Ex-DeepMind Researcher's Framework Bans Autonomous Targeting and Mass Profiling

TurnTrout, known for his recent departure from Google DeepMind, has published a detailed proposal for governing government AI contracts. The framework centers on two hard red lines: Standard 1 requires meaningful human control over any AI-enabled targeting and use of force, with exemptions only for anti-munition defense, intelligence analysis under Standard 2, logistics, and R&D. Standard 2 prohibits untargeted AI profiling that converts bulk data into individualized intelligence on non-suspects. These standards apply whether the AI is directly provided or used as a component in a larger pipeline, and include a right to legal transparency and independent auditing.

The oversight mechanism is a Review Body chaired by the Chief Scientist, with seven seats (including two for the Cloud business) and attorney-client privilege for deliberations. While the Body cannot block contracts, it issues non-compliance findings that leadership can override—but overrides are disclosed in a yearly transparency report to all AI employees. TurnTrout designed the framework to be robust against loopholes: dissolving the Body requires advance notice and disclosure of outstanding violations. He also included negotiable provisions (e.g., escalation to Alphabet's board) to protect load-bearing terms. The framework aims to scale from internal corporate governance to potential congressional legislation.