The argument inside DeepMind spilled into public this week. The BBC talked to current and former employees who do not think Google has a clear view of how the US military uses the company’s AI. At least two researchers have left over it in October.
Who left, and what they raised
Sofia Liguori, a research engineer, resigned this week. Andreas Kirsch, a researcher, had already gone earlier in the month. The concern, as the BBC reported it, is visibility and control. Not whether a contract exists. Whether Google can tell what the Pentagon does with the systems after the contract is signed.
That is a harder problem than a policy paragraph. A model or a platform sold for one classified purpose can be reused for another. The people building the model often cannot see the reuse. Several of them have decided that gap is a reason to walk.
Google’s answer
The company said public-sector contracts are checked against its AI principles and its acceptable-use rules. That is the process answer. It does not, by itself, tell a researcher what a program office did with a system six months later.
For anyone buying or regulating this technology, the episode is a reminder that “the lab objected” and “the contract was signed” can both be true. The principles document is not the deployment.