The AI Safety Index is a recurring report card, published by the nonprofit Future of Life Institute (FLI), that grades the world’s leading artificial intelligence developers on how seriously they manage the risks of the systems they build. Each company receives a school-style letter grade, from A to F, based on a panel of independent experts scoring public evidence and, where companies choose to respond, their own survey answers.

What It Measures

The index does not just ask whether a company talks about safety — it scores actual practices across six domains:

  • Risk assessment — whether a company systematically tests its models for dangerous capabilities before release.
  • Current harms — how well a company limits real-world problems already happening today, such as biased outputs or misuse.
  • Safety frameworks — whether published safety policies include measurable thresholds, not just general promises.
  • Existential safety — planning for scenarios where highly capable future systems could cause catastrophic, hard-to-reverse harm.
  • Governance and accountability — internal oversight structures, board authority, and whistleblower protections.
  • Information sharing — how transparent a company is with regulators, researchers, and the public about what it knows.

Each domain is broken into specific indicators — dozens in total — so a company’s final grade reflects many individual judgment calls rather than one vague impression.

How the Grading Works

FLI recruits a small panel of independent academics — specialists in machine learning, AI governance, and risk analysis from universities including UC Berkeley, Oxford, and the University of Montreal — who score each domain using a standard US-style GPA scale (A = 4.0 down to F = 0). Their domain scores are averaged into an overall grade per company. Because the panel works from public documents, model cards, and policy pages rather than confidential internal access, the index measures what a company demonstrably discloses and does — not what it privately intends. FLI has published the index roughly twice a year since 2024, allowing grades to be tracked over time as companies’ practices (or lack of progress) become visible.

Why It Matters

AI companies frequently publish their own safety pledges and blog posts, which makes it hard for outsiders to compare one company’s actual practices against another’s marketing. An independent, recurring, methodologically consistent scorecard gives regulators, investors, journalists, and the public a common yardstick. It also creates a form of reputational pressure: no major AI lab has ever scored above a C+ on the index, and the existential-safety domain in particular has consistently drawn the lowest marks industry-wide, which the panel has pointed to as evidence that safety planning is not keeping pace with how fast model capabilities are advancing.

In the News

The Summer 2026 edition of the index, published in July, graded nine companies — Anthropic, OpenAI, Google DeepMind, Meta, Z.ai, Alibaba Cloud, xAI, DeepSeek, and Mistral — and none scored above a C+. See our report: Future of Life Institute: No AI Firm Scores Above a C+. For the broader field this index sits within, see our explainer on what AI safety actually covers.

FAQ

Is the AI Safety Index a government requirement?
No. It’s independent research from a nonprofit advocacy organization, not a binding regulation — though regulators and journalists frequently cite it.

Do companies have to participate?
No. FLI scores public evidence regardless of participation, but companies can also submit survey responses and supporting documents that the panel factors in.

How often is it updated?
Roughly twice a year, allowing grades to shift as companies change their practices or as new evidence becomes public.

Why hasn’t any company scored an A?
The panel has said no company has published safety frameworks with the measurable thresholds and independent audits it considers necessary for a top grade, particularly around lower-probability but catastrophic risks.

Sources: Future of Life Institute — AI Safety Index, Summer 2026