FOR EDUCATIONAL AND KNOWLEDGE SHARING PURPOSES ONLY. NOT-FOR-PROFIT. SEE COPYRIGHT DISCLAIMER.

I’m really excited today because the new AI safety index after a ton of work is is finally done and the results are really exciting. The control problem is simply the problem that if we build AI that can do everything better than us, we’ve basically not built just a new technology but a new species that can make robot factories that can make new robot factories. And as Alan Turing said already in 1951, you know, the default outcome of that is that the machines take control just like we humans have taken control over the other animals on this planet. We have these seven companies who all have as a stated goal to build machines that can outthink us humans in every way, which would be the biggest uh change for humanity in its history. And they’re all saying, “Oh, well safety, we got it.” What the safety index does is take a hard look at the actual facts and see if they’re just talking the talk or whether they’re actually walking the walk. The AI safety index is like a scorecard for companies trying to build powerful AI where a massive amount of information is first gathered about what they actually do and have done and then a bunch of independent experts review this and assign letter grades to them. At the high level, they all got very bad grades, C’s and D’s and Fs, particularly in the category of how you they can control the AGI and the super intelligence that that they’re trying to build where almost all of them got Fs and no nobody got a passing grade. The most striking is having some kind of plan for how the AGI and super intelligence that they’re all saying they want to build is actually going to stay under control. The reviewers looked at whatever plans or or ideas the companies had posted and written about and nobody got a passing grade. And we saw for example among the the top three rated an interesting change. Last year it was Anthropic first, Google Deep Mind second, OpenAI third and this time um DeepMind actually got overtaken. A key reason why uh Deep Mind fell down was because OpenAI put a big effort into transparency. They publicly posted their whistleblowing policy that they have now and shared a bunch of information with the index. Uh Google Deep Mind has so far refused to do both of those. The obvious takeaway from this is that the self-governance just doesn’t work. you know, you can’t let the fox guard the hen house. This shows that you really need governments to step in and say, “Here are the safety standards. These are binding rules and you can’t just voluntarily opt out of them.” The first thing companies should do to improve their uh safety performance is read very carefully the section about their company in this report. And the goal of this index is not to shame companies, but to incentivize the people in the companies who are pushing for more safety work and empower them against the people in the same companies who maybe think safety is stupid. This is by far the most comprehensive assessment of what the companies are actually doing on safety. It’s also quite unique in that it’s not run by the companies. It’s truly independent just like any kind of grading effort has to be.

AI Safety Index by FLI. Summer 2025.

AI experts rate leading AI companies on key safety and security domains. 17 July 2025

Key findings

Top takeaways from the index findings:
  • Anthropic gets the best overall grade (C+). The firm led on risk assessments, conducting the only human participant bio-risk trials, excelled in privacy by not training on user data, conducted world-leading alignment research, delivered strong safety benchmark performance, and demonstrated governance commitment through its Public Benefit Corporation structure and proactive risk communication.
  • OpenAI secured second place ahead of Google DeepMind. OpenAI distinguished itself as the only company to publish its whistleblowing policy, outlined a more robust risk management approach in its safety framework, and assessed risks on pre-mitigation models. The company also shared more details on external model evaluations, provided a detailed model specification, regularly disclosed instances of malicious misuse, and engaged comprehensively with the AI Safety Index survey.
  • The industry is fundamentally unprepared for its own stated goals. Companies claim they will achieve artificial general intelligence (AGI) within the decade, yet none scored above D in Existential Safety planning. One reviewer called this disconnect “deeply disturbing,” noting that despite racing toward human-level AI, “none of the companies has anything like a coherent, actionable plan” for ensuring such systems remain safe and controllable.
  • Only 3 of 7 firms report substantive testing for dangerous capabilities linked to large-scale risks such as bio- or cyber-terrorism (Anthropic, OpenAI, and Google DeepMind). While these leaders marginally improved the quality of their model cards, one reviewer warns that the underlying safety tests still miss basic risk-assessment standards: “The methodology/reasoning explicitly linking a given evaluation or experimental procedure to the risk, with limitations and qualifications, is usually absent. […] I have very low confidence that dangerous capabilities are being detected in time to prevent significant harm. Minimal overall investment in external 3rd party evaluations decreases my confidence further.”
  • Capabilities are accelerating faster than risk-management practice, and the gap between firms is widening. With no common regulatory floor, a few motivated companies adopt stronger controls while others neglect basic safeguards, highlighting the inadequacy of voluntary pledges.
  • Whistleblowing policy transparency remains a weak spot. Public whistleblowing policies are a common best practice in safety-critical industries because they enable external scrutiny. Yet, among the assessed companies, only OpenAI has published its full policy, and it did so only after media reports revealed the policy’s highly restrictive non-disparagement clauses.
  • Chinese AI firms Zhipu.AI and Deepseek received failing overall grades. However, the report scores companies on norms such as self-governance and information-sharing, which are far less prominent in Chinese corporate culture. Furthermore, as China already has regulations for advanced AI development, there is less reliance on AI safety self-governance. This is in contrast to the United States and United Kingdom, where the other companies are based, and which have, as yet, passed no such regulation on frontier AI.
Note: the scoring was completed in early July and does not reflect recent events such as xAI’s Grok 4 release, Meta’s superintelligence announcement, or OpenAI’s commitment to sign the EU AI Act Code of Practice.

FOR EDUCATIONAL AND KNOWLEDGE SHARING PURPOSES ONLY. NOT-FOR-PROFIT. SEE COPYRIGHT DISCLAIMER.