Open questions

LD-01: What kind of thing is generative AI, and does the answer predict anything?

Research opportunities

• Which observations does each account of generative AI forbid?Which observations would one account predict and another forbid, stated before the data arrive?
• Do the failure patterns of 2023–2024 language models persist in reasoning models?Do the failure patterns of 2023–2024 models hold for later reasoning and agentic systems?
• Do products built on the same base model fail on the same people and cases?Are failures and biases correlated across products built on the same base models, as the infrastructure account predicts?
• When AI output goes wrong at work, who answers for it?What does the artificial-communication account predict about accountability and uptake, and can it be measured?
• Does the EU's training-compute criterion track how general a model is?Does training compute track the generality that the EU's criterion is meant to capture?
• Does concentration at the model layer make apps fail alike? A test on China's registerIn China's register of generative-AI services, where every launched application must display the filed service or model it uses, how concentrated is applications' dependence on a few base models, and on a fixed set of test cases are the errors, refusals and biased outputs of applications that share a base model more strongly correlated than those of applications built on different models, beyond the correlation already found between models from different providers?

LD-02: Aggregate null or distributional shift? Employment

Claims recorded as unknown

• The current R05 evidence base does not establish a causal pass-through from AI-induced productivity gains into worker wages.

Research opportunities

• Why do Nordic registers and US data disagree about young workers in AI-exposed jobs?Why are the broad young-worker effects near zero in Norway while US datasets show large divergences?
• Does the young-worker hiring gap hold up with survey-measured AI use instead of exposure?Are current results robust when actual telemetry-measured use replaces exposure indices?
• Do AI-exposed graduates' early penalties fade or persist beyond the first two years?What happens to displaced or downshifted entrants over five to ten years?
• Are juniors who still get hired at AI-adopting firms promoted and trained less?Will entry-level learning ladders erode if firms hire fewer juniors?
• How much of a firm's AI productivity gain reaches the wages of workers who stay?Will productivity gains pass through into wages, profits, prices or output?
• Do French firms that sign AI agreements then hire and train differently?How do collective bargaining, worker voice and labour regulation alter adoption strategies?

LD-03: Organizational absorption: Why micro gains do not become macro outcomes

Claims recorded as unknown

• The current R05 evidence base does not establish a causal pass-through from AI-induced productivity gains into worker wages.

Research opportunities

• Does deploying AI to whole teams change the work that individual licences did not?Does deploying AI to whole teams, with redesigned workflows, change coordination-dependent work where individual provision did not?
• Where does the time saved with generative AI go?Where does saved time go: more output, new tasks, checking, or leisure? No study follows saved time directly; Danish register data and one Korean survey come closest.
• How much of the time AI saves its users do colleagues spend checking and fixing?How much of the measured gain is spent on checking and repairing AI output elsewhere in the organization?
• Is there a generative-AI productivity J-curve inside firms?Are the positive industry-level correlations causal, and do they survive longer series?
• Do generative-AI deployments in government improve services, not just staff time?Do public-sector deployments improve service outcomes? The UK and Australian government trials could not be read for this version.
• Do organizations' AI policies match what their staff actually do with AI?Is "digital formalism", a literature in Chinese public administration on technology adopted for show, a useful lens for generative-AI deployment, and does it describe what happens elsewhere?
• Is the time AI saves paid for with weaker checking?How much of the measured gain is spent on checking and repairing AI output elsewhere in the organization?
• In small Greek firms, does generative AI stay a personal tool or change how teams work?In small firms in Greece outside tourism (for example accounting offices, shops, small manufacturers and software companies), who uses generative AI and for which tasks, and has any firm changed how work is divided, checked or handed over because of it?

LD-04: Levelling or hollowing? Skill compression and the reproduction of expertise

Research opportunities

• Is the time AI saves paid for with weaker checking?How much of the measured gain is spent on checking and repairing AI output elsewhere in the organization?
• Do professionals trained with AI lose the skill without it a year later?Do people who learn a task with AI become less able to do it without AI over months or years ("never-skilling")?
• Does retained skill depend on how routine the task is and who is learning?Does retained skill depend on how routine the task is and on prior expertise, as the split in tool-removed studies suggests?
• Do newcomers get less feedback and real work once AI coding tools arrive?Does AI reduce the real work, feedback and supervised struggle that seniors hand to juniors?
• What AI set-ups do employers actually give trainees and junior staff?What do employers actually deploy for juniors: guardrails, learning modes, or unrestricted tools?
• Does skill compression reverse when AI acts as an agent rather than an assistant?Will compression reverse as AI agents act autonomously rather than assist?
• Do experienced clinicians lose unaided skill after routine AI use?Does routine AI use deskill experienced professionals in the skills they no longer exercise?
• Does handing work to the model cause both skill loss and sameness?When people learning a writing or analysis task are randomly assigned to use a language model that does the work for them, a model that only explains while they do the work, or no model, do the same participants who lose more skill once the model is removed also produce work more similar to each other's and to the model's, and does the way the model is used account for both?
• What work do junior staff in Greek firms still learn by doing once AI arrives?In Greek accounting offices, law firms and software companies that use generative AI, which tasks that juniors used to learn on are now done with AI or by seniors, and how do juniors and their supervisors describe the feedback and practice juniors now get?

LD-05: Discretion curtailed, relocated or encoded?

Research opportunities

• Is the time AI saves paid for with weaker checking?How much of the measured gain is spent on checking and repairing AI output elsewhere in the organization?
• Outside corrections, how often do officials depart from algorithmic recommendations?How often do public-sector workers actually depart from algorithmic recommendations across comparable systems?
• Does telling reviewers a case was picked by a risk model change what they decide?Under which organisational conditions does human review genuinely correct upstream model errors?
• What does being flagged by a fraud-risk model cost claimants before any decision?How common are selection-stage burdens that are difficult to challenge because no formal adverse decision has yet occurred?
• How many AI transcription errors survive social workers' checks into case records?How frequently do uncorrected AI hallucinations enter public-sector case files?
• In contracts for public decision systems, who controls thresholds, variables and updates?How much substantive discretion sits with vendors and procurement contracts rather than public agencies?
• Who first detects failures in public-sector algorithmic systems, and how late?How often do internal monitoring and frontline staff catch system failures before regulators, courts, journalists or advocacy groups do?

LD-06: The evaluation apparatus: Can benchmarks, leaderboards and audits carry the weight placed on them?

Research opportunities

• Do benchmark scores predict what models achieve in real deployments?Do benchmark scores across many models predict measured outcomes in deployment, in any domain?
• How faithfully do indices and policy documents relay AI evaluation results?How often do indices, procurers and regulators rely on developer-reported scores, and how often do relays distort them?
• How much does knowing it is tested change a model's measured safety?How much does evaluation awareness change measured safety behaviour?
• How is external evaluation under the EU AI Code of Practice carried out?How are the EU code's external-evaluation requirement and its exceptions applied, and what access do evaluators get?
• Who builds and checks China's compliance test banks for generative AI?Who builds the test banks under China's standard, and how are they checked?
• How many published AI evaluations outside medicine can no longer be repeated?How far can published evaluations of closed models be repeated after the model is retired? (Field Map: independent evaluation under API retirement.)
• What do evaluators, researchers and regulators actually get to see under EU law?For general-purpose AI providers that also run platforms or search engines designated under the Digital Services Act, what does each legal route for outside scrutiny deliver in practice (external evaluation under the GPAI Code of Practice, which applies only to models with systemic risk, vetted-researcher data access under the Digital Services Act, and the AI Office's information requests and model evaluations), recorded for each route as what object is reached (model, platform data or documents), who can start it, how long it takes and how deep the access goes?

LD-07: AI in the research apparatus

Research opportunities

• Do pinned model snapshots keep social-science conclusions stable over time?How stable are 2025–2026 model snapshots over time when used for the same social-science measurement task?
• Does quantization or the serving engine change social-science findings?For open-weight models, do deployment choices such as quantization materially change social-science reliability metrics?
• How much AI-assisted or agent responding reaches probability panels?How common are fully autonomous agents in probability-based or high-assurance survey panels?
• Can synthetic respondents recover strong treatment-effect heterogeneity?Can synthetic populations ever recover subgroup treatment effects and causal mechanisms out of distribution?
• Does post-training change what an LLM measures when it codes text?How much do jurisdictional alignment regimes causally change social-science measurement, rather than merely refusal behavior?
• Do AI literature-review tools change the conclusions of social-science reviews?How often does retrieval-augmented synthesis change which social-science evidence enters a review and therefore changes its conclusion?
• How do research ethics boards judge synthetic participants and AI-interpreted data?How should research ethics boards evaluate delegated interpretation, synthetic participants and confidential-data use?
• Do silicon samples work outside the North Atlantic, and in which language?Using Afrobarometer, Arab Barometer and Latinobarómetro items fielded after model training cutoffs, does synthetic-sample accuracy at country, subgroup and individual level depend on interview language (for example Swahili or Hausa versus English, Arabic versus English), and do LLM codes of political constructs in low-resource languages pass a replacement test against trained in-country coders?
• Do LLM judges inflate the quality of AI-written deliberation?On a corpus mixing human-written and LLM-written deliberative contributions, do LLM judges assign higher construct scores (for example on the Discourse Quality Index or argument quality) to LLM-authored text than expert coders blind to authorship, and is the gap larger when the judge model also wrote the text?
• Does "LLMs can replace coders" depend on which humans you compare with?Holding items and codebook fixed, how often does a formal replacement-test verdict flip depending on whether the comparator pool is crowdworkers, trained research assistants, the original study's coders or domain experts, and does the flip rate track construct ambiguity?
• Does asking in Greek change how AI "silicon samples" answer as Greek respondents?When a language model answers a few European Social Survey Round 11 questions as Greek respondents of a given age, sex, education and region, do its answers come closer to or move further from those of the Greek respondents when the prompt is in Greek rather than in English, overall and across groups?

LD-08: Who can know about AI? Epistemic dependence and jurisdiction over the field

Research opportunities

• Does industry funding or data access change what research on AI's social effects finds?Does industry funding, data or access change what research on AI's social effects concludes?
• How tied to industry are researchers of AI's social effects outside computer science?How large are industry ties among researchers who publish on AI's social effects in social-science journals, books and policy reports?
• Do developers' usage statistics hold up when someone else classifies the conversations?How stable are vendor usage measures when the classifier, the validation and the data owner change?
• Does the EU's researcher-access law actually reach AI features, and who gets in?Will Article 40 access be granted in practice, and does it reach AI features?
• Who writes research on AI's social effects outside arXiv?Is computer science absorbing research on AI's social effects outside technical preprint servers?
• Whose conversations do data donations capture?Are donated or public conversation datasets representative enough to check vendor usage statistics?
• What do evaluators, researchers and regulators actually get to see under EU law?For general-purpose AI providers that also run platforms or search engines designated under the Digital Services Act, what does each legal route for outside scrutiny deliver in practice (external evaluation under the GPAI Code of Practice, which applies only to models with systemic risk, vetted-researcher data access under the Digital Services Act, and the AI Office's information requests and model evaluations), recorded for each route as what object is reached (model, platform data or documents), who can start it, how long it takes and how deep the access goes?

LD-09: Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike?

Research opportunities

• Does handing work to the model cause both skill loss and sameness?When people learning a writing or analysis task are randomly assigned to use a language model that does the work for them, a model that only explains while they do the work, or no model, do the same participants who lose more skill once the model is removed also produce work more similar to each other's and to the model's, and does the way the model is used account for both?
• Does language-model use narrow what whole fields study, including the social sciences?Does language-model use narrow what fields study, measured as section 1 measured AI methods in the natural sciences?
• Does the convergence caused by writing with AI persist or compound over months of use?Does convergence persist or compound over months and years of use, and over generations of students trained with AI?
• Is the diversity of what reaches the market falling, or only diversity inside experiments?Does market-level diversity (books, apps, published papers) fall, or only diversity within experiments?
• Does pooling different models keep restoring diversity as models converge?Does pooling models and prompts keep working as models' errors and outputs converge?
• Does AI writing help pull writers toward Western styles in languages other than English?Is there convergence outside English and outside the US, UK and India?
• In science, is it the wording that converges or the ideas?When writing converges, is it the wording or the ideas?
• Does concentration at the model layer make apps fail alike? A test on China's registerIn China's register of generative-AI services, where every launched application must display the filed service or model it uses, how concentrated is applications' dependence on a few base models, and on a fixed set of test cases are the errors, refusals and biased outputs of applications that share a base model more strongly correlated than those of applications built on different models, beyond the correlation already found between models from different providers?

LD-10: Persuasion, Intermediation & the Information Order

Research opportunities

• Do AI answers in search cut clicks to outside sites on phones and outside the US?mobile and non-U.S. AI-search RCTs
• Are AI answer layers narrowing or widening the range of sites people visit?long-run source-diversity change across platforms
• Are AI search answers less faithful to their cited sources outside US English?stronger non-English empirical coverage
• Does weeks of everyday AI-answer use make people check sources less?repeated everyday-assistant effects on epistemic dependence
• Do AI-content labels change belief and sharing on a live social platform?natural-field effects of AI labelling
• An updated meta-analysis of AI persuasion that separates attitudes from behavioura newer meta-analysis of frontier-model persuasion

LD-11: Concentration, Stack & Sovereignty

Claims recorded as unknown

• NVIDIA states that U.S. export rules effectively foreclosed it from competing in China's data-center compute market; the scale of Chinese domestic substitution remains unresolved in the R06 evidence base.

Research opportunities

• Does public research compute change what academic AI researchers can do without industry?how much public research compute changes bargaining power
• How large is open-weight model use beyond a single routing platform?global open-weight share beyond OpenRouter
• Did UK public bodies switch cloud providers after egress fees were waived?measured post-remedy cloud switching
• Do AI startups backed by a cloud giant end up running on that giant's cloud?a quantified AI-specific bundling/default foreclosure effect
• How much of the fall in AI prices is competition rather than cheaper technology?the causal decomposition of inference-price decline
• Which layers stay foreign when sovereign AI compute projects go live?operational EU/Gulf public or sovereign compute at material scale
• Does concentration at the model layer make apps fail alike? A test on China's registerIn China's register of generative-AI services, where every launched application must display the filed service or model it uses, how concentrated is applications' dependence on a few base models, and on a fixed set of test cases are the errors, refusals and biased outputs of applications that share a base model more strongly correlated than those of applications built on different models, beyond the correlation already found between models from different providers?

LD-12: Regulation-in-Practice: Rights, Compliance & Developmental Steering

Claims recorded as unknown

• China's State Council set policy targets for application penetration of new-generation intelligent terminals and AI agents above 70% by 2027 and above 90% by 2030; these are targets, not measured outcomes.
• EN 18286:2026 has been published by CEN/CENELEC as the first European AI Act-supporting quality-management standard, but HQ did not locate an OJEU reference giving it presumption of conformity at the evidence cut.

Research opportunities

• Does concentration at the model layer make apps fail alike? A test on China's registerIn China's register of generative-AI services, where every launched application must display the filed service or model it uses, how concentrated is applications' dependence on a few base models, and on a fixed set of test cases are the errors, refusals and biased outputs of applications that share a base model more strongly correlated than those of applications built on different models, beyond the correlation already found between models from different providers?
• How often do people use GDPR rights over automated decisions, and what happens?remedy-use rates under GDPR automated-decision rights
• How many staff and how much money do national AI Act enforcers actually have?adequate staffing of national authorities
• Did signing the EU's voluntary AI code change what model providers actually do?substantive risk reduction from voluntary-code adherence
• China's AI filing register: how often are filings sent back, and whose models do apps use?Chinese rejection or sanction rates under the filing regime
• Do EU delays of AI products line up with the AI Act's dates or with other EU rules?causal product-withholding effects attributable to a specific EU AI law
• Do AI hiring-audit rules cost small employers more than large ones?robust SME compliance costs or incumbent advantage
• What do evaluators, researchers and regulators actually get to see under EU law?For general-purpose AI providers that also run platforms or search engines designated under the Digital Services Act, what does each legal route for outside scrutiny deliver in practice (external evaluation under the GPAI Code of Practice, which applies only to models with systemic risk, vetted-researcher data access under the Digital Services Act, and the AI Office's information requests and model evaluations), recorded for each route as what object is reached (model, platform data or documents), who can start it, how long it takes and how deep the access goes?