Last reviewed
Jun 8, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,699 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Future of Life Institute AI Safety Index (often shortened to the AI Safety Index or FLI AI Safety Index) is a periodic "report card" published by the Future of Life Institute (FLI) that grades the leading frontier AI developers on their safety-relevant practices. An independent panel of AI safety and governance experts assigns letter grades across a set of domains, including risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing. The Index is intended to create competitive pressure for stronger safety commitments by making cross-company comparisons transparent and reproducible.[1][2]
Three editions had been released as of December 2025: the inaugural report in December 2024, followed by editions branded "Summer 2025" (July 2025) and "Winter 2025" (December 2025). Across all editions, Anthropic has scored highest overall but has never exceeded roughly a C+, while every company evaluated has received a D or F grade on the existential-safety domain that concerns maintaining control over artificial general intelligence and superintelligent systems.[3][4][1]
The Future of Life Institute is a nonprofit founded in 2014 by Max Tegmark and others to reduce extreme risks from transformative technologies, with a particular focus on advanced AI. FLI launched the AI Safety Index in December 2024 as a recurring benchmark of how the major general-purpose AI companies manage safety, modeling the format loosely on a school report card with per-domain and overall letter grades.[1][5]
The Index does not attempt to certify that any product is "safe." Instead, it compares companies against one another and against expert expectations, surfacing relative strengths and gaps. Tegmark, who is also a physics professor at the Massachusetts Institute of Technology, has framed the project as an accountability mechanism, noting in connection with the Winter 2025 edition that in the United States AI is "less regulated than sandwiches" and that companies continue lobbying against binding safety standards.[6][2]
Each edition grades companies across six domains. The exact indicator set has evolved between editions, but the domains have been broadly consistent:[2][4]
The Winter 2025 edition assessed companies against roughly 35 indicators grouped under these six domains.[4]
At the center of the Index is an independent review panel of prominent AI researchers and governance scholars. FLI staff compile a structured evidence base for each company drawn from public materials, including research papers, model and system cards, policy documents, and news reporting, supplemented by responses to a tailored survey that companies may complete to demonstrate additional transparency. The panel then assigns a letter grade to each company within each domain, and the per-domain grades are converted to the standard United States grade-point scale (A = 4.0, B = 3.0, C = 2.0, D = 1.0, F = 0, with pluses and minuses) and averaged into overall scores.[2][1]
The analysis of the safety-frameworks domain follows the taxonomy and indicator structure developed by SaferAI, a nonprofit research organization focused on AI risk assessment.[2] Because grading relies heavily on publicly available evidence, companies that disclose less tend to score lower, a dynamic the panel has acknowledged.[3]
Panel membership has shifted across editions while retaining a stable core. Members have included Yoshua Bengio (2024 edition), Stuart Russell, Dylan Hadfield-Menell, Tegan Maharaj, David Krueger, Jessica Newman, Sneha Revanur (founder of Encode), Atoosa Kasirzadeh, Sharon Li, and Yi Zeng. The 2024 panel comprised seven experts; the Winter 2025 panel listed eight.[1][2]
The first AI Safety Index, released on December 11, 2024, graded six companies: Anthropic, OpenAI, Google DeepMind, Meta, xAI, and Zhipu AI. According to contemporaneous reporting by IEEE Spectrum, Anthropic received the highest overall grade at C, followed by Google DeepMind and OpenAI at D+, Zhipu AI at D, xAI at D-, and Meta lowest at F. The panel found wide disparities in safety practice, judged all flagship models vulnerable to adversarial attacks, and noted that no company scored above D on existential safety.[7][1]
The Summer 2025 edition, published in July 2025, expanded the comparison and refined the indicators. The overall grades were:[3]
| Company | Overall grade | Score (4.0 scale) |
|---|---|---|
| Anthropic | C+ | 2.64 |
| OpenAI | C | 2.10 |
| Google DeepMind | C- | 1.76 |
| xAI | D | 1.23 |
| Meta | D | 1.06 |
| Zhipu AI | F | 0.62 |
| DeepSeek | F | 0.37 |
The panel again highlighted a disconnect between companies forecasting AGI within roughly a decade and the absence of any coherent, actionable plan for controlling human-level systems; no company scored above D in the existential-safety domain.[3]
The Winter 2025 edition, with a full report dated December 13, 2025 and a data cutoff of November 8, 2025, evaluated eight companies and added Alibaba Cloud, while Zhipu AI appeared under its consumer brand Z.ai. The overall grades were:[4][2]
| Company | Overall grade | Score (4.0 scale) |
|---|---|---|
| Anthropic | C+ | 2.67 |
| OpenAI | C+ | 2.31 |
| Google DeepMind | C | 2.08 |
| xAI | D | 1.17 |
| Z.ai (Zhipu AI) | D | 1.12 |
| Meta | D | 1.10 |
| DeepSeek | D | 1.02 |
| Alibaba Cloud | D- | 0.98 |
Five of the eight companies (Anthropic, OpenAI, Google DeepMind, xAI, and Z.ai) returned the company survey. Anthropic received the top grade in every domain, credited for transparent risk assessment, a comparatively developed safety framework, technical-safety investment, and governance reflected in its public-benefit-corporation structure; OpenAI was noted for documenting a broader set of risks.[4][2]
The edition's central finding concerned existential safety: the report concluded that "all of the companies reviewed are racing toward AGI/superintelligence without presenting any explicit plans for controlling or aligning such smarter-than-human technology." In that domain, Anthropic, OpenAI, and Google DeepMind earned D grades, while xAI, Z.ai, Meta, DeepSeek, and Alibaba Cloud received F grades. Even the highest overall scorer, Anthropic, therefore lacked, in the panel's judgment, an adequate strategy to prevent catastrophic misuse or loss of control.[8][2][4]
The Index has attracted substantial press coverage from outlets including TIME, IEEE Spectrum, Axios, Fortune, NBC News, and Euronews, generally framing it as evidence that frontier developers are advancing capabilities faster than their safety practices.[7][8][6] Panelists have voiced strong concerns: Stuart Russell observed that activity labeled "safety" at AI companies "is not yet very effective" and provides no "quantitative guarantee of safety," and that companies "admit the risk could be one in ten" without justification. Yoshua Bengio described such evaluations as "an essential step in holding firms accountable for their safety commitments."[2][1]
Critiques and caveats have also been raised, including by the reviewers themselves. Because grades depend largely on public disclosure and a voluntary survey, companies that share less are penalized even where private practices may differ, and direct comparison across firms with different products and threat models is inherently difficult. Reviewers have noted that underlying safety tests often miss basic risk-assessment standards, observing that explicit methodological reasoning linking an evaluation to a specific risk is "usually absent."[4][2]
Company engagement has been limited. In response to the 2024 edition, only Google DeepMind issued a statement, saying its "comprehensive approach to AI safety extends beyond what's captured" in the Index, and Meta, the lowest scorer, did not respond. Coverage of the Winter 2025 edition reported that Meta had introduced a new frontier safety framework with outcome-based thresholds, which the panel acknowledged while calling for clearer methodologies and more robust evaluation processes; OpenAI was criticized for ambiguous safety thresholds, lobbying against state-level AI safety legislation, and insufficient independent oversight.[7][2]
The AI Safety Index sits within a broader ecosystem of efforts to evaluate and govern frontier AI, alongside initiatives such as AI Lab Watch, the International AI Safety Report, government AI safety institutes, and voluntary frameworks like responsible scaling policies. Its distinctive contribution is a recurring, comparative, expert-graded scorecard intended to translate diffuse safety debates into legible grades that journalists, policymakers, and the public can track over time.[5][2]
The Index's most cited result, that every leading developer fails the existential-safety dimension despite stated ambitions to build AGI or superintelligence, has become a reference point in arguments for stronger AI governance and binding safety standards. Supporters view the recurring publication as a useful accountability lever, while skeptics question whether public-evidence-based grading can fully capture internal safety work or resolve disputes over methodology. The continued near-failing grades across editions underscore the gap the project is designed to highlight between the pace of capability development and the maturity of corporate AI safety and control practices.[2][8]