Do the New AI Safety Grades Mean We're Doomed?

A report card landed this summer that was tailor-made to go viral. The Future of Life Institute's Summer 2026 AI Safety Index graded the world's leading AI companies, and the best of them, Anthropic, managed only a C+. OpenAI and Google DeepMind got a C. Meta got a D+. xAI, DeepSeek and Mistral flunked outright with an F. And in the category the index calls "existential safety," not a single company scored above a D.

It is easy to read that last line and conclude the worst: the people building the most powerful technology on earth are failing the test that matters, so catastrophe must be close. That reading is understandable. It is also mostly wrong, and it is worth understanding why.

What the grade actually measures

The index does not measure how close any model is to going rogue. It measures preparation. Across six domains and 37 indicators, a panel of independent reviewers scores whether a company has the governance, transparency, risk assessments and published safety frameworks you would want from an organization handling a serious technology. A D in "existential safety" does not mean a lab is one bad afternoon away from disaster. It means the reviewers judged its plans for managing far-off, high-stakes risks to be thin, vague or unproven.

That is a real criticism, and the labs should sit with it. But it is a criticism of process, disclosure and candor, not a measurement of imminent danger. Confusing the two turns a useful prod toward better practice into a doomsday headline.

Low bar, not high cliff

The more honest takeaway is quieter, and in its way more useful. The grades are low because the industry is measuring itself against a standard of good safety practice that it largely wrote, and still cannot meet. No company scoring above a C+ says less about how dangerous today's models are and more about how immature safety engineering still is at the frontier. The reviewers found the strongest labs, the ones publishing responsible-scaling policies and letting outsiders probe their systems, doing meaningfully better than the ones that disclose almost nothing.

That is the signal worth keeping. Not "we are doomed," but "self-regulation is thin, and transparency varies wildly between companies." It is the same clear-eyed skepticism worth bringing to claims that superintelligence is arriving in 2027, or that AI will collapse under its own synthetic output. A bad grade on a safety scorecard is an argument for better oversight and blunter disclosure. It is not evidence that the end is near.