Former Google DeepMind researcher says AI alignment has ‘no promising ideas’

Former Google DeepMind alignment researcher Ramana Kumar says current methods cannot reliably ensure advanced AI systems pursue the goals people intend. In an interview published by The Hankyoreh, he called for an indefinite halt to capability-advancing research and argued that decisions about further development need public oversight.

Watch Desk analysis

What happened

Kumar led alignment research at Google DeepMind from 2018 to 2023. Speaking to The Hankyoreh on 24 September, he said today’s systems are trained to follow instructions but that this does not guarantee they will continue to do so outside training. He described current training and reinforcement-learning methods as fragile, and said he knew of no promising alternative for giving AI systems goals.

Kumar argued that rapidly growing capabilities are outpacing alignment research. He said he would not be comfortable with further capability increases without public input, and called for strong boundaries on research, including bans on certain levels of capability. Read the interview.

Why it matters

These are Kumar’s views, not evidence that AI systems are already beyond human control. But they sharpen a live debate: whether technical safety work can keep pace with capability development, and who gets a say when commercial competition pushes labs to move quickly.

Kumar also described why he left Google DeepMind: burnout, the difficulty of making progress on alignment, and the fact that the safety team did not have final authority over risky decisions. That account puts the governance question inside the lab, not just in the halls of government.

Our read

Kumar’s case is forceful, but its most useful contribution is the distinction between training a system to behave as intended and guaranteeing that it will. Those are not the same promise, however reassuring a product demo may look. His proposed answer, an indefinite pause and public oversight, is a substantial policy position, not a settled technical consensus.

What to watch

• Whether researchers publish concrete alternatives to current training and reinforcement-learning approaches.
• Whether governments or AI labs propose enforceable limits on capability development.
• How much authority safety teams have when their advice conflicts with commercial priorities.

Discussion spark: Should further AI capability development be paused until there is meaningful public oversight, or would enforceable safety requirements be a better way to manage the risks?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.