high olume o low-quali y "Accep s") o i i is being o ganically ejec ed by he
"C ew" ( us ed mode a o s).
3. The Backend: WHO (Iden i ies)
The "Who" is he iden i y ex ac ion laye . This backend p ocess di es in o he use -le el da a
o es ablish c edibili y. I looks a :
•App Usage: Does he use beha e like a human (sc olling, pausing) o a sc ip (API
calls)?
•His o y: A e he e p e ious lags on his use ? Wha is hei e i ied ie ?
This allows MindWa e o s ip away he anonymi y ha bo - a ms ely on. I alida es ha he
"Who" behind he in e ac ion is a legi ima e ac o wi hin he ecosys em.
4. The Backend: WHEN (Con ex ual Timing)
The "When" analyzes he ime o in e ac ions ela i e o ou side e en s. Bo - a ms o en swa m
immedia ely a e a igge ing e en (e.g., a b eaking news s o y o a speci ic p ess elease).
MindWa e asks: *Based on ou side e en s, is his su ge in ele ance logical?* I a hash ag
spikes in a ic h ee seconds a e a news e en , he "When" lag is aised, sugges ing coo -
dina ion a he han o ganic eac ion.
The Golem: The Unde lying Le els o E alua ion
Unde nea h he 4Ws lies he Golem, which comp ises specialized AI agen s ha p ocess
he 4W da a o assess ele ance and au hen ici y, iden i ying lags o au o-mode a ion. These
agen s unc ion as a “T ansac i e Memo y Sys em”—a collec i e b ain o he pla o m.
1. Technological AI (The Fo malis )
This agen ope a es a he o malis le el. I analyzes he ajec o y o use decision ou pu s.
I looks a he eac ion a he use le el—speci ically, Social Valida ion. I asks: "Does he
ma hema ical ajec o y o hese likes and sha es ma ch he expec ed pa e n o o ganic social
alida ion?". I is he s a is ical backbone ha alida es i he numbe s ( he ’Wha ’) align wi h
eali y o i hey a e being spoo ed.
2. Decision Making (Mode a o Agen s Le el)
This agen ope a es on local con ex . I is inspi ed by he concep o C ew Knowledge in XAI
sys ems [3]. We ain hese agen -mode a o s o unde s and he speci ic no ms, slang, and
his o y o he communi y. I looks a pas decisions made by human mode a o s o in e how
o handle cu en con en . I he "C ew" usually ejec s con en om a speci ic sou ce, his
5
agen applies ha heu is ic. I coun e s he Simila i y Heu is ic used by bo s by p o ing ha
ue simila i y equi es deep, aci cul u al knowledge ha bo s canno ake.
3. O ganiza ional Le el AI
This is he highes le el o abs ac ion. I deals wi h he Me a-Knowledge o O ganiza ional
P ac ices. This agen unde s ands he b oade ules o he pla o m and he o ganiza ion. I
ensu es ha decisions align no jus wi h he local "C ew" bu wi h he mac o-le el policies o
he pla o m. I is he "Seam ul" agen ha unde s ands he in as uc u al limi a ions and ules,
ensu ing ha he mode a ion doesn’ jus happen in a acuum bu espec s he o ganiza ional
a chi ec u e.
Conclusion o P oposal
By in eg a ing he 4W F amewo k wi h he Golem, MindWa e c ea es a de ense-in-dep h s a -
egy. While i may no silence he bo s; i makes hei lack o Social T anspa ency isible. I
e eals ha while hey may ha e he bandwagon, hey lack he Knowledge; while hey may
ha e he olume, hey lack he o ganic Why.MindWa e ende s he in isible coo dina ion
o bo - a ms isible, allowing human mode a o s (and humans) o eclaim he u h.
Re e ences
[1] Human-cen e ed Explainable AI: Towa ds a Re lec i e Socio echnical App oach. In e na-
ional Con e ence on Human-Compu e In e ac ion.
[2] Explainabili y Pi alls: Beyond Da k Pa e ns in Explainable AI. a Xi p ep in a Xi :2109.12480.
[3] Expanding Explainabili y: Towa ds Social T anspa ency in AI Sys ems. P oceedings o he
2021 CHI Con e ence on Human Fac o s in Compu ing Sys ems.
[4] The Who in XAI: How AI Backg ound Shapes Pe cep ions o AI Explana ions. a Xi p ep in
a Xi :2107.13509.
[5] Unpack ha Twee : A T aceable and In e p e able Cogni i e Modeling Sys em. P oceedings
o he ICCC.
[6] Seam ul XAI: Ope a ionalizing Seam ul Design in Explainable AI. a Xi p ep in a Xi :2211.06753.
6
