Google DeepMind's Broken Promises

AI company Google DeepMind’s release of Gemini 2.5 Pro violated their commitments to governments and the public, as established in the Frontier AI Safety Commitments

.

AI Safety Summits

• In 2023, the UK held the first AI Safety Summit
  at Bletchley Park, bringing together governments and AI developers to discuss AI safety for the first time.
• In 2024, the UK co-hosted the AI Seoul Summit
  . Tech companies, including Google, signed the Frontier AI Safety Commitments
  , setting out an approach for developing frontier AI systems more safely.

The Frontier AI Safety Commitments

The commitments lay out an initial framework for AI developers to take basic precautions against the increasing risks of frontier AI models. They were voluntary commitments, not enforced in any jurisdiction.

Two of the commitments are as follows.

• “I: Assess the risks posed by their frontier models […] before deploying that model. They should also consider results from […] external evaluations as appropriate.”
• “VIII: Explain how, if at all, external actors, such as governments […] are involved in the process of assessing the risks of their AI models.”

Google effectively ignored these commitments for their release of Gemini 2.5 Pro, setting a dangerous precedent for future models.

We believe that they should be held accountable for this failure in order to maintain the legitimacy of the international AI Safety Summits and to facilitate the passage of future AI regulation.

Timeline and details of Google’s violation

25 March 2025

• Gemini 2.5 Pro Experimental
  becomes available for anyone to access for free. It is arguably the most capable AI model released by any company at this point.
• No information
  about safety testing is published.

3 April 2025

• Google’s head of product for Gemini tells TechCrunch
  the company hasn’t published a ‘model card’ (ie. safety report) for Gemini 2.5 Pro “because it considers the model to be an ‘experimental’ release”.

9 April 2025

• In correspondence with Fortune magazine
  , Google does not answer “direct questions” about the involvement of the UK AI Security Institute in the testing process for Gemini 2.5 Pro. However, a spokesperson for Google says that the model has undergone pre-release testing.

16 April 2025

• Gemini 2.5 Pro Preview is now available (essentially the same as the Experimental model).
• Google publishes its ‘model card’
  (safety report) for Gemini 2.5 Pro Preview.
• The testing report makes no mention of external testing.

28 April 2025

• Google updates
  its model card to include mention of “third party external testers”, but no detail about who they are.

April - June 2025

• Further technical reports
  have since been published. None of them name the third party external testers or state whether governments have been involved in testing.

Conclusion

• Google violated the spirit of commitment I by publishing its first safety report almost a month after public availability and not mentioning external testing in their initial report.
• Google explicitly violated commitment VIII by not stating whether governments are involved in safety testing, even after being asked directly by reporters.

Protest

In response, PauseAI will be holding our biggest protest to date, with more than 100 attendees, on Monday, 30th of June.

Sign up here.