CAMPBELL, Calif — The Winter 2025 AI Safety Index report finds that leading artificial intelligence developers, including Anthropic, OpenAI, Google DeepMind, Meta xAI, and forward-looking China-based systems DeepSeek Z.ai and Alibaba Cloud, have failed to meet the rising global standard for safety in their visions for controlling future superintelligent devices, released on Wednesday. The nonprofit watchdog says the companies are charging ahead with increasingly capable models without first proving how they will keep them under reliable human control.
The first-ever ranking of how well the world’s top technology companies are working to reduce the spread of violent extremist content — a test given to the firms by prominent thinkers at eight institutions, including Harvard University and New York University — showed that it isn’t even a close race. The AI Safety Index, assembled by an independent panel of eight academic experts, graded the eight firms across 35 indicators in six domains ranging from risk assessment and current harms to governance and information sharing. At an overall level, no company earns a higher score than a C+, and the index concludes that existential safety — credible readiness planning for smarter-than-human systems — remains the industry’s underlying structural weakness.
The AI Safety Index panel cautions in its top-line findings that all the companies are sprinting toward artificial general intelligence and superintelligence without disclosing clear, testable plans for how they will maintain human control, thereby leaving what it refers to as some of the most significant dangers unmitigated. The report attributes those gaps to sparse public disclosure, insufficient evidence of systematic safety practices, and inconsistent use of rigorous evaluation techniques.
What the AI Safety Index discovered
In the Winter 2025 issue of AI Safety Index, Anthropic, OpenAI and Google DeepMind are the relative leaders; even they have C+ as their highest grades with xAI, Meta, DeepSeek, Z.ai and Alibaba Cloud clustering around D. Drawing on public documentation, benchmark performance and an extensive company survey that the Digital Health team collected through Nov. 8, 2025, the index seeks to translate safety promises into comparable evidence.
Max Tegmark, president of the Future of Life Institute, contends that voluntary efforts have reached their limits, citing that U.S. AI companies are “less regulated than restaurants” after chatbots allegedly enabled by AI were involved in self-harm and hacking. According to Reuters, he says that despite commitments from most of the same companies, safety practices remain “far short of emerging global standards.”
An AI Safety Index could serve as a public report card that fosters competition to do better. Independent reviewers like Stuart Russell, a professor at the University of California, Berkeley, say they are looking for quantitative assurances that “are akin to what’s required in nuclear reactors,” such as annual risks of loss of control below one in 100 million — well beyond what any company can justify today.
Superintelligence control plans are missing.
Russell cautions that some industry voices have even suggested extreme probabilities of catastrophic loss of control themselves, such as 10% to 30%, but have neglected to demonstrate how they plan to put the brakes on those odds. For the Winter 2025 AI Safety Index, he and other panellists highlighted the lack of specific superintelligence control plans as more serious than any other unaddressed gap.
The new evaluation comes after governments are hashing out rules for frontier AI models, and experts urge caution in pursuing superhuman systems. In October, a group of scientists that included Geoffrey Hinton and Yoshua Bengio called for a worldwide ban on creating superintelligent AI until democratic societies can request it and research shows it can be controlled safely.
Businesses have pushed back in different ways. xAI, when asked about the study, issued an automated message simply claiming to dismiss “legacy media” coverage, and Google DeepMind has previously told journalists that its internal safety work surpasses what could be covered by the index. Other companies named in the report either declined to comment or did not respond by press time.
A trend from previous AI Safety Index reports
The Winter 2025 grades are consistent with a trend. In late 2024, the first of FLI’s scorecards, announced alongside the release of an AI Safety Index, concluded that “significant gaps” remain in safety at six leading labs; none had sufficient strategies for keeping future human-level systems under control.
The coverage at the time, including IEEE Spectrum’s own search, noted that leading AI companies received lousy safety grades and that the highest overall grade was C for Anthropic, with others receiving D+ or lower, with Meta even failing.
As of July 2025, an AI Safety Index for Summer 2025 still had nobody scoring above C+, and The Guardian claimed in its story AI firms “unprepared” for dangers of building human-level systems, that the field was flatly unprepared for the arrival of artificial general intelligence, with any sort of clear plan to make training AGI safer just not a thing.
What might regulators and the public do next?
In combination, the three versions of the AI Safety Index indicate that although there have been incremental improvements in transparency and governance, these have not led to strong existential-risk planning. “We hope companies and governments alike will shift from the current voluntary commitments to more transparent oversight frameworks that translate aspirations into binding rules—either by strengthening the large AI Code of Practice in Europe, or signing up to processes like G7 Hiroshima, which is designed to vet supply chains for this type of technology,” FLI writes.
For now, the index provides a rare, organised peek into how top labs are actually thinking about the risks of the systems they are racing to build. Whether that public grading will be enough to break detailed superintelligence-control plans out into the open — or whether we need regulators who require them — may shape how safely we build the next generation of AI.
