Your ethics policy ends where your customer’s deployment begins
Every supplier of AI, data or cloud services faces the question Google answered in April, whether it realises that or not.
Sign up for my Substack daily AI newsletter https://johanosteyn.substack.com
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
In October 2026, two researchers resigned publicly from Google DeepMind over the company's military work, and the resignations made headlines around the world. The more consequential fact had been settled six months earlier and attracted far less attention. When Google extended its Pentagon agreement to classified work in April 2026, the amendment allowed the Department to use its AI for any lawful government purpose, required Google to help adjust its AI safety settings at the government's request, and left the company with no right to veto how its technology would be used. Google's published AI Principles promise human oversight and alignment with international law. The contract removed the means to insist on either. For any organisation that buys or supplies artificial intelligence, that gap is the lesson, and it has very little to do with defence.
CONTEXT AND BACKGROUND
The amendment extended Gemini from unclassified to classified work. More than 600 Google employees signed a letter urging the chief executive to reject it. Google says it remains committed to the position that AI should not be used for domestic mass surveillance or autonomous weaponry without human oversight. Whether that commitment appears in the contract itself has not been established.
Google abandoned its pledge not to build AI for weapons and surveillance in February 2025, and in May 2026 some 98% of Google DeepMind workers backed a union recognition request aimed at ending military use of their technology, which the company declined to grant voluntarily. Google says it does not reprimand employees for expressing views or for protected union activity.
INSIGHT AND ANALYSIS
A principle is only a control if somebody can act on it. Google's AI Principles describe what the company believes its technology should be used for. The contract determines what it can require. Andreas Kirsch, who resigned in October, said Google appears to have limited visibility into how its AI is deployed and no right to veto how the Pentagon uses it. He was describing the agreement, not levelling an accusation. Sofia Liguori, who resigned after three years, says internal posts of hers were deleted and that the company created a pervasive sense of fear which induces self-censorship.
The strongest objection deserves stating plainly. No private company should hold a veto over the operational decisions of a democratically elected government's armed forces. Civilian control of the military belongs to elected representatives, not technology vendors, and a corporate veto would be its own governance failure. That objection is correct, and it does not rescue the supplier's position. A supplier never needed a veto. Its leverage was the decision whether to sign. The failure is not the absent veto. It is publishing principles while agreeing terms that make them unenforceable.
I have previously written about this in When The AI Builders Grade Their Own Homework, where the argument was that a supplier's safety claims are not evidence until an independent party verifies them. This is the same problem one link further along the chain. There the question was whether anyone checks a vendor's account of its own testing. Here it is whether anyone, the vendor included, can see what happens after handover.
IMPLICATIONS
South Africa has spent years arguing the opposite case. It voted for United Nations General Assembly resolutions on autonomous weapons in December 2023, October 2024 and December 2025, and the African Group has stated that weapons systems must remain under the direct, meaningful control and supervision of humans at all times. The agreements that decide how military AI behaves are written elsewhere, and they say the supplier has no veto.
South Africa already imposes the missing control in another industry. A foreign buyer of controlled goods must sign an end-user certificate confirming the item is for their sole use and will not be transferred or re-exported without the written consent of the South African government, and the exporter must obtain it. Software is sold with no comparable undertaking. A defence analysis argues the answer is sovereign assurance, a national capability in testing and verification despite heavy dependence on foreign technology.
For boards the questions are identical whether buying or supplying. What does the contract entitle us to see about downstream use, and who can stop a use we never sanctioned. Does our liability persist when a customer adjusts safety settings or fine-tunes the model. Finally, do our reporting channels carry independent investigative authority and a route to the board, or terminate inside human resources.
CLOSING TAKEAWAY
The hardest objection is that foundation models are general-purpose infrastructure, and drawing clean contractual lines between commercial optimisation and military logistics may be impossible, which pushes vendors into a gatekeeping role they have no mandate to hold. Nobody should pretend the answer is simple. The test, however, is. An ethics policy that cannot be enforced against anyone is a marketing document. Before the next AI contract is signed, somebody should be able to name the clause that gives effect to the values in the brochure. Where nobody can point to one, the values are not in the agreement.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence's impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan's work at https://www.aiforbusiness.net
Comments