⚡ Quick Summary
Published by the Future of Life Institute, this report presents the AI Safety Index Winter 2025, an expert-led comparative assessment of safety practices at eight frontier-model developers: Anthropic, Alibaba Cloud, DeepSeek, Google DeepMind, Meta, OpenAI, xAI, and Z.ai. It assesses 35 indicators of responsible conduct across six domains: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing. The index uses US GPA letter-grade boundaries and displays both overall grades and numerical scores.
The assessment combines publicly available company material with a tailored industry survey, and an independent panel of eight technical and governance experts provides grades, brief justifications, and recommendations for improvement. The summary identifies Anthropic, OpenAI, and Google DeepMind as the leading performers, with overall grades of C+, C+, and C respectively. It concludes that existential safety is the sector's central structural weakness: none of the companies is presented as having an explicit plan to control or align smarter-than-human systems. It also finds uneven implementation of safety practices relative to emerging global standards, including the EU AI Code of Practice.
🧩 What’s Covered
The two-page summary covers the index's comparative results, rationale, and assessment approach:
- Overall comparison: A scorecard gives each of the eight named companies an overall letter grade and numerical score. Anthropic receives the highest displayed score, 2.67 and C+, followed by OpenAI at 2.31 and C+, while Alibaba Cloud is shown at 0.98 and D-.
- Six assessment domains: The chart reports domain grades for risk assessment (six indicators), current harms (seven), safety frameworks (four), existential safety (four), governance and accountability (four), and information sharing (10). It also records whether each company submitted a survey response.
- Grading convention: The document explains that letter grades use US GPA boundaries, with values ranging from A+ at 4.3 through F at 0. This establishes the basis for the displayed numerical overall scores.
- Executive findings: The summary distinguishes the three top-performing companies from the other five and attributes major gaps in risk assessment, safety frameworks, and information sharing to limited disclosure, weak evidence of systematic safety processes, and uneven evaluation practices.
- Existential-safety finding: It states that all reviewed firms are pursuing AGI or superintelligence without explicit plans for controlling or aligning smarter-than-human technology, leaving consequential risks unaddressed in the index's assessment.
- Methodology and evidence: The index draws on public research papers, policy documents, news articles, and industry reports, alongside a tailored survey intended to increase transparency about safety practices, processes, and structures. The document says the full list of indicators and collected evidence appears in the full report.
- Independent review panel: Eight named volunteers with technical and governance expertise are listed: Stuart Russell, David Krueger, Dylan Hadfield-Menell, Jessica Newman, Sharon Li, Sneha Revanur, Tegan Maharaj, and Yi Zeng.
💡 Why it matters?
The index offers policymakers, the public, and those responsible for AI governance a structured way to compare disclosed safety practices across major frontier-model developers. Its six-domain design separates immediate harms, organisational governance, disclosure, and longer-term controllability rather than treating safety as a single claim.
The survey and public-evidence approach also makes transparency a practical part of the assessment: firms could provide additional information about their safety-related practices and structures. The document explicitly positions this visibility and comparison as a counterweight to competitive pressures that can reward profits over safety, while linking the desired rigor, measurability, and transparency to emerging standards such as the EU AI Code of Practice.
❓ What’s Missing
As a two-page summary, the document does not reproduce the 35 indicators, the company-specific evidence base, reviewers' brief justifications, or their recommendations for improvement; it says these are in the full report. It gives domain grades but does not explain in this summary how individual indicator assessments are aggregated into domain grades or overall numerical scores beyond the GPA grade boundaries. The summary identifies gaps in companies' practices but does not provide detailed company-by-company examples supporting each grade. It also names the survey-response status without setting out the survey questions, response content, or how survey information affected particular assessments.
👥 Best For
This resource is best suited to policy teams, AI-governance leads, researchers, and civil-society users seeking a rapid comparative view of frontier developers' disclosed safety practices. It is particularly useful for identifying which of the six domains warrant closer examination before consulting the referenced full report and underlying evidence.
📄 Source Details
AI Safety Index Winter 2025 is a two-page English document published by the Future of Life Institute in December 2025. No individual authors are credited; it lists an eight-member independent review panel. The document prints a route to a fuller report as futureoflife.org/index, but does not print a direct HTTP(S) URL to this PDF. The supplied PDF is complete at two pages.
