Documentation template for the use of Generative AI in research

Documen a ion empla e o GenAI Documen a ion empla e o he use o Gene a i e AI in esea ch Me ada a and decision-making c i e ia By Thijs De iend , PhD – Science Policy O ice Open Science & Gene a i e AI V ije Uni e si ei B ussel, Belgium EUTOPIA-MORE T ain- he-T aine ma e ial doi: 10.5281/zenodo.17243261 Documen a ion empla e o GenAI 1 Table o Con en s Table o Con en s ...................................................................................................................................................................... 1 1. In oduc ion ............................................................................................................................................................... 2 2. When o use he empla e? ..................................................................................................................................... 2 3. Sec ions o he documen ......................................................................................................................................... 3 3.1 Explana ion ................................................................................................................................................................ 3 3.2 Decision-making c i e ia ............................................................................................................................................ 3 3.3 Me ada a .................................................................................................................................................................... 3 4. Examples o gene a i e AI applica ions .................................................................................................................... 5 4.1 Uploading long o ma i ems in gene a i e AI ......................................................................................................... 5 4.2 Assigning ca ego ies o esea ch pape s .................................................................................................................. 6 4.3 Conduc ing a sys ema ic e iew ............................................................................................................................... 6 5. Conclusion .................................................................................................................................................................. 6 Documen a ion empla e o GenAI 2 1. In oduc ion Gene a i e AI is being inc easingly in eg a ed in o esea ch p ocesses. These include bu a e no limi ed o li e a u e explo a ion, he gene a ion o code, and he execu ion o speci ic analyses (e.g., sen imen analysis). Despi e his in eg a ion, he me ada a ela ed o use o gene a i e AI is no always being ully o co ec ly epo ed. Some esea ch s udies omi en i ely essen ial de ails o wha models we e used o gene a e esponses. A common mis epo ed elemen is he in e mixing o Cha GPT ( he se ice o e ed by OpenAI ha is buil on LLM echnology) and he unde lying base models, such as GPT-4o. This leads o con using con o ions such as Cha GPT-4o. O he elemen s a e also o en missing, which impac s he c edibili y o he esea ch and he deg ee o which esul s can be ep oduced. A V ije Uni e si ei B ussels (VUB), we de eloped a s anda d o epo ing he use o gene a i e AI in esea ch p ocesses (= he documen a ion empla e). The in en ion was o make VUB esea che s awa e o he ype o in o ma ion hey need o epo when using gene a i e AI. The empla e was designed o be easy o use, and o be applicable o use o cha bo se ices in a esea ch con ex . Fu he mo e, pa icula a en ion was paid o commonly mis epo ed elemen s in scien i ic li e a u e. 2. When o use he empla e? One ongoing deba e and commonly asked ques ion is when use o gene a i e AI is subs an ial enough o be epo ed in a esea ch publica ion. While we canno p o ide de ini i e answe s o his ques ion ha would be applicable o all possible scena ios ac oss all esea ch ields, we p o ide esea che s wi h h ee c i e ia o guide hei decision o epo use o gene a i e AI. All hese c i e ia mus apply o he use o he empla e o be highly ecommended: - Use o gene a i e AI mus be o a esea ch-speci ic pu pose (e.g., gene a ion o hypo heses) and no gene ic (e.g., making Powe poin -slides, p epa ing a speech). - Use o gene a i e AI mus impac an in ellec ual s ep wi hin esea ch. This includes da a analysis, da a in e p e a ion, design o an empi ical s udy, building a gumen a ion... The e m "in ellec ual" should be b oadly in e p e ed and unde s ood wi hin he con ex o he esea ch domain. Fo ins ance, in esea ch ields ha do no adi ionally use empi ical esea ch me hods and a e la gely ex -based (e.g., philosophy), easoning will be he p ima y in ellec ual s ep. A simple example o a s ep ha is no conside ed in ellec ual is he w i ing down o s udy esul s in he Resul sec ion o he pape . - Use o gene a i e AI mus be mo e subs an ial han me ely p o iding mino assis ance o esea che -led asks. The use mus pu he b un o he in ellec ual wo k o a speci ic esea ch p ocess on he model i sel a he han on he esea che . Use wi hin a s uc u ed amewo k (e.g., as in ica e pa o a s udy design, sys ema ic li e a u e e iew, o o mal analysis pipeline) is conside ed mo e subs an ial han ad hoc, uns uc u ed use (e.g., spo adic inqui ies o con enience). We p o ide examples: Wi hin scope (i in ellec ual s ep & subs an ial use) Ou side o scope Sys ema ic hypo hesis gene a ion G amma and s yle checking (modi ica ions a e manually e iewed and no au oma ically inse ed) Da a analysis au oma ion (e.g., sen imen & hema ic analysis) Li e a u e explo a ion (in o mal, no pa o s uc u ed s udy) Expe imen al design se up Code debugging Gene a ing a gumen s Assis ing in li e a u e e iew design Li e a u e syn hesis Ci a ion o ma ing Da a cleaning & ans o ma ion C ea ing non- esea ch g aphs De eloping in e iew ques ions Gene a ing summa ies & abs ac s (uns uc u ed use, no pa o me hod) W i ing en i e pa ag aphs, including ci a ions Answe ing ques ions based on pape Documen a ion empla e o GenAI 3 3. Sec ions o he documen 3.1 Explana ion The documen is spli in o wo sec ions. The i s sec ion includes in o ma ion (called “decision-making c i e ia”) ha may help esea che s discuss whe he gene a i e AI should be employed. These may be epo ed wi hin he me hodology sec ion. The second sec ion includes in o ma ion (called “me ada a”) ha should be epo ed as pa o he publica ion i sel . 3.2 Decision-making c i e ia In o ma ion p o ec ion #1 A e you aking measu es o be su e ha any da a a e going o be p ocessed esponsibly and in acco dance wi h ele an legisla ion? You may wan o hink abou (a) con ac ual a angemen s; (b) license condi ions; and (c) da a p o ec ion egula ions applicable o you egion. Con ex window #2 A e you su e ha he p ocessing o in o ma ion you inpu going o ake place wi hin he con ex window*o models? Please be awa e ha , e en wi hin he con ex window, models do no always " ac o in" all in o ma ion o he same deg ee o make p edic ions ("lead bias"). *Con ex window e e s o he maximum amoun o okens ha he model can p ocess o o mula e an answe . Ve i ica ion #3 Wha measu es will you ake o e i y ou pu s? I you a e no aking measu es, does empi ical e idence exis on accu acy o eliabili y o LLM use in simila se ings? Explainabili y #4 I explainabili y* is ele an o you use o models, a e you going o ake s eps o imp o e i ? *Explainabili y e e s o he abili y o explain why a model p o ided a ce ain esponse. Accoun abili y #5 A e you going o keep logs o all (o some) o you in e ac ions wi h he models? T anspa ency #6 Wha pa s o in e ac ions, i any, a e you going o epo wi hin you publica ion? Plagia ism #7 A e you going o ake s eps o ensu e p ope a ibu ion o ideas and sou ces when engaging in idea ion? I so, how? Resea ch in eg i y #8 Could he use o LLMs po en ially con lic wi h esea ch in eg i y (e.g., when using LLMs o explo a o y da a analysis)? D awbacks #9 Could he d awbacks o he models unde mine he in ended pu pose o you esea ch (e.g., in oducing bias, andomness p e en s ep oducibili y...)? I so, how? Ha ms #10 Could he e po en ially be disp opo iona e socie al cos s associa ed wi h using he models o hese pu poses (e.g., clima e impac , g oup ha m due o biases...)? Me hod #11 Could he e be po en ial di e ences be ween ou pu s c ea ed by humans e sus ou pu s c ea ed by models? I so, a e hese di e ences p oblema ic? (e.g., hemes eme ging om hema ic analysis) 3.3 Me ada a Ca ego ies Ques ion No. Ques ion De ini ions & Examples Expec ed o ma Documen a ion empla e o GenAI 4 Basic in o ma ion #1 Wha se ice a e you using (e.g., Cha GPT, Claude, Gemini, API, so wa e o un models locally…)? Example: Cha GPT cha bo in e ace, API o OpenAI models, models locally un on my compu e … Name o se ice #2 Wha model and model e sion a e you accessing unde his se ice? Check he websi e o he p o ide . I he e sion name is no indable, epo he da e o use. Example: GPT-4o-2024-05-13, GPT-4- 1106-p e iew… Model e sion Technical cha ac e is ics #3 A e you using an unmodi ied model (i.e., base model) o a e you wo king wi h a modi ied e sion (e.g., ine- uned)? Unmodi ied e sion = The model as i was o iginally eleased (e.g., GPT-4, Llama 3 8b) Modi ied = The model as i was o iginally eleased plus modi ica ions ha we e made o he pa ame e s (e.g., h ough ine- uning). Fine- uned = Fu he adjus ing model pa ame e s based on speci ic da ase s o ailo models be e o one speci ic ask Ca ego y #4 A e da a p ocessed locally o elsewhe e? Local p ocessing = The da a does no lea e you compu e . The compu ing powe comes om you RAM memo y o a local sou ce . No local p ocessing = The da a lea es he compu e . Fo ins ance, i migh be p ocessed on egional da a se e s when you en e p omp s in o cha bo s. A e wa d his, he da a will also be s o ed o limi ed du a ion on o he se e s, o en loca ed ou side he Eu opean Union. Ca ego y #5 Wha is he cu -o da e o he model? Cu -o da e = The mos ecen da e ha he aining da a has. Fo ins ance, i he aining da a con ains no da a mo e ecen han 1s o Ap il, his will be he cu -o da e. Da e #6 Wha is he con ex window o he model? Con ex window = The maximum amoun o okens ha he model can p ocess o p edic he nex oken. No e: This is only impo an o da a analysis and long ex p ocessing (e.g., legal ex s, books…). Numbe Wo k low #7 Wha ype o da a was inpu ed (e.g., audio, abula da a, p omp )? Inpu da a = The da a you en e as a use o he model. I ex , his is called a "p omp ". Inpu ype #8 Wha ype o da a was ou pu ed (e.g. ex , audio, abula da a)? Ou pu da a = The da a ha is gene a ed. This may be na u al language, images, code, ideo… Ou pu ype Documen a ion empla e o GenAI 5 #9 We e speci ic pa ame e s manually se when wo king wi h he model o no ? Only answe o he inal con igu a ion. Pa ame e s = Va iables ha you may gi e a ce ain alue when wo king wi h he model (e.g., empe a u e, Top P...) Ca ego y #10 I e e , which pa ame e s did you se manually? Only desc ibe o he inal con igu a ion. Wha is he a ionale i you used pa icula se ings o pa ame e s? Desc ibe his in you me hodology sec ion. Pa ame e s, Tex #11 Did you euse exis ing p omp s o use ce ain p omp ing echniques (e.g., ew-sho p omp ing, Chain-o -Though )? I so, e e o hese in you me hodology sec ion. Only desc ibe he inal con igu a ion. P omp s = An inpu , o en in na u al language, ha elici s a esponse by he model P omp ing echniques = Techniques which a e known h ough ei he o mal esea ch o empi ical es ing o elici mo e desi able ou pu s (e.g., highe quali y). Ca ego y, Re e ence #12 How was he model applied and in eg a ed wi hin you esea ch wo k low? Desc ibe his in you me hodology sec ion. Only desc ibe he inal con igu a ion. Me hod: Desc ibe he speci ic p ocedu e, echnique, o app oach used o apply he model. I a p e- exis ing me hod was used, e e ence i wi hin he me hodology sec ion. Depic he p ocedu e in a lowcha . Examples: Gene a ion ollowed by manual e i ica ion, e ie al- augmen ed gene a ion... Ca ego y, Re e ence 4. Examples o gene a i e AI applica ions 4.1 Uploading long o ma i ems in gene a i e AI Imagine ha you upload an 800-page book on he his o y o in e na ional law. You would like o know wha speci ic his o ical case law you could e e ence ela ed o speci ic cases ha you a e looking in o. A book has abou 300 wo ds pe page, which o als abou 240.000 wo ds o he en i e book. This means he model has o p ocess abou 320.000 okens (coun ing ha in gene al he amoun o okens mul iplied by 0.75 gi es he amoun o wo ds). Fo each case you p esen , you ge a lis o his o ical e en s. He e is why epo ing documen a ion o gene a i e AI ma e s: • Models ha e a maximum amoun o okens ha hey can p ocess be o e hey s a “ o ge ing” hings ( his maximum is called he con ex window) • I is known ha he e is al eady deg ada ion o pe o mance aking place e en wi hin he con ex window. I you do no epo his in o ma ion, o he esea che s canno e i y whe he you use o he model made sense me hodologically. I you op ed o a model wi h e y limi ed con ex window (e.g., ~200.000 okens), he model will no be able o gi e answe s ha ake he en i e book in o accoun . Documen a ion empla e o GenAI 6 4.2 Assigning ca ego ies o esea ch pape s Imagine you a e doing a scien ome ic s udy. As pa o he me hodology, you a e using gene a i e AI o assign ca ego ies o in o ma ion. (You don’ wan o ely on p e-exis ing classi ica ion schemes ha a e known o con ain many laws). You a e assigning classi ie s o en i e pape s based on hei con en s. You decide o use one o OpenAI’s models ia hei API. You lea e he hype pa ame e s in hei de aul alues as hey a e because you don’ unde s and hem. He e is why epo ing documen a ion o gene a i e AI ma e s: • The andomness o esponses is in luenced by se e al hype pa ame e s, including he empe a u e a iable. The s anda d se ing migh be a alue be ween 0.8 and 1. A se ing 1, he model will occasionally pick “unlikely” okens o he p edic ion. A se ing 0, he model will always pick he mos likely oken e e y ime, leading o ep oducible ou pu s. • The ca ego ies ha you assigned will be subjec o andomness! Doing he same un wice will esul in di e en ca ego ies assigned. • You jus used gene a i e AI as one s ep in an en i e wo k low. You need o documen and illus a e you en i e wo k low in a lowcha . I you do no epo his in o ma ion, o he esea che s canno e i y whe he you use o he model made sense me hodologically. You assignmen o ca ego ies will be i ep oducible by o he esea che s i you jus use de aul alues o hype pa ame e s. Addi ionally, he me hodological alidi y o you use o gene a i e AI will depend on he wo k low i sel (i.e., does you use make sense). 4.3 Conduc ing a sys ema ic e iew Imagine ha you a e conduc ing a sys ema ic e iew. You decide o use wo ways o iden i ying a icles. Fi s , you will design sea ch s ings ha you un in academic da abases. You ead he abs ac s and con en s o he pape s. Based on inclusion c i e ia, you hen decide whe he o include he a icles. Second, you decide o also ga he a icles ia Deep Resea ch Modules o e ed by AI- ools. The Deep Resea ch Modules au ogene a e sea ch s ings and un hem agains di e en academic da abases and sea ch engines. The Modules ead h ough he a icles and decide wha o include in he inal epo . You e i y whe he hese a icles mee you inclusion c i e ia. I so, you ex ac he ci a ions and use hese a icles in you e iew. He e is why epo ing documen a ion o gene a i e AI ma e s: • Using a Deep Resea ch Module is a comple ely di e en hing om jus using a s anda d La ge Language Model. I is a wo k low ha includes mul iple s eps, such as sea ch s ing gene a ion, unning he sea ch s ings agains academic da abases, eading h ough he collec ed a icles and d a ing an en i e epo on he opic. This needs o be epo ed co ec ly. • Deep Resea ch Modules canno do comp ehensi e sea ches as he e a e limi a ions o how many sea ch s ings hey un, how many da abases hey can access and how many pape s hey ead. The numbe o un sea ch s ings, accessed da abases and ead a icles mus also ideally be epo ed. A u he p oblem is ha he ou pu o a Module is no ep oducible because he c ea ion o sea ch s ings gene a ed is subjec o andomness. I you do no epo his in o ma ion, o he esea che s canno de e mine whe he you me hodology is sound. They may also no unde s and he sho comings o he Deep Resea ch Module ha you used. I you ha e no e lec ed on he limi a ions, you may w ongly po ay he Deep Resea ch Module as highly eliable and ep oducible. 5. Conclusion This documen a ion empla e helps esea che s epo accu a e me ada a on he use o gene a i e AI in esea ch. This boos s anspa ency o e he analy ical choices made, and ep oducibili y o he esea ch s udy. We encou age esea che s o inco po a e he empla e in o hei wo k low (inso a no domain-speci ic s anda d al eady exis s). Jou nals may also d aw inspi a ion om he epo ing s anda ds desc ibed in his documen .