Claims
The 295 claims the debates rest on, grouped by debate. Each claim is listed once, under the debate its ID names. 13 of them are also cited by another debate. Their row says which, and the Debate filter finds them for every debate that cites them.
Each row shows what kind of claim it is (fact, estimate, interpretation or unknown) and how mature its evidence is (emerging, rapidly maturing, mature or contested). Each claim page also says what the claim does not license and which sources it rests on.
Showing all 295 claims.
Showing 37 of 295 claims.
Showing 18 of 295 claims.
Showing 32 of 295 claims.
Showing 28 of 295 claims.
Showing 17 of 295 claims.
Showing 29 of 295 claims.
Showing 34 of 295 claims.
Showing 25 of 295 claims.
Showing 21 of 295 claims.
Showing 19 of 295 claims.
Showing 193 of 295 claims.
Showing 8 of 295 claims.
Showing 4 of 295 claims.
Showing 20 of 295 claims.
Showing 22 of 295 claims.
Showing 15 of 295 claims.
Showing 6 of 295 claims.
Showing 12 of 295 claims.
Showing 58 of 295 claims.
No claim matches all these choices. Set Evidence maturity or Debate back to Any.
Showing 7 of 295 claims.
Showing 5 of 295 claims.
Showing 2 of 295 claims.
Showing 10 of 295 claims.
Showing 3 of 295 claims.
Showing 1 of 295 claims.
Showing 89 of 295 claims.
Showing 13 of 295 claims.
Showing 14 of 295 claims.
No claim matches all these choices. Set Kind of claim or Debate back to Any.
Showing 50 of 295 claims.
No claim matches all these choices. Set Kind of claim, Evidence maturity or Debate back to Any.
Showing 11 of 295 claims.
Showing 23 of 295 claims.
Showing 167 of 295 claims.
Showing 30 of 295 claims.
Showing 123 of 295 claims.
Showing 16 of 295 claims.
No claim matches all these choices. Set Debate back to Any.
Showing 67 of 295 claims.
Showing 24 of 295 claims.
Showing 39 of 295 claims.
No claim matches all these choices. Set Kind of claim or Evidence maturity back to Any.
No claim matches all these choices. Set at least two of the filters back to Any.
No claim matches all these choices. Set Kind of claim back to Any.
- LD-01: What kind of thing is generative AI, and does the answer predict anything?
- LD-02: Aggregate null or distributional shift? Employment
- LD-03: Organizational absorption: Why micro gains do not become macro outcomes
- LD-04: Levelling or hollowing? Skill compression and the reproduction of expertise
- LD-05: Discretion curtailed, relocated or encoded?
- LD-06: The evaluation apparatus: Can benchmarks, leaderboards and audits carry the weight placed on them?
- LD-07: AI in the research apparatus
- LD-08: Who can know about AI? Epistemic dependence and jurisdiction over the field
- LD-09: Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike?
- LD-10: Persuasion, Intermediation & the Information Order
- LD-11: Concentration, Stack & Sovereignty
- LD-12: Regulation-in-Practice: Rights, Compliance & Developmental Steering
LD-01: What kind of thing is generative AI, and does the answer predict anything?
- Narayanan and Kapoor describe AI as “normal technology” and predict that slow diffusion “will continue to be the norm in high-consequence tasks”; they argue that diffusion is limited by the speed at which individuals, organizations and institutions adapt, and that for past general-purpose technologies it “occurs over decades, not years.” Interpretation · contested evidence LD01-SL-C01
- Narayanan and Kapoor predict that AI will not “meaningfully outperform trained humans (particularly teams of humans and especially if augmented with simple automated tools)” at forecasting geopolitical events or at persuading people to act against their own self-interest, and that “an increasing percentage of human jobs and tasks will be related to AI control.” Interpretation · contested evidence LD01-SL-C02
- Reading a 2024 study that found 40% of US adults used generative AI, Narayanan and Kapoor note that infrequent use translated into only 0.5%–3.5% of work hours, and argue that, depending on how adoption is measured, generative AI adoption may have been much slower than personal-computer adoption. Interpretation · contested evidence LD01-SL-C03
- Yiu, Kosoy and Gopnik describe large language models as “cultural technologies” and “efficient imitation engines,” and predicted that they would be almost as capable as humans at linking familiar objects but fail at tool innovation; in their study, children chose the functionally right novel tool 85.2% of the time and adults 95.7%, while the models tested were “much less capable.” Estimate and interpretation · emerging evidence LD01-SL-C04
- Mitchell and Krakauer report that in a 2022 survey of natural-language-processing researchers, 51% of 480 respondents agreed and 49% disagreed that a generative model trained only on text could, with enough data and compute, understand natural language “in some non-trivial sense”; they argue for an extended science of intelligence that recognizes distinct modes of understanding. Estimate and interpretation · contested evidence LD01-SL-C05
- Shanahan, McDonell and Reynolds argue that a dialogue agent is best understood as role-play by a simulator, not as an entity with its own agenda: an agent frozen before the 2022 World Cup that names France as the current champion behaves as a simulator role-playing a knowledgeable person would, not as a helpful person would. Interpretation · emerging evidence LD01-SL-C06
- Millière and Buckner take the view that language models are “merely sophisticated mimics” as a null hypothesis to be tested; they find that models show “at least local task generalization, if not broad task generalization,” but “often seem to have a hard time pushing beyond the range of tasks well-represented in their training data,” and say their part in cumulative cultural learning “appears contingent on further advancements” that “might lie beyond the reach of current architectures.” Interpretation · contested evidence LD01-SL-C07
- Esposito argues that comparing algorithms with human intelligence is misleading and proposes thinking of them “not in terms of artificial intelligence but in terms of artificial communication”: if machines contribute to social intelligence, “it will not be because they have learned how to think like us but because we have learned how to communicate with them.” Interpretation · emerging evidence LD01-SL-C08
- Drawing on Luhmann, Esposito, Wittgenstein and Brandom, Fenoglio argues that what language models lack is a position in the practices of commitment and assessment, not a behavioural capacity, so that “no behavioral evidence can satisfy a positional criterion, and this is by design rather than by evasion.” Interpretation · emerging evidence LD01-SL-C09
- Chan and colleagues treat agency in algorithmic systems as a matter of degree, raised by four characteristics: underspecification, directness of impact, goal-directedness and long-term planning; they stress that recognizing such agency “does not absolve or shift” human responsibility. Interpretation · emerging evidence LD01-SL-C10
- Kolt notes that the US Restatement (Third) of Agency holds that a computer program “is not capable of acting as a principal or an agent as defined by the common law” and that “at present, computer programs are instrumentalities” of their users, and he uses agency law and principal–agent economics to analyse AI agents’ problems of information asymmetry, discretionary authority and loyalty. Fact and interpretation · emerging evidence LD01-SL-C11
- Kasirzadeh and Gabriel characterize AI agents along four dimensions, autonomy, efficacy, goal complexity and generality, each with gradations, and draw “agentic profiles” for systems including AlphaGo, ChatGPT-3.5, Claude 3.5 Sonnet with tool use and a self-driving car. Interpretation · emerging evidence LD01-SL-C12
- Anthropic distinguishes “workflows,” in which language models and tools are orchestrated through predefined code paths, from “agents,” in which models “dynamically direct their own processes and tool usage”; it says agents are “typically just LLMs using tools based on environmental feedback in a loop.” Fact · emerging evidence LD01-SL-C13
- Eloundou, Manning, Mishkin and Rock conclude that language models “exhibit traits of general-purpose technologies,” estimating that about 80% of the US workforce could have at least 10% of their tasks affected and about 19% at least half; they “do not make predictions about the development or adoption timeline.” Estimate and interpretation · emerging evidence LD01-SL-C14
- In nationally representative US surveys from late 2024, nearly 40% of adults aged 18–64 used generative AI; 27% of workers used it for work two years after ChatGPT’s release, against 25% using computers at work in 1984, three years after the IBM PC, while only 1%–5% of all work hours were assisted by it. Estimate · emerging evidence LD01-SL-C15
- A GPT-style model trained only on sequences of legal Othello moves, with no rules or board given, made illegal moves 0.01% of the time; its activations encoded the board state, and editing that internal state changed its predicted moves as the edited board implied. Estimate · emerging evidence LD01-SL-C16
- Linear probes on Llama-2-70B’s activations predicted the locations of world places with out-of-sample R² of 0.911 and the death years of historical figures with 0.835; the authors say these representations “do not constitute a dynamic causal world model in their own right.” Estimate · emerging evidence LD01-SL-C17
- McCoy and colleagues predicted that, because language models are trained to predict the next word, their accuracy should depend on how probable the task, input and output are, even where probability should not matter; GPT-4 decoded a simple cipher with 51% accuracy when the output was a high-probability word sequence and 13% when it was low-probability. Estimate and interpretation · emerging evidence LD01-SL-C18
- Models fine-tuned on facts of the form “A is B” failed to answer “B is A”; GPT-4 named the parent of 79% of 1,000 celebrities, but on the 1,573 child–parent pairs this produced it named the celebrity from the parent only 33% of the time. Estimate · emerging evidence LD01-SL-C19
- Adding a clause that seems relevant but does not change the answer to grade-school maths problems cut accuracy by up to 65 percentage points across 25 models; GPT-4o scored 95.2% on the full GSM8K test set and 63.1% with the clause added, and OpenAI’s o1-preview 94.9% and 77.4%. Estimate · emerging evidence LD01-SL-C20
- Off-the-shelf ChatGPT and GPT-4 reached only 55% and 59% accuracy on 3-digit by 3-digit multiplication, and accuracy on compositional tasks fell towards zero as complexity grew; the authors hypothesize that transformers reduce multi-step reasoning to “linearized path matching.” Estimate and interpretation · emerging evidence LD01-SL-C21
- On 11 tasks rewritten with changed premises, such as arithmetic in base 9 or 1-based indexing in code, language models performed above chance but their performance “substantially and consistently degrades” relative to the familiar versions, which the authors attribute partly to overfitting to the default tasks. Estimate and interpretation · emerging evidence LD01-SL-C22
- METR researchers estimate that the length of software tasks, measured in human time, that AI agents complete with 50% success doubled about every 207 days (95% interval 166–240) from 2019 to 2025, while performance was “much lower on less structured, ‘messier’ tasks.” Estimate · emerging evidence LD01-SL-C23
- After METR revised its task suite and moved its evaluation software from Vivaria to Inspect in January 2026, its estimate of the post-2023 doubling time of AI time horizons moved from 165 to 131 days; METR says the trend “is somewhat sensitive to task composition” and that it had measured human baseline times for only 5 of its 31 tasks of eight hours or more. Estimate · emerging evidence LD01-SL-C24
- In a US online sample of 300 adults surveyed in July 2023, 67% attributed some possibility of conscious experience to ChatGPT; people who had used it gave higher ratings (mean 29.59 on a 1–100 scale) than those who had not (19.37). Estimate · emerging evidence LD01-SL-C25
- Comparing 15,881 human–ChatGPT dialogues with 10,784 human–human dialogues, Mitiaeva and Xiao report “a consistent dissociation: AI produces the surface features of cooperative communication without the underlying social architecture”; for example, hedging and softening that go with more accommodation between people go with less alignment when the AI produces them. Estimate and interpretation · emerging evidence LD01-SL-C26
- In a Stanford audit of 844 tasks in 104 US occupations, workers expressed positive attitudes toward automation by AI agents for 46.1% of their tasks, most often to free time for high-value work, but “generally prefer higher levels of human agency” than experts deemed “technologically necessary.” Estimate · emerging evidence LD01-SL-C27
- Bommasani and colleagues named “foundation models” for their “critically central yet incomplete character,” argued that their significance lies in “emergence and homogenization,” warned that homogenization “creates single points of failure,” and predicted that the barrier to entry for developing them “will continue to rise.” Interpretation · emerging evidence LD01-SL-C28
- Van der Vlist, Helmond and Ferrari describe “Big AI” as the structural convergence of AI and Big Tech, arguing that AI is “an actual technology stack encompassing infrastructure, models, applications” and that Amazon, Microsoft and Google have integrated all three into their cloud offerings. Interpretation · emerging evidence LD01-SL-C29
- Luitse and Denkena examine large language models as “infrastructural components for AI research and development,” arguing that they are intertwined with big technology companies’ business models and, using GPT-3, describing “corporate efforts to commodify” them “through paid API access and exclusive licensing.” Interpretation · emerging evidence LD01-SL-C30
- Krokowski and Hirsch-Kreinsen treat AI as “a paradigmatic case of a promising technology” whose future-oriented narratives mobilize resources while application problems and functional limits are “frequently overlooked,” and suggest that its future “could be characterized more by” a path of “routinization” than by the vision of artificial general intelligence. Interpretation · emerging evidence LD01-SL-C31
- Comparing scientists’ citation networks with market data for the dot-com era (1994–2001) and the AI era (2017–2024), Chelikavada and Bennett found that dot-com patterns “did not definitively predict the rise and fall of an AI bubble,” and that “the data seems to suggest two possibilities”: an unprecedented kind of bubble or no bubble. Interpretation · emerging evidence LD01-SL-C32
- The Bank for International Settlements wrote in June 2024 that adoption of general-purpose technologies “typically follows a J-curve: it is slow at first, but eventually accelerates,” that AI shows features suggesting “an even steeper J-curve,” and that if AI is a true general-purpose technology that raises “total factor productivity in all industries to a similar extent,” “the demand for labour is set to increase across the board.” Interpretation · emerging evidence LD01-SL-C33
- The BIS found in December 2025 that the “Magnificent 7” technology stocks had grown to nearly 35% of S&P 500 market capitalization, from about 20% in November 2022, and called the recent rally in their prices “dissimilar to the dotcom bubble” because it rested on realized earnings growth as well as optimistic expectations, while noting price–earnings multiples approaching the top 10% of their history and concerns about stretched valuations. Estimate and interpretation · emerging evidence LD01-SL-C34
- China’s State Council set targets in August 2025 that the penetration rate of new-generation smart terminals and “intelligent agents” (智能体) exceed 70% by 2027 and their adoption rate exceed 90% by 2030, and said that implementing “AI+” will bring “new infrastructure” and accelerate “new quality productive forces” (新质生产力) (CSET translation). Fact · emerging evidence LD01-SL-C35
- The European Commission’s guidelines note that the AI Act’s definition of a “general-purpose AI model” “does not set out specific criteria” and adopt an indicative criterion: training compute above 10²³ FLOP and the ability to generate language, images or video, while calling compute “an imperfect proxy for generality and capabilities.” Fact · emerging evidence LD01-SL-C36
- Testing OpenAI’s o1, a model trained to reason step by step, the authors of “Embers of Autoregression” found that it substantially outperformed earlier models “in many cases” but “still displays the same qualitative trends”: its accuracy at decoding a shift cipher ranged from 47% in the lowest-probability case to 92% in the highest, and it used more reasoning tokens on low-probability examples. Estimate · emerging evidence LD01-SL-C37
LD-02: Aggregate null or distributional shift? Employment
- Exposure, adoption, actual use, task productivity, firm productivity, labour demand, hiring, employment stock and wages are distinct empirical objects; evidence at one stage does not establish the next. Interpretation · rapidly maturing evidence LD02-HQ-C01 Also cited by LD-03.
- In 11 AI-exposed Danish occupations, Humlum and Vestergaard estimate precise null effects on earnings and recorded hours two years after ChatGPT, ruling out effects larger than 2%. Estimate · rapidly maturing evidence LD02-HQ-C02 Also cited by LD-03, LD-04, LD-08.
- Danish firms that began using AI in 2023 had roughly 11% lower relative employment growth by late 2025 than comparable non-adopters, mainly through lower recruitment, while economy-wide employment remained relatively unaffected. Estimate · emerging evidence LD02-HQ-C03
- By June 2026, US workers aged 22–25 in highly AI-exposed occupations were about 19% below the employment path of similarly aged workers in less-exposed occupations, with the divergence driven mainly by reduced hiring rather than separations. Estimate · contested and rapidly maturing evidence LD02-HQ-C04 Also cited by LD-04, LD-08.
- In US Census administrative data, regression-adjusted employment of workers aged 22–24 in the most AI-exposed industry-state cells fell 12% over 10 quarters after ChatGPT, with reduced hiring the primary mechanism. Estimate · emerging evidence LD02-HQ-C05
- US graduates from the most AI-exposed tenth of college majors experienced a 5 percentage-point decline in initial employment and a 13% decline in initial earnings relative to less-exposed majors after ChatGPT. Estimate · emerging evidence LD02-HQ-C06
- In a 41-country working paper, AI-adopting multinational affiliates shift toward more senior workforces: junior share falls by 1.9 percentage points, driven mainly by 6.7% senior employment growth rather than a statistically clear decline in junior headcount. Estimate · emerging evidence LD02-HQ-C07
- In a UK exposure-based working paper, highly LLM-exposed firms experienced about 4.5% lower employment, junior employment about 5.8% lower, and highly exposed occupations about 23.4% fewer job postings over the study period. Estimate · emerging evidence LD02-HQ-C08
- Norwegian private-sector registers show near-zero employment differences between the most- and least-AI-exposed occupations since late 2022, and the young-worker gap across full exposure quintiles is an imprecise zero. Estimate · emerging evidence LD02-HQ-C09 Also cited by LD-04.
- In the US Census 2026 AI supplement, 18% of firms reported using AI in a business function, while only 2% self-reported AI-related employment decreases during the reference period. Fact · rapidly maturing evidence LD02-HQ-C10 Also cited by LD-03.
- Instrumented AI adoption among more than 12,000 firms raises labour productivity by about 4% with no adverse short-run firm-level employment effect. Estimate · emerging evidence LD02-HQ-C11 Also cited by LD-03.
- In the final QJE study of one customer-support firm, AI assistance raised issues successfully resolved per hour by 15% on average, with substantially larger gains for lower-skill and less-experienced agents. Estimate · mature evidence LD02-HQ-C12 Also cited by LD-03, LD-04.
- Across three company field experiments with 4,867 software developers, access to an AI coding assistant increased completed tasks by 26.08% on average; less-experienced developers adopted the tool more and gained more. Estimate · mature evidence LD02-HQ-C13 Also cited by LD-03, LD-04.
- In a preregistered experiment with 758 BCG consultants, GPT-4 users completed 12.2% more inside-frontier tasks, worked 25.1% faster and produced roughly 30–34% higher human-graded quality, but were 19 percentage points less likely to be correct on a task outside the tested frontier. Estimate · mature evidence LD02-HQ-C14 Also cited by LD-03, LD-04.
- Peer-reviewed online-platform evidence shows that freelancers in occupations highly exposed to generative AI experienced short-run declines in employment and earnings after model releases. Estimate · rapidly maturing evidence LD02-HQ-C15 Also cited by LD-04.
- US work-use intensity differs sharply by measure: in Q2 2026 about 45.2% of employed adults reported current generative-AI use for work, 39.2% reported use in the last week and 13.7% reported use every workday. Estimate · rapidly maturing evidence LD02-HQ-C16 Also cited by LD-03.
- The current R05 evidence base does not establish a causal pass-through from AI-induced productivity gains into worker wages. Unknown and interpretation · emerging evidence LD02-HQ-C17 Also cited by LD-03.
- As of September 2026, Budget Lab monitoring finds no clear broad US employment, unemployment or occupational-mix footprint attributable to AI, while explicitly warning that effects concentrated in narrow subgroups may be missed by broad CPS measures. Estimate · rapidly maturing evidence LD02-HQ-C18
LD-03: Organizational absorption: Why micro gains do not become macro outcomes
- In S&P 500 annual reports, 11% of firms had AI deeply integrated into their business processes in 2025 and a further 10% used it in production, against 5% in 2022; among non-technology firms profit margins dipped at the pilot stages and rose with deep integration, while revenue per employee showed no difference at the advanced stages. Estimate · emerging evidence LD03-SL-C01
- In nationally representative US surveys to late 2024, workers' self-reported time savings from generative AI came to 1.4% of all work hours, which the authors translate into a potential aggregate productivity gain of about 1.2%; they note that worker adoption is running ahead of firm adoption. Estimate · rapidly maturing evidence LD03-SL-C02
- In May 2026, 45% of US workers used generative AI for work, but adoption was widespread and shallow: within most occupations and tasks fewer than half of workers used it, no task exceeded 70% adoption, and mean weekly hours among work users had stayed flat since late 2024. Estimate · rapidly maturing evidence LD03-SL-C03
- In a survey of nearly 6,000 senior executives in the US, UK, Germany and Australia, 69% of firms actively used AI, but nine in ten executives reported no impact on employment or productivity over the previous three years; they expected AI to raise their firm's productivity by 1.4% over the next three. Estimate · emerging evidence LD03-SL-C04
- In a survey of nearly 750 corporate executives, the authors find positive labour-productivity gains from AI, reflecting revenue-based total factor productivity and expected to strengthen in 2026, but report a productivity paradox: executives perceive larger gains than are measured, which the authors attribute to a delay in revenue realizations. Estimate · emerging evidence LD03-SL-C05
- In a six-month randomized trial of an integrated AI assistant across 66 firms and 7,137 knowledge workers, workers who used it spent about two fewer hours a week on email, but the authors detect no shift in the quantity or composition of workers' tasks from individual-level provision, and no change in time spent in meetings. Estimate · emerging evidence LD03-SL-C06
- In 27 interviews in regional Chinese newsrooms, journalists used generative AI frequently in their own tasks, but its use stayed disconnected from team workflows, hindered by structural barriers and a reluctance to share practices, and no formal mechanism coordinated individual use. Fact · emerging evidence LD03-SL-C07
- In a case study of three German organizations, generative AI was used mainly as a personal assistant rather than changing team processes, the effort of validating its output partly offset efficiency gains, and policies written in response to regulation without regard to actual use led to gaps between policy and practice and to shadow IT. Fact · emerging evidence LD03-SL-C08
- In paired interviews with 54 developers in 27 teams at a large software company committed to AI adoption, the authors identify a 'Productivity Pressure Paradox': management expected rapid productivity gains without supporting learning, which left developers without the time to build the skills that would save time. Fact and interpretation · emerging evidence LD03-SL-C09
- The authors of a survey of 576 public managers in seven EU member states describe public servants using generative AI to draft reports, refine policy proposals and generate content, often without explicit approval or institutional policies, which they call 'shadow use'; the abstract does not say how much of this the survey itself measured. Estimate · emerging evidence LD03-SL-C10
- In 806 open-source projects that adopted the Cursor AI editor, compared with 1,380 matched projects, development velocity rose sharply but only for one or two months, while static-analysis warnings rose 30% and code complexity 42% and stayed higher; the added complexity and warnings in turn slowed later development. Estimate · emerging evidence LD03-SL-C11
- On 18 real tasks from two open-source repositories, an AI agent's code passed the maintainers' tests 38% of the time, but none of the 15 pull requests METR reviewed by hand could be merged as is; METR roughly estimated 26 minutes of human repair even for those that passed, about a third of the original human time. Estimate · emerging evidence LD03-SL-C12
- A METR note shows that when workers shift towards the tasks AI speeds up, time saved measured on the new mix of tasks overstates the value gained, which lies between the speedup on the old task mix and on the new one under the note's assumptions; the authors suggest this may help explain why reported gains exceed aggregate data. Interpretation · emerging evidence LD03-SL-C13
- In a METR survey of 349 technical workers in early 2026, the median respondent reported that AI tools made them about 3 times faster but changed the value of their work by only 1.4 to 2 times; METR warns that such self-reports have overstated measured effects in its own trials. Estimate · emerging evidence LD03-SL-C14
- In randomized experiments run by a large cross-border online retail platform in 2023–2024, building generative AI into seven business workflows changed sales by between nothing detectable and 16.3%, depending on how much the AI added to the firm's existing practice; the four workflows with gains imply about $5 a year per consumer. Estimate · emerging evidence LD03-SL-C15
- In a preregistered field experiment with 791 professionals at Procter & Gamble, individuals working with AI matched the performance of teams working without it on product-innovation challenges, and AI helped bridge the gap between R&D and commercial staff. Estimate · emerging evidence LD03-SL-C16
- A task-based macroeconomic model projects that AI will raise US total factor productivity by no more than 0.66% over ten years, and by less than 0.53% once harder-to-learn tasks are accounted for. Interpretation · contested evidence LD03-SL-C17
- Brynjolfsson, Rock and Syverson model how general-purpose technologies such as AI need large, poorly measured intangible investments, so measured productivity growth is first understated and later overstated (a 'productivity J-curve'); adjusting for intangibles tied to computers put US total factor productivity 15.9% above official measures by the end of 2017. Interpretation · mature evidence LD03-SL-C18
- Reviewing evidence on computers, Brynjolfsson and Hitt argue that organizational investments, such as new work practices, strongly shape the value of information-technology investments, and that the benefits are often intangible and hard to measure. Interpretation · mature evidence LD03-SL-C19
- Across European and US industries, a 10-percentage-point rise in AI adoption is associated with 2 to 5 points of extra cumulative productivity growth in Europe to 2024 and about 3 points relative to trend in the US since 2022, with no such association in the European placebo period (2015–2019); the authors stress that these correlations are not causal. Estimate · emerging evidence LD03-SL-C20
- Across 56 economies and 16 industries, industries more exposed to generative AI grew faster in 2022–23 where countries were better prepared for AI, which the authors read as generative AI likely benefiting advanced economies more than emerging ones and widening global income disparities in the near term. Estimate · emerging evidence LD03-SL-C21
- In 22 US state agencies using chatbots, the factors that drove adoption (ease of use, leadership, external shocks such as COVID-19) overlapped only partly with those that shaped implementation (building and maintaining the knowledge base, skills, resources, cross-agency coordination, external shocks and citizens' expectations). Fact · emerging evidence LD03-SL-C22
- In a field study of three radiology departments in a major US hospital using diagnostic AI, only one department consistently incorporated AI results into professionals' final judgements; in the others, opaque AI results that diverged from experts' own views were often not engaged with. Fact · mature evidence LD03-SL-C23
LD-04: Levelling or hollowing? Skill compression and the reproduction of expertise
- In a preregistered online experiment with 453 college-educated professionals doing occupation-specific writing tasks, ChatGPT access cut time taken by 40% and raised output quality by 18%, and inequality between workers decreased. Estimate · rapidly maturing evidence LD04-SL-C01
- In a field experiment with 640 Kenyan small-business owners, a GPT-4 business adviser on WhatsApp had no average effect on revenue or profit, but its effect for owners who started below the median was more than 0.20 standard deviations lower than for those above it. Estimate · emerging evidence LD04-SL-C02
- In an online experiment with 1,174 Argentine adults, AI access shrank the performance gap between education groups on a business problem-solving task from 0.548 to 0.139 standard deviations; in a follow-up module without AI in the same session, a gap of 0.200 standard deviations remained. Estimate · emerging evidence LD04-SL-C03
- In a preregistered online experiment with a sample of the UK working-age population, ChatGPT use raised productivity on every task tested, more on complex and less ambiguous tasks, but compression between tasks was limited and ChatGPT use did not change productivity differences between gender, age, education or occupation groups. Estimate · emerging evidence LD04-SL-C04
- In two experiments on creative tasks (flash fiction and song composition), generative AI raised the returns to general human capital such as cognitive ability and education and reduced the value of domain-specific expertise. Estimate · emerging evidence LD04-SL-C05
- In an experiment where participants studied a technical domain with or without an LLM, gains were highly uneven and were predicted not by grades or prior knowledge but by skill at eliciting, filtering and verifying AI output; participants low on that skill saw limited or negative returns, and a concept-map scaffold reduced the spread. Estimate · emerging evidence LD04-SL-C06
- In a 2025 randomized trial with 16 experienced open-source developers working on 246 tasks in their own projects, allowing AI tools increased completion time by 19%, although the developers estimated it had cut their time by 20%; METR later called its 2026 follow-up data an unreliable signal because of selection. Estimate · contested evidence LD04-SL-C07
- A 2026 meta-analysis of 23 programming studies found an overall productivity effect of g = 0.33, large in controlled experiments (g = 0.73) but not distinguishable from zero in enterprise or open-source settings, and no significant effect on learning (g = 0.14). Estimate · emerging evidence LD04-SL-C08
- In the customer-support firm where an AI assistant raised productivity most for less-skilled agents, agents who had used it still outperformed their own pre-AI baseline during software outages when it gave no suggestions, more so after longer exposure and closer adherence to its suggestions. Estimate · emerging evidence LD04-SL-C09
- In a three-month randomized trial with 133 patent lawyers, AI assistance raised graded drafting quality by 0.38 standard deviations at 90 days; on a later task done without AI, treated seniors scored 0.45 standard deviations higher than control seniors, treated juniors no higher (−0.03), and the senior–junior difference was statistically significant. Estimate · emerging evidence LD04-SL-C10
- In a field experiment with nearly 1,000 Turkish high-school students, practising with a standard GPT-4 interface raised practice scores by 48% but left students 17% worse than controls on a closed-book exam later in the same session, while a tutor version with safeguards raised practice by 127% and left exam scores indistinguishable from controls. Estimate · emerging evidence LD04-SL-C11
- In a randomized experiment with 52 Python programmers learning an unfamiliar library, those given an AI assistant scored 17% lower on a later quiz of concepts, code reading and debugging, with no significant speed gain; participants who delegated code to the AI scored lowest. Estimate · emerging evidence LD04-SL-C12
- In two preregistered lab experiments and a field study, LLM access had no overall effect on learning: students who used it to generate solutions covered more topics but understood each less, students who asked it for explanations understood more, and the gap between students with low and high prior knowledge widened. Estimate · emerging evidence LD04-SL-C13
- In an observational study at four Polish centres, experienced endoscopists’ adenoma detection in colonoscopies done without AI fell from 28.4% to 22.4% in the three months after an AI detection system was introduced; the study is observational, and its authors list differences between the cohorts and a near doubling of colonoscopy volume after AI was introduced among its limitations. Estimate · contested evidence LD04-SL-C14
- In US résumé data covering 65 million workers, firms that began integrating generative AI saw junior employment fall by about 9% relative to senior employment after six quarters, compared with non-adopters, mainly because they hired fewer juniors, while junior promotion rates rose slightly. Estimate · emerging evidence LD04-SL-C15
- In Swedish job-ad and register data, the fall in job postings since 2022 tracks monetary tightening rather than AI, but within the same employers, workers aged 22–25 in highly AI-exposed occupations were 5.5% lower by early 2025 than in less-exposed occupations, while workers over 50 were 1.3% higher. Estimate · emerging evidence LD04-SL-C16
- Across 205 US occupations, those more exposed to AI already had higher experience premiums, and since 2022 higher AI exposure has gone with weaker wage growth only in occupations where the experience premium is low. Estimate · emerging evidence LD04-SL-C17
- A two-year comparative ethnography of robotic and open surgery, with interviews at 13 US teaching hospitals, found that approved routes for learning failed in robotic surgery; only a minority of trainees came to competence, through tolerated but unapproved practices, and these practices led to hyperspecialisation and a shrinking supply of experts relative to demand. Fact and interpretation · mature evidence LD04-SL-C18
- A theoretical model in which novices acquire tacit knowledge by working with experts shows that automating entry-level work can raise output yet lower long-run growth, even without cutting entry-level employment, by moving novices away from the most productive experts. Interpretation · emerging evidence LD04-SL-C19
- A model of knowledge work predicts that AI which only assists (a copilot) benefits the least knowledgeable workers, while autonomous AI benefits the most knowledgeable, so the direction of skill compression can reverse as AI becomes more autonomous. Interpretation · emerging evidence LD04-SL-C20
- In a 2022 experiment run by Microsoft and GitHub researchers, freelance developers randomly given GitHub Copilot completed a standard programming task 55.8% faster than those without it; the authors report larger gains for less experienced developers, but that difference is not statistically significant at the 5% level. Estimate · emerging evidence LD04-SL-C21
LD-05: Discretion curtailed, relocated or encoded?
- The presence of a human reviewer does not establish meaningful oversight; effective discretion depends on what the human can see, whether they can depart, how often they do so, and the organisational cost of disagreement. Interpretation · rapidly maturing evidence LD05-HQ-C01
- Human override has no fixed fairness direction: it can reduce or amplify disparity depending on the human baseline, institutional setting and the margin being overridden. Interpretation · rapidly maturing evidence LD05-HQ-C02
- In Poland’s unemployment profiling system, formal profile changes were extremely rare—0.58% in the cited early-period data—while discretion persisted through interviewing, interpretation and data entry. Fact · rapidly maturing evidence LD05-HQ-C03
- In Allegheny County, the authors’ corrected analysis supports the conclusion that worker judgement reduced racial disparity relative to an AFST-only counterfactual; supervisors also screened out children marked as mandatory screen-ins (21.0% in the original analysis; the authors later said all of that analysis’s numbers changed and gave no corrected share). Fact · rapidly maturing evidence LD05-HQ-C04
- Kentucky’s mandatory risk-assessment policy changed bail practice but its effects eroded over time; separate working-paper evidence indicates that departures toward harsher conditions were more common for Black defendants. Fact · rapidly maturing evidence LD05-HQ-C05
- DWP’s Universal Credit Advances model selects claims for additional fraud review while the final decision remains human; the reviewer is not shown the risk rating or told that the referral was model-generated. Fact · rapidly maturing evidence LD05-HQ-C06
- Non-UK nationals were over-referred by the DWP Universal Credit Advances model in both audited periods; some older groups were over-referred in 2024–25, creating additional verification burden before the final decision. Fact · rapidly maturing evidence LD05-HQ-C07
- DUO’s risk-based selection for home visits indirectly discriminated against students with a migration background; the risk-profile selection was abandoned and current home-visit selection is random. Fact · rapidly maturing evidence LD05-HQ-C08
- The Dutch Tax Administration’s risk-classification model used Dutch/non-Dutch nationality status as a selection indicator during the verified period; the regulator later imposed a €2.75 million fine for unlawful and discriminatory nationality processing. Fact · mature evidence LD05-HQ-C09
- Algorithmically mediated selection can create material scrutiny and administrative burden before a final legal decision exists. Interpretation · rapidly maturing evidence LD05-HQ-C10
- Försäkringskassan withdrew an ML risk profile for parental-benefit controls during Swedish data-protection supervision in 2025; the regulator did not make a discrimination finding. Fact · emerging evidence LD05-HQ-C11
- CNAF’s DMDE 2026 is a targeting model rather than an automatic final decision-maker; its source code and model documentation are public and the redesign excludes several sensitive/profile variables. Fact · emerging evidence LD05-HQ-C12
- Robodebt shows that formal review institutions can exist without reliable system-level repair; the Royal Commission recommended stronger escalation, review visibility, debt-review protections and automated-decision governance. Fact and interpretation · mature evidence LD05-HQ-C13
- An Austrian administrative court (BVwG, September 2025) assigned the substantive final AMAS assessment to counsellors and rejected classification of the process as solely automated; this legal finding does not establish effective discretion in practice. Fact and interpretation · rapidly maturing evidence LD05-HQ-C14
- A 2020 judicial-review challenge alleging race discrimination and feedback-loop effects was followed by the Home Office’s withdrawal of its original visa Streaming Tool before a merits ruling; successor routing systems later remained in production. Fact · rapidly maturing evidence LD05-HQ-C15
- Qualitative evidence from UK social care shows that generative-AI transcription and summarisation can introduce hallucinations, omissions and changes in wording into official records, shifting checking and accountability work onto social workers. Fact · emerging evidence LD05-HQ-C16
- For AI-mediated administrative processes, effective contestability depends partly on process provenance—knowing whether a system selected the case, influenced routing, or generated part of the evidentiary record. Interpretation · emerging and rapidly maturing evidence LD05-HQ-C17
LD-06: The evaluation apparatus: Can benchmarks, leaderboards and audits carry the weight placed on them?
- The 2026 AI Index reports that a review found invalid-question rates in nine benchmarks “ranging from 2% on MMLU Math to 42% on GSM8K”, citing Truong et al. (2025); in that paper’s figure, 2 and 42 are the numbers of confirmed invalid questions among the 50 that a statistical screen flagged in each benchmark (“2% corresponds to one question”, so each question is 2 points on that scale), and the paper puts GSM8K’s benchmark-wide share of invalid questions at about 5%, citing Vendrow et al., while its own Platinum-based labelling marks 88 of 997 (8.8%) invalid. Fact · emerging evidence LD06-SL-C01
- The 2026 AI Index states that the benchmark scores in its technical-performance chapter come from leaderboards, public repositories and company disclosures, including papers, blog posts and product releases, and that it “assumes that company-reported results are accurate”, while its introduction says that “independent testing does not always confirm what developers report”. Fact · mature evidence LD06-SL-C02
- Re-checking 15 popular benchmarks wherever a frontier model failed an item, MIT researchers found many “riddled with errors (e.g., 5% of GSM8K)”, and for most benchmarks more than half of model failures could be traced to label noise rather than to the model; they describe their counts as a lower bound. Estimate · emerging evidence LD06-SL-C03
- Re-annotating 5,700 questions across all 57 subjects of MMLU, a widely used knowledge benchmark, researchers estimated that 6.49% of its questions contain errors, 57% of those analysed in the Virology subset, and that the errors change how models rank. Estimate · emerging evidence LD06-SL-C04
- In a review by 29 experts of 445 language-model benchmarks from leading machine-learning and NLP venues, only 16.0% used any statistical testing and about half (53.4%) presented evidence that they measure what they claim to measure; 12.3% relied wholly and another 27.0% partly on convenience sampling. Estimate · rapidly maturing evidence LD06-SL-C05
- Scoring 24 AI benchmarks against 46 quality criteria, BetterBench rated MMLU lowest (5.5) and GPQA much higher (11.0), yet model announcements for GPT-4, Claude 3 and Gemini reported both without acknowledging their limitations or quality differences. Estimate · emerging evidence LD06-SL-C06
- On GSM1k, a newly written set of grade-school maths problems built to mirror GSM8K, some model families showed accuracy drops of up to 8% against GSM8K, consistent with partial memorization, but many models, especially frontier ones, showed minimal signs of overfitting. Estimate · emerging evidence LD06-SL-C07
- An audit of agentic benchmarks found flaws in task setup or scoring, for example insufficient test cases in SWE-bench Verified and empty responses counted as successes in TAU-bench, that can under- or overestimate agents’ performance by up to 100% in relative terms. Estimate · emerging evidence LD06-SL-C08
- Meaning-preserving changes to prompt formatting changed few-shot accuracy by up to 76 percentage points for LLaMA-2-13B, and which formats work best correlated only weakly between models, which the authors say calls into question comparing models on one fixed prompt format. Estimate · mature evidence LD06-SL-C09
- Using item response theory, a method from psychometrics, researchers showed that 100 curated items suffice to estimate a model’s performance on MMLU’s 14,000, and released small versions of several popular benchmarks that reproduce the original results. Estimate · emerging evidence LD06-SL-C10
- An analysis of MMLU found that 28% of its questions require culturally sensitive knowledge and that 84.9% of its geography questions concern North America or Europe; model rankings change depending on whether the culturally sensitive questions are included, so machine-translated versions distort multilingual comparisons. Estimate · emerging evidence LD06-SL-C11
- When four maintainers of three repositories reviewed 296 AI-written pull requests that had passed SWE-bench Verified’s automated tests, roughly half would not have been merged; on average, maintainer merge rates, scored relative to the 68% at which the same maintainers would merge the human-written patches actually merged, ran about 24 percentage points below the benchmark scores. Estimate · emerging evidence LD06-SL-C12
- In interviews with 19 people who use or decided against using AI benchmarks, academics generally saw public benchmarks as suitable for tracking research progress, while in product and policy work benchmarks, even internally built ones, were often found inadequate for substantive decisions. Estimate · emerging evidence LD06-SL-C13
- The International AI Safety Report 2026 finds an “evaluation gap”: performance on pre-deployment tests does not reliably predict real-world utility or risk, models “now regularly identify evaluation prompts as tests”, and although 12 companies published or updated frontier safety frameworks in 2025, there are no standardised external audits of their compliance. Interpretation · rapidly maturing evidence LD06-SL-C14
- An audit of Chatbot Arena’s data from January 2024 to April 2025 found undisclosed private testing (27 private Meta variants before the Llama 4 release), data skewed towards proprietary providers (Google and OpenAI about 20% each), and 205 of 243 public models silently deprecated, and concluded that rankings partly reflect adaptation to the Arena rather than general model quality. Estimate · contested evidence LD06-SL-C15
- Responding to “The Leaderboard Illusion”, LMArena said providers may test several private variants, promised to mark scores provisional until 2,000 fresh votes when more than 10 variants are tested, and disputed the paper’s figures, estimating the gain from private testing at about 11 Elo points; the paper’s 8.8% open-source figure that it disputed was a share of data, not of models. Fact and interpretation · contested evidence LD06-SL-C16
- The 2025 Foundation Model Transparency Index found that the average transparency score of major AI developers fell from 58 in 2024 to 40 out of 100, and that no company adequately disclosed how far its training data overlap with the evaluations it is tested on; the authors note that changed indicators and companies may contribute to the decline. Estimate · rapidly maturing evidence LD06-SL-C17
- Across 20 datasets with human annotations, the best language-model judge (GPT-4o) reached an average Spearman correlation of only 0.50 with human judgements on the datasets with graded annotations (categorical ones were scored by Cohen’s kappa), and agreement varied widely by dataset, by the property judged and by whether humans or models wrote the text; the authors advise validating judges against task-specific human judgements before use. Estimate · rapidly maturing evidence LD06-SL-C18
- The team behind Chatbot Arena found that GPT-4 as a judge agreed with human preferences over 80% of the time, the same level as humans with each other, but also that it gave consistent verdicts when answer order was swapped in only 65% of cases; it gave its own answers a 10% higher win rate, though the authors say their limited data cannot determine whether this is a self-enhancement bias. Estimate · rapidly maturing evidence LD06-SL-C19
- On a benchmark of 1,000 transcripts from 61 datasets, frontier models distinguished evaluation transcripts from real use well above chance (Gemini 2.5 Pro: AUC 0.83, against 0.92 for two of the paper’s authors rating every transcript, a baseline the authors treat as an approximate upper bound), more so in agentic settings than in chat. Estimate · emerging evidence LD06-SL-C20
- Testing an anti-scheming training method on OpenAI models across 26 evaluations, Apollo Research and OpenAI found covert actions fell from 13% to 0.4% for o3, but the models’ reasoning often showed awareness of being evaluated, they report causal evidence that awareness reduced covert behaviour, and the authors could not exclude that the reduction was partly driven by it. Estimate · emerging evidence LD06-SL-C21
- In an auditing game in which a red team trained some models to underperform deliberately on evaluations, the blue team could not reliably tell these sandbagging models from benign ones; black-box methods were defeated by imitation of a weaker model, and training-based elicitation recovered full performance but sometimes also raised benign models’ scores. Estimate · emerging evidence LD06-SL-C22
- In their joint pre-deployment test of OpenAI’s o1, the US and UK AI Safety Institutes stated that their findings were preliminary and partial, that they tested an early version without the full mitigations after a limited period of access and after fixing tool-use problems with their own scaffolding, and that the results should not be read as showing whether the model was safe or appropriate for release. Fact · emerging evidence LD06-SL-C23
- The EU’s Code of Practice for general-purpose AI models with systemic risk requires state-of-the-art model evaluations with internal validity, external validity and reproducibility, elicitation that minimises under-elicitation and sandbagging, and independent external evaluation, unless the model is “similarly safe or safer” or no suitable evaluator is found after a public call such as one open for 20 business days; the Model Report with evaluation results goes to the AI Office, signatories commit to publish summarised versions of their framework and Model Reports if and insofar as necessary to assess or mitigate systemic risks, and external evaluators keep control over publishing their own findings. Fact · emerging evidence LD06-SL-C24
- China’s national standard GB/T 45654-2025 for generative AI services sets a pass rate: at least 90% of sampled output must be free of 31 risk types; its reference appendix specifies a test bank of at least 2,000 questions, refusal of at least 95% of should-refuse questions, and refusal of no more than 5% of should-not-refuse questions. Fact · emerging evidence LD06-SL-C25
- In a 2022 field scan of algorithmic auditors (survey of 152 and 10 interviews), only 7% said their audit process used a standardized framework, 80% had recommended changes to AI systems that were not implemented, and half of third-party auditors reported lacking access, against none of the first-party auditors. Estimate · emerging evidence LD06-SL-C26
- Surveying 104 papers and six industry cases, researchers found that AI red-teaming is “an unstructured procedure with undefined scope” with no unified protocols for reporting results, and argued that treating it as a cure for every risk verges on security theater. Estimate and interpretation · emerging evidence LD06-SL-C27
- Reviewing 341 academic AI audit studies and non-academic audits, researchers found that only a subset led to accountability outcomes, that many prominent academic audits were collaborations with large technology companies examining their own systems, and that journalistic audits resulted in the most impactful outcomes. Estimate and interpretation · emerging evidence LD06-SL-C28
- Of 8,931 mentions of the 50 most used language models in 5,242 biomedical publications from 2022 to March 2026, 77.7% were of closed commercial models, and 42% involved a model already retired when the paper was published or scheduled to retire within two years. Estimate · emerging evidence LD06-SL-C29
LD-07: AI in the research apparatus
- High annotation accuracy does not by itself guarantee valid downstream inference; treating imperfect predicted labels as observed variables can bias estimates and invalidate confidence intervals, while design-based validation with a probability-sampled human-labelled subset can restore valid inference under stated assumptions. Fact · rapidly maturing evidence LD07-HQ-C01
- In a large 2025 preprint spanning 37 annotation tasks from 21 published studies, plausible choices of LLM and annotation configuration produced incorrect inferential conclusions in about one third of tested hypotheses for state-of-the-art models, with higher rates for smaller models and near conventional significance thresholds. Estimate · rapidly maturing evidence LD07-HQ-C02
- Whether an LLM can replace human annotation is a task-specific empirical proposition that can sometimes be justified with an explicit statistical replacement test; it is not a general property of LLMs. Fact and interpretation · rapidly maturing evidence LD07-HQ-C03
- Recent LLMs can be observationally equivalent to expert coders on some standardized binary political-text classification tasks, particularly when the codebook is clear; this does not imply that substituting LLM labels preserves historical published analyses. Fact · emerging evidence LD07-HQ-C04
- LLM simulations can usefully forecast the relative pattern of treatment effects in some US survey-experiment archives, while systematically overstating effect magnitudes. Fact · contested evidence LD07-HQ-C05
- Synthetic respondents can match selected survey averages while failing to reproduce real-world variance, regression relationships, individual structure and subgroup heterogeneity. Fact and interpretation · contested evidence LD07-HQ-C06
- Statistical realism of synthetic responses is not a reliable proxy for treatment-effect accuracy. Fact · emerging evidence LD07-HQ-C07
- Chinese-language silicon-sample studies benchmarked against CGSS data find partial macro-level reproduction alongside representational bias, micro/macro/meso instability and cross-model differences; current systems do not warrant replacement of human respondents. Fact · emerging evidence LD07-HQ-C08
- Autonomous LLM agents can complete online surveys while passing conventional attention and logic checks at very high rates under tested conditions. Fact · mature evidence LD07-HQ-C09
- AI-mediated responding is measurable in current online research, but self-reported assistance, overt outsourcing, AI-test failure and autonomous-agent detection are different estimands and cannot be pooled into one prevalence rate. Interpretation · emerging evidence LD07-HQ-C10
- Provider and jurisdictional context can be associated with materially different refusal and accuracy patterns on politically sensitive prompts; existing evidence does not isolate regulation as the causal mechanism. Fact and interpretation · emerging evidence LD07-HQ-C11
- Model identity is time-indexed: proprietary services can change materially across deployed snapshots, so exact version and date are part of the measurement protocol. Fact · emerging evidence LD07-HQ-C12
- LLM judges can correlate with human evaluations while exhibiting systematic evaluator-specific biases, including position bias and self-preference; task-specific validation is therefore required. Fact and interpretation · rapidly maturing evidence LD07-HQ-C13
- LLMs can support codebook-driven and deductive qualitative coding in researcher-supervised workflows, reducing coding burden without establishing autonomous interpretive validity. Fact and interpretation · emerging evidence LD07-HQ-C14
- Human–LLM agreement is not a sufficient decision rule for qualitative coding quality, and current evidence does not establish LLM replacement of researcher interpretation in inductive qualitative analysis. Fact and interpretation · emerging evidence LD07-HQ-C15
- In a conditionally accepted AJPS manuscript re-running 14 published political-science annotation pipelines, substituting LLM-generated labels changed the substantive statistical conclusion in approximately 35–42% of estimates across the tested prompt/in-context-learning settings. Estimate · rapidly maturing evidence LD07-HQ-C16
- In the study that introduced "silicon samples", GPT-3 conditioned on the demographic and political backstories of real US survey respondents reproduced their two-party presidential vote closely (tetrachoric correlations of 0.90, 0.92 and 0.94 for 2012, 2016 and 2020) and the pattern of associations among 2016 survey answers; the authors called this "algorithmic fidelity" and said researchers can use such samples once it has been established for a given model and domain. Estimate · contested evidence LD07-SL-C01
LD-08: Who can know about AI? Epistemic dependence and jurisdiction over the field
- Among notable machine-learning models in the top quarter by training compute, the share built by academic laboratories fell from about 65% in the early 2010s to about 10% in the early 2020s; across about 1.2 million publications, a 10% rise in compute-related keywords in a topic went with a 1.73% fall in the share of academic-only papers. Estimate · emerging evidence LD08-SL-C01
- In 171,394 papers from 57 computer-science conferences (2000–2019), after deep learning took off in 2012 large technology firms published 44 more papers a year per AI conference than a synthetic counterfactual predicts and elite universities 40 more, while mid-tier and lower-tier universities published fewer; the authors attribute the gap to unequal access to compute. Estimate · emerging evidence LD08-SL-C02
- In 786,118 AI papers (2000–2020), researchers moved from academia to industry in growing numbers, and 25% of the AI researcher transitions from the top five Nature Index institutions went to industry; deep-learning specialists and more highly cited researchers were more likely to move. Estimate · emerging evidence LD08-SL-C03
- In 100 highly cited NeurIPS and ICML papers, the share with ties to large technology firms (author affiliation or funding) rose from 21% in 2008/09 to 66% in 2018/19, and the share with any corporate tie from 45% to 79%; 15% of the papers justified their work by a societal need and 1% discussed potential negative effects. Estimate · emerging evidence LD08-SL-C04
- Among faculty authors at ACL 2022 whose funding could be traced, 66% had received industry funding: 72% in the US, 69% in China and 38% in Europe; among student authors 74% had an industry scholarship or internship, including 65% in Europe. Estimate · emerging evidence LD08-SL-C05
- In a sample of 100 EMNLP 2022 papers, 44% had at least one industry-affiliated author, against 24% reported for 2019–2021 by earlier work, and citations of industry-made datasets, pre-trained models and previous best results were at least three times what the industry share of the field would predict; the authors state that this “is not proof of industry capture.” Estimate · emerging evidence LD08-SL-C06
- Of about 49,800 papers at ten top AI conferences, about one in ten of those published from 2018 to 2022 acknowledged funding from large technology firms; 12% of these reached high citation rates, against 4% of papers with other funding and 2% of unfunded papers, and their reference lists were more insular. Estimate · emerging evidence LD08-SL-C07
- Among computer-science faculty at four North American universities who had published on AI ethics or fairness between January 2015 and April 2020 and whose funding sources were known, 58% (19 of 33) had at some point been directly funded by large technology firms, and 97% (32 of 33) had a financial tie to them at some stage of their careers, including graduate funding or past employment; 10 of 17 position papers on AI’s ethical or societal impact published in Nature and Science since 2015 had at least one author with such ties. Estimate · emerging evidence LD08-SL-C08
- In 10,772 NeurIPS, ICML and CVPR papers (2015–2019), only three contained an explicit conflict-of-interest statement, although 15–20% of purely academic papers named a company in their acknowledgements; industry and academic papers differed by only 3.1% in their use of social-impact terms, academic papers using them slightly more. Estimate · emerging evidence LD08-SL-C09
- Among AI papers in an arXiv corpus of 1.84 million articles (1986 to August 2020), AI research involving the private sector was less thematically diverse and more highly cited than academic AI research, and private-sector research tended to specialise in data- and compute-intensive deep learning at the expense of research on the societal and ethical implications of AI. Estimate · emerging evidence LD08-SL-C10
- Roughly 15% or less of peer-reviewed responsible-AI publications each year had a co-author affiliated with industry; from 2018 to 2022 industry’s participation at responsible-AI conferences stayed between about 20% and 23% while at leading AI conferences it rose from 29% to 37%; 34.5% of 519 AI-publishing firms published any responsible-AI research. Estimate · emerging evidence LD08-SL-C11
- In about 102,000 arXiv AI papers (2014–2024), the share of societally oriented sentences written by computer-science-only teams rose from 49.0% to 71.2%, while the share written by teams including social scientists or humanists fell from 25.7% to 3.8%, although such teams wrote the most societally oriented papers (20.9% of sentences, against 7.8% for computer-science-only teams). Estimate · emerging evidence LD08-SL-C12
- Of the papers published at the FAccT conference from 2018 to 2021, 61.3% were labelled STEM and 29.7% humanities, social sciences and arts; summarising a survey of conference affiliates, the authors list concerns including “out-sized industry connections.” Estimate · emerging evidence LD08-SL-C13
- Of 84 AI ethics guidelines collected worldwide by 2019, private companies issued the largest share (19, or 22.6%), ahead of governmental agencies (18) and academic and research institutions (9). Fact · mature evidence LD08-SL-C14
- Green argues that technology ethics is “a terrain of contestation where the central debate is not whether ethics is desirable, but what ‘ethics’ entails and who gets to define it,” and that current approaches “are poised to enable” technology companies to label themselves ethical without substantively altering their practices. Interpretation · contested evidence LD08-SL-C15
- Describing Clio, its system for analysing Claude conversations, Anthropic writes that AI providers face “competitive pressures not to release usage data even if it would be in the public interest”; it validated Clio internally, for example reconstructing the topic distribution of 19,476 synthetic transcripts with 94% accuracy. Fact · emerging evidence LD08-SL-C16
- Anthropic’s first Economic Index report classified Claude conversations as 57% augmentation and 43% automation; the classifications were made by Claude and checked by the authors on 150 hand-validated conversations drawn from those for which users had submitted feedback, with 90.7% agreement on the augmentation–automation label and 86% to 95.3% acceptable task assignments by level of detail. Estimate · emerging evidence LD08-SL-C17
- In its September 2025 Economic Index report, Anthropic reported that the share of automation in Claude.ai conversations depended in part on which model did the classifying: re-running the same August 2025 data with Claude Sonnet 3.7 instead of Claude Sonnet 4 gave 45% automation instead of 49%, though the rise in directive use remained. Fact · emerging evidence LD08-SL-C18
- Anthropic’s January 2026 Economic Index report added measures such as “task success,” which “measures Claude’s assessment of whether Claude completes tasks successfully”; for Claude.ai the classifiers were checked against a human researcher on a small set of transcripts for which users had given feedback, and the report calls them “directionally accurate”; it found augmentation rising to 52% of Claude.ai conversations. Fact · emerging evidence LD08-SL-C19
- Microsoft researchers who measured the applicability of generative AI to occupations from Bing Copilot conversations state that they “cannot release conversational data due to legal and privacy obligations” and published aggregated occupation-level metrics instead. Fact · emerging evidence LD08-SL-C20
- A study of ChatGPT use co-written by OpenAI staff reports that ChatGPT had been adopted by around 10% of the world’s adult population by July 2025 and that non-work messages grew from 53% to more than 70% of use, from conversations classified by a privacy-preserving automated pipeline; the code is “available on request.” Estimate · emerging evidence LD08-SL-C21
- Analysing 37,845 Australian conversations from WildChat, a public dataset of ChatGPT logs collected outside OpenAI, university researchers classed 61.1% as work-related against 37.3% in OpenAI’s published global figures; to classify them they adopted the classifier prompts and categories of OpenAI-co-authored work. Estimate · emerging evidence LD08-SL-C22
- Researchers studied ChatGPT health advice without the developer’s involvement by collecting 202,590 donated conversations from 1,252 users in India, Nigeria, Brazil and Pakistan (December 2022 to February 2026); an AI judge applying a published rubric rated roughly one in five health conversations high-risk. Estimate · emerging evidence LD08-SL-C23
- On 2 July 2025 the European Commission adopted a delegated act on data access under the Digital Services Act that lets vetted researchers obtain internal data from very large online platforms and search engines to study systemic risks; applicants must show affiliation to a research organisation, “independence from commercial interests” and the ability to keep the data secure. Fact · emerging evidence LD08-SL-C24
- As of 19 May 2026, national Digital Services Coordinators had received 49 applications for vetted-researcher access under the Digital Services Act, mostly for social-media data and on risks including illegal content, advertising transparency and “AI features”; the coordinators were assessing the applications, and their update reported no access granted. Fact · emerging evidence LD08-SL-C25
- On 5 December 2025 the European Commission fined X €120 million under the Digital Services Act, in part for failing to give researchers access to public data: X’s terms of service prohibited eligible researchers from independently accessing its public data, including by scraping. Fact · emerging evidence LD08-SL-C26
- On 16 July 2026 the European Commission accepted X’s action plan, under which X will speed up screening of researchers’ applications, give eligible researchers access to public data free of charge and stop contractually prohibiting them from scraping. Fact · emerging evidence LD08-SL-C27
- Twenty-three researchers argued that the terms of service and enforcement strategies of prominent AI companies discourage good-faith safety evaluation and that researcher access programmes “often do not notify researchers of rejections”; they proposed voluntary legal and technical safe harbours. Interpretation · emerging evidence LD08-SL-C28
- Anthropic’s External Researcher Access Program gives successful applicants API credits, typically $1,000, for AI safety and alignment topics the company considers high priority; it gives no access to non-public models, and Anthropic reserves the right to reject submissions for any reason without explanation. Fact · emerging evidence LD08-SL-C29
- After Twitter ended its academic access programme in 2023 and set its Enterprise API price at $42,000 a month, the number of studies using Twitter data, which had grown by a median 25% a year, rose 0.6% in 2023 to 4,249 and fell 13% in 2024 to 3,712. Estimate · emerging evidence LD08-SL-C30
- Bak-Coleman, O’Connor, Bergstrom and West argue that technology firms shape research through control of data and funding: in the Meta 2020 US election studies “all experimental designs had to be approved by Meta,” and while Meta lists more than a thousand people it funded through open awards or fellowships, OpenAlex shows only 472 publications by independent academics acknowledging Facebook or Meta as a funder, a gap the authors say should not be read as a deliberate attempt to obscure funding. Interpretation · contested evidence LD08-SL-C31
- In July 2026 Anthropic committed $200 million to an Economic Futures Research Fund for external research on interventions to prepare society for AI’s economic impacts, mainly in projects of $5–30 million, supporting research “on interventions proposed in the EPF,” its own Economic Policy Framework, and on other open questions; its public page says it wants partners “willing to share what they’re learning publicly at key milestones” and states no terms on publication rights or review. Fact · emerging evidence LD08-SL-C32
- Anthropic’s Economic Futures Program offered research awards of $10,000 to $50,000 plus $5,000 in Claude API credits and “consultation on analyzing Anthropic Economic Index data,” for empirical work completed within six months, asked recipients to present and share their findings within six months, and encouraged economists with strong quantitative backgrounds to apply. Fact · emerging evidence LD08-SL-C33
- Microsoft’s AI Economy Institute offered researchers $75,000 grants that require participation in biweekly virtual convenings “that may include interim deliverables, including data analyses, initial findings, and working papers,” and scored proposals partly (20%) on how well they align with the institute’s mission. Fact · emerging evidence LD08-SL-C34
- In April 2025 Anthropic set up an Economic Advisory Council of “distinguished economists” to advise it on AI’s economic effects and to “inform the research agenda for the Anthropic Economic Index.” Fact · emerging evidence LD08-SL-C35
LD-09: Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike?
- In 41.3 million natural-science papers (1980–2025), researchers who used AI as a research method published 3.02 times as many papers a year, received 4.84 times as many citations and became research project leaders 1.37 years sooner than researchers who did not. Estimate · emerging evidence LD09-SL-C01
- In the same natural-science corpus, AI-augmented research was associated with a 4.63% smaller median collective knowledge extent and 22% less follow-on engagement among papers citing the same work, while each AI paper’s own family of citing papers was 3.46% more spread out; the authors link the contraction to AI work clustering where data are abundant. Estimate · emerging evidence LD09-SL-C02
- Before generative AI, as journals moved online (1998–2005), papers cited more recent work, fewer journals and fewer articles, and concentrated their citations on fewer of them; the author argued that online search “may accelerate consensus and narrow the range of findings and ideas built upon”. Estimate · mature evidence LD09-SL-C03
- Messeri and Crockett argue that AI tools promise scientists productivity and objectivity but can produce “illusions of understanding” that hide the formation of “scientific monocultures, in which some types of methods, questions and viewpoints come to dominate alternative approaches”, risking a phase of science “in which we produce more but understand less”. Interpretation · emerging evidence LD09-SL-C04
- Across arXiv, bioRxiv and SSRN preprints (2018–2024), authors whose abstracts first showed signs of language-model writing then posted 36.2%, 52.9% and 59.8% more papers respectively; for their manuscripts, more complex writing went with lower odds of publication, the reverse of the pattern for human-written ones; and adopters drew on “more diverse prior work”, being 11.9% more likely to cite books and citing younger documents. Estimate · contested evidence LD09-SL-C05
- A comment on the preprint-productivity study shows that dating adoption to an author’s first flagged abstract can by itself create post-adoption gains: placebo flags gave mean effects of 0.37 and 0.30 against 0.36 in the replicated baseline; the original authors reply that the artefact is “real but bounded” and that the association persists in designs it cannot bias. Estimate and interpretation · contested evidence LD09-SL-C06
- Among 775,323 scientists with full-text papers in PubMed Central, individual research portfolios spread across more fields after 2022: field entropy rose from 0.55 in 2022 to 0.71 in 2025 and the share entering a new field from 30.4% to 35.5%, more so for scientists in non-English-speaking low- and middle-income countries; the authors call the design descriptive. Estimate · emerging evidence LD09-SL-C07
- Across 74.6 million publications from 1960 to 2019, papers that use AI were more likely to be among the top 5% most cited in their field and year for a majority of disciplines, a premium the authors find stronger in disciplines with a lower propensity to use AI, while disciplines with more women or Black scientists showed less potential benefit. Estimate · emerging evidence LD09-SL-C08
- By February 2024 an estimated 17.5% of sentences in computer-science abstracts on arXiv were substantially modified by language models; the share was higher in papers closer to their nearest peer, a proxy for more crowded research areas (22.2% against 14.7% for less similar papers). Estimate · rapidly maturing evidence LD09-SL-C09
- In a preregistered online experiment, writers given one or up to five story ideas from GPT-4 wrote stories rated 5.4% and 8.1% more novel and 3.7% and 9.0% more useful than writers without AI, with the largest gains for less creative writers; but their stories were 5.2% and 5.0% more similar to the AI ideas and more similar to each other. Estimate · emerging evidence LD09-SL-C10
- On a large art-sharing platform, artists who adopted text-to-image AI raised their output by 25% and the favourites their works received per view by 50%; average content novelty and visual novelty fell over time, while peak content novelty rose and favourites became less concentrated among adopters. Estimate · emerging evidence LD09-SL-C11
- Re-analysing a published set of five brainstorming experiments, the authors found ChatGPT-aided idea sets less diverse than comparison sets in all five (significantly in four), and lower on all of their alternative diversity metrics, with 37 of 45 comparisons statistically significant; in one task 6% of ChatGPT-only ideas were unique against 100% of human-only ideas. Estimate · emerging evidence LD09-SL-C12
- In a writing experiment with Indian and American participants, inline AI suggestions cut task time by 35% and 30% respectively, but Indian participants gained less per suggestion, and the similarity between Indian and American essays rose from 0.48 to 0.54 as Indian writing moved towards Western styles. Estimate · emerging evidence LD09-SL-C13
- In argumentative essays co-written with a language model, only the instruction-tuned model (InstructGPT) made essays more alike; the base model (GPT-3) did not, and the reduction in diversity came from the text the model contributed, while the users’ own text remained unaffected. Estimate · emerging evidence LD09-SL-C14
- In a within-subjects ideation study, ideas produced with ChatGPT were more similar to each other across users than ideas produced with a non-AI creativity aid, but there was no difference at the level of the individual user; the authors conclude the effect comes from the model giving different users similar ideas. Estimate · emerging evidence LD09-SL-C15
- In a large dynamic experiment where each participant saw example ideas from ChatGPT or from earlier participants before brainstorming, high exposure to AI ideas did not make individual ideas more creative but increased the amount and rate of change of collective idea diversity: in the authors’ words, “AI made ideas different, not better”. Estimate · emerging evidence LD09-SL-C16
- Ideas generated by language models for a new product had higher average purchase intent than human ideas and were 7 times more likely to rank in the top 10%, but were less diverse; pooling ideas across vendors, chain-of-thought prompting and varied personas or constraints raised their diversity “almost to the level of human idea generation”. Estimate · emerging evidence LD09-SL-C17
- In two preregistered experiments, language-model assistance gave short-term boosts during assisted tasks only in some measures (in the divergent-thinking task, originality during exposure was similar across conditions), did not raise originality or fluency in later unassisted tasks, and participants who had received strategy coaching from the model produced more similar ideas afterwards. Estimate · emerging evidence LD09-SL-C18
- In a blind review by 79 NLP researchers, research ideas from a language-model agent were rated more novel than ideas from experts (5.64 against 4.84 on a 10-point scale) and slightly weaker on feasibility; but of 4,000 seed ideas the agent generated per topic, only about 200 were not duplicates. Estimate · emerging evidence LD09-SL-C19
- When 43 researchers each spent over 100 hours carrying out randomly assigned ideas from the earlier study, review scores of the AI-generated ideas fell significantly more than those of expert ideas on novelty, excitement, effectiveness and overall, and on many measures human ideas then ranked higher. Estimate · emerging evidence LD09-SL-C20
- In a formal model of firms hiring from a shared pool, there are conditions under which adopting a shared, more accurate screening algorithm is the dominant choice for both firms in a two-firm model yet lowers overall welfare compared with each firm using its own noisier judgement; under one standard ranking model (Plackett–Luce) the effect disappears. Interpretation · contested evidence LD09-SL-C21
- A 2026 philosophical and formal analysis, co-written by an author of the monoculture model, concludes that algorithmic “monoculture is less problematic than its critics have supposed: commonly cited objections fail, and while other objections have some force, they are not decisive against monoculture in general”. Interpretation · contested evidence LD09-SL-C22
- Across commercial machine-learning services tracked from 2020 to 2022, cases that every system got wrong were more common than independence would predict, and on average only 10% of one system’s improvement fell on cases all the other systems misclassified, though such cases made up 27% of the room for improvement; in a dermatology case, human dermatologists’ outcomes were more homogeneous than the models’. Estimate · emerging evidence LD09-SL-C23
- In field data from one hiring-software vendor covering 3 million applicants, 4% of applicants who applied to 10 positions were recommended for rejection by all of them, a rate the authors report as higher than expected by chance. Estimate · emerging evidence LD09-SL-C24
- Across more than 350 language models, when two models both answer a leaderboard question wrongly they give the same wrong answer 60% of the time on one dataset, and larger, more accurate models have highly correlated errors even across different architectures and providers. Estimate · emerging evidence LD09-SL-C25
- How much “excess agreement” among models is measured depends on the null model chosen and on which models and items are compared; a null model that allows for item difficulty gives results that differ “drastically” from earlier work. Interpretation · emerging evidence LD09-SL-C26
- Within six months of ChatGPT’s release, activity on Stack Overflow fell by 25% relative to Russian- and Chinese-language counterparts where access to ChatGPT was limited, an estimate the authors call a lower bound; votes on posts showed no large change in quality. Estimate · rapidly maturing evidence LD09-SL-C27
- After ChatGPT’s release, Stack Overflow saw significant declines in visits and questions, especially on topics where ChatGPT does well and among newer users, while Reddit developer communities showed no evidence of decline, which the authors suggest reflects their social fabric. Estimate · rapidly maturing evidence LD09-SL-C28
- A 2026 study finds that ChatGPT reduced Stack Overflow question volume by about 14% on average (up to 27.9% over time), with larger declines in mid- to low-quality questions and among less experienced users, and improved question quality; the authors read this as “niche partitioning rather than full displacement”. Estimate · rapidly maturing evidence LD09-SL-C29
- In laboratory experiments, indiscriminate training of models over generations on data produced by earlier models caused “irreversible defects” in which the tails of the original data distribution disappear; in the preprint’s language-model experiment, keeping 10% of the original data led to only minor degradation. Estimate · contested evidence LD09-SL-C30
- In laboratory experiments, replacing real training data with each generation’s synthetic data tends towards model collapse, but accumulating synthetic data alongside the original real data avoids it; for a sequence of linear models, the authors prove that test error then has a finite upper bound however many generations pass. Estimate · contested evidence LD09-SL-C31
- Across 27 language models, 155 topics and 12 countries, the epistemic diversity of models’ answers rose substantially over three years, but every model was less diverse than a web-search baseline, larger models were less diverse than smaller ones, and for country-specific topics models’ built-in knowledge reflected English more than local-language knowledge. Estimate · emerging evidence LD09-SL-C32
- Kaggle competition code from 2019 to mid-2026 has become more alike in its literal syntax, but the author finds no evidence of homogenization in the approach or intent of the code. Estimate · emerging evidence LD09-SL-C33
LD-10: Persuasion, Intermediation & the Information Order
- Persuasion capability is not the same as realized social influence; population effects require evidence on exposure, per-contact effect, persistence and behavioural transmission. Interpretation · mature evidence LD10-HQ-C01
- In preregistered UK experiments, post-training and prompting increased measured conversational persuasiveness by up to 51% and 27% relative, and higher persuasion was associated with lower factual accuracy. Estimate · rapidly maturing evidence LD10-HQ-C02 Also cited by LD-09.
- In one UK experiment, about 36%–42% of GPT-4o's immediate attitude effect remained at a one-month recontact. Estimate · emerging evidence LD10-HQ-C03
- In the 2024 U.S. election study, AI advocacy shifted candidate preference by 2.879 points on a 0–100 scale in the preregistered specification; post-hoc stated vote-choice effects were a few percentage points. Estimate · rapidly maturing evidence LD10-HQ-C04
- Parallel 2025 experiments in Canada and Poland found larger candidate-preference shifts than the U.S. study, around 10 points on the study's 0–100 scale in preregistered specifications. Estimate · emerging evidence LD10-HQ-C05
- Personalized GPT-4 outperformed human debate opponents in the Salvi et al. experiment, but a September 2026 Author Correction says the direct personalized-vs-non-personalized GPT-4 comparison did not reach conventional significance (P=.07). Estimate · rapidly maturing evidence LD10-HQ-C06
- A 2025 meta-analysis of seven earlier studies found no average persuasiveness difference between LLMs and humans, with substantial heterogeneity. Estimate · emerging evidence LD10-HQ-C07
- In four preregistered 2025–26 UK experiments, frontier AI out-persuaded strong human text persuaders; its advantage over coached elite debaters disappeared when AI was constrained to the debaters’ message length and writing speed. Estimate · emerging evidence LD10-HQ-C08
- In two preregistered UK experiments, brief AI conversations increased real petition signing by 12.8 and 19.7 percentage points, while attitude effects did not predict behavioural effects. Estimate · emerging evidence LD10-HQ-C09
- A 2026 peer-reviewed study estimated LLM persuasion at about US$48–$75 per persuaded voter under its simulation, but concluded traditional campaign methods were currently easier to scale. Estimate and interpretation · emerging evidence LD10-HQ-C10
- A widely cited 2024 conspiracy-belief experiment has been under a Science Editorial Expression of Concern since June 2026; its substantive effect estimate should not be treated as settled pending resolution. Fact · contested evidence LD10-HQ-C11
- In a U.S. desktop field experiment, an AI Overview reduced outbound organic clicks by 39.8% relative when it appeared and increased zero-click searches by 34.5%, without a measured user-experience gain. Estimate · emerging evidence LD10-HQ-C12
- In a preregistered one-week U.S. field experiment, forced Google AI Mode reduced external click-through by 18.8 percentage points and lowered trust in information on Google. Estimate · emerging evidence LD10-HQ-C13
- In March 2025 U.S. browsing data, users clicked traditional results in 8% of Google visits with an AI summary versus 15% without; 18% of studied searches generated an AI summary. Fact · rapidly maturing evidence LD10-HQ-C14
- A 2026 Google AI Overviews audit found 13.7% overall activation, 64.7% on question-form queries, nearly 30% of cited domains absent from the co-displayed first results page, and 11.0% of atomic claims unsupported by cited pages. Estimate · emerging evidence LD10-HQ-C15
- U.S. panel evidence through 2024 found LLM adoption associated with more unique websites visited and no greater concentration of LLM-routed traffic than search traffic. Estimate · emerging evidence LD10-HQ-C16
- A 2025 EBU/BBC-led audit across 18 countries and 14 languages found significant issues in 45% of AI-assistant news answers, serious sourcing problems in 31% and major accuracy issues in 20%. Estimate · rapidly maturing evidence LD10-HQ-C17
- In the 2026 Digital News Report, weekly use of standalone AI chatbots for news reached 10% across surveyed markets, up from 7% a year earlier, but only 1% named AI their main news source. Estimate · rapidly maturing evidence LD10-HQ-C18
- Across 27 markets in DNR 2026, 4% of all respondents said they always/often click through from AI chatbot news answers; among chatbot-news users, the corresponding rate was about 42%. Estimate · rapidly maturing evidence LD10-HQ-C19
- In a preregistered U.S. experiment, citations increased trust in AI-search answers even when incorrect or hallucinated, while uncertainty highlighting reduced trust and sharing. Estimate · emerging evidence LD10-HQ-C20
- CETaS identified 16 viral AI-enabled disinformation/deepfake cases in the 2024 UK election and 11 across EU/French elections, and found no evidence that the detected cases meaningfully affected election results. Estimate and interpretation · emerging evidence LD10-HQ-C21
- Within a large fact-check dataset, generative-AI media rose relatively recently while simple context manipulation remained the majority through November 2023. Estimate · emerging evidence LD10-HQ-C22
- China-originating LLMs showed higher refusal rates, shorter responses and more inaccurate answers on 145 China-politics questions than non-China models; the study is observational and does not establish that regulation causes the gap. Estimate · emerging evidence LD10-HQ-C23
- EU AI Act Article 50 transparency obligations applying from 2 August 2026 require disclosure for chatbot interaction and clear labelling in key cases including deepfakes and AI-generated/manipulated public-interest text; generative AI systems already on the market before that date have until 2 December 2026 to meet the Article 50(2) machine-readable marking obligation; the voluntary Code provides compliance guidance. Fact · rapidly maturing evidence LD10-HQ-C24
- Label effects are context-dependent: one 2026 experiment found simple AI-authorship labels did not reduce persuasion of policy messages, while experiments on misleading synthetic images found warning labels reduced belief. Estimate and interpretation · emerging evidence LD10-HQ-C25
LD-11: Concentration, Stack & Sovereignty
- Concentration claims in AI must identify the stack layer, market definition, metric, geography and period; concentration at one layer does not establish concentration or power at another. Interpretation · mature evidence LD11-HQ-C01
- ASML is currently the world’s only manufacturer of EUV lithography systems, used for critical layers in the most advanced chips. Fact · mature evidence LD11-HQ-C02
- TrendForce estimated that TSMC captured 72.5% of revenue among the world’s top ten contract foundries in Q2 2026. Estimate · rapidly maturing evidence LD11-HQ-C03
- Counterpoint estimated Q2 2026 HBM revenue shares of SK hynix 50%, Samsung 33% and Micron 18%. Estimate · rapidly maturing evidence LD11-HQ-C04
- Epoch AI estimated that five U.S. hyperscalers owned about 71% of cumulative global AI compute in Q4 2025, up from about 63% in Q1 2024, measured in H100-equivalents. Estimate · rapidly maturing evidence LD11-HQ-C05
- Epoch AI estimated Google held about one quarter of global cumulative AI compute in Q4 2025, driven primarily by roughly 3.81 million H100-equivalents of custom TPUs. Estimate · emerging evidence LD11-HQ-C06
- Synergy estimated AWS 28%, Microsoft 20% and Google 15% of global cloud infrastructure services revenue in Q2 2026, a combined 63%. Estimate · mature evidence LD11-HQ-C07
- The UK CMA found AWS and Microsoft hold significant market power in UK cloud and identified egress fees, interoperability barriers and Microsoft software licensing as constraints on switching and multi-cloud. Fact · rapidly maturing evidence LD11-HQ-C08
- On 25 June 2026 the European Commission preliminarily considered AWS and Azure important cloud gateways that should be designated under the DMA, citing lock-in, switching costs and the growing role of AI in cloud procurement. Fact · rapidly maturing evidence LD11-HQ-C09
- Under the April 2026 Microsoft–OpenAI terms, Microsoft remains OpenAI’s primary cloud partner, but OpenAI can serve products across clouds and Microsoft’s model/product IP licence through 2032 is non-exclusive. Fact · rapidly maturing evidence LD11-HQ-C10
- FTC staff found that major cloud–AI-developer partnerships included equity, revenue-sharing and varying consultation, control and exclusivity rights, while not establishing antitrust liability. Fact · emerging evidence LD11-HQ-C11
- Anthropic reports that it trains or serves Claude on AWS Trainium, Google TPU and NVIDIA GPU infrastructure and distributes Claude across AWS, Google Cloud and Microsoft Azure. Fact · emerging evidence LD11-HQ-C12
- Menlo Ventures estimated late-2025 enterprise LLM API shares of Anthropic 40%, OpenAI 27% and Google 21%, compared with OpenAI 50% and Anthropic 12% in 2023. Estimate · contested evidence LD11-HQ-C13
- OpenRouter and a16z report that open-weight models reached approximately one-third of token usage on OpenRouter by late 2025; an NBER working paper using OpenRouter and Microsoft Azure usage data reports a lower open-source share of OpenRouter tokens, about 25% by December 2025 after a peak near 45% in August. Estimate · emerging evidence LD11-HQ-C14
- Epoch AI's September 2026 multi-benchmark analysis estimates that the price of reaching fixed LLM capability levels has fallen about 13× per year since 2023. Estimate · rapidly maturing evidence LD11-HQ-C15
- In November 2025 U.S. Commerce authorized G42 and Humain to purchase the equivalent of up to 35,000 NVIDIA Blackwell GB300 chips each, subject to security and reporting requirements. Fact · rapidly maturing evidence LD11-HQ-C16
- As of 25 September 2026, Mubadala described Stargate UAE as under construction, with its initial 200MW cluster due to begin operations in late 2026. Fact · emerging evidence LD11-HQ-C17
- The EU's Cloud and AI Development Act was still a proposal on 25 September 2026, while the AI Gigafactories initiative was in procurement for up to seven sites rather than operating capacity. Fact · rapidly maturing evidence LD11-HQ-C18
- NSF reported that NAIRR had supported over 800 research projects by 1 September 2026, while the programme relies materially on private-sector in-kind resources. Fact · emerging evidence LD11-HQ-C19
- 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. Fact and unknown · contested evidence LD11-HQ-C20
- NVIDIA disclosed US$279B in supply and capacity commitments as of 26 July 2026, primarily for memory and manufacturing facilities. Fact · rapidly maturing evidence LD11-HQ-C21
LD-12: Regulation-in-Practice: Rights, Compliance & Developmental Steering
- The EU AI Act is in force, but Annex III high-risk obligations apply from 2 December 2027 and Annex I product-embedded high-risk obligations from 2 August 2028. Fact · mature evidence LD12-HQ-C01
- On the Commission's September 2026 list, 9 EU Member States had a final AI Act market-surveillance single point of contact, 12 were pending final adoption and 6 had none listed. Fact · emerging evidence LD12-HQ-C02
- The EU GPAI Code of Practice is voluntary, but Commission guidance recognizes adherence as a route to demonstrate compliance; non-signatories must demonstrate compliance by other adequate means. Fact · rapidly maturing evidence LD12-HQ-C03
- The EU AI Office began using information-request powers after its GPAI enforcement powers became applicable; the exact report that more than 30 providers received initial RFIs remains press-relayed. Fact · emerging evidence LD12-HQ-C04
- In the 2025 IAPP/Credo AI survey, 77% of surveyed organizations reported working on AI governance and only 10 of 671 expected to need no additional AI-governance staff within a year. Estimate · emerging evidence LD12-HQ-C05
- In an audit-style study of 391 NYC employers after Local Law 144 took effect, 18 posted bias-audit reports and 13 posted transparency notices; employer scope discretion means those counts are not a violation rate. Fact · mature evidence LD12-HQ-C06
- The New York State Comptroller concluded NYC's enforcement of Local Law 144 was ineffective; 75% of test calls were misdirected, and DCWP identified one issue where auditors identified at least 17 potential issues in the same 32-company review. Fact · emerging evidence LD12-HQ-C07
- The CJEU held in C-203/22 that an automated-decision subject is entitled under GDPR to meaningful information about the procedure and principles actually applied; merely handing over a complex formula or algorithm is insufficient. Fact · mature evidence LD12-HQ-C08
- Italy's €15 million GDPR fine on OpenAI was annulled by the Court of Rome in March 2026 on competence grounds under the GDPR one-stop-shop; the court did not resolve the substantive allegations. Fact · emerging evidence LD12-HQ-C09
- Colorado's original 2024 AI framework was delayed, stayed from enforcement before its operative date, and replaced in 2026 by a narrower disclosure/ADMT regime whose core duties begin in 2027. Fact · emerging evidence LD12-HQ-C10
- U.S. Executive Order 14365 directed a federal AI Litigation Task Force to challenge selected state AI laws and directed Commerce to condition some non-deployment BEAD funding, to the maximum extent allowed by law, on the absence of state AI laws identified as onerous; the order itself does not automatically preempt state law. Fact and interpretation · emerging evidence LD12-HQ-C11
- China's CAC reported 1,112 generative-AI services filed and 731 applications/functions registered by 31 August 2026, up from 988 and 598 at the end of June. Fact · rapidly maturing evidence LD12-HQ-C12
- China's CAC notices require launched generative-AI applications or functions to display the filed/registered service or model they use, including the model name and filing or launch number. Fact and interpretation · emerging evidence LD12-HQ-C13
- 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. Fact and unknown · emerging evidence LD12-HQ-C14
- South Korea's AI Basic Act and implementing decree are in force; an at-least-one-year grace policy deferring most fact-finding and administrative fines is reported by the U.S. International Trade Administration as MSIT implementation policy. Fact · emerging evidence LD12-HQ-C15
- South Korea is using staged public support and competitive evaluation to narrow a sovereign-foundation-model field; official Phase 2 results advanced Upstage, SK Telecom and LG AI Research to the next round. Fact · emerging evidence LD12-HQ-C16
- The EU AI Gigafactory programme is at procurement stage: the formal call opened in July 2026, closes in November 2026 and contemplates up to seven facilities; operational capacity is not yet established. Fact · emerging evidence LD12-HQ-C17
- 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. Fact and unknown · rapidly maturing evidence LD12-HQ-C18
- Across the examined jurisdictions, the operative function of AI regulation differs by instrument and stage: rights, compliance infrastructure, market organization and developmental steering coexist rather than forming one regime. Interpretation · rapidly maturing evidence LD12-HQ-C19