Future of Life Institute's first AI Safety Index gives every major AI lab a C+ or below, and the evaluator's verdict is that the four named labs are 'moving the goalposts' on prior safety pledges.
Anthropic, OpenAI, Google DeepMind, and Meta all scored C+ or below on the Future of Life Institute's inaugural Summer 2026 AI Safety Index, released this week. FLI says the labs are "moving the goalposts" on safety commitments they made publicly just two years ago. No lab cleared a C+.
The index is not a peer-reviewed study. It is a third-party grades-and-rankings exercise from a long-standing AI-safety nonprofit, designed to track whether companies are keeping the voluntary safety pledges they made during the 2023–2024 model launch cycle. FLI's verdict: the gap between rhetoric and practice has grown since the index's Winter 2025 baseline.
The "goalposts" critique is concrete. FLI's methodology scores labs on specific commitments: publishing model evaluations before deployment, conducting third-party red-teaming of frontier systems, disclosing training-compute thresholds, and maintaining internal frontier-risk policies tied to capability milestones. The Summer 2026 index finds that several labs have either narrowed the scope of those commitments, pushed disclosure to after release, or stopped publishing the underlying evaluations. The pattern is retreat by redefinition: a company can keep the words of a 2024 pledge and shift what those words cover through narrower scope, later disclosure, or removed evaluations.
Three outlets have reported on the index: WBUR's Here & Now interview with Axios chief technology correspondent Ina Fried, TIME, and Axios itself. Fried said the companies have not disavowed their original safety commitments; they have narrowed what those commitments cover. TIME and Axios's reporting aligns with FLI's finding: the gap between stated safety policy and operational practice is widening, not closing. WFAE and Iowa Public Radio versions of this story are syndicated pickups of the WBUR segment, not separate reporting.
The public record so far is FLI's graded assessment, sympathetic but limited mainstream-press coverage, and silence or non-response from the labs themselves. No company has published a public rebuttal with the same specificity as FLI's methodology. Until this index, the loudest public signal on whether AI labs were keeping their safety pledges was the labs' own announcements. Frontier AI development is concentrated in four private labs, none of which face a regulator with jurisdiction to enforce those pledges. The FLI index is the first external, comparable grade to substitute for that self-reporting void.
The Summer 2026 edition is the baseline. The next FLI update will say whether the labs recover, hold, or keep eroding on the specific commitments the index now tracks. Each company now has three public options: publish the evaluations FLI says are missing, narrow its public commitments to match its actual practice, or contest FLI's methodology with the same level of detail FLI has published. Each path is now a public test of the original commitments.