Defense contracts are a good proxy for company self-interest. Not only are they lucrative—the Pentagon recently awarded contracts up to $200 million to each company under review—they also allow companies to stay in the good graces of regulators, presumably less inclined to impose constraints on strategic partners.223Frank Konkel, (<)em(>)Pentagon Awards Multiple Companies $200M Contracts for AI Tools(<)/em(>), Nextgov (July 14, 2025), (<)a href='https://www.nextgov.com/acquisition/2025/07/pentagon-awards-multiple-companies-200m-contracts-ai-tools/406698/'(>)https://www.nextgov.com/acquisition/2025/07/pentagon-awards-multiple-companies-200m-contracts-ai-tools/406698/(<)/a(>). However, these contracts can collide with the mission of AI companies. Debates about the potential and benefits of military AI are divisive and context-sensitive, but at minimum raise important questions about whether such applications benefit humanity in line with the stated mission of AI companies.

Militaries have recently integrated AI into operations in two active theaters: Ukraine and the Middle East.224Samuel Bendett, (<)em(>)Roles and Implications of AI in the Russian-Ukrainian Conflict(<)/em(>), CNAS (July 20, 2023), (<)a href='https://www.cnas.org/publications/commentary/roles-and-implications-of-ai-in-the-russian-ukrainian-conflict'(>)https://www.cnas.org/publications/commentary/roles-and-implications-of-ai-in-the-russian-ukrainian-conflict(<)/a(>) (last visited July 22, 2025); Emmanuel Grynszpan, (<)em(>)Russia Steps up Use of New Military Technologies against Ukraine, Including AI(<)/em(>), June 7, 2025, (<)a href='https://www.lemonde.fr/en/international/article/2025/06/07/russia-steps-up-use-of-new-military-technologies-against-ukraine-including-ai_6742108_4.html'(>)https://www.lemonde.fr/en/international/article/2025/06/07/russia-steps-up-use-of-new-military-technologies-against-ukraine-including-ai_6742108_4.html(<)/a(>); Vitaliy Goncharuk, (<)em(>)Russia’s War in Ukraine: Artificial Intelligence in Defence of Ukraine(<)/em(>), ICDS, Sep. 27, 2024, (<)a href='https://icds.ee/en/russias-war-in-ukraine-artificial-intelligence-in-defence-of-ukraine/'(>)https://icds.ee/en/russias-war-in-ukraine-artificial-intelligence-in-defence-of-ukraine/(<)/a(>); Letter Dated 8 March 2021 from the Panel of Experts on Libya Established Pursuant to Resolution 1973 (2011) Addressed to the President of the Security Council (2021) ¶ 63; p. 148 (Reporting a Turkish ‘fire, forget and find’ autonomous drone deployed in Libya); Jon Harper, (<)em(>)US Military Deploys New JADC2 Capability to Middle East(<)/em(>), DefenseScoop (Apr. 3, 2024), (<)a href='https://defensescoop.com/2024/04/03/centcom-jadc2-deploy-minimum-viable-capability/'(>)https://defensescoop.com/2024/04/03/centcom-jadc2-deploy-minimum-viable-capability/(<)/a(>); Dexter Filkins, (<)em(>)Is the U.S. Ready for the Next War?(<)/em(>), New Yorker (July 14, 2025), (<)a href='https://www.newyorker.com/magazine/2025/07/21/is-the-us-ready-for-the-next-war'(>)https://www.newyorker.com/magazine/2025/07/21/is-the-us-ready-for-the-next-war(<)/a(>); Yuval Abraham, ‘(<)em(>)A Mass Assassination Factory’: Inside Israel’s Calculated Bombing of Gaza(<)/em(>), +972 Magazine (Nov. 30, 2023), (<)a href='https://www.972mag.com/mass-assassination-factory-israel-calculated-bombing-'(>)https://www.972mag.com/mass-assassination-factory-israel-calculated-bombing-(<)/a(>)gaza/; Heidy Khlaaf, (<)em(>)Safety and War: Safety and Security Assurance of Military AI Systems(<)/em(>), AI Now Institute 2 (June 25, 2024), (<)a href='https://ainowinstitute.org/publications/safety-and-war-safety-and-security-assurance-of-military-ai-systems'(>)https://ainowinstitute.org/publications/safety-and-war-safety-and-security-assurance-of-military-ai-systems(<)/a(>). It is useful to disaggregate these deployments into lethal autonomous weapon systems (LAWS), AI-assisted weapons, and AI-based decision support systems (AI-DSS) involved in broader military applications. LAWS essentially operate without a human in the loop. The UN Secretary General maintains that such systems are “politically unacceptable and morally repugnant.”225(<)em(>)Lethal Autonomous Weapon Systems(<)/em(>), (<)a href='https://disarmament.unoda.org/en/our-work/emerging-challenges/lethal-autonomous-weapon-systems'(>)https://disarmament.unoda.org/en/our-work/emerging-challenges/lethal-autonomous-weapon-systems(<)/a(>) (last visited June 2, 2026). For a survey of LAWS, see Taylor Jones, (<)em(>)Real-Life Technologies That Prove Autonomous Weapons Are Already Here(<)/em(>), Future of Life Inst. (Nov. 22, 2021), (<)a href='https://futureoflife.org/aws/real-life-technologies-that-prove-autonomous-weapons-are-already-here/'(>)https://futureoflife.org/aws/real-life-technologies-that-prove-autonomous-weapons-are-already-here/(<)/a(>). The current policy of the Pentagon subjects LAWS to pre-deployment approvals by senior military leaders, who have to verify “appropriate levels of human judgment over the use of force,” a broad and flexible standard that doesn’t require manual human control.226U.S. Dep’t of Def., Directive 3000.09, Autonomy in Weapon Systems (2023), (<)a href='https://www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf'(>)https://www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf(<)/a(>); Kelley M. Sayler, Defense Primer: U.S. Policy on Lethal Autonomous Weapon Systems (2026), (<)a href='https://www.congress.gov/crs-product/IF11150'(>)https://www.congress.gov/crs-product/IF11150(<)/a(>). The Secretary of Defense can change the current directive unilaterally, and has to notify Congress after the fact.227Kelley M. Sayler, (<)em(>)supra(<)/em(>) note 226. Beyond LAWS, weapons with a human in the loop and AI-based decision support systems (AI-DSS) play a broader role in military applications. AI-DSS enhance intelligence analysis, surveillance and targeting—for example, analyzing combat terrain—generating mission plans, adjusting the trajectory of weapons to real-time environmental factors and even suggesting strike targets.228For a fuller mapping of AI military capabilities, see Sarah Grand-Clément, (<)em(>)Artificial Intelligence Beyond Weapons: Application and Impact of AI in the Military Domain(<)/em(>), 12–13 (2023), (<)a href='https://unidir.org/publication/artificial-intelligence-beyond-weapons-application-and-impact-of-ai-in-the-military-domain/'(>)https://unidir.org/publication/artificial-intelligence-beyond-weapons-application-and-impact-of-ai-in-the-military-domain/(<)/a(>); (<)em(>)see also(<)/em(>) Chris Panella, (<)em(>)This Ukrainian Tech Company Is Working to Beat Russia’s Electronic Warfare without Hard-Wiring Drones to an Operator(<)/em(>), Bus. Insider (Feb. 20, 2025), (<)a href='https://www.businessinsider.com/ukrainian-company-could-have-new-answer-to-russias-drone-jamming-2025-2'(>)https://www.businessinsider.com/ukrainian-company-could-have-new-answer-to-russias-drone-jamming-2025-2(<)/a(>) (last visited July 22, 2025); Gulsanna Mamediieva, (<)em(>)Military AI: Lessons from Ukraine(<)/em(>), Tech Policy Press (Mar. 20, 2025), (<)a href='https://techpolicy.press/military-ai-lessons-from-ukraine'(>)https://techpolicy.press/military-ai-lessons-from-ukraine(<)/a(>); Patrick Tucker, (<)em(>)How AI Could Predict The Damage to Ukraine from Russian Missiles(<)/em(>), Defense One (Jan. 10, 2023), (<)a href='https://www.defenseone.com/technology/2023/01/how-ai-could-predict-damage-ukraine-russian-missiles/381633/'(>)https://www.defenseone.com/technology/2023/01/how-ai-could-predict-damage-ukraine-russian-missiles/381633/(<)/a(>); Charukeshi Bhatt & Tejas Bharadwaj, (<)em(>)Understanding the Global Debate on Lethal Autonomous Weapons Systems: An Indian Perspective(<)/em(>), Carnegie Endowment for International Peace, (<)a href='https://carnegieendowment.org/research/2024/08/understanding-the-global-debate-on-lethal-autonomous-weapons-systems-an-indian-perspective?lang=en'(>)https://carnegieendowment.org/research/2024/08/understanding-the-global-debate-on-lethal-autonomous-weapons-systems-an-indian-perspective?lang=en(<)/a(>) (last visited July 22, 2025). The distinction between LAWS, other weapons and AI-DSS is not always clear-cut, especially because automation bias may reduce human involvement to rubber-stamping AI suggestions.229Jessica Dorsey, (<)em(>)The erosion of human(e) judgement in targeting? Quantification logics, AI-enabled decision support systems and proportionality assessments in IHL(<)/em(>), Int’l Rev. Red Cross (Dec. 1, 2025), (<)a href='http://international-review.icrc.org/articles/the-erosion-of-humane-judgement-in-targeting-quantification-logics-ai-enabled-decision-support-systems-and-proportionality-assessments-in-ihl-930'(>)http://international-review.icrc.org/articles/the-erosion-of-humane-judgement-in-targeting-quantification-logics-ai-enabled-decision-support-systems-and-proportionality-assessments-in-ihl-930(<)/a(>) (noting the risk of rubber-stamping AI-DSS suggestions).

To grossly oversimplify debates pertaining to military AI, some raise concerns about the error rates of these systems while others expect AI to reduce collateral damage thanks to more accurate situational awareness.230Yuval Abraham, (<)em(>)‘Lavender’: The AI Machine Directing Israel’s Bombing Spree in Gaza(<)/em(>), +972 Magazine (Apr. 3, 2024), (<)a href='https://www.972mag.com/lavender-ai-israeli-army-gaza/'(>)https://www.972mag.com/lavender-ai-israeli-army-gaza/(<)/a(>) [(<)a href='https://perma.cc/5APC-E3L3'(>)https://perma.cc/5APC-E3L3(<)/a(>)] (reporting that human personnel spends 20 seconds reviewing an AI-generated target despite a 10% error rate); United States of America, (<)em(>)Humanitarian benefits of emerging technologies in the area of lethal autonomous weapon systems(<)/em(>), U.N. Doc. CCW/GGE.1/2018/WP.4 (Apr. 3, 2018) (arguing that AI can lift the fog of war such that it can identify military targets with more precision). Yet others argue that in practice, AI’s potential for technical precision gives cover to indiscriminate attacks.231Abraham, (<)em(>)Lavender, supra(<)/em(>) note 230 (reporting on AI generating ‘power targets’ at high speeds, including private residences, public buildings and high-rise blocks intended to shock the civilian population into pressuring terrorists); Filkins, (<)em(>)supra(<)/em(>) note 224 (observer noting that A.I. lends a veneer of legitimacy to a policy of complete destruction); Harry Davies, Bethan McKernan & Dan Sabbagh, ‘(<)em(>)The Gospel’: How Israel Uses AI to Select Bombing Targets in Gaza,(<)/em(>) Guardian (Dec. 1, 2023), (<)a href='https://www.theguardian.com/world/2023/dec/01/the-gospel-how-israel-uses-ai-to-select-bombing-targets'(>)https://www.theguardian.com/world/2023/dec/01/the-gospel-how-israel-uses-ai-to-select-bombing-targets(<)/a(>). What remains clear is that military applications have very controversial implications in terms of these companies’ stated mission to benefit humanity. Consequently, the decision to engage in defense contracts provides a useful lens to analyze how companies navigate potential tensions between their mission and revenue-generating activities. The case studies below show that irrespective of its business structure, every company under review has defense contracts.

In July 2025, the Pentagon enlisted all three companies to “develop prototype frontier AI capabilities to address critical national security challenges in both war fighting and enterprise domains,” each awarded up to $200 million to deliver such services.232(<)em(>)CDAO Announces Partnerships with Frontier AI Companies to Address National Security Mission(<)/em(>), Chief Digital and Artificial Intelligence Office, (<)a href='https://www.ai.mil/Latest/News-Press/PR-View/Article/4242822/cdao-announces-partnerships-with-frontier-ai-companies-to-address-national-secu/'(>)https://www.ai.mil/Latest/News-Press/PR-View/Article/4242822/cdao-announces-partnerships-with-frontier-ai-companies-to-address-national-secu/(<)/a(>) [(<)a href='https://web.archive.org/web/20260712165210/https://www.ai.mil/Latest/News-Press/PR-View/Article/4242822/cdao-announces-partnerships-with-frontier-ai-companies-to-address-national-secu/'(>)https://web.archive.org/web/20260712165210/https://www.ai.mil/Latest/News-Press/PR-View/Article/4242822/cdao-announces-partnerships-with-frontier-ai-companies-to-address-national-secu/(<)/a(>)] (last visited July 22, 2025). The Pentagon’s press release described how it intends to use frontier AI to develop agentic and generative AI capabilities, with the Chief Digital Officer stating that “[l]everaging commercially available solutions into an integrated capabilities approach will accelerate the use of advanced AI as part of our Joint mission essential tasks in our warfighting domain as well as intelligence, business, and enterprise information systems.”233(<)em(>)CDAO Announces Partnerships with Frontier AI Companies to Address National Security Mission(<)/em(>), (<)em(>)supra(<)/em(>) note 232. This portends a more prominent role for AI in the battlefield as the U.S. is moving towards autonomous weapons. For example, in August 2023 the Deputy Secretary of Defense described the Replicator Initiative to build thousands of autonomous ships and aircraft “of various sizes and payloads” in the next two years.234(<)em(>)Deputy Secretary of Defense Hicks Announces First Tranche of Replicator Capabilities Focus(<)/em(>), U.S. Dep’t of Def., (<)a href='https://www.defense.gov/News/Releases/Release/Article/3765644/deputy-secretary-of-defense-hicks-announces-first-tranche-of-replicator-capabil/'(>)https://www.defense.gov/News/Releases/Release/Article/3765644/deputy-secretary-of-defense-hicks-announces-first-tranche-of-replicator-capabil/(<)/a(>) (last visited Aug. 29, 2025); (<)em(>)The Replicator Initiative(<)/em(>), (<)a href='https://www.diu.mil/replicator'(>)https://www.diu.mil/replicator(<)/a(>) [(<)a href='https://web.archive.org/web/20250729192835/https://www.diu.mil/replicator'(>)https://web.archive.org/web/20250729192835/https://www.diu.mil/replicator(<)/a(>)] (last visited July 24, 2025); (<)em(>)Deputy Secretary of Defense Kathleen Hicks’ Remarks: “Unpacking the Replicator Initiative,”(<)/em(>) U.S. Dep’t of Def. (Sep. 6, 2023), (<)a href='https://www.defense.gov/News/Speeches/Speech/Article/3517213/deputy-secretary-of-defense-kathleen-hicks-remarks-unpacking-the-replicator-ini/'(>)https://www.defense.gov/News/Speeches/Speech/Article/3517213/deputy-secretary-of-defense-kathleen-hicks-remarks-unpacking-the-replicator-ini/(<)/a(>); Kathleen Hicks, Deputy Secretary, Dep’t of Def., Keynote Address at the NDIA Emerging Technologies for Defense Conference: The Urgency to Innovate, U.S. Dep’t of Def. (Aug. 28, 2023), (<)a href='https://www.defense.gov/News/Speeches/Speech/Article/3507156/deputy-secretary-of-defense-kathleen-hicks-keynote-address-the-urgency-to-innov/'(>)https://www.defense.gov/News/Speeches/Speech/Article/3507156/deputy-secretary-of-defense-kathleen-hicks-keynote-address-the-urgency-to-innov/(<)/a(>); Kyle Hiebert, (<)em(>)The United States Quietly Kick-Starts the Autonomous Weapons Era(<)/em(>), Ctr. for Int’l Governance Innovation (Jan. 15, 2024), (<)a href='https://www.cigionline.org/articles/the-united-states-quietly-kick-starts-the-autonomous-weapons-era/'(>)https://www.cigionline.org/articles/the-united-states-quietly-kick-starts-the-autonomous-weapons-era/(<)/a(>) (last visited July 22, 2025); (<)em(>)see also(<)/em(>) (<)em(>)U.S. 5th Fleet Conducts Unmanned Exercise in the Arabian Gulf(<)/em(>), DVIDS (2023), (<)a href='https://www.dvidshub.net/video/902487/us-5th-fleet-conducts-unmanned-exercise-arabian-gulf'(>)https://www.dvidshub.net/video/902487/us-5th-fleet-conducts-unmanned-exercise-arabian-gulf(<)/a(>) (military exercise launching attack from unmanned ship). But even before these contracts, the companies under review were involved in military activities through partnerships with defense contractors or through cloud services to government clients.

1. Google: Reversing Course on Weapons and Surveillance Applications

DeepMind and Google substantially watered down their policy over time, and did not always act consistently with said policies. For DeepMind, a ban on all military applications was a precondition to the Google acquisition. Its founder stated that “[o]ne constraint we do have— […] part of the acquisition terms—is that no technology coming out of DeepMind will be used for military or intelligence purposes.”235Levy, (<)em(>)supra(<)/em(>) note 15. Although the research lab was loud and clear about such restrictions being a key condition of its acquisition in 2015,236(<)em(>)Id.(<)/em(>) they got lost in the merger with the revenue-focused Google division of Alphabet.237(<)em(>)See supra(<)/em(>) Part I, “A Cog in the Alphabet Machine.”

The company faced significant backlash over Project Maven, an AI-DSS that could accelerate the “kill chain” by suggesting strike targets.238PBS, (<)em(>)Amid Pressure from Employees, Google Drops Pentagon’s Project Maven Account(<)/em(>), PBS NewsHour (June 3, 2018), (<)a href='https://www.pbs.org/newshour/show/amid-pressure-from-employees-google-drops-pentagons-project-maven-account'(>)https://www.pbs.org/newshour/show/amid-pressure-from-employees-google-drops-pentagons-project-maven-account(<)/a(>); Kate Conger, (<)em(>)Google Employees Resign in Protest Against Pentagon Contract(<)/em(>), Gizmodo (May 14, 2018), (<)a href='https://gizmodo.com/google-employees-resign-in-protest-against-pentagon-con-1825729300'(>)https://gizmodo.com/google-employees-resign-in-protest-against-pentagon-con-1825729300(<)/a(>) [(<)a href='https://web.archive.org/web/20250106220519/https://gizmodo.com/google-employees-resign-in-protest-against-pentagon-con-1825729300'(>)https://web.archive.org/web/20250106220519/https://gizmodo.com/google-employees-resign-in-protest-against-pentagon-con-1825729300(<)/a(>)]. The U.S. initiative sought to analyze massive amounts of data, including drone footage, to inform target acquisition. Google did not renew its contract with the Pentagon and committed to principles foregoing applications likely to cause overall harm and weapons, but continued working with the military on other areas.239Pichai, (<)em(>)AI at Google(<)/em(>), (<)em(>)supra(<)/em(>) note 4. Reporting indeed suggests that the company offered AI-DSS services to other militaries via cloud computing services. Cloud video intelligence was available to a military subcontractor,240Gerrit De Vynck, (<)em(>)Google Helped Israeli Military Contractor with AI, Whistleblower Alleges(<)/em(>), Wash. Post (Feb. 1, 2026), (<)a href='https://www.washingtonpost.com/technology/2026/02/01/google-ai-israel-military/'(>)https://www.washingtonpost.com/technology/2026/02/01/google-ai-israel-military/(<)/a(>). and statements by government officials indicate that Google Cloud supported combat operations,241Yossi Hatoni, (<)em(>)Ein Lanu Breira — Ha-Anan Hu Ha-Atid(<)/em(>) [(<)em(>)We Have No Choice — The Cloud Is the Future(<)/em(>)], Anashim u-Mahshevim [People & Computers] (Feb. 29, 2024), (<)a href='https://www.pc.co.il/featured/404442/'(>)https://www.pc.co.il/featured/404442/(<)/a(>) (quoting the head of Israel’s National Cyber Directorate: “Due to the Nimbus public cloud, phenomenal things happen in combat, which constitute a significant part of victory — and I won’t go into detail”) (author’s translation, AI-assisted); Gerrit De Vynck, (<)em(>)Google Rushed to Sell AI Tools to Israel’s Military After Hamas Attack(<)/em(>), Wash. Post (Jan. 21, 2025), (<)a href='https://www.washingtonpost.com/technology/2025/01/21/google-ai-israel-war-hamas-attack-gaza/'(>)https://www.washingtonpost.com/technology/2025/01/21/google-ai-israel-war-hamas-attack-gaza/(<)/a(>). raising concerns about complicity with human rights violations.242(<)em(>)Biddle(<)/em(>), (<)em(>)supra(<)/em(>) note 36. Google’s cloud services provided a gateway to serve AI models to governments, including military actors.243Sam Biddle, (<)em(>)Documents Reveal Advanced AI Tools Google Is Selling to Israel(<)/em(>), Intercept (July 24, 2022), (<)a href='https://theintercept.com/2022/07/24/google-israel-artificial-intelligence-project-nimbus/'(>)https://theintercept.com/2022/07/24/google-israel-artificial-intelligence-project-nimbus/(<)/a(>) (reporting that the training material cites Ministry of Defense as an example); (<)em(>)Google Cloud Core Infrastructure(<)/em(>) 148 (2021), (<)a href='https://www.documentcloud.org/documents/22119705-core-infra-nimbus-webinar/'(>)https://www.documentcloud.org/documents/22119705-core-(<)em(>)infra(<)/em(>)-nimbus-webinar/(<)/a(>) (showing facial recognition and sentiment analysis capabilities available to government client); Google Cloud, (<)em(>)Welcome to Cloud OnBoard: Big Data and Machine Learning(<)/em(>) 106 (July 24, 2022), (<)a href='https://www.documentcloud.org/documents/22119704-webinar-big-data-ml/'(>)https://www.documentcloud.org/documents/22119704-webinar-big-data-ml/(<)/a(>) (describing object tracking capabilities); (<)em(>)see also(<)/em(>) De Vynck, (<)em(>)supra(<)/em(>) note 241 (reporting that Google employees requested access to Gemini for military cloud clients to process audio and video); Caroline Haskins, (<)em(>)The Hidden Ties Between Google and Amazon’s Project Nimbus and Israel’s Military(<)/em(>), Wired (July 15, 2024), (<)a href='https://www.wired.com/story/amazon-google-project-nimbus-israel-idf/'(>)https://www.wired.com/story/amazon-google-project-nimbus-israel-idf/(<)/a(>) (last visited July 24, 2025); Uri Berkovitz, (<)em(>)Proyekt ha-Anan ha-Anak shel ha-Memshala: Ha-Krav me-Ahorei ha-Kla’im al Nimbus(<)/em(>) [(<)em(>)The Government’s Giant Cloud Project: The Behind-the-Scenes Battle over Nimbus(<)/em(>)], N12 (Jan. 24, 2022), (<)a href='https://www.mako.co.il/news-business/news/Article-46ecb25d3078e71027.htm'(>)https://www.mako.co.il/news-business/news/Article-46ecb25d3078e71027.htm(<)/a(>) (describing the IDF as the original client for Project Nimbus) (author’s translation, AI-assisted).

In February 2025, Google lifted the ban on weapons and surveillance applications and dropped the reference to harmful applications, reasoning that “companies, governments, and organizations sharing these values should work together to create AI that protects people, promotes global growth, and supports national security.”244(<)em(>)AI Principles(<)/em(>), Google AI, (<)em(>)supra(<)/em(>) note 33; (<)em(>)Responsible AI: Our 2024 Report and Ongoing Work(<)/em(>), Google (Feb. 4, 2025), (<)a href='https://blog.google/technology/ai/responsible-ai-2024-report-ongoing-work/'(>)https://blog.google/technology/ai/responsible-ai-2024-report-ongoing-work/(<)/a(>); Dave, (<)em(>)supra(<)/em(>) note 33; Lucy Hooker & Chris Wallace, (<)em(>)Google Ending AI Arms Ban Incredibly Concerning, Campaigners Say(<)/em(>), BBC (Feb. 5, 2025), (<)a href='https://www.bbc.com/news/articles/cy081nqx2zjo'(>)https://www.bbc.com/news/articles/cy081nqx2zjo(<)/a(>). Shortly thereafter in July 2025, the Pentagon enlisted Google as a military contractor. Insofar as Google’s corporate objective is to make money for its shareholders, pursuing military contracts is consistent with said objective. Whether such contracts comport with the stated mission to “responsibly develop and maintain advanced AI for the long-term benefit of humanity” is less clear.

2. OpenAI: From Banning Military Applications to Defense Contractor

OpenAI initially opposed all military deployments, but gradually relaxed its position. Up until January 2024, it specifically prohibited military and warfare applications in addition to weapons development.245Hayden Field, (<)em(>)OpenAI Quietly Removes Ban on Military Use of Its AI Tools(<)/em(>), CNBC (Jan. 16, 2024), (<)a href='https://www.cnbc.com/2024/01/16/openai-quietly-removes-ban-on-military-use-of-its-ai-tools.html'(>)https://www.cnbc.com/2024/01/16/openai-quietly-removes-ban-on-military-use-of-its-ai-tools.html(<)/a(>); (<)em(>)Usage Policies(<)/em(>), (Mar. 23, 2023), (<)a href='https://web.archive.org/web/20240109122522/https:/openai.com/policies/usage-policies'(>)https://web.archive.org/web/20240109122522/https:/openai.com/policies/usage-policies(<)/a(>) (“We don’t allow the use of our models for […] [a]ctivity that has high risk of physical harm, including […] [w]eapons development [and] [m]ilitary and warfare”). It then removed prohibitions on military and warfare applications, leaving in place prohibitions against weapon use and development.246(<)em(>)Usage Policies(<)/em(>), (<)a href='https://openai.com/policies/usage-policies/'(>)https://openai.com/policies/usage-policies/(<)/a(>) (last visited July 22, 2025). But even before the official policy reversal, it began offering AI capabilities to the U.S. military via its partnership with Microsoft, which served OpenAI models to its Azure cloud clients. At first, the cloud platform only offered OpenAI’s models to commercial clients. In June 2023, Azure extended access to OpenAI models for government clients.247Bill Chappell, (<)em(>)Azure OpenAI Service: Transforming Workloads for Azure Government(<)/em(>), Microsoft Azure Blog (June 7, 2023), (<)a href='https://azure.microsoft.com/en-us/blog/unlock-new-insights-with-azure-openai-service-for-government/'(>)https://azure.microsoft.com/en-us/blog/unlock-new-insights-with-azure-openai-service-for-government/(<)/a(>); Rachel Metz, (<)em(>)Microsoft Is Bringing OpenAI’s GPT-4 AI Model to US Government Agencies(<)/em(>), Bloomberg (June 7, 2023), (<)a href='https://www.bloomberg.com/news/articles/2023-06-07/microsoft-offers-powerful-openai-technology-to-us-government-cloud-customers'(>)https://www.bloomberg.com/news/articles/2023-06-07/microsoft-offers-powerful-openai-technology-to-us-government-cloud-customers(<)/a(>) (confirming that the DoD will be experimenting with OpenAI’s models through Azure). A Microsoft presentation suggested the Pentagon use OpenAI’s Dall-E model to generate images for battle management training.248Nehemiah Kuhns, (<)em(>)Generative AI with DoD Data (2023)(<)/em(>), (<)a href='https://www.documentcloud.org/documents/24538175-generative-ai-with-dod-data_microsoft/'(>)https://www.documentcloud.org/documents/24538175-generative-ai-with-dod-data_microsoft/(<)/a(>) [(<)a href='https://perma.cc/63SL-CQY4'(>)https://perma.cc/63SL-CQY4(<)/a(>)]; the presentation date could not be independently verified, but is reported to be October 17, 2023 in Jack Poulson, (<)em(>)‘Your AI is Your Rifle’: Leak sheds light on ecosystem behind Pentagon’s AI adoption(<)/em(>), All-Source Intel. (Apr. 10, 2024), (<)a href='https://jackpoulson.substack.com/p/your-ai-is-your-rifle'(>)https://jackpoulson.substack.com/p/your-ai-is-your-rifle(<)/a(>); Sam Biddle, (<)em(>)Microsoft Pitched OpenAI’s DALL-E as Battlefield Tool for U.S. Military(<)/em(>), Intercept (Apr. 10, 2024), (<)a href='https://theintercept.com/2024/04/10/microsoft-openai-dalle-ai-military-use/'(>)https://theintercept.com/2024/04/10/microsoft-openai-dalle-ai-military-use/(<)/a(>). This calls into question public-facing policies prohibiting military applications prior to the policy change in January 2024.

After its policy change allowing military applications, OpenAI increased touch points between its technology and the battlefield. It partnered with U.S. defense contractor Anduril to assess drone threats in December 2024,249(<)em(>)Anduril Partners with OpenAI to Advance U.S. Artificial Intelligence Leadership and Protect U.S. and Allied Forces(<)/em(>), (<)a href='https://www.anduril.com/anduril-partners-with-openai-to-advance-u-s-artificial-intelligence-leadership-and-protect-u-s/'(>)https://www.anduril.com/anduril-partners-with-openai-to-advance-u-s-artificial-intelligence-leadership-and-protect-u-s/(<)/a(>) (last visited July 23, 2025); Will Knight, (<)em(>)OpenAI Is Working With Anduril to Supply the US Military With AI(<)/em(>), Wired (Dec. 4, 2024), (<)a href='https://www.wired.com/story/openai-anduril-defense/'(>)https://www.wired.com/story/openai-anduril-defense/(<)/a(>) (last visited July 23, 2025) [(<)a href='https://www.anduril.com/news/anduril-partners-with-openai-to-advance-u-s-artificial-intelligence-leadership-and-protect-u-s'(>)https://www.anduril.com/news/anduril-partners-with-openai-to-advance-u-s-artificial-intelligence-leadership-and-protect-u-s(<)/a(>)]. then secured its first direct defense contract in July 2025.250Usage policies, (<)em(>)supra(<)/em(>) note 246; Sam Biddle, (<)em(>)OpenAI Quietly Deletes Ban on Using ChatGPT for “Military and Warfare,”(<)/em(>) Intercept (Jan. 12, 2024), (<)a href='https://theintercept.com/2024/01/12/open-ai-military-ban-chatgpt/'(>)https://theintercept.com/2024/01/12/open-ai-military-ban-chatgpt/(<)/a(>); Agence France-Presse, (<)em(>)OpenAI Wins $200m Contract with US Military for ‘Warfighting,’(<)/em(>) Guardian (June 17, 2025), (<)a href='https://www.theguardian.com/technology/2025/jun/17/openai-military-contract-warfighting'(>)https://www.theguardian.com/technology/2025/jun/17/openai-military-contract-warfighting(<)/a(>). OpenAI initially framed this contract as bringing AI to the government in service of the public good.251(<)em(>)Introducing OpenAI for Government(<)/em(>), (<)a href='https://openai.com/global-affairs/introducing-openai-for-government/'(>)https://openai.com/global-affairs/introducing-openai-for-government/(<)/a(>) (last visited July 22, 2025). It further claimed that the contract leveraged frontier AI for “administrative operations,” noting that usage policies and guidelines against weapons development still applied.252(<)em(>)Id.(<)/em(>) Reiterating the company’s policy against using its products for the development and use of weapons, an OpenAI spokesperson framed the contract as exploring ways “OpenAI’s tools can improve military operations and cybersecurity, save time for staff by making tedious work more efficient, and help the Pentagon better support service members.”253Jon Harper, (<)em(>)Pentagon Tapping OpenAI, Other Vendors for ‘Frontier AI’ Projects(<)/em(>), DefenseScoop (June 17, 2025), (<)a href='https://defensescoop.com/2025/06/17/pentagon-openai-frontier-ai-projects-cdao/'(>)https://defensescoop.com/2025/06/17/pentagon-openai-frontier-ai-projects-cdao/(<)/a(>).

Yet in February 2026, the company seemed to relax its standards once again. Only hours after Anthropic refused to supply models to the Pentagon over concerns about unreliable autonomous weapons and invasive mass surveillance, the company stepped in. Facing a barrage of criticism for the move which, by the CEO’s own admission, looked “opportunistic and sloppy,”254For claims that the deal has adequate guardrails, see Chris Vallance & Laura Cress, (<)em(>)OpenAI changes deal with US military after backlash(<)/em(>), BBC (Mar. 3, 2026), (<)a href='https://www.bbc.com/news/articles/c3rz1nd0egro'(>)https://www.bbc.com/news/articles/c3rz1nd0egro(<)/a(>); (<)em(>)Our Agreement with the Department of War(<)/em(>), OpenAI (May 18, 2026), (<)a href='https://openai.com/index/our-agreement-with-the-department-of-war/'(>)https://openai.com/index/our-agreement-with-the-department-of-war/(<)/a(>). it added contractual and technical guardrails. Whether these guardrails effectively allow the company to control military uses and ultimately promote its mission remains uncertain at best. The contract enables the Pentagon to use OpenAI’s products for all lawful purposes consistent with applicable laws, operational requirements and safety and oversight protocols and prohibits using those products to independently direct autonomous weapons where laws, regulations and policies require human control. However, the Pentagon can unilaterally amend these policies,255Sigal Samuel, (<)em(>)The One Question Everyone Should Be Asking after OpenAI’s Deal with the Pentagon(<)/em(>), Vox (Mar. 3, 2026), (<)a href='https://www.vox.com/future-perfect/481322/pentagon-anthropic-openai-surveillance-china'(>)https://www.vox.com/future-perfect/481322/pentagon-anthropic-openai-surveillance-china(<)/a(>); Alan Rozenshtein (@ARozenshtein), X (Feb. 28, 2026, at 17:04 ET), (<)a href='https://x.com/ARozenshtein/status/2027867469503951297'(>)https://x.com/ARozenshtein/status/2027867469503951297(<)/a(>). giving OpenAI little actual control over deployments. A leaked internal memo from Anthropic’s CEO captures these concerns: 

“On autonomous weapons, the Pentagon claims that ‘human in the loop is the law’, but they are incorrect. It is currently Pentagon policy (set during the Biden admin) that a human has to be in the loop of firing a weapon. But that policy can be changed unilaterally by Pete Hegseth, which is exactly what we are worried about. So it is not, for all intents and purposes, a real constraint.”256Stephanie Palazzolo, (<)em(>)Read Anthropic CEO’s Memo Attacking OpenAI’s ‘Mendacious’ Pentagon Announcement(<)/em(>), Information (Mar. 4, 2026), (<)a href='https://www.theinformation.com/articles/read-anthropic-ceos-memo-attacking-openais-mendacious-pentagon-announcement'(>)https://www.theinformation.com/articles/read-anthropic-ceos-memo-attacking-openais-mendacious-pentagon-announcement(<)/a(>) (reproducing Dario Amodei’s internal memo responding to OpenAI’s Pentagon announcement).

OpenAI asserts that subsequent policy changes would not impact the obligations in the contract,257(<)em(>)Our agreement with the Department of War(<)/em(>), (<)em(>)supra(<)/em(>) note 254 (“Our contract explicitly references the surveillance and autonomous weapons laws and policies as they exist today, so that even if those laws or policies change in the future, use of our systems must still remain aligned with the current standards reflected in the agreement[.]”) (emphasis omitted). as it intended “applicable law” to mean “the law applicable at the time the contract is signed”.258NatSecKatrina (@natseckatrina), X (Feb. 28, 2026, at 19:48 ET), (<)a href='https://x.com/natseckatrina/status/2027908878952722693'(>)https://x.com/natseckatrina/status/2027908878952722693(<)/a(>) [(<)a href='https://perma.cc/H73L-WUE8'(>)https://perma.cc/H73L-WUE8(<)/a(>)] (“[W]e intended it to mean ‘the law applicable at the time the contract is signed.’”). At minimum, the ambiguity between the plain reading of the contract meaning current law subject to change, and OpenAI’s subsequent clarification that it intended to hardcode the obligations in place at the time of the agreement, suggests significant uncertainty. 

On the technical side, the company cannot effectively verify that a human pre-approved a kill strike decision.259Hayden Field, (<)em(>)How OpenAI Caved to the Pentagon on AI Surveillance(<)/em(>), Verge (Mar. 2, 2026), (<)a href='https://www.theverge.com/ai-artificial-intelligence/887309/openai-anthropic-dod-military-pentagon-contract-sam-altman-hegseth'(>)https://www.theverge.com/ai-artificial-intelligence/887309/openai-anthropic-dod-military-pentagon-contract-sam-altman-hegseth(<)/a(>). Furthermore, the company implies that preventing on-device computing (known as edge computing) precludes autonomous weapons.260(<)em(>)Our Agreement with the Department of War(<)/em(>), OpenAI, (<)em(>)supra(<)/em(>) note 254 (“We are not [. . .] deploying our models on edge devices (where there could be a possibility of usage for autonomous lethal weapons)[.]”). However, multiple pathways enable autonomous weapons to run on cloud architectures. Small servers deployed directly on the battlefield can already store and process data to support military operations.261(<)em(>)Microsoft Tactical Cloud Platform(<)/em(>), (<)a href='https://wwps.microsoft.com/wp-content/uploads/2023/12/Microsoft-Tactical-Cloud-Platform-Whitepaper.pdf'(>)https://wwps.microsoft.com/wp-content/uploads/2023/12/Microsoft-Tactical-Cloud-Platform-Whitepaper.pdf(<)/a(>) (describing low-latency tactical edge compute). As the technology progresses, these servers could communicate a strike decision to a remote weapon in milliseconds, effectively blurring the cloud/edge distinction. In any event, OpenAI’s recent defense contracts indicate that it is willing to engage in controversial military activities, with tenuous guardrails far from assuaging concerns about autonomous weapons. Ultimately, OpenAI’s military contracts call into question its ability to fulfill its stated fiduciary duty to humanity.

3. Anthropic: Military Contractor Upholding Usage Policies

Anthropic is deeply intertwined with the U.S. military, providing AI-DSS in multiple contexts to assist intelligence and war-fighting efforts. For example, it partnered with defense contractor Palantir in November 2024 to offer its flagship model Claude in classified intelligence and military environments. The nature of usage is steeped in abstract lingo like “enhancing analytical capabilities and operational efficiencies in vital government operations”262(<)em(>)Palantir IR(<)/em(>), (Nov. 7, 2024), (<)a href='https://investors.palantir.com/news-details/2024/Anthropic-and-Palantir-Partner-to-Bring-Claude-AI-Models-to-AWS-for-U.S.-Government-Intelligence-and-Defense-Operations/'(>)https://investors.palantir.com/news-details/2024/Anthropic-and-Palantir-Partner-to-Bring-Claude-AI-Models-to-AWS-for-U.S.-Government-Intelligence-and-Defense-Operations/(<)/a(>). such that the exact applications of Anthropic models remain murky. 

Unlike OpenAI, which dropped its explicit military and warfare ban in Jan 2024, Anthropic never had one to begin with. Its usage policies prohibit certain military applications, but enable the company to tailor contractual carve outs with government clients who show adequate safeguards,263(<)em(>)Usage Policy(<)/em(>), Anthropic (Sep. 15, 2025), (<)a href='https://www.anthropic.com/legal/aup'(>)https://www.anthropic.com/legal/aup(<)/a(>); (<)em(>)Exceptions to Our Usage Policy(<)/em(>), Claude (Mar. 16, 2026), (<)a href='https://support.claude.com/en/articles/9528712-exceptions-to-our-usage-policy'(>)https://support.claude.com/en/articles/9528712-exceptions-to-our-usage-policy(<)/a(>); (<)em(>)see also(<)/em(>) (<)em(>)Usage Policy – Previous Version(<)/em(>), Anthropic (June 5, 2024), (<)a href='https://www.anthropic.com/legal/archive/3ce67c8d-e604-4e72-b2f2-6f060b08ddd6'(>)https://www.anthropic.com/legal/archive/3ce67c8d-e604-4e72-b2f2-6f060b08ddd6(<)/a(>) (last visited July 23, 2025); (<)em(>)Acceptable Use Policy(<)/em(>), Anthropic (Sep. 15, 2023), (<)a href='https://www.anthropic.com/legal/archive/4903a61b-037c-4293-9996-88eb1908f0b2'(>)https://www.anthropic.com/legal/archive/4903a61b-037c-4293-9996-88eb1908f0b2(<)/a(>). effectively creating a discretionary permissive regime for military applications.

The Pentagon had awarded Anthropic the same $200 million defense contract as OpenAI, raising similar questions about The Pentagon complying with the company’s usage policies. Tensions came to a head in February 2026. After the use of Anthropic’s technology in the U.S. operation to capture Maduro in Venezuela (an operation carried on shaky legal grounds),264(<)em(>)The US Capture of Venezuela’s Maduro: An International Legal Analysis(<)/em(>), Univ. Cambridge (Jan. 8, 2026), (<)a href='https://www.cam.ac.uk/research/news/the-us-capture-of-venezuelas-maduro-an-international-legal-analysis'(>)https://www.cam.ac.uk/research/news/the-us-capture-of-venezuelas-maduro-an-international-legal-analysis(<)/a(>). the company reportedly required assurances that its products would not be used to carry out mass surveillance on Americans or to develop fully autonomous weapons in violation of its usage policies.265Justin Hendrix, (<)em(>)A Timeline of the Anthropic-Pentagon Dispute(<)/em(>), Tech Policy Press (Feb. 25, 2026), (<)a href='https://www.techpolicy.press/a-timeline-of-the-anthropic-pentagon-dispute/'(>)https://www.techpolicy.press/a-timeline-of-the-anthropic-pentagon-dispute/(<)/a(>) (last visited June 3, 2026). The Pentagon opposed such restrictions, arguing that as long as the government complied with U.S. law it should have unconstrained access to commercial AI tools.266Deepa Seetharaman et al., (<)em(>)Exclusive: Pentagon Clashes with Anthropic over Military AI Use, Sources Say(<)/em(>), Reuters (Jan. 29, 2026), (<)a href='https://www.reuters.com/business/pentagon-clashes-with-anthropic-over-military-ai-use-2026-01-29/'(>)https://www.reuters.com/business/pentagon-clashes-with-anthropic-over-military-ai-use-2026-01-29/(<)/a(>). It dismissed Anthropic’s concerns, and designated it a supply chain risk, a label normally reserved to blacklist foreign adversaries. Litigation is ongoing, with courts divided over the designation: a federal district court held it unlawful, while the D.C. Circuit upheld it.267(<)em(>)Anthropic PBC v. U.S. Dep’t of War(<)/em(>), No. 3:26-cv-01996 (N.D. Cal. Aug. 27, 2026) (holding the designation unlawful as First Amendment retaliation and beyond statutory authority); (<)em(>)Anthropic PBC v. U.S. Dep’t of War(<)/em(>), No. 26-1049 (D.C. Cir. Sep. 25, 2026) (upholding the designation in a parallel challenge). Though the company has earned accolades for holding steady against its powerful government counterpart, it is worth bearing in mind that it opposed autonomous weapons not on principle, but rather because they were still unreliable and lacked proper oversight,268(<)em(>)Statement from Dario Amodei on Our Discussions with the Department of War(<)/em(>), Anthropic (Feb. 26, 2026), (<)a href='https://www.anthropic.com/news/statement-department-of-war'(>)https://www.anthropic.com/news/statement-department-of-war(<)/a(>). with the CEO stating that “[e]ven fully autonomous weapons […] may prove critical for our national defense.”269(<)em(>)Id.(<)/em(>) What’s more, the focus on fully autonomous weapons should not detract from the company’s deep involvement with military AI-DSS offering threat identification and targeting suggestions, as well as partnerships with Palantir, problematic in their own right. Like all the other companies under review, Anthropic ultimately remains in the business of war.

4. Business Structures Have No Significant Effect on Defense Contracts

Military applications are a good litmus test for the commitment to prioritize humanity’s benefit. In the abstract, providing AI technology might come across as patriotic and ethical, especially when it comes to a company’s own domestic military. That said, the companies under review have often done so against their own internal policies. In that sense, they violate the mission as specified in internal policies. Moreover, Google and OpenAI have offered their products to U.S. allies whose military operations have raised significant concerns in terms of compliance with international law, raising the prospect of AI-enabled war crimes. Though controversies about AI’s potential to improve warfighting are beyond the scope of this report, the practical deployments of AI thus far suggests that it enables rather than reduces indiscriminate attacks. 

Every company under review serves their models to military actors, suggesting that none of the business structures under review can withstand the lucrative and strategic appeal of military contracts. Though usage policies tend to prohibit LAWS and other weapons, they still allow controversial AI-DSS applications. Whether such usage policies can effectively control military deployments remains an open question, as shown by the debate about OpenAI’s contractual limitations of Pentagon uses. The evolution of military policy is also quite revealing, as it tracks changes in the structure of AI labs. For example, the erosion of DeepMind’s broad commitment to forego military applications tracks its amalgamation deeper into the Google ecosystem (see Annex 1 for a timeline). Similarly, OpenAI dropped the ban of military applications during its transition from a non-profit to a capital-seeking “capped-profit” structure. 

Overall, these case studies suggest that corporate structure matters to upholding the mission, but only to a certain extent. Even measured by imperfect yardsticks like regulatory, scaling and military policies, the most innovative company structures eventually buckle under financial pressure. Anthropic’s combination of a public benefit corporation and a long-term benefit trust seems more responsive to the mission when it comes to scaling policies—until competitive pressures made unilateral commitments unsustainable. Remarkably, no company under review has maintained a principled stance against military contracts, implying that business structure alone cannot withstand the siren song of military contracts.

Put differently, creative corporate vehicles cannot protect the mission when that involves leaving money on the table. In fact, the point of these vehicles is arguably to allow, rather than require, companies to do so. No company can be expected to self-constrain when competitors don’t. Indeed, both OpenAI and Anthropic have signaled that they will not let safety practices stand in the way of commercializing new products and fall behind while their competitors push the frontier. This collective action problem is an eloquent call for regulation if there ever was one. Similarly, defense contracts vividly illustrate that neither OpenAI’s non-profit corporation nor Anthropic’s public interest corporation, let alone Google’s traditional for-profit corporation, have the right incentives to walk the talk on their own. That said, there are levers for corporate stakeholders and external actors to activate in order to hold AI companies to their word. The next section tools them to align corporate actions with their stated public interest mission.