Google DeepMind has launched an institute to explore how society should prepare for artificial general intelligence, starting with proposals for worker protections and AI oversight.
Shane Legg announced the institute Sept. 16. He directs it alongside James Manyika and Demis Hassabis, bringing together researchers from Google and beyond.
The directors’ launch essay says current AI lacks the consistency and creativity required for full AGI. They expect those gaps to close soon, without setting a date.
Worker protections that scale with disruption
Economists Julian Jacobs and Alex Imas evaluate 11 policy options, arguing that governments should prepare flexible responses tied to evidence about employment and wages.
For milder disruption, they favor expanded unemployment insurance, earned income tax credits and employer-led retraining. If unemployment persists and wages fall, they suggest a negative income tax, which provides cash to people below an income threshold. If economic gains shift more fundamentally from labor to capital, they propose considering universal basic capital: giving people assets or equity stakes.
Their analysis combines literature reviews and surveys with 51 AI agents modeled on real economists’ survey responses. The authors describe the approach as exploratory and call for further empirical validation. The proposals depend on how disruption unfolds, making better labor-market data central to deciding when to act.
Keeping AI reasoning readable
In a separate essay on reasoning transparency, Rohin Shah and Anca Dragan argue for preserving models’ readable intermediate reasoning, often called chain of thought. These written steps can help researchers investigate failures and detect signs of deception.
They propose measuring how informative those traces are, retaining model designs that support readable reasoning, and auditing training incentives. Penalizing a model for writing down a troubling plan, they warn, could teach it to hide the plan rather than abandon it.
The authors also acknowledge a limit: more capable models might learn to obscure their reasoning despite those precautions. Transparency is one safety tool, not a guarantee. If developers adopt less readable methods, they argue, they should demonstrate that the replacement can be monitored comparably well.
The institute’s publication disclaimer makes the status of these proposals explicit: essays express their authors’ ideas and should not be read as Google’s official position.
