Stay Human.
Today is my last day at Google DeepMind.
I recognise that I was lucky to end up working here for a bit, and that it gave me opportunities I never expected to have in life. That said, it's been my worst work experience and I am leaving with a heavy sense of disillusionment. I won't miss the breakfasts. I can no longer enter the office without an oppressive feeling of disgust.
I have been on leave since Google signed the contract with the DoW, after a year of intense activism trying to break the silence around Google's complicity in the Gaza genocide[1] and the increasing integration of our technology into military and surveillance systems, from Lockheed Martin[2], to DHS/ICE/CBP[3], to the recent DoW contract[4]. For more details, see Google's applied principles, a portal documenting these partnerships.
The modus operandi tested in Gaza with Nimbus[5] was legitimised as the standard with April's DoW classifiedWhat "classified" meansInformation the U.S. government restricts for national-security reasons, at levels such as Confidential, Secret and Top Secret. Classified AI workloads run on separate government networks, often in isolated or air-gapped environments with no connection to the public internet; DoD Impact Level 6 covers data up to Secret. Only cleared personnel can access them, so the AI provider typically has little or no visibility into how its models are used, cannot monitor prompts or outputs as it would on its commercial cloud, and relies on the government for any oversight. deal: contracts that allow the integration of AI into military and surveillance systems while imposing no restrictions on permissible uses and leaving the company no monitoring power.
No ethical principles can be taken seriously if we operate in this way. It is also a very convenient contract design for executives to escape accountability. Moreover, the lack of transparency around contract terms creates a layer of opacity to prevent workers' pushback, to keep us in the maybe, in a half-light where reassuring narratives can feed our innate need to absolve ourselves and think well of the technology we help to build.
We are willing to believe that it is just for administrative purposes, that it's just standard enterprise access, and that classified does not imply military (sure, but when the partner is a Department of War that openly frames its AI strategy around aggressive military use, what's the deal really for?).
I no longer believe that our technology will have a net positive impact on society. I think there's a fantasy world where AI can help decarbonise grids, cut waste, protect ecosystems, and benefit humanity. But it's not this world, and it's not this type of general-purpose AI we are racing to develop.[6] Racing to deploy agentic AI in society as it is now and in this historical moment is a terrible idea, and will definitely be a net evil.
Technology is NOT a neutral tool. This assessment assumes we can separate technology from the context it's deployed into, but we should not: we have a responsibility to assess the tool in the context in which it's going to be used. It also assumes we can separate the technology from who owns it, builds it, controls it, and benefits from it.
Agentic AI in this society, built by a very small oligarchy of tech giants with enormous economic and political power, is NOT neutral. It's going to increase inequality, authoritarianism (and actually threaten the very foundations of democracy), surveillance, war and misinformation, and disrupt education, jobs, and the very way we learn, think and interact. I don't want to be part of a company that is ruining my daughter's future.
The risks agentic AI poses to society are already here, in our present. They stem from AI's unrestrained, largely unregulated integration into military systems, surveillance networks, financial systems, critical infrastructure, and civilian data. The question of losing control over models is relevant. But it should never suggest that AI risk comes only from error, from agents slipping out of our control. Unfortunately, that is the message most often conveyed, because AI safety tends to be discussed in isolation. It is cut off from the risks that AI integration already poses, and from a broader reflection on which applications we consider acceptable and what our goals should be. I rarely see anyone question whether we need, or should, pursue "superintelligent AI" at all. It is taken for granted that we must.
When we frame the discussion solely around agents going rogue, we imply that if control could be guaranteed, forcing AI into every sector of society would be acceptable. It is not.
I see a parallel in how AI in warfare is discussed. AI's unreliability is clearly a problem, as the Minab school strike[7] made clear to the world. But focusing on unreliability alone implies that if AI made no mistakes, its use for targeting would be acceptable. We end up treating the error as the problem, rather than the standard. We end up forgetting that "Where's Daddy?"[8] is a system designed to wipe out entire families at scale, and civilian deaths are not an unfortunate accident.
In talking to people this year, I mostly found dissatisfaction with the company's direction, and general agreement on principles. Many people have good values, in theory. But good values not acted upon are just like Google's good principles never enforced.
It is this passive, fearful attitude that allows evil to spread on a mass scale.
This is happening in our workplace. This is our labour. It is within my scope of action. It is within yours.
For me, this year we should have stopped working. Maybe without striking, maybe just as I did in February, taking days off and demonstrating in front of the office, maybe in some other way we still need to figure out together.[9] But we should have said a much louder NO to our tech being used in the kill chain, a much louder NO to normalising what was tested in Gaza.
When the DoW contract was signed, the day after our letter was delivered, I thought "I'm done". It was very clear it was coming: the genAI.mil contract[10][11] in December, accompanied by Google Cloud ramping up its security-level authorisations[12][13], made it very foreseeable. At each of these steps, we could have stopped it, and for me, we really should have.
It is in these kinds of situations that personal values are tested, that we are tested as people. We need to be very alert to avoid becoming complicit in mass-scale evil. We need to keep watching the unpleasant truth of people being killed with AI tools even when it becomes ordinary and repetitive, and think 'we are part of it, and if not in this specific instance, that's where we're going as a company'. We need to avoid falling into the temptation of accepting reassuring narratives to stay in our comfort zone and do nothing.
One of the most powerful of these reassurances is compartmentalisation. Many of us don't work on projects directly related to warfare, myself included. And even people working on Gemini can take comfort in the broad range of its applications, or detach their work from how it is concretely used by focusing on the research or technical side. It's easy to push any ethical consideration to the side.
But we can't accept ‘My project is not related to that’, ‘It can also be used for good’, ‘I just make it run fast’, ‘If not us, someone else, it makes no difference’.
Compartmentalisation in this context is just wrong. It's easy and natural given the high pace of our jobs and all the difficulties of our personal lives. But it's wrong, and accepting it is a destructive process for ourselves.
As I wrote in another letter to my team when I came back from maternity leave:
What kind of mental barriers and compartmentalisation are we forcing upon ourselves to find a way to accept this business as usual (in SILENCE AS USUAL) and still watch ourselves in the mirror?
We cannot preserve our integrity while compartmentalising this deeply. Integrity means wholeness. Compartmentalisation, past a certain point, is fragmentation, and as such, it opposes integrity. Some separation of contexts is healthy, but when it becomes a way to avoid seeing contradictions, it starts to resemble Orwell's doublethink[14]. The scope within which our values, our empathy, and even our rational thought (which we can apply so masterfully to scientific questions) are allowed to operate begins to narrow. We stop criticising; we are no longer indignant at plain lies and incoherence. The best part of us has less and less room to exist.
We also need to strongly reject this corporate culture of censorship and intimidation. Everyone knows, without needing to say it, that there are topics we shall not discuss loudly. Military contracts are among them. Some topics are simply forbidden. There's a pervasive, constant fear. If we talk about it in the office, we whisper, and look around for cameras. Aside from selected (and moderated) channels, we don't raise these issues with the teams, even when they are clearly in scope. If you do, as I did, you get called into HR meetings on the pretext of having breached some internal policy (but to be honest, it's just intimidation, and little concrete action follows). We end up preemptively refraining from mentioning some topics, heaven forbid it should affect our careers. In fact, they rely on this induced self-censorship; that's why they are so subtle with intimidation.
If you challenge that, there's more freedom than you'd expect. There's a danger to who we are in simply accepting these rules. Let the fear drive you and the space to operate closes in on you, until your very self begins to shrink. Maybe in an era of increasing surveillance and authoritarianism that's something to reflect on: how living in a repressive and surveilled environment changes us, and how we internalise these restrictions and become diminished versions of ourselves if we fail to challenge them. I have felt freer since I found the courage to speak out and take some risks. It's empowering.
We need to stay whole to preserve our humanity, and to be able to care about humans. Losing this ability takes away a fundamental part of our own identity. Accepting the silence lessens you. Accepting oppression lessens you.
It took me a while to realise that this was the pressure I was feeling at work: a pressure to deny who I am, or compartmentalise it into my private sphere.
If this is the price of staying, I don't want this job. No job is worth that price.
"Stay Human", Vittorio Arrigoni said.[15] I repeat his call to tech workers today, as the technology we build powers brutal wars.
Join your colleagues who oppose unethical use of AI
I don't think resigning is an effective method, certainly not as an individual act. I believe we have incredible power as tech workers to influence the future of this technology from the inside. And for people fighting from the outside, it matters to know that employees are resisting too. I tried to stay and fight for as long as I could, and I am leaving the company, but I'm not giving up activism. The need to fight back against unethical use of AI is something that is staying with me.
There is a growing network of workers organising, and if you stay, you should join.
- GDM Union (UTAW / Unite) – UK GDM employees are organising via UTAW and Unite unions. Our demands include enforceable red lines on military/surveillance applications, increased transparency around contracts, the right to abstain on ethical grounds and a stronger sustainability commitment. I'm not naive: these demands won't be easy to win, and to have effective leverage we need a lot of workers' support and, most of all, readiness to act. But it is a viable path that gives legal strength to the demands. Moreover, being a member of a union gives us the power to add our voices to union consultations with the government and regulators. This is crucial at a moment when the UK is designing the shape of an AI regulatory framework. If you care, join. You can stay anonymous. If you are too scared to join, be ready to vote YES to the statutory recognition ballot when it happens.
- No Tech for Apartheid – Worker-led labour organisation of Google workers organising to hold Google accountable for creating positive impact in the world and challenging complicity in militarism and surveillance. Complicity in providing the infrastructure for genocide and war must stop. Google has been formally warned of legal exposure over Project Nimbus, and is putting the company and its workers at risk through its unethical contracts. I still hope history will hold this company accountable for its horrific complicity. It will be too late, but it's much needed. If you want an overview of Nimbus, AI systems deployed in Gaza, and NoTA, have a look at this talk, from minute 29. A former colleague gives a good overview.
- Googlers Against ICE – A campaign of Google workers organising around ICE/CBP/DHS partnerships. The campaign is being coordinated and run by workers with No Tech for Apartheid. The relationship deepened very recently, with Google Public Sector LLC becoming a direct primary cloud provider, alongside AWS and Oracle, under 'Cumulus', DHS's department-wide cloud initiative.
- Coalition of Concerned AI Staff – Cross-lab effort of employees (Anthropic, OpenAI, GDM) sharing growing concerns about how AI is being developed and used in the real world, with focus on military and surveillance applications. They coordinate cross-lab action, connect people with legal & whistleblowing advice and funnel people into workstreams. They also organise very informative webinars with experts.