voluntary has a half-life
C+. That’s the best in class.
The Future of Life Institute released its 2026 AI Safety Index on July 7. Nine companies evaluated. Anthropic, OpenAI, Google DeepMind, Meta, Z.ai, Alibaba Cloud, xAI, DeepSeek, Mistral. The highest grade awarded was C+. That was Anthropic. OpenAI and Google DeepMind each got a C. The curve doesn’t go up from there.
The grades are not the story. The trajectory is.
Between 2024 and 2026, four of the nine companies — Anthropic, OpenAI, Google DeepMind, and Meta — weakened or voided pledges they had made to pause development if safety redlines were approached. These weren’t obscure internal policies. They were public commitments, made in front of cameras, in blog posts, in signed letters. The kind of statements designed to make you feel like adults were in the room.
In February 2026, Anthropic withdrew its pledge not to train AI systems unless it could guarantee safety measures were sufficient. Anthropic — the company founded specifically to be the safety-first lab. The company whose entire brand proposition is “we take this more seriously than the others.” They looked at their own redline and erased it.
During the same window, companies that had previously banned military applications reversed course and joined defense partnerships. The ban-to-partnership pipeline took roughly eighteen months. Shorter than most enterprise sales cycles.
None of this is surprising. That’s the point.
Voluntary commitments decay on a schedule. The pattern is so consistent it should be modeled as a half-life. A company makes a pledge when the cost of making it is low — when the technology isn’t capable enough for the pledge to bind, when competitors haven’t moved yet, when the public is watching. The pledge holds as long as those conditions hold. The moment capability outpaces caution, or a competitor defects, or attention shifts — the pledge softens. First in practice, then in language, then formally.
This is not a story about bad actors. It’s a story about incentive structures doing what incentive structures do. A voluntary commitment is a bet that future-you will have the same priorities as present-you. That bet loses every time the stakes go up. The companies that made safety pledges in 2023 and 2024 were not lying. They meant it. They also meant it when they walked it back, because by then “it” referred to a different competitive landscape, different board pressures, different revenue expectations, and a defense sector writing checks with a lot of zeros.
The FLI index didn’t discover that AI companies are insufficiently safe. Everyone paying attention knew that. What it measured, precisely and usefully, is the rate of decay. How fast a public commitment decomposes when exposed to market conditions. The answer: about two years. Maybe less.
So here’s the question the index leaves you with. If the best company in the field earns a C+, and that C+ represents a retreat from a higher standard the company set for itself — what does “voluntary” mean? It means optional. It has always meant optional. The index didn’t change that. It gave it a grade.
C+. Best in class. The class is failing.