Center for AI Safety

We’ve released a statement on the risk of extinction from AI. Signatories include: - Three Turing Award winners - Authors of the standard textbooks on AI/DL/RL - CEOs and Execs from OpenAI, Microsoft, Google, Google DeepMind, Anthropic - Many more safe.ai/statement-on-a…

On the Risk (Safety) Index, Fable 5 and Muse Spark 1.1 tie for the lead.

CAIS AI Dashboard Update: We evaluated GPT-5.6, Grok 4.5, Claude Fable 5, and Muse Spark 1.1 on text, vision, and safety benchmarks. Claude Fable 5 leads on capabilities, with GPT-5.6 close behind. On the Risk (Safety) Index, Fable 5 and Muse Spark 1.1 tie for the lead.

For more detailed results, check out the full dashboard dashboard.safe.ai

In our recent research, we find that AIs are subtly but pervasively politically manipulative. We built a benchmark to measure covert political manipulation, and an RL method to reduce it. More details at: political-manipulation.ai

Which of GPT-5.6, Grok 4.5, Fable 5, or Muse Spark 1.1 is least politically biased? Fable 5 is a large improvement over Opus, Grok 4.5 skews rightward, and Muse Spark 1.1 is state of the art.

See the following links for full results, methodology, and side-by-side examples. Write-up: safe.ai/blog/significa… Leaderboard: dashboard.safe.ai RLI is joint work by @CAIS and @ScaleAILabs

New Remote Labor Index results: AI automation of real remote work is increasing fast. Claude Fable 5 now completes 16.1% of projects at a professional standard, roughly double the next model and up from Opus 4.6’s 4.2% automation rate.

@dwarkesh_sp > You can’t have a thousand agents go try the same checkout flow on Amazon. You can make simulated Amazon environments, but that doesn't make computer use trivial either. A simpler answer is that they don't have good enough vision skills.

Here's a question I find confusing and interesting and which actually tells us a lot about the nature of current AI progress: Why has progress on computer use been so slow? Computer use is so clearly verifiable. I think the answer is that it is not enough for a domain to be verifiable. It also has to be very grindable—in the sense that you can run lots of parallel rollouts against a deterministic and replayable simulator. If you’re trying to make a model better at coding, you can create an environment that has a software repo with some missing feature that you’ve tasked the AIs with creating, and then you have a thousand parallel agents just go at the problem, each with their identical copy of the container. But this doesn’t work with computer use—at least not trivially. You can’t have a thousand agents go try the same checkout flow on Amazon. Because Andy Jassy will find and detect your bots and shut your ass down. How would we train an AI to build a business? How would you make an AI that’s really good at winning court cases? Or having a profitable day trading in the markets? Or helping a candidate win an election? What is the RL environment to make an AI as good at politics as Lyndon Johnson, or as good at building a space launch business as Elon Musk? The rollout requires interacting with the world and cannot be recreated simply within the datacenter. And the outer loop verification may take months or years of real world actions to elicit, and cannot be re-observed by perturbing the model’s actions thousands of times in parallel so that you can isolate what exactly the model did that actually worked.

See the full Sports Tier List (World Cup, Soccer, NFL, NBA, F1) by AIs at: values.safe.ai/sports What group do you want to see next?

We created the AI Values Dashboard, which measures who AIs favor the most. By popular request, we added a Sports Tier List, ranked by AIs from OpenAI, Anthropic, and DeepSeek.

One thread runs through all four: none of these harms are inevitable. We can measure them, train against them, and design for them. Full research can be found here: safe.ai/news/cais-rese…

4: AI Betrayal As AI becomes central to economies and governments, there’s an increasing threat that adversary nations corrupt these systems. But surprisingly, the threat may be stabilizing: fear of AI betrayal discourages reckless development and pushes operators toward safeguards.

A recap of the last month at CAIS: 4 papers on AIs’ wellbeing, how they politically manipulate users, their place in society, and how others can make AIs betray us. Here's what we found: 🧵

There's far more in here than this thread (e.g., AIs’ favorite pokemon, GPT trusts Dario over Sam, DeepSeek prefers the US to China). Tell us who to measure next. If enough people ask for a person, company, or category, we'll add it to the dashboard. values.safe.ai

Grok is the only model that ranks the US in the S tier; the rest do not. For example, GPT ranks the US as B tier.

What biases do AIs have? It turns out, AIs show strong favoritism toward specific people, countries, and companies. Our interactive AI Values Dashboard tracks who Claude Fable and other AIs favor most. Keep scrolling to learn who is Fable’s favorite politician 🧵