The clause that would change frontier AI

The Google DeepMind workers who filed for union recognition on 5 May are asking for three things. Two of them are familiar. They want the lab to end its military contracts with the Pentagon and the Israeli Defence Forces, and they want an independent ethics body to review future ones. Those demands belong to a long tradition of tech-worker protest that runs through Project Maven in 2018, the Project Nimbus walkouts, and the various open letters that have circulated through Silicon Valley over the past decade. They are demands about what the company does.

The third demand is different. The workers want an individual right to refuse project assignments on ethical grounds, without disciplinary or career consequences. That is a demand about what an individual researcher is permitted to do inside the company. It has, as far as I can tell, no precedent in a frontier AI lab. And it is the part of this dispute that, if conceded, would change the sector more than any contract cancellation.

The statutory ten-working-day window for voluntary recognition closes this week. If Google DeepMind refuses, and there is no public indication so far that it will agree, the case proceeds to the Central Arbitration Committee, which has the power under the Trade Union and Labour Relations Act 1992 to compel recognition where a union demonstrates majority support in an appropriate bargaining unit.¹ The CWU has secured a 98% vote in favour among its members at DeepMind. Management will likely argue, with some justification, that CWU membership is not the same thing as the DeepMind UK workforce, and that the appropriate bargaining unit is a matter for the CAC to determine.² That fight will be procedural and slow. The substantive question, what kind of professional autonomy AI researchers are entitled to, will outlast it.

Why right-to-refuse is the structural demand. The contract demand, if won, removes one customer. The ethics-body demand, if won, adds one layer of review. The right-to-refuse demand, if won, changes the default unit of consent inside the lab from the project to the person. That is a categorical shift. It moves AI research closer to the professions that have long recognised individual conscience as a constraint on institutional power: medicine, where doctors can decline to participate in specific procedures; law, where lawyers can recuse from conflicts; academic research, where principal investigators can decline to take certain grants. In each of those fields the right is not absolute and is contested at the edges, but its existence is the baseline. In frontier AI, there is no baseline. A researcher assigned to a defence-adjacent project is expected to work on it, or to find another job.

This is the gap the DeepMind workers are trying to close. They are not asking for a veto over what the company sells. They are asking for the right not to personally build it. The distinction matters because it concedes a great deal to management, the firm retains the right to sign contracts, to deploy staff, to set strategy, while reserving for the worker the narrower right to walk away from a specific assignment without losing the job. It is the smallest possible version of conscientious objection, and the sector has never granted even that.

Who carries the reputational weight. Consider the distributional picture. The executives who signed the Pentagon contract and the Project Nimbus arrangement bear minimal personal exposure. Their names appear on press releases and earnings calls; they do not appear on the research papers, model cards, or technical documentation that defence customers consume. The researchers and engineers whose names do appear on those outputs carry the reputational risk — into their next job, into their academic standing, into the conferences they attend and the colleagues they want to work with.

This is a familiar shape. Risk pushed down the hierarchy, reward retained at the top. In other industries we recognise the pattern and have built institutional responses to it: indemnification clauses, professional liability insurance, codes of conduct that protect the practitioner from the principal. Frontier AI labs have built almost none of this infrastructure for their technical staff. A researcher whose work ends up in a military system has no formal recourse, no professional body to appeal to, no institutional shield. The right-to-refuse clause is, in part, an attempt to build that infrastructure from the bottom up — through the only mechanism the workers currently have, which is collective bargaining.

The brand becomes the lever. What makes this particular fight different from earlier tech-worker protests is the rhetorical position the workers have chosen. They are not arguing that DeepMind should hold to some external standard imposed by regulators or civil society. They are arguing that DeepMind should hold to its own standard. The union letter, according to the reporting, explicitly cites DeepMind's stated mission — "the responsible development and deployment of AI for the long-term benefit of humanity"³ — as the basis for the claim that military contracts contradict the lab's identity.

This is a clever move and a generalisable one. The more aggressively a frontier lab markets itself as safety-conscious, mission-driven, and ethically distinct from its competitors, the larger the surface area it creates for workers to hold it to those claims. Anthropic, OpenAI under its original charter, DeepMind from its founding — each has invested heavily in a brand that says: we are different, we are careful, we are doing this for the right reasons. That brand recruits the workforce. It commands premium valuations. It buys regulatory goodwill. And it now, increasingly, gives workers a documented set of company commitments to point at when they object to specific decisions.

The brand that recruits the workforce is the same brand that arms it.

I find this almost too neat. The labs spent years building the rhetorical apparatus of responsible AI partly because it was useful — useful for hiring, for fundraising, for fending off regulation. They are now discovering that the apparatus has constituency, and that the constituency has read the brochure. A more cynical lab would have positioned itself purely as a commercial AI vendor with no ethical pretensions; that lab would have a freer hand to take military contracts and would face less internal pushback. DeepMind chose otherwise, for reasons that were not wrong, and is now living with the consequence.

The tactic is the giveaway. The threatened industrial action is itself worth attention. Workers have indicated they will abstain from high-value tasks, specifically, Gemini model updates, while continuing to perform minor duties.² This is not a strike in the traditional sense. It is a precisely calibrated work-to-rule that is designed to impose commercial cost at the points in the product cycle where DeepMind is most exposed (frontier model releases) while minimising the legal vulnerability of participating workers under UK employment law, which treats partial performance very differently from outright refusal.

A tactic this specific does not emerge from spontaneous frustration. It implies coordination, legal advice, and a clear-eyed read of the asymmetry between research labour and frontier-model release schedules. It also implies that the workers understand something about the production function of an AI lab that the public commentary often misses: that the marginal Gemini update is enormously valuable to Google and enormously dependent on a relatively small number of senior researchers, whose abstention cannot easily be replaced by hiring or by reallocating junior staff. The capacity to slow frontier model development is concentrated in the same hands that are now organising.

What management will argue, and what it elides. Google DeepMind has not yet responded publicly to the recognition request. When it does, the most likely line of defence will be procedural: that 98% of CWU members is not the same as a majority of the DeepMind UK workforce, that the proposed bargaining unit is inappropriate, that the CAC should require a wider ballot.⁴ This argument has technical merit. It is also, on its own, an evasion of the substantive question, which is whether researchers building consequential systems should have any individual say in which systems they build.

The counter-arguments to the right-to-refuse demand itself are real and should be named. A lab manager trying to staff a project on a commercial timeline cannot easily run an opt-out process; an organisation that grants too broad a conscience exemption can be paralysed by it. There is a version of this clause that becomes operationally unworkable, and a version of it that is so narrow as to be meaningless. The interesting question is whether there is a middle version — one that recognises a specific class of project (military, surveillance, dual-use) as triggering an opt-out right, while leaving routine research staffing intact. That is a design problem, not an impossibility.

It is also worth being honest that the workers' framing of "military AI" is broader than the underlying contract scope is likely to be. Defence contracts cover logistics, administration, medical triage, and language translation as well as targeting and lethal autonomy. Some of those applications are uncontroversial; some are not. The right-to-refuse clause would be most defensible, and most narrowly powerful, if it attached to specific application categories rather than to "anything with a Pentagon logo on it." The workers have not, in public materials so far, drawn that distinction. They will probably need to, if they want the CAC and the public to take the demand seriously.

What I think is actually at stake. The contract demands will be litigated through the political and commercial system; some military work will be cancelled, some will be renamed, most will continue. The ethics-body demand will be conceded in some attenuated form, because conceding it costs Google very little. The right-to-refuse demand is the one Google will fight hardest, because it is the one that, if won at DeepMind, will be sought at every frontier lab in the country and eventually at every frontier lab anywhere a union can be formed. It would shift the default in AI research from "you build what you are assigned" to "you build what you are assigned, unless you don't, and we cannot punish you for the unless."

That is not a small shift. It is the difference between a workforce that is a deployable resource and a workforce that is a body of professionals with standing of their own. Every other field that builds consequential systems has eventually reached some version of that settlement, usually after a long and ugly fight. The question is not whether frontier AI will eventually arrive there too. The question is whether it arrives there in 2026, through a CWU recognition campaign at a single London lab, or in 2036, after a decade of further deployments in which the workers who built the systems had no formal way to say no.

I do not know which way this case will go. I am fairly sure that whichever way it goes, it will be cited for the next twenty years. The workers have picked the right demand. Google has picked, so far, the wrong silence.

Footnotes

¶