Search the evidence

Find the signal.

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

15 stories found

Two small paper cabinets hold legal files and pigment swatches with visible gaps.
Technical failuresGlobal+2 clusters01

Who gets left out of an AI test? Two studies count the gaps

A million images sounds like a generous foundation for skin AI. What if the patient is a child? Researchers who examined 29 public dermatology image collections started with more than 1.1 million images. After quality filtering and removal of near duplicates, they estimated that 2.3 percent of images were pediatric and 11.0 percent fell in the two darkest Fitzpatrick categories. Some missing demographic labels were estimated, not recorded at the bedside. These are shares of images, not percentages of patients misdiagnosed. A separate study asked what happens when safety benchmarks look beyond a familiar set of obvious requests. Researchers tested ten language models on 5,747 instructions derived from real legal cases. The collection spanned 269 offense types and 50 intents under Chinese and US law. Scores describe how those models responded under that evaluation. They do not measure how often users act on an answer, and they are not a judicial determination of any particular response. The connection between the studies is a practical one for anyone buying an AI system: ask to see what is in the test. An image library can grow while scarcely broadening the ages or skin tones it contains. A new set of cases can change what a safety score means even if the model stays the same. Neither audit alone can say whether a particular clinic or service is safe. Both tell its evaluators where to look next.

3 min
Linocut workers walking together toward an unfinished data centre, with a work glove and blank payslip in the foreground.
Work & marketsEurope+1 clusters02

The AI buildout has workers. Their pay cannot be optional.

The people building the cloud should not disappear behind it. Investigate Europe's reporting describes disputed pay, unsafe conditions and obstacles to worker representation on European data-centre construction projects. These are specific reported experiences and allegations, not a finding that every facility or contractor behaves the same way. The companies quoted say they require safety and labour standards. In a separate Belgian accident case, the labour inspectorate closed its inquiry without finding wrongdoing; that contrary evidence belongs in the account too. There is also a development worth welcoming. Investigate Europe and Swedish labour publications report an agreement between Benbau and Solidariska byggare concerning work on Microsoft's Swedish data-centre construction. It includes 15 million Swedish kronor in back pay and damages, alongside improved terms. Investigate Europe reports that it takes effect on November 1. Agreement is not proof that every payment has already reached a worker. Our concern is the distance between a customer's standards and the conditions experienced through subcontracting. A policy can be clear while a worker's route to challenge a deduction or raise a safety concern remains difficult. That gap cannot be measured by the size of a construction budget. The settlement offers a more practical story than despair: workers can organize and negotiate a different distribution of the gains. The next test is implementation. Verify payment, access to representation and freedom to raise concerns. The technology being built may be advanced; paying people correctly is not an advanced requirement.

4 min
An election-ad editing bay reveals synthetic layers behind apparently ordinary political video frames.
Law & informationUnited States+2 clusters03

AI-touched election ads passed $137 million, but most lacked an AI label

The Wesleyan Media Project identified 220 U.S. campaign ads made at least partly with AI and estimated more than $137 million in spending on them through October 11. It found such ads in 34 states and said only 27% of identified ads contained a disclaimer. The spending figure comes from AdImpact with Wesleyan's analysis. This is a count of identified AI-assisted ads, not a measure of all campaign spending or of voter persuasion. The researchers say it is almost certainly an undercount because subtle AI use can escape detection. Their comparison with less than $50,000 identified in 2024 is striking, but detection methods, markets and election cycles should not be treated as identical. The project's sample also found a partisan imbalance: 81% of the identified AI ads were sponsored by Republican candidates or groups, and 12% by Democrats. That describes this dataset, not every political message voters see. The real civic problem is that a voter often cannot know whether an image, voice or scene was synthetic at the moment it makes an impression. The FEC's 2024 interpretation said existing fraudulent-misrepresentation rules can apply regardless of technology, but it did not create a general AI-ad disclosure rule. Spending and disclaimers are observable; changed votes are not shown here. Our test for the remaining campaign is whether clear, consistent labeling catches up with the scale of use.

6 min
Separate legal case folders sit beside a chatbot phone screen and an unresolved balance scale.
Law & informationUnited States+3 clusters04

Thirteen new complaints against OpenAI test what AI safety promises mean in court

Tech Policy Press catalogued 13 complaints filed against OpenAI on October 13 and 14: eight separate suits by state attorneys general and five private complaints. The state filings generally allege misleading safety representations, while different cases raise issues involving minors' data, chatbot behavior, alleged psychological or physical harm and agent activity. The private complaints include a proposed class action about children's data and four individual suits. These are claims to be tested, not court findings. The complaints have different plaintiffs, legal theories and requested remedies; bundling them into one verdict would mislead readers. Iowa's attorney general, for example, publicly framed its filing under the state's Consumer Fraud Act. Tech Policy Press reports that OpenAI says it has built age-specific protections and wants pragmatic industry-wide policies. The public interest lies in the evidence the cases may force into view: what the company knew about particular risks, what users were told, whether safeguards worked in practice and what causal link exists between design and alleged harm. Litigation can create accountability, but it can also take years and produce narrow rulings. Courts may reject some claims while accepting others. A large complaint count is a signal of legal exposure, not a probability of liability. We will track pleadings, judicial decisions and independently verified safety outcomes rather than treating accusation as conclusion.

6 min
A neutral investigator examines two opaque AI systems and their surrounding safety records under a forensic light without any symbol of guilt or verdict.
Law & informationUnited States+2 clusters05

The FTC can demand AI safety evidence that voluntary pledges do not provide

One day after leading AI companies signed a voluntary White House accord built around internal controls, outside evaluation, and board oversight, the United States' consumer-protection agency confirmed that it is investigating AI companies. The Associated Press says an FTC spokesperson acknowledged an investigation involving OpenAI, Anthropic, and other companies but declined to provide its scope. Reuters, Axios, CBS News, and other outlets report that civil investigative demands may seek documents, testimony from executives, and information from independent evaluators. Those details remain reported rather than published by the agency. No company has been found liable, and an investigation is not proof that a safety claim was deceptive or a product harmed consumers. The agency does, however, possess an AI-specific compulsory-process resolution adopted in 2023, allowing staff to issue demands for documents, information, and testimony in consumer-protection or competition investigations. It has also used Section 6(b) orders to study AI partnerships and companion chatbots, a form of fact-finding that need not allege a law violation. The distinction matters because “probe” can describe very different processes. The public does not yet know the targets, legal theory, questions, time period, deadlines, or whether demands have been served. The real significance is the evidence boundary: voluntary auditors review what an agreement defines, while a regulator may compel records the company would not otherwise publish. Accountability begins when safety claims can be tested against the files behind them.

6 min
A university promotional banner emerges from an AI editing station with one student silhouette replaced while an unsigned consent form remains in the foreground.
PrivacyCalifornia, United States+3 clusters06

Stanford’s AI-edited banner replaced a real student and exposed a consent failure

Stanford University has acknowledged that a campus dining operation used generative AI to alter real students in a promotional photograph and published the result without disclosure. The original image was taken during a 2024 Lunar New Year dinner and had already appeared in university material. In the new banner, one Hispanic male student was replaced by a synthetic Black woman; reporting also found that two students’ faces or body shapes were changed and their clothing was converted into Stanford merchandise. The banner appeared in student housing before being removed. Stanford said both the alteration and lack of disclosure violated university rules and promised additional training and review. Its current communications guidance already contains the relevant protections: staff must obtain written permission before publishing an individual’s likeness, clearly identify materially manipulated media when omission could mislead, and may not create synthetic depictions of real people without explicit consent. The document also says a human must approve any automated workflow that produces public-facing content. That makes this more than an image-generation mistake. It is a control failure between policy and publication. The university has not publicly identified which tool was used, who approved the prompt or edit, whether the original releases permitted synthetic alteration, or how the banner passed review. The incident also exposes a crude temptation in institutional communications: instead of representing the people who are present, generative tools can manufacture the appearance an organization wants. Removing the banner addresses distribution. Rebuilding trust requires an auditable consent record, a review owner, and a way for people to know when their bodies or identities have been digitally changed before the file leaves the workflow.

9 min
Multiple international control lines converge on an independently operated frontier-model inspection gate inside a diplomatic chamber.
Law & informationGlobal+3 clusters07

Leaders from 20 countries call for independent control of frontier AI

An international appeal launched by Finland's president and Norway's prime minister has brought together 22 leaders and senior officials from 20 countries around a direct proposition: frontier AI must remain under human direction, oversight, and control. The signatories call for transparent company safety protocols, mandatory predeployment testing, independent evaluation with sufficient access, coordinated government standards, shared reporting of serious incidents, and scientific capacity that is not confined to wealthy states. They also ask UN members to explore an international institution that could set standards, enable verification, and convene governments when capability thresholds are crossed. The coalition is geographically broader than many earlier frontier-safety initiatives, spanning Europe, Africa, Asia, the Middle East, and North America. That breadth matters because AI failures and benefits cross borders while evaluation capacity remains concentrated. But this is an open political statement, not a treaty, enforcement body, budget, or agreed threshold. It does not specify who qualifies as an independent evaluator, what model access is mandatory, which incidents trigger reporting, or what happens when a company or state refuses. The signal is therefore political alignment around verification, not operational control. Its credibility will depend on whether endorsers convert the appeal into domestic access rights, common incident categories, funded evaluation institutions, and a process that can impose consequences when a frontier system fails a test.

8 min
Precision measurement instruments from multiple jurisdictions align around one frontier-AI calibration frame while a separate approval lever remains outside it.
Law & informationGlobal+4 clusters08

OpenAI proposes common frontier standards without global prerelease approval

OpenAI is proposing a U.S.-led international standards network for frontier AI, automated research, and recursive self-improvement. The company argues that shared measurements should cover capability evaluation, risk assessment, safeguard sufficiency, human oversight of automated research, and common severity levels for alignment incidents. It points to the existing international network created through the U.S. Center for AI Standards and Innovation as an institutional base. NIST says that network already includes government bodies from ten jurisdictions and has published consensus areas for automated evaluations. OpenAI draws a careful boundary around the proposal: the standards would not themselves be licenses, mandatory prerelease reviews, or approvals. National governments would decide whether and how to incorporate them into law. The post also says fully autonomous recursive self-improvement is not happening today and should not be pursued until it can be done safely. This is a consequential shift from general principles toward common technical definitions, but it also preserves national discretion and avoids a global permission system. A frontier developer has an obvious interest in standards that prevent fragmentation without slowing releases through external approval. That interest does not invalidate the proposal; it makes governance of the standard-setting process central. Credibility will depend on transparent methods, equal access for independent experts and open-model developers, declared conflicts, field validation, and evidence that a failed measurement changes what a laboratory is allowed to do.

9 min
A globe-shaped assembly table links an independent evidence panel to a ring of national seats, with one open gap in the global AI guardrail.
Law & informationGlobal+3 clusters09

The UN links scientific evidence to a global dialogue on AI rules

UN News describes a governance structure intended to match artificial intelligence's cross-border effects. Under the Global Digital Compact, member states created an Independent International Scientific Panel on AI and an annual Global Dialogue on AI Governance. The panel is meant to assess what is known and unknown about capabilities, opportunities, and risks; the dialogue gives governments and other stakeholders a place to compare approaches and coordinate. A preliminary panel report identified rapid progress in reasoning, coding, and science alongside misinformation, discrimination, privacy violations, cyberattacks, and possible future loss of control. The secretary-general argues that national action remains essential but that isolated, uneven, or unverifiable voluntary slowdowns will not be enough if risks rise. He has also called for child-safety commitments, support for developing countries, and contact between leading AI powers to avoid a race to the bottom. These mechanisms do not create a world regulator. The dialogue cannot automatically bind a frontier laboratory or a state, and geopolitical rivals may resist common restrictions precisely when they matter most. Yet the design contains an important principle: independent evidence should precede political bargaining, and countries outside the frontier race need standing in decisions whose effects cross their borders. Success should be measured by whether the panel can publish contested findings, whether the dialogue produces interoperable safeguards, and whether agreed evidence activates action rather than another declaration.

7 min
A mechanical confidence dial controls an answer gate while a separate correctness marker remains visibly misaligned.
Technical failuresGlobal+1 clusters10

Language models use internal confidence to decide when to abstain

A peer-reviewed study has moved the debate about AI uncertainty beyond asking whether a model can produce a confidence score. Across four language models, researchers used a four-phase experiment to test whether confidence-related internal states actually drive the decision to answer or abstain. Confidence strongly predicted refusal behavior. More importantly, activation steering that boosted or suppressed confidence changed abstention rates, and instructions that altered the decision threshold changed behavior without fundamentally changing the underlying confidence representation. That is causal evidence for a two-stage control process: an internal confidence signal and a policy that decides how much confidence is enough. The safety opportunity is real. Systems could be engineered to defer, verify, or request human review when their own uncertainty crosses a tested boundary. The warning is just as important. Verbal confidence independently influenced abstention even though it was less effective than calibrated token probabilities at distinguishing correct from incorrect answers. A model can therefore act on a confidence signal that is behaviorally powerful but imperfectly connected to truth. This is not evidence of consciousness, and the experiment does not show that open-ended agents can reliably monitor long reasoning chains. It used factual multiple-choice questions without chain-of-thought instructions. The practical lesson is narrower and more useful: confidence is a control surface. High-stakes deployment must validate both the internal signal and the threshold policy under real costs, because a model that knows when it feels unsure can still be confidently wrong about whether to proceed.

5 min
A private phone line connects a corporate tower and Washington above competing blueprints for a national AI regulator.
Law & informationUnited States+1 clusters11

A private call exposes the fight over who should regulate frontier AI

The fight over a national AI regulator has moved behind closed doors. Politico reports that Meta's chief executive told President Trump in a private call that a proposed FINRA-style AI body was a flawed idea and could be vulnerable to regulatory capture. The model under discussion reportedly involved an independent organization operating with government oversight and industry membership or funding. Supporters could argue that one technically specialized body would reduce the conflict among state rules, concentrate expertise, and update standards faster than Congress. Critics can reasonably worry that the largest companies would finance the institution, shape its membership, control access to evidence, and write compliance standards that smaller rivals cannot afford. The report relies on anonymous sourcing and no transcript of the call is public. A second person familiar with the conversation told Politico that the executive did not ask the president to change his stance. Those limits matter, especially when the headline involves private influence. The larger governance question is still visible: whether AI oversight should be led by a public agency, an industry self-regulator, or a hybrid. The answer should not be inferred from the word independent. It should be tested through appointments, funding, statutory authority, public representation, disclosure, audit access, enforcement power, and appeal rights. A regulator can coordinate a market or entrench it. Its institutional design decides which.

5 min
An uncertainty-aware AI map narrows hundreds of possible chemistry experiments to one illuminated vial while a laboratory counter records fewer physical trials.
Social good & healthGlobal+2 clusters12

A language model learned uncertainty and reached results with 41 percent fewer experiments

A Nature Machine Intelligence study introduces GOLLuM, a framework that trains language models through the probabilistic objective used in Gaussian-process Bayesian optimization. Instead of treating a language model as a confident generator of experimental suggestions, the method reshapes its internal representation using observed outcomes and calibrated uncertainty so it can help decide which experiment to run next. Starting from ten low-performing experiments, GOLLuM ranked first on average across 23 tasks spanning organic synthesis, process chemistry, materials, catalysis, and molecular design. It matched traditional Bayesian optimization's final performance with a median 41 percent fewer iterations. In a Buchwald–Hartwig reaction benchmark, the approach nearly doubled the discovery rate for high-performing conditions compared with expert quantum-chemical descriptors and state-of-the-art language models, 43 percent versus 24 to 25 percent. The result matters because laboratory time, materials, and failed experiments are expensive. It also shows that uncertainty can be part of a model's training objective rather than a confidence label added afterward. The evidence comes from benchmarked experimental-design tasks, not unrestricted autonomous laboratories. Domain review, physical safety limits, dataset quality, secondary objectives, replication, and transparent decision records remain necessary before an optimization gain becomes a discovery system people can trust.

6 min
A brutalist paper polygraph confidently identifies identical masks but falters when an unfamiliar mask enters the test chamber.
Technical failuresGlobal+2 clusters13

Anthropic's lie detector scored 0.95 at home and stumbled outside the test

Anthropic's Alignment Science team trained lie detectors using roughly 200,000 labeled examples from 12 settings and eight model families. In-distribution performance rose from an AUROC of 0.60 to 0.95, but cross-category transfer reached only about 0.70 to 0.75, and larger models prompted as judges often beat the fine-tuned detectors. The research also exposes a label problem: about one quarter of labels changed during a GPT-5-assisted cleaning process, particularly around ambiguous behavior such as sycophancy. Third-person monitoring worked better than asking a model to report on itself. The team released its datasets and explicitly limits its conclusion to controlled settings rather than production behaviors such as alignment faking or reward hacking. The result is a valuable negative finding. A detector that excels only on familiar lies is not a universal truth machine, and institutions must not convert an uncertain score into punishment without evidence and appeal.

5 min
A barred campaign podium casts a synthetic AI-avatar silhouette toward a Brazilian ballot box while a legal boundary flashes red.
Cognition & learningBrazil+2 clusters14

A barred politician’s AI avatar is testing Brazil’s election rules

An AI-generated avatar of former Brazilian president Jair Bolsonaro appeared at the launch of his son Flávio’s presidential campaign while the elder Bolsonaro remains under house arrest and barred from public political communication. The video disclosed that it was AI-generated, but leftist parties challenged it in court, arguing that synthetic media can influence voters and potentially route around judicial restrictions. The episode expands the election-integrity problem beyond deceptive deepfakes: a recognizable digital surrogate can reproduce the political force of someone legally unable to campaign.

3 min
Technical failuresAustralia+1 clusters15

Klindt et al., “A unifying framework from neural superposition to sparse interpretable codes”

Researchers from Australian National University, UC Santa Barbara and partner institutions address the problem of neural networks representing more concepts than they have individual neurons, making internal representations difficult to interpret. Their proposed framework combines identifiability theory, sparse coding and behavior-grounded metrics to determine whether extracted model features correspond to meaningful concepts.

2 min