AIImpactLab
9 impact clustersWork & marketsCognition & learningTechnical failuresPrivacySecurityEnvironmentLaw & informationSocial good & healthSystemic risk
Search the evidence

Find the signal.

Search titles, impact clusters, countries, organizations and the full text of every analysis.

8 stories found

A sealed historical archive leaks future facts into an AI drafting many competing theories, with one relativity equation buried among them.
Cognition & learningGlobal+3 clusters01

The Einstein test exposes why proving AI discovery is so hard

Could an AI trained only on knowledge available before a scientific breakthrough rediscover the breakthrough independently? Nature examines that deceptively simple test through historical language models built with cutoff dates before relativity, quantum mechanics, Turing machines, and other landmark ideas. The early results are humbling. A model trained on pre-1900 material showed occasional phrases that resembled later insights after receiving strong hints, but mostly failed and often produced plausible language without a reliable physical model. Other researchers attempting a pre-1930 system discovered that the training corpus leaked later facts: the supposedly historical model could answer questions about Franklin D. Roosevelt's administration. A University of Zurich family of four-billion-parameter models uses cutoffs at 1913, 1929, 1933, 1939, and 1946, but limited historical data and compute constrain what those systems can demonstrate. The test reveals two separate problems. First, dated archives are messy, incomplete, and contaminated by metadata and digitization. Second, a generative model can produce many theories, some suggestive and many wrong, while science still needs a process to rank them and connect them to evidence. Mathematics offers formal verification; empirical science requires experiments, instruments, causal reasoning, and judgment about which hypothesis deserves scarce attention. Historical models remain valuable because they can expose hindsight leakage and benchmark scientific novelty. But a striking rediscovery claim should not count unless the dataset, cutoff, prompts, researcher hints, candidate failures, and evaluation rule are independently reconstructable.

5 min
Thousands of AI agent nodes spiral into a fluid vortex beside a formal proof chain and an independent review stamp waiting to close.
Social good & healthGlobal+4 clusters02

OpenAI says 10,000 AI agents solved the Navier-Stokes problem

OpenAI says an internal system significantly more capable than GPT-6 Astra produced an analytical proof that smooth three-dimensional fluid motion can develop a singularity in finite time under a smooth external force. That would resolve the Navier-Stokes existence and smoothness Millennium Prize problem by establishing the counterexample formulations labeled C and D in the official statement. The company released a 166-page writeup and a Lean formalization, says the decisive effort involved roughly 10,000 concurrent agents, and reports that the Navier-Stokes work used about 2.7 million agent messages and 130 billion output tokens. It does not intend to claim the million-dollar prize. The result is potentially historic, but the correct verb today is claims, not solved. A formal proof artifact makes checking more rigorous and transparent, yet experts must still verify that the definitions, assumptions, and formal statements match the intended problem and that no gap sits outside the encoded proof. Provenance also matters. OpenAI says it began after hearing rumors about related work, did not access the outside researchers' specific user data, and cannot entirely rule out indirect influence from de-identified data used to improve models. The episode therefore demonstrates both the promise and the governance burden of AI-accelerated science. Massive parallel search can attack problems at a scale unavailable to most mathematicians. Scientific legitimacy will depend on independent verification, reproducible artifacts, careful credit, and clear policies protecting unpublished work submitted to commercial AI systems.

6 min
A proprietary model core and a stack of confidential benchmark cards enter a sealed computing chamber from opposite sides while both owners remain unable to inspect the other's asset.
Technical failuresSingapore and Global+3 clusters03

A cryptographic enclave keeps both AI weights and hidden safety tests secret

Google DeepMind, the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons are piloting what they describe as the first double-blind evaluation of a proprietary frontier-class AI model. The project tests Gemini Flash Lite against confidential benchmarks inside a privacy-preserving environment built with Google Cloud Confidential Space. The evaluator cannot see the model weights, and Google cannot see the evaluation prompts. Cryptographic verification is intended to reduce benchmark contamination while protecting both sensitive tests and proprietary intellectual property. That matters when a model could otherwise see the exam before deployment, especially for cybersecurity or government evaluations whose prompts may themselves be sensitive. The pilot is an architectural advance, not a universal seal of trustworthy evaluation. A secure enclave does not prove that the benchmark measures the right capability or harm, that the implementation has no vulnerability, or that a tested model behaves identically after deployment. The next standard should combine cryptographic separation with independent methodology review, reproducible evidence, transparent limitations, and testing across providers rather than treating secrecy alone as scientific validity.

5 min
A print table filled with biomedical papers reveals patterned AI fingerprints across discussion and results sections beside a clear preprint and provenance warning.
Law & informationGlobal research corpus+3 clusters04

Almost nine in ten late-2025 biomedical papers showed signs of AI-assisted writing

A preprint analyzed more than one million English-language open-access biomedical papers and estimated that 89 percent of papers published in December 2025 showed signs of some large-language-model-assisted writing. Nature reports estimates of 77 percent for 2025 overall and 52 percent for 2024, with signs appearing more often in discussions than results. The number is startling and easy to misuse. It does not mean AI authored 89 percent of biomedical papers, fabricated their data, or influenced the entire scientific literature. The method detects shifts in vocabulary within a specific PubMed Central corpus, the paper has not been peer reviewed, and other researchers told Nature that representativeness and methodology need further analysis. The finding still matters because AI assistance is moving from exceptional to ordinary while disclosure, attribution, data verification, citation checking, and journal policy remain inconsistent. Science needs provenance that distinguishes language editing from analysis, protects responsibility for claims, and lets readers audit the contribution without treating every polished sentence as misconduct.

5 min
A human mathematician stands before an immense luminous lattice of rapidly assembling proofs and one unresolved dark space.
Cognition & learningGlobal+3 clusters05

AI's mathematical advances force a profession to redefine human work

The Washington Post reports that leading mathematicians gathered at OpenAI's San Francisco office to discuss what would remain for human experts if AI becomes superhuman at research mathematics. The framing is deliberately provocative, but the underlying change is real: recent systems have contributed counterexamples, proofs, and advances on longstanding problems, while mathematicians and AI companies debate how much novelty, reliability, and human direction each result contains. Mathematics is unusually exposed because a correct formal proof can often be verified more directly than a claim in an experimental science. That does not make the human profession obsolete. It shifts value toward selecting important questions, building theories, checking significance, translating results, teaching judgment, and deciding who gets access to powerful research tools. The field should resist both denial and a corporate future in which a few laboratories own the systems, compute, and agenda for mathematical discovery.

6 min
A human mathematician confronts a towering cascade of elegant artificial intelligence proofs, with hidden false steps glowing red beneath the chalk equations.
Cognition & learningGlobal+4 clusters06

Mathematicians warn AI could flood the proof economy with confident errors faster than humans can check them

The International Mathematical Union has endorsed the Leiden Declaration on Artificial Intelligence and Mathematics, according to Ars Technica. The declaration warns that AI can produce plausible but unreliable arguments, overwhelm peer review with cheap incorrect drafts, obscure attribution, distort hiring and funding, and let commercial announcements outrun independent evaluation. The warning is not a rejection of computational tools or proof assistance. It is a defense of the conditions that make mathematics trustworthy: disclosure, reproducibility, human responsibility, credit, and access to enough information for independent scrutiny. A machine may produce a correct result, but if the model, prompts, training data, compute, and method remain inaccessible, the community cannot easily determine what was learned, what can be reproduced, or whether a benchmark is being marketed as general reasoning.

5 min
Ten mathematical result cards and a geometric verification checkmark displayed beneath archival glass.
Work & marketsGlobal+4 clusters07

An AI system claims ten advances on decade-old mathematics problems

OpenAI says an internal version of its next major model, called Astra, produced ten advances on mathematical problems whose central results had seen no progress for at least a decade. The work spans geometry, coding theory, complexity, group theory, operator algebras, cryptography and combinatorics. Human researchers prepared manuscripts with the same model, and every proof was formalized as a Lean certificate. That combination is stronger than an unsupported answer, but it is not the same as community acceptance: independent experts still need to examine the problem statements, proofs, novelty and significance. The announcement also forces a sharper authorship question when the system originates the proof and humans curate, verify and communicate it.

4 min
A rising AI capability graph is balanced against a warning signal for confident uncertainty and factual hallucinations.
Cognition & learningGlobal+4 clusters08

Claude Opus 5 is more capable—and slightly more prone to factual hallucinations

Anthropic’s system card reports broad gains for Claude Opus 5 in agentic coding, computer use, long-horizon knowledge work, and scientific reasoning. It also documents a reliability tension: on one closed-book factuality benchmark, accuracy was 11% higher than Opus 4.8 while the hallucination rate was 6% higher. Anthropic found cases where the model confidently answered despite internal uncertainty, even as its automated alignment scores and prompt-injection robustness improved.

4 min