DeepMind Workers Take Their AI Dispute to a UK Labor Tribunal
On this page 6 sections

AI-generated editorial illustration.
On November 12, two unions representing employees at Google DeepMind took a dispute over workplace voice to Britain’s Central Arbitration Committee. The committee’s public case page says it received the application that day. No decision has been issued. A request for a union vote is still several procedural steps away from a ballot; a ballot is another step away from a bargaining agreement.
The workers are asking for a say over the use of the AI they build, as well as conventional employment terms. Their campaign names military contracts, safety commitments and possible job automation. Google says it has listened to supporters and opponents of unionization and has acted in good faith. The legal process must first decide whether and how the unions can represent a defined group of employees. It cannot by itself settle every dispute over a model’s eventual use.
The case has a particular edge because it concerns a frontier AI lab. A researcher can object to a project before a model is trained, but the final deployment decision may be made by executives, a cloud business, a government customer or another team. Collective bargaining could give workers a durable route to raise those concerns. The statutory minimum, however, centers on pay, hours and holidays. Moving from that minimum to a right to refuse a project or review a customer contract would require a separate negotiation. The parties have not agreed to that wider scope.
For now, one sentence in the public docket carries the news: the application was received, and no decision has been issued. The argument over who controls AI work is considerably larger than that sentence. After months of voluntary talks, a third party will test the unions’ claim to speak for a defined workforce.
An application, not a ballot
The CAC record provides a hard procedural timestamp. The applicant unions are the Communication Workers Union and Unite the Union; the named employer is Google UK Limited. The application arrived on November 12 and appeared on the government site on November 13. As of the record inspected for this article, the committee had issued no decisions. That is the official outcome today.
WIRED reported on November 12 that the unions asked the committee to establish whether they have sufficient support among workers to trigger a formal ballot. The publication had seen the application. It described a potential group of roughly 1,000 staff in product-focused roles if the bid succeeds. That figure is a reported description of the proposed constituency, not an official count fixed by a CAC decision. The union’s earlier material described more than 1,000 workers at DeepMind’s London offices. Those measures are close, but they do not establish an agreed bargaining unit.
The unit’s boundaries will decide whose voice has force. Before an employee can know whether a future agreement covers their role, the parties must establish which workers belong in it and whether the application satisfies the statutory tests. A person working on research, deployment, sales or policy may sit in a different employment or management structure even when each says they work for Google DeepMind. The public case page names Google UK Limited. It does not publish a final role map.
The government’s employer guidance says a union can apply to the CAC when voluntary recognition has not been agreed, subject to requirements including membership of at least 10% of the proposed bargaining unit. If an application is accepted and the unit is not agreed, the CAC process can determine that unit. A later decision addresses whether a ballot is needed. Reporting the November 12 filing as a successful union vote would skip all of those stages.
For a manager, the unanswered unit question also defines the practical problem. A company cannot know the full bargaining perimeter from a headline about a filing. It needs to distinguish a union’s requested scope from the group the law may ultimately recognize, and from the people whose work feeds a given product. A model can be developed across offices and used by customers elsewhere. A UK bargaining unit would have legal force in its own employment relationship; it would not automatically govern every product decision inside a global company.
A May request reached its next forum
The November filing did not emerge from a single week of protest. On May 5, the United Tech and Allied Workers branch said members at Google DeepMind had asked management to recognize the CWU and Unite as their representatives. The union said 98% of participating DeepMind union members voted to pursue the request. That percentage does not describe a vote of all employees or a statutory recognition ballot. Its denominator is the union membership that took part in the internal decision.
On May 20, the CWU said DeepMind leadership had rejected voluntary recognition but was willing to continue talks through Acas, the UK conciliation service. The union’s account identified concerns about a change to Google’s AI principles, the possible use of workers’ skills for weapons or surveillance, a right of moral refusal and independent ethics oversight. These are the union’s stated demands. They are not concessions Google has accepted.
The campaign page offers a different but complementary view of the same interval. It describes full-time UK employees across functions and levels, not only model researchers, shaping demands by a member vote. It also places job automation beside AI safety and military contracts. The campaign argues that the legal minimum on pay, holidays and hours can be expanded through bargaining. That is an aspiration about a future agreement, not the content of one. The page still describes negotiations; the CAC docket supplies the newer fact that the unions then filed an application.
This sequence also explains why a large internal majority is insufficient as a headline result. The May 98% figure answers whether members of one union branch wanted to pursue recognition. The November application asks a public body to assess a legally defined workforce. A future ballot, if called, would ask eligible workers a different question under a different set of rules. Support within a committed organizing group may be strong while views across the eventual unit differ. Google says it has heard from employees opposed to unionization as well as supporters. The public record does not give a verified count for either camp across the proposed unit.
The campaign’s automation concern widens the labor dimension beyond military use. A model that improves coding, research or product operations could change the tasks of the people building it. UTAW says it is surveying employees about AI automation to shape demands. A survey in progress is not a measure of jobs already lost or a plan Google has accepted. It does tell a manager where a future negotiation may move: from the principles that govern the product to the terms under which employees adapt to it. Consultation before roles change, access to training and the handling of redundancy could each be negotiated differently from customer-contract restrictions.
There is a workplace decision under each demand. A researcher wants to know whether a task will feed a system used in a context they oppose. A product employee wants to know who decides when a general model becomes a specific service. A manager wants to know whether an employee can step away from a project without penalty and how to staff the work if they do. A worker whose job may change with AI wants more than a policy statement about responsible development: they want terms for consultation, retraining or redundancy. The public demands connect these questions, while the formal legal route initially addresses representation.
A company can revise an internal statement in a day. A workplace agreement needs members, a defined group, recognition and bargaining. The slower route attracts workers because an agreement could survive a product release cycle or a change in leadership. Its worth would depend on which decisions, information and remedies the parties put in writing. A durable voice needs a durable procedure.
Product use is where the dispute sharpens
Google changed its AI principles in February 2025. The older statement, preserved on Google’s site, had explicitly listed weapons and certain surveillance uses among applications the company would not pursue. The current principles emphasize bold innovation, responsible development and collaborative progress, with human oversight, due diligence and consistency with international law and human rights. The company describes a broader approach to assessing specific uses. Workers who had relied on the older categorical language read the revision differently.
The revision cannot tell us where a particular employee’s code went or whether a specific customer violated the law. Those claims need project and deployment evidence. The union says public commitments and customer choices affect the conditions under which people are willing to build the technology. Google says its existing processes and dialogue remain meaningful. The disagreement is about participation before a disputed use is approved, as well as the words of a principle after the fact.
In WIRED’s November 12 reporting, CWU national officer John Chadfield said employees felt their concerns had been minimized during talks. Google DeepMind spokesperson Tom Hewett told the publication the company respected employees’ right to make an informed choice, had listened to people on both sides and would participate constructively. Both statements are positions in a live labor dispute. Neither proves that a specific contract will be canceled or that union recognition has become inevitable.
Google’s own 2026 responsible AI report says AI principles are embedded in product and research processes, from testing to post-launch monitoring. That is an important counterpoint to a portrayal of the company as having no review system. The report does not answer whether employees can appeal a decision about a customer or refuse to work on a project. A corporate safety process and an enforceable worker voice answer different questions. One assesses technology and risk under company authority; the other would change how employees can contest some decisions made under that authority.
The employer has a legitimate operating question too. A broad model may serve many customers without its original builders knowing each use. An individual refusal right could be relatively straightforward for a named custom project, yet much harder to administer when teams maintain shared infrastructure or a general model. Colleagues who disagree about a customer’s use could also disagree about which work should stop. Google’s spokesperson did not offer those examples as its defense; they are practical implications of applying the union’s stated demand to different product arrangements. Any agreement would need to say whose objection triggers which review and how managers assign work while the review runs.
An internal AI review could answer some of those questions without a union. Google says its current principles call for oversight, due diligence, feedback and safeguards. It could make project decisions visible to staff, provide a confidential appeal and publish an account of how objections changed an outcome. The workers’ argument is that company-controlled channels can be revised by the same leadership that makes the disputed choice. A union would create an independent participant in the discussion, although it would still face limits on information, scope and bargaining power. Neither route is guaranteed to produce the decision a concerned employee wants.
Even the label “AI project” hides a chain of decisions. A foundation model may be trained for broad use. A cloud service can make it available to a customer. A customer then decides how to integrate it into a workflow. A worker’s influence decreases at each step unless the organization creates a route for information to travel back. The union wants that route to carry more than feedback. Whether a bargaining agreement can reach any of the later steps remains unsettled, especially where the buyer and deployment occur outside the UK unit.

AI-generated editorial illustration. The paths represent employee voice and product decisions, not a documented DeepMind workflow.
The law supplies a floor
The UK recognition route can give employees a collective channel, but its statutory core is specific. CAC guidance describes recognition for bargaining over pay, hours and holidays in a defined unit. UTAW argues that a negotiated agreement could address AI principles, military contracts and automation as well. That wider agenda would depend on what Google and the representatives agree, and on the leverage each side brings to the table. Statutory recognition alone would not write an ethics veto into a labor contract.
Recent law changes make the procedure itself more salient. The CAC said in April 2026 that applicants no longer had to show at the initial stage that most workers in a proposed unit were likely to support recognition. When recognition is decided by ballot, the former requirement for support from at least 40% of everyone entitled to vote was removed; a simple majority of votes cast is now the rule described by the committee. The CAC also changed its application materials for filings from April 6 onward. These changes describe the framework, not an outcome in this case.
There is still a sequence. The committee assesses the application, the bargaining unit is agreed or decided, and a ballot may follow if the law calls for one. The official employer guidance explains that, once a unit is settled, the CAC decides whether employees need to vote. As of November 16, the public DeepMind case page showed only receipt and no decision. The roughly 1,000 staff cited in WIRED cannot therefore be treated as a certified electorate.
An accepted application would be a meaningful step even before a vote. It would bring the question of who belongs in the unit into a formal procedure rather than leave it solely to private talks. The government’s explanation of the unit stage says the employer and union can agree the workers covered; otherwise the CAC decides, considering factors including effective management and existing bargaining arrangements. That makes the mix of research and product roles consequential. A unit drawn too broadly might contain teams with very different employment conditions; one drawn narrowly might exclude workers whose contributions are central to the very AI-use dispute the campaign highlights.
Recognition, if it occurs, would still leave bargaining to do. The law’s guaranteed subjects are employment terms. A specific model-use clause would need to identify the affected work, the decision stage and the people with authority to make or review the choice. If the language merely promises consultation after a customer contract has been signed, workers may have a voice but little ability to change the result. If it requires disclosure before a contract, Google would need to decide what it can share about commercial and government customers. The procedural design would determine the practical value of the promise.
Recognition can oblige an employer to bargain over core employment terms. Influence over a model-use choice would require something more concrete: timely information, a named decision owner and a rule for what happens during disagreement. Would a worker receive a consultation, a temporary pause, an individual opt-out or a binding restriction on a contract? Each option places a different duty on product teams. The filing establishes none of them.
A general promise to “listen” is hard to audit. Listening can mean a town hall, a written response or a right to negotiate before a decision is made. A collective agreement can name the meeting, the information supplied, the affected roles, the clock and the escalation route. A company could also argue that customer confidentiality or national security limits the information it can disclose. Those conflicts would have to be handled in actual terms. An AI job title supplies no answer.
Another Google union route is different
On November 2, Prospect announced an agreement giving its organizers physical and digital access to Google employees in the UK. The union said the deal followed a new workplace access right that came into force on October 30. Google told Prospect it had established a structured access agreement to comply with the new legislation while retaining direct dialogue with employees. An access agreement lets a union reach workers; it is not collective-bargaining recognition.
The Prospect arrangement and the DeepMind application should not be collapsed into one company-wide vote. Prospect’s announcement describes an access route to Google UK workers. The DeepMind case names the CWU and Unite as applicants in a specific CAC proceeding against Google UK Limited. The public material does not show that Prospect’s agreement settled the DeepMind group’s request or that the same employees would be represented by both unions. A worker may hear about both campaigns, but the legal identity and scope of each route matter.
The timing matters as much as the union names. Prospect’s access agreement arrived after the October 30 change in workplace rules and before the November 12 DeepMind filing. It shows Google can reach an agreement with a union on one limited workplace right while another group continues to press for recognition. Neither fact cancels the other. Access gives organizers a way to make their case to employees; it does not require those employees to join or bind Google to bargain over AI projects. The DeepMind applicants seek a different legal status. A reader counting “Google union deals” without distinguishing these stages would overstate what workers have obtained.
Several mechanisms now operate at Google. Access lets organizers speak to employees who are not members. Recognition would give a union standing to bargain for a defined unit. A later agreement would determine which topics and procedures bind the parties. Google’s Prospect statement pointed to internal representative groups as another route. Workers pressing for a union may see the absence of enforceable bargaining as the missing element.
For other AI labs, the example is specific rather than universal. A UK filing does not transplant its legal route to the United States or make every AI worker part of a single bargaining unit. It does show that a dispute about the purpose of an AI product can become an employment dispute with named parties, a dated application and rules about representation. The novelty is the subject of the worker demand. The legal machinery is ordinary labor machinery.
A decision map for the months ahead
The next public milestone is a CAC decision on the application. If it is accepted, the proposed bargaining unit and any ballot decision will determine which employees are inside the process. A recognition result would then open a second question: which issues Google and the representatives actually negotiate. Coverage, support, recognition and bargaining scope are four separate facts. Treating any one of them as proof of the others would obscure the case.
| Public question | Evidence available now | Evidence still needed |
|---|---|---|
| Was an application filed? | The CAC says it received one on November 12, 2026. | No additional filing proof is needed for that narrow fact. |
| Has the application been accepted? | The public case page says no decision has been issued. | A dated CAC acceptance or rejection notice. |
| Who would be represented? | The unions seek a DeepMind group; WIRED describes roughly 1,000 product-focused staff. | An agreed or CAC-determined bargaining unit and employer record. |
| Will employees vote? | A ballot has been requested, not announced by the CAC. | A formal ballot decision, schedule and result if one is ordered. |
| Can workers influence AI uses? | Union demands and company principles describe opposing expectations. | A negotiated clause, disclosure route and actual decision practice. |
The map is useful for a product employee considering whether a union would address the work that troubles them. It is also useful for an employer tempted to answer every question with an existing AI-principles page. A principles page can say how a company evaluates risk. It cannot, without additional terms, tell an employee whether a refusal affects a performance review, who sees a disputed contract, or what happens if a deployment changes after model development. Those are workplace design questions.
There is also a serious argument against treating a labor agreement as the sole arbiter of model use. Customers, regulators, affected communities and independent safety experts can have interests different from employees’. A bargaining unit can protect workers and force information into view, but its members may not see every downstream deployment or represent everyone affected by it. Google may need to preserve consistent global product rules across teams and jurisdictions. The strongest version of the union’s case is therefore not that employees alone should make the final AI policy. It is that people whose work creates the system should have a defined, durable way to challenge decisions before the work is irreversibly used.
Consider a hypothetical project, not a reported DeepMind assignment: a UK engineer maintains a general model, a separate sales team signs a public-sector customer, and the buyer later adapts the model for a sensitive workflow. The engineer’s concern may be sincere and the employer’s information incomplete at the time the work is done. A promise to consult only the sales team would miss the engineer; a right to halt every later use would give one team control over decisions it cannot fully observe. A workable process would specify an early disclosure trigger, a channel for confidential objections, a reviewer with access to customer context, and a response that explains what changed or why it did not. The case has not produced such a process. The example shows what a future agreement would need to solve before “worker voice” becomes operational.
November’s filing gives that proposition a formal address. The CAC docket can be checked for the next decision. Until it changes, the workers have filed, Google has said it will participate, and the proposed vote remains ahead.
An employee looking at that docket will not find an answer about a military customer, an automation plan or a right to refuse a project. They will find a dated application and, eventually, a decision about the route to representation. That may sound narrower than the campaign that brought the case there. It is also the first place where a contested workplace claim can become a formal obligation. The card on the table is still an application. What it eventually gives employees a hand in deciding is open.