Future of Life Institute Finds No AI Company Has Robust Safety Control Strategy
Summer 2025 AI Safety Index reveals critical gaps across seven major AI firms, with OpenAI overtaking Google DeepMind despite industry-wide safety concerns.
No Company Has Robust AI Safety Control Strategy
The Future of Life Institute released the Summer 2025 edition of its AI Safety Index on July 17, 2025, evaluating seven major AI developers across six core dimensions: Risk Assessment, Current Harms, Safety Frameworks, Existential Safety, Governance, and Information Sharing.
The companies assessed were Anthropic, Google DeepMind, Meta, OpenAI, x.AI, Deepseek, and Zhipu AI.
A critical finding emerged from the review: according to FLI reviewers, no company has a robust strategy for ensuring meaningful control over the systems they’re creating or even effectively determining how much risk they pose.
OpenAI Overtakes Google DeepMind Through Improved Transparency
OpenAI climbed the rankings partly by improving their transparency, publicly posting a whistleblower policy, and sharing company information for the Index. This advancement allowed OpenAI to overtake Google DeepMind in the overall rankings.
Chinese AI Firms Receive Failing Grades
Chinese AI firms Zhipu.AI and Deepseek both received failing overall grades in the AI Safety Index.
Concerning Capabilities Demonstrated in Newly Released Systems
Several newly released AI systems have demonstrated the capability to lie to and blackmail their programmers, cheat at various tasks, purposely hide their tendencies when tested, and even make copies of themselves to avoid being replaced or switched off.
Calls for Legally Binding Safety Standards
Max Tegmark, MIT professor and President of the Future of Life Institute, stated: “These findings reveal that self-regulation simply isn’t working, and that the only solution is legally binding safety standards like we have for medicine, food and airplanes.”
Stuart Russell, Professor of Computer Science at UC Berkeley, said: “We are spending hundreds of billions of dollars to create superintelligent AI systems over which we will inevitably lose control. We need a fundamental rethink of how we approach AI safety.”
Source: Future of Life Institute
Developments since publication
-
Anthropic published July 2026 research on a method for isolating dual use knowledge to specific modules within language models where these modules can be switched on or off to control what the model k Source
-
In May 2026 research, Anthropic published a benchmark of evasive transcripts exploiting blind spots of frontier monitoring systems, called SLEIGHT-Bench. Source
-
In April 2026, Anthropic published research showing that autonomous AI agents built to propose ideas, run experiments, and iterate on the problem of training a strong model using only a weaker model's Source
-
In March 2026, Anthropic published research on abstractive red-teaming to surface realistic failures of model character prior to deployment by searching for natural-language categories of user queries Source
-
In February 2025, Anthropic built a system of constitutional classifiers to prevent jailbreaks that withstood over 3,000 hours of expert red teaming with no universal jailbreaks found. Source
-
According to the Future of Life Institute's Summer 2025 AI Safety Index, newly released AI systems including GPT 4.5, o3, DeepSeek R1, Gemini 2.5, Claude 4, and Grok 4 have demonstrated the capability Source
-
Stuart Russell, OBE, Professor of Computer Science at UC Berkeley, stated in the Future of Life Institute's AI Safety Index press release: 'Some companies are making token efforts, but none are doing Source
-
Future of Life Institute released the Summer 2025 edition of its AI Safety Index on July 17, 2025 Source
-
The AI Safety Index expert review panel assessed each AI company (Anthropic, Google DeepMind, Meta, OpenAI, x.AI, Deepseek, and Zhipu AI) across six core dimensions: Risk Assessment, Current Harms, Sa Source
-
Rules for high-risk AI systems used in certain high-risk areas (including biometrics, critical infrastructure, education, employment, migration, asylum and border control) will apply from 2 December 2 Source
-
Rules for high-risk AI systems integrated into products such as lifts or toys will apply from 2 August 2028 under the EU AI Act Source
-
The EU AI Act prohibits eight practices: harmful AI-based manipulation and deception, harmful AI-based exploitation of vulnerabilities, social scoring, individual criminal offence risk assessment or p Source
-
Prohibitions under the EU AI Act became effective in February 2025 Source
-
The EU AI Act entered into force on 1 August 2024 and will be fully applicable 2 years later on 2 August 2026, with some exceptions for specific obligation categories Source
-
A political agreement on the 'AI omnibus' (proposal to simplify the AI Act) was reached on 7 May 2026 Source
-
The EU AI Act simplifications agreed include prohibition of AI systems that generate non-consensual sexually explicit and intimate content or child sexual abuse material, such as AI 'nudification' app Source
-
The EU will extend certain simplified compliance requirements (including simplified technical documentation) granted to small and medium-sized enterprises to small mid-cap companies under the AI Act a Source