Google DeepMind resignations show where human oversight of AI breaks down

The BBC reported on 21 October that several researchers have resigned from Google DeepMind, or are in the process of leaving, because of Google's military AI contracts. Sofia Liguori, a London-based research engineer, has resigned after three years at the company. Andreas Kirsch, an AI researcher, left earlier in October. Google signed its Pentagon contract earlier this year. It allows Google's models to be used in classified work for "any lawful government purpose". A Google spokesperson told the BBC that the company follows "a consistent process for reviewing and entering these contracts" and has channels for staff to raise concerns without retaliation.

The BBC covered this as a story about ethics. For an operator it is also a story about what the phrase "human oversight" actually means once a system runs faster than the person reviewing it.

The human in the loop is on a clock

The whole contract depends on one clause. Google's models "should not be used for domestic mass surveillance or autonomous weapons (including target selection) without appropriate human oversight and control".

Sarah Shoker, formerly head of OpenAI's Geopolitics team, told the BBC what that oversight looks like under pressure: "When you have hundreds of targets that are selected by AI and a short amount of time, it is almost impossible for a human operator to meaningfully review the AI's outputs for errors."

Retired Lieutenant General Jack Shanahan, former director of Project Maven, described the philosophy as "Go as fast as possible to deploy [AI technology] as widely as possible, potentially at the expense of the rigour and discipline of test and evaluation." The Pentagon's own line comes from Under Secretary Emil Michael: "Speed defines victory in the AI era."

The organisation wants speed, the system supplies it, and the human reviewer is the one part that cannot speed up. Once that happens, oversight becomes a job title with no time to do the job.

The same thing happens in ordinary offices. AI drafts the contract, the credit decision or the code change, and a person signs it off. When the volume doubles and the deadline stays put, the review turns into a formality. The signature still carries that person's name.

Kirsch also identified a second break in the chain. "Google seems to have no right to control or veto how the Pentagon uses its AI and also seems to have limited visibility into how it's being deployed," he told the BBC. Accountability that stops at the vendor's door does not reach the place where decisions get made.

Liguori spoke to the BBC even though her contract restricts her from doing so. "In principle I'm not allowed to criticise Google for the duration of my life," she said. She described staff who are "scared of retaliation and putting their career at risk".

A current DeepMind employee, speaking anonymously, told the BBC there is "a pipeline of people at the start of the careers who could fill those positions."

That is the price of dissent inside a large organisation: one replaceable seat. An individual's leverage lies elsewhere. It lies in the terms they agree before they sign, and in having a record of what they actually reviewed and what they refused.

AI does the routine work. You do the thinking. That only holds while the thinking has the time and the authority to change the outcome.

Your Next Move

1. Audit your sign-offs this week. List every AI-assisted output you approve. For each one, note how long you have to review it and whether you could reject it in practice. Where meaningful review is impossible, say so in writing and ask for more time, lower volume or a narrower scope. A signature without a review is a liability with your name on it.

2. Read your confidentiality and non-disparagement clauses now. Raise anything open-ended, such as a restriction with no end date, at your next contract renewal or before you accept your next offer.

3. Write down your lines in advance. Name two or three uses of your work you will not support. Decide while the choice still costs little.