OpenAI announces results on local smoothing, a disproof of the exponential time hypothesis, and the construction of quantum field theories. The source presents these as research results to be understood with academic groups: physical artifacts are not yet public, and independent verification is not established here.
Technology: OpenAI internal models
Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Announcement only
- Test type: prospective or independent check
- Tested: The local-smoothing proof was formalized in Lean the day after the AI output; the refutation of the Exponential Time Hypothesis was verified in Lean and is being checked by an external group led by Impagliazzo, Paturi and Williams.
- Still unknown: The physics result was not released; the external groups had not reported conclusions on the page.
- AI role: OpenAI internal models produced the proofs; external experts were asked to shepherd and explain the results.
Evidence sources: https://openai.com/index/building-scientific-understanding-with-academic-communities/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 5/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Google's Empirical Research Assistance (ERA), a Gemini-based tree-search system, reaches expert-level performance on six benchmarks spanning genomics, public health, satellite imagery, neuroscience, time-series forecasting and mathematics. Its state-level U.S. hospital admission forecasts for flu, COVID-19 and RSV rank at or near the top of CDC leaderboards. Results are computational and checked against public benchmarks and leaderboards, with no wet-lab validation; exact benchmark scores were not visible in the sources read.
Technology: Empirical Research Assistance (ERA), Gemini-based tree search
Family Gemini: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: ERA-generated methods and models were compared against human-developed methods on a public leaderboard for bioinformatics and against the CDC ensemble and other individual models for COVID-19 hospitalization forecasts.
- Still unknown: Not reported.
- AI role: A Gemini-based large language model with tree search generated scientific software, including 40 single-cell data analysis methods and 14 forecasting models.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-026-10658-6&resultType=core&format=json
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
In three large experiments with 76,977 participants, 19 LLMs argued 707 British political positions in dialogue. Persuasiveness was driven mainly by information density, while tailoring to user traits added little, and post-training for persuasion raised effectiveness by up to 50%. The most persuasive models also produced the most inaccurate claims. Caveat: the persuasion effects are measured in survey experiments, and accuracy was tested on a subset of claims.
Technology: 19 open and closed LLMs (including GPT-4.5, GPT-4o, Grok-3)
Family GPT-4: profile · models page
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: Three large experiments with 76,977 participants tested 19 LLMs arguing 707 British political positions in dialogue, and the factual accuracy of claims was checked on a subset.
- Still unknown: Persuasion effects are measured in survey experiments, and accuracy was tested only on a subset of claims.
- AI role: LLMs, some post-trained for persuasion, were the persuaders in dialogue.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/science.aea3884&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
A collection of case studies with universities and national labs in mathematics, physics, biology, computer science, astronomy, and materials. Examples: it helped complete a proof of an Erdős problem with Sawhney and Sellke, reconstructed the SL(2,R) symmetry of the Kerr wave equation, and proposed an immunology mechanism later confirmed by Unutmaz's lab. Reported by OpenAI; experts supervised the work.
Technology: GPT-5 and GPT-5 Pro
Family GPT-5: profile · models page
Publication: Publication status unspecified in annotation · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Not reported
- Test type: direct test
- Tested: The OpenAI page lists case studies, for example a mannose rescue experiment in T cells whose results matched GPT-5 predictions, and Erdos Problem 848, where authors turned a GPT-5 density estimate into a full proof.
- Still unknown: OpenAI says the cases are curated illustrations, not a systematic sample. GPT-5 can invent citations, mechanisms or proofs, and expert oversight is essential. One case reproduced an earlier published proof without citing it.
- AI role: GPT-5 suggested ideas, mechanisms, literature links and experiments; scientists set questions, verified results and ran lab work. OpenAI says the model is not autonomous.
Evidence sources: https://openai.com/index/accelerating-science-gpt-5/
The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Each run reads about 1,500 papers and executes 42,000 lines of analysis code; outside scientists judged 79.4% of its conclusions accurate. Of seven discoveries, three reproduce results already obtained by humans and four are new. Limitations include spurious correlations, a cost of about $200 per run, and a six-month work gain estimated by beta users.
Technology: Kosmos (agents with a structured world model)
Family Kosmos: profile · models page
Publication: Preprint · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: prospective or independent check
- Tested: Independent scientists judged 79.4% of Kosmos's conclusions accurate; three discoveries reproduced findings from preprints or unpublished manuscripts that Kosmos did not access at runtime; the four new discoveries are wet-lab validation in progress, with one checked in an independent scRNA-seq dataset.
- Still unknown: The authors state that Kosmos chases statistically significant but irrelevant findings, that about one in five statements is not accurate, and that the human time-saving estimates are intrinsically suspect.
- AI role: An AI scientist runs about 12-hour cycles of data analysis and literature search (about 1,500 papers and 42,000 lines of code per run), linked by a structured world model, and generates traceable reports; humans set the objectives.
Evidence sources: https://advances.edisonscientific.com/research/announcing-kosmos/ · https://arxiv.org/abs/2511.02824
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: blog. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Fine-tuning Llama 3.1 70B on Psych-101, a dataset of about 10 million human choices from 160 psychology experiments, produced a model that predicts and simulates human behaviour in experiments expressed in natural language, including tasks it had not seen. The authors report that its internal representations align better with human neural activity after fine-tuning. Caveat: a 2025 critique argues that its language comprehension is limited and that it fails to follow altered instructions.
Technology: Centaur (Llama 3.1 70B fine-tuned on Psych-101)
Family Llama: profile · models page
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: retrospective or benchmark
- Tested: Centaur, a language model fine-tuned on Psych-101, was evaluated on held-out participants and on generalization to unseen cover stories, modified task structures and new domains, and its representations were checked against human neural activity.
- Still unknown: The abstract states no limitations; the authors present the case study as a demonstration.
- AI role: A large language model was the base, fine-tuned on about 10 million human choices from 160 experiments.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-025-09215-4&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 227 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
LLM agents built from participants' full interview transcripts reproduced their General Social Survey answers at about 85% of the rate at which participants reproduced their own answers two weeks later, a normalized score. They also matched personality and economic-game results (normalized correlations of about 0.80 and 0.66) and replicated four of five behavioural experiments. Caveat: arXiv preprint, not peer reviewed, and the results are normalized rather than raw accuracy.
Technology: Generative agent architecture with LLM and full interview transcripts
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Preprint
- Test type: retrospective or benchmark
- Tested: Agents built from interview and survey data of 1,052 Americans were scored on held-out General Social Survey items against participants' own two-week test-retest consistency, and on personality, economic-game and experiment replications.
- Still unknown: Preprint, not peer reviewed; there is no dedicated limitations section, and combining interviews and surveys gave only modest gains over either source.
- AI role: An LLM was the engine of agents that answer questions on behalf of individuals, grounded in their self-report data.
Evidence sources: https://arxiv.org/abs/2411.10109
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review scope and date are not recorded; source checked does not establish full-text review.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.
Across experiments with more than 2,000 participants, three dialogues with GPT-4 reduced belief in the targeted conspiracy theory by about 20%, and the effect persisted for two months and spilled over to unrelated conspiracies. A professional fact-checker rated 99.2% of 128 sampled chatbot claims as true. Caveats: lab setting, the chatbot did not reduce belief in true conspiracies, and most participants still held the hoax belief.
Technology: GPT-4 (DebunkBot dialogue system)
Family GPT-4: profile · models page
Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: direct test
- Tested: More than 2,000 conspiracy believers held three personalized, evidence-based dialogues with GPT-4 Turbo, and belief change was measured, including persistence at two months and spillover to unrelated conspiracies.
- Still unknown: The abstract does not describe a control group and states no limitations; the page metadata lists a 2026 expression of concern and erratum.
- AI role: GPT-4 Turbo delivered the personalized counterevidence in the dialogues.
Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/science.adq1814&resultType=core&format=json
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 529 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
No follow-up recorded; this is not a claim that no follow-up exists.
Follow-up review scope and date are not recorded.
No source-supported relationships recorded.