The Future of Life Institute graded nine leading AI companies on safety and not one of them scored above a C+. Anthropic took the top spot with a C+ and an overall score of 2.66 out of 4; xAI, DeepSeek and Mistral each received a failing grade. The report's sharpest finding is not the low ceiling but the direction of travel: several labs have quietly weakened the safety commitments they made when the stakes were lower.
Key facts
- The headline number: The best grade in the entire industry was a C+ (Anthropic, 2.66 out of 4). Three of nine companies failed outright.
- When: Published July 14, 2026.
- Who: The Future of Life Institute, with an independent panel of expert reviewers grading across six domains.
- Primary source: The Summer 2026 AI Safety Index.
The index works like a report card. An independent review panel scores each company across six domains -- among them risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing -- using public evidence and company disclosures. It is the closest thing the field has to a standardized audit, in an industry with no mandatory one.
What the grades say
Anthropic finishes first and the report is specific about why: "Anthropic again earns the highest overall grade and leads five of six domains via relatively strong transparency, a comparatively established safety framework, technical research, and governance." Note the hedging in the report's own words -- "relatively," "comparatively." It is winning a race in which everyone is walking.
At the bottom, the failures are geographically even-handed. As the index's summary sheet puts it: "Three companies receive failing grades, one each from the US (xAI), China (DeepSeek), and Europe (Mistral)." That distribution matters, because it undercuts the convenient story in which safety is a Western virtue and recklessness is imported. It isn't. Every jurisdiction produced a lab that couldn't clear the bar.
The weakest domain across the whole industry was existential safety, where most companies scored D or below. This is the category covering whether a company has any credible technical plan for keeping control of a system meaningfully smarter than its operators. The reviewers' verdict is that essentially nobody does -- including the labs whose leaders talk most about it.
The backsliding
The finding with the longest shelf life is about commitments already made and then unmade. The index states it directly: "Even industry leaders in safety practices are retreating from prior commitments, despite calling publicly for a pause."
Anthropic, OpenAI, Google DeepMind and Meta have all weakened or voided pledges to stop unilaterally if capability redlines were approached. Anthropic's is named specifically -- the reviewers call for reversing the "RSP 3.0 walk-back" on pause commitments and replacing qualitative safety thresholds with quantitative, risk-tied ones. That recommendation lands awkwardly next to Anthropic's own recent argument that the industry needs a pause button the world can verify. The company asking for a verifiable brake is, per the index, among those loosening its own.
The reviewers also flag a gap between rhetoric and behavior at Google DeepMind, OpenAI and xAI, where leadership messaging about caution runs against the companies' commercial conduct and their positions in legislatures.
Here is the mechanism worth understanding. A voluntary safety commitment is a promise made in a moment of low cost -- before the capability exists, before a competitor is shipping it, before a board is watching the revenue line. The commitment gets tested precisely when it becomes expensive, which is exactly when the incentive to reinterpret it peaks. A qualitative threshold ("if capabilities approach dangerous levels") is reinterpretable by construction. A quantitative one tied to a measured risk is not, which is why the reviewers want the swap and why it is not happening.
Why it matters, and the caveat
This is the argument for external structure rather than pledges, and it is the same argument Demis Hassabis made in proposing a FINRA-style regulator for frontier AI. Self-regulation that erodes under competitive pressure is not regulation; it is a press release with a decay function. The index's value is that it makes the decay legible over time -- you can compare this edition to the last one and see who moved.
The honest caveat is that grading is judgment, not measurement. FLI is an advocacy organization with a stated position on existential risk, its panel is assembled rather than randomly drawn, and the domains are weighted by choices someone made. A company that discloses more can score worse than one that discloses nothing interesting -- transparency is partly self-punishing here. The grades are best read as a structured expert opinion with its work shown, which is more than the industry offers about itself, and less than an audit.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)