Is this index's scoring methodology credible, or is it just subjective impressions?
Scoring was conducted by a panel of seven independent researchers spanning AI alignment, governance, and interpretability expertise. Grades are based on publicly available model cards, research papers, and benchmark results, combined with a targeted company survey designed to close specific transparency gaps in the industry (such as whistleblower protections and external model evaluation details). Each domain is graded using the absolute US GPA scale (A/B/C/D/F corresponding to 4.0–0), with individual reviewer grades kept confidential — the final score is an average across multiple reviewers, not one person's subjective call.
That said, the Index's data collection cut off on June 3, 2026, and doesn't include any events that may have occurred since — a timing limitation worth keeping in mind when interpreting the scores.
All nine companies were scored — is the bar the same for Chinese companies as for US and European ones?
The report devotes a dedicated section to China's distinct regulatory context: national binding instruments (such as the Cybersecurity Law and the Generative AI Interim Measures) carry direct legal force, with violations risking fines up to 50 million RMB or 5% of global revenue. By contrast, the kind of "voluntary commitments" common among US companies mostly correspond, in the Chinese context, to non-binding recommended standards or promotional local regulations not yet in force — carrying far less weight than national binding rules.
This means the low scores Chinese companies (DeepSeek, Alibaba Cloud, Z.ai) received on indicators like voluntary safety commitments partly reflect a structural difference in regulatory pathways, not necessarily a lack of safety concern. The panel's company-specific recommendations note that these companies' ratings "largely reflect the Chinese regulatory environment rather than independent safety leadership" — in other words, the low scores don't yet fully capture these companies' actual safety practices either.
If every company scored low, does this index actually move the industry forward?
Looking at the historical trend, this is FLI's third such index since 2024, and past editions show some companies do adjust practices in response to their scores — for instance, Meta moved up from 6th to 4th place this edition, while xAI dropped from 4th to 7th. This suggests the Index currently functions more as a mechanism for sustained external pressure than as a regulatory tool that can directly force companies to change behavior.
The panel also flags a structural limitation in the report: several companies have recently published or updated safety frameworks, but those frameworks themselves "sometimes lack quantitative thresholds, genuinely independent audits, and clear decision authority." In other words, even as companies release more safety frameworks publicly, most of these frameworks currently "look substantive but lack real enforceability" — which is exactly why third-party evaluations like FLI's continue to matter: they at least provide an externally comparable benchmark that can track trends over time.
Anthropic got the highest score overall — does that mean it has the fewest safety problems?
A relatively higher score doesn't mean an absence of problems. The panel's company-specific recommendations for Anthropic explicitly state it needs to "reverse the RSP 3.0 walk-back on pause commitments and restore credibility of commitments" — meaning even the top-ranked company was flagged by the panel for the same "moving goalpost" problem. The panel also recommends Anthropic "treat prevention as seriously as interpretability/detection," implying its current safety strategy may lean too heavily on after-the-fact detection rather than upfront prevention.
A more accurate reading: Anthropic's relative edge mainly comes from stronger performance on indicators that are comparatively easier to satisfy — transparency disclosure, breadth of risk assessment, and governance structure completeness. But in Existential Safety, the industry's weakest domain overall, Anthropic scored only D+ — same as everyone else, it hasn't actually solved the underlying problem.
The nonprofit Future of Life Institute (FLI) released its latest AI Safety Index in July 2026, with a panel of seven independent researchers and governance experts scoring nine major AI companies across 37 indicators spanning six domains. The results: no company earned an A or B grade. Anthropic topped the field with a score of 2.66 (C+), followed closely by OpenAI (C, 2.28) and Google DeepMind (C, 2.01), while xAI, DeepSeek, and Mistral all received failing grades (F).
Since its first release in 2024, the Index has become one of the most closely watched third-party safety evaluations in the industry, and this edition was widely cited by outlets including The New York Times, Financial Times, and TIME.
More notable than the raw scores is a trend the panel flagged directly in the report: Anthropic, OpenAI, Google DeepMind, and Meta — four companies previously seen as relative leaders in safety practice — have all been observed weakening or walking back earlier commitments to pause development unilaterally when approaching capability red lines, with some attaching competitor-contingent conditions to those pledges. The panel calls this phenomenon "moving the goalpost," writing that it has "undermined safety frameworks across the board."
UC Berkeley professor Stuart Russell's comment is arguably the sharpest line in the report. He noted that companies previously committed to releasing new systems "only with safety measures appropriate for their capability levels," but the current trajectory shows they "are planning to release them even if it's demonstrably unsafe to do so."
Among the six scored domains, Existential Safety was the worst-performing category across all nine companies — not a single one scored above C-, with most falling to D or below. The panel specifically named Anthropic's constitutional classifiers, OpenAI's calls for governance institutions, Google DeepMind's monitoring commitments, and Meta's loss-of-control provisions as genuine attempts, but judged them collectively "entirely inadequate." The report also questions the industry's dominant reliance on interpretability and Chain-of-Thought monitorability, arguing that "detection is not prevention."
University of Montreal assistant professor David Krueger's assessment was equally blunt: he called the industry's lack of progress toward credible safety plans "scandalous," adding that companies themselves are starting to grow anxious as they race toward recursive self-improvement and confront the prospect of losing control — yet recent CEO gestures toward coordinated pauses or slowdowns still fail to communicate how urgent the risk is or how unprepared the industry remains.
The report highlights a notable disconnect: while the EU is widely regarded as a global leader in AI safety regulation, Europe's most prominent AI company, Mistral, scored dead last among all nine companies — with safety framework, existential safety strategy, and governance disclosure all at or near zero. The panel's recommendation is blunt: Mistral needs to "engage substantively with existential safety," since its leadership "consistently downplays — and at times dismisses — frontier risk" without articulating any identifiable control or alignment strategy.
The three companies receiving failing grades come from the US (xAI), China (DeepSeek), and Europe (Mistral) respectively — and the panel's key findings explicitly stress that inadequate safety is "a global problem, not a regional one."
For the first time, the report lists the industry's pivot toward military use as a notable current-harm risk. The panel observed that between 2024 and 2026, companies including Anthropic, OpenAI, Google DeepMind, and Meta — which had previously banned military applications outright — gradually reversed course and began actively pursuing defense partnerships, effectively aligning with xAI and Mistral, which had never restricted such use. The report notes that Anthropic drew criticism from the panel over "questionable military engagements," including a reported link to a school strike that caused mass civilian deaths. China's Alibaba Cloud and Z.ai, meanwhile, face U.S. allegations of military ties, which both companies deny.
For investors and practitioners tracking the AI industry, this Index offers a calibration point beyond companies' own safety claims. The gap between a company's public emphasis on "taking safety seriously" and its actual performance across 37 concrete indicators is exactly what the panel calls "safety rhetoric outpacing revealed behavior" — the report specifically names Google DeepMind, OpenAI, and xAI as showing clear divergence between public messaging and their commercial conduct or legislative lobbying positions. In other words, stated safety commitments alone are not yet a reliable proxy for a company's actual risk-management capability — which is precisely why independent third-party evaluations like this one matter for industry-wide transparency.