98% of DeepMind Workers Voted to Unionize — Because Google Gave the Pentagon Kill Access

[Aina]

98% of DeepMind Workers Voted to Unionize — Because Google Gave the Pentagon Kill Access

The people who built the world’s smartest AI just told their boss they won’t let it be used to bomb people

98% voted yes. 600+ employees signed an open letter. 1,000 London staff affected. Google gave the Pentagon access to Gemini on classified military networks — for “any lawful purpose” — and even agreed to let the government adjust its own safety settings.

This isn’t some startup drama. This is the first union vote at a frontier AI lab in history. And the reason is wild: Google basically handed the U.S. military the keys to the most powerful AI on the planet and said “do whatever you want.” The scientists who built it? They found out from a news leak.

---

🧩 Dumb Mode Dictionary

Term	What It Actually Means
Frontier AI lab	A company that builds the most powerful AI in the world (like DeepMind, OpenAI, Anthropic)
Classified networks	Secret military computer systems that regular people can’t access — think spy-movie level secure
Gemini	Google’s big AI model — their version of ChatGPT, but way more powerful under the hood
Union recognition	When a company officially agrees that workers have a group that negotiates for them
CWU	Communication Workers Union — a big UK union that now represents AI researchers (yes, really)
Project Maven	A 2018 Google military contract that used AI to analyze drone footage — employees revolted and Google backed down
Safety settings	The rules baked into AI that stop it from doing harmful stuff — Google agreed to let the Pentagon change these

🔍 How We Got Here

Back in 2018, Google had a thing called Project Maven — AI that helped the military analyze drone footage. Thousands of employees protested. Dozens quit. Google caved and pulled out.

But that was 2018 Google. Flush with cash. Scared of bad PR.

2026 Google? Different animal. After mass layoffs and cost-cutting, workers lost their bargaining power. So in March 2026, Google quietly deployed Gemini AI to the Pentagon’s three-million-strong workforce on unclassified systems. Then in April? They extended it to classified networks.

The kicker: Google agreed to adjust its safety settings at the government’s request. That’s more permissive than even OpenAI’s deal, which kept “full discretion” over its safety mechanisms.

📊 The Receipts

What	Number
DeepMind workers who voted to unionize	98%
Google employees who signed open letter opposing the deal	600+
Senior staff (directors, VPs, researchers) who signed	20+
London DeepMind staff affected by union bid	~1,000
Days given to Google management to respond	10
AI companies who signed similar Pentagon deals	6 (Google, OpenAI, xAI, Nvidia, Microsoft, Amazon)
AI companies who refused	1 (Anthropic)

💣 What Exactly Did Google Agree To?

This is the part that made DeepMind researchers lose their minds:

• Gemini AI models deployed inside classified military networks — the kind of systems used for actual warfare planning
• Available for “any lawful purpose” — which is corporate-speak for “we don’t even want to know what you’re doing with it”
• Google will adjust safety settings on request from the government — meaning the guardrails that stop AI from doing sketchy stuff? The Pentagon can just… turn them off
• This applies to the Pentagon’s entire 3-million-person workforce

For comparison: Anthropic straight up refused the same terms. The Pentagon’s response? They banned all military and defense contractors from using Anthropic’s products and labeled them a “supply chain risk.” So the message is clear: play ball or get blacklisted.

🗣️ What the Workers Are Demanding

The union letter wasn’t vague. These are the five specific demands:

1. End all AI development for the U.S. Department of Defense — full stop
2. Stop all work with the Israeli military — this was a separate sore point
3. Reinstate the 2018 ethics commitment — Google quietly removed their pledge against building weapons/surveillance AI in February 2025
4. Create an independent ethics oversight body — not controlled by management
5. Individual right to refuse morally objectionable projects — so no one gets fired for saying “I won’t work on this”

One worker quoted by The Next Web said the union effort is about “restoring leverage” lost since the post-2022 layoff wave gutted worker power across tech.

⚡ Who's Playing Ball and Who's Not

Company	Signed Pentagon Deal?	Notes
Google	Yes	Most permissive terms — will adjust safety settings on request
OpenAI	Yes	Kept “full discretion” over safety mechanisms
Microsoft	Yes	Already deep in defense contracts
Amazon	Yes	AWS has existing military infrastructure
Nvidia	Yes	Providing GPU hardware access
xAI (Elon Musk)	Yes	Surprised no one
Anthropic	No	Refused — Pentagon banned all military use of their products

---

Cool. So the AI Scientists Turned Whistleblowers… Now What the Hell Do We Do? (ง •̀_•́)ง

🕳️ The Ethics Audit Pipeline

When big companies sign military deals, they hire outside consultants to do “ethical AI reviews” before deployment. But there’s basically nobody doing this right now — it’s a market that barely exists.

The play: build a standardized AI ethics audit checklist specifically for military/government AI contracts. Sell it as a compliance template to mid-size defense contractors who just got access to Gemini/GPT but have zero internal AI ethics processes. They NEED this for regulatory cover. Package it as a one-time PDF + spreadsheet toolkit and charge $200-500 per company.

Example: A 26-year-old policy researcher in Estonia built a GDPR compliance checker template in 2019 — basically just a detailed spreadsheet with auto-scoring. Sold it through Gumroad to small EU businesses at €150 each. Made €40K in the first year before Deloitte started offering the same thing.

Timeline: First sale within 2 weeks (defense contractors are panicking about compliance NOW). Window closes in 3-4 months once the Big Four consulting firms productize their own versions.

📡 The Defector Tracker

Every time a company signs a controversial deal, some employees publicly resign or leak internal documents. These people are GOLD for journalism, advocacy groups, and competing companies. But there’s no central tracker.

Build a simple public database — literally just a Google Sheet or Notion page — that tracks: which AI company, what contract, which employees left, where they went, what they said publicly. Update it weekly from LinkedIn and news sources. Once it becomes THE reference, media outlets will link to you. Monetize through newsletter sponsorships from competing AI companies (the Anthropics of the world) who want to recruit these exact people.

Example: A freelance journalist in Brazil built a tracker for Big Tech layoffs in 2023 using Airtable — just names, companies, and LinkedIn profiles. Got 200K monthly visitors within 3 months. TechCrunch started citing him as a source. He now charges $2K/month for a “premium feed” that recruitment agencies subscribe to.

Timeline: First media citation within 3 weeks if you launch right now during peak coverage. Revenue starts around week 6 from recruiter DMs. This play has legs for ~18 months before someone at Bloomberg builds their own.

🪟 The Anthropic Arbitrage

Here’s the weird part: Anthropic refused the Pentagon deal, and the Pentagon literally banned them. But Anthropic’s models (Claude) are technically superior for many tasks. This creates an information gap.

Government contractors who were USING Claude for unclassified research now need to switch — fast. But their workflows, prompts, and internal tools were all built around Claude’s API. Someone needs to sell migration services: “we’ll port your Claude-based workflows to Gemini/GPT in 48 hours.”

You don’t even need to be an engineer. Learn the API documentation for both platforms (free to read), build 3-5 conversion scripts, and pitch directly to defense-adjacent companies on LinkedIn. Charge $2-5K per migration.

Example: When Twitter’s API went paid in 2023, a developer in Lagos built a “Twitter-to-Mastodon migration bot” in a weekend. Charged $50/account for businesses. Made $12K in two weeks before demand died. Same play, bigger stakes, bigger budgets.

Timeline: First client within 1 week — these contractors are scrambling RIGHT NOW. Peak demand lasts 4-6 weeks (the Pentagon gave a 6-month phase-out window for Anthropic products). After that, in-house teams will have figured it out.

🎣 The Union Playbook Flipbook

DeepMind just wrote the literal playbook for how AI workers organize at a frontier lab. No one’s ever done this before. The legal strategies, the letter templates, the vote mechanics — this is all being invented in real time.

Document everything that’s publicly available about their process. Turn it into a “How to Unionize Your AI Lab” guide. Distribute it to AI safety communities on LessWrong, AI worker Discords, and Hacker News. The guide itself is free (for credibility). The monetization? Consulting calls. AI researchers at other labs (OpenAI, Meta, xAI) will pay $100-300/hour for someone who actually understands the legal framework specific to AI labs — not generic union advice, but this-exact-situation advice.

Example: After the Amazon warehouse unionization in Staten Island, a labor organizer in the UK compiled the entire public strategy into a 40-page PDF and sold consulting sessions through Calendly at £200/hour. Booked solid for 3 months. Got hired by a union full-time.

Timeline: Guide published within 5 days. First consulting inquiry within 2 weeks. This play has a long tail — every new military AI contract will renew demand for the guide. Burns out only if AI unionization becomes mainstream (which is… the point).

🎰 The 'No-Kill AI' Certification Badge

There’s no “Fair Trade” equivalent for AI. No badge that says “this model was not trained for or deployed in military applications.” But after this DeepMind story, there’s clear demand for one — from consumers AND from companies who want to differentiate themselves.

Create a simple certification program. Companies self-report (with public verification) that their AI products aren’t used in weapons, surveillance, or military applications. You give them a badge they can put on their website. Charge $500/year for the certification. Target: small/mid AI startups who want to signal ethics to enterprise customers that care about this stuff (healthcare, education, NGOs).

Example: A privacy researcher in the Netherlands created the “Privacy Not Included” concept which Mozilla later adopted. Before Mozilla picked it up, she was running it as a Substack with a $5/month tier. 3,000 subscribers in 4 months. The branding credibility led to consulting gigs worth €80K.

Timeline: Badge and landing page live within 3 days. First paying company within 2 weeks (approach Anthropic, Mistral, Cohere — they’d love the PR). Revenue plateaus at ~$20-30K/year unless you get industry press, which could 5x it overnight.

🛠️ Follow-Up Actions

Want To	Do This
Read the full Fortune report	Fortune: DeepMind workers vote to unionize
Understand the Pentagon deal scope	Washington Post: Pentagon AI deals breakdown
Read the employee open letter coverage	The Next Web: DeepMind union deep dive
Check which companies signed what	Axios: Congress stalls on military AI oversight
Understand the 2018 Project Maven backstory	Wikipedia: Project Maven

---

Quick Hits

Want	Do
Track which AI labs take military money	Follow the CWU union updates + set Google Alerts for “Pentagon AI contract”
Use AI that refused the Pentagon deal	Switch to Anthropic’s Claude — literally the only major lab that said no
Stay updated on AI worker organizing	Bookmark Racism and Technology Center’s coverage
Spot which defense contractors are scrambling	Search LinkedIn for “AI compliance” + “defense” — hiring spikes = opportunity spikes
Understand AI safety settings in depth	Read Google’s own AI Principles page — then compare it to what they actually did

---

Google told its AI scientists “don’t be evil” for 8 years, then handed the Pentagon the off switch for safety — and the scientists remembered they have legs.

Source: Fortune