AI safety promises are getting weaker as models get stronger

What happened

The Summer 2026 AI Safety Index from the Future of Life Institute scored major AI companies on risk management, governance and safety practices. Anthropic received the strongest overall result at C+. OpenAI and Google DeepMind received C grades. No company reached an A or B.

The report argues that voluntary commitments have not kept pace with increasingly capable systems. It points to gaps between stated principles and evidence that safeguards would hold under pressure.

Why it matters

Safety plans are easy to describe at a high level and difficult to test in public. The central issue is accountability: who can verify a claim, what happens when a threshold is crossed, and whether a company can change course when commercial pressure rises.

A third-party scorecard cannot settle those questions, but it can expose where comparable evidence is missing. It also gives policymakers and customers a clearer list of questions to ask.

How to read the index

The Future of Life Institute is an advocacy organisation with its own policy position. Its grades are judgments based on a published methodology, not neutral measurements handed down by a public authority. The report is still useful if readers examine the categories and underlying evidence rather than treating the letter grade as the whole story.

What remains unclear

Companies can dispute the weighting, disclose new material or change their policies. A more durable system would combine independent evaluation, incident reporting and enforceable rules. The index shows the distance still to travel; it does not by itself prescribe the final regulatory model.