Back to Resources
AI Governance12 min read
AI in Military Operations: The Ethical Divide Reshaping the Industry
How the split between Anthropic's refusal and OpenAI's embrace of defence contracts is defining the future of military AI.
The Moment That Split the AI Industry
On February 26, 2026, Anthropic CEO Dario Amodei walked out of a face-to-face meeting with U.S. Defense Secretary Pete Hegseth at the Pentagon and issued a statement that would fracture the AI industry: "We cannot in good conscience accede to their request."
The Pentagon had demanded that Anthropic remove contractual prohibitions on two uses of its AI: mass surveillance of civilians and fully autonomous weapons systems. Amodei refused. Within 24 hours, Trump banned all federal agencies from using Anthropic products, and Hegseth designated the company a "supply-chain risk to national security" — a label historically reserved for foreign adversaries like Huawei.
Hours later, OpenAI announced its own deal with the Pentagon for classified military use. CEO Sam Altman later admitted the negotiations were "definitely rushed" and that the deal "looked opportunistic and sloppy."
This is not just a policy dispute. It is the most consequential test case yet for whether AI companies can maintain ethical red lines when governments demand compliance — and it has direct implications for every business choosing an AI vendor.
The Pentagon's AI-First Strategy
Understanding the Anthropic standoff requires context on just how aggressively the U.S. military is pursuing AI. In January 2026, Hegseth issued the AI Acceleration Strategy, mandated by President Trump, positioning the U.S. military as an "AI-first" force built to outpace rivals in speed, scale, and decision-making. The Pentagon's budget expanded to $1 trillion for FY2026, with the FY2026 NDAA authorising $900.6 billion — $8 billion above the presidential budget request — with heavy emphasis on AI.
The strategy directed all DoD AI contracts to incorporate "any lawful use" language within 180 days — the mandate that directly triggered Anthropic's refusal.
Key milestones in the Pentagon's AI push include:
• AI Rapid Capabilities Cell (December 2024): Launched with $100 million in funding, replacing Task Force Lima, to accelerate adoption of large language models and generative AI across military operations
• Classified AI deployment (March 2026): AI models including Claude are already used in classified settings. The Pentagon has reached agreements with OpenAI and xAI to operate models in classified environments, with training models on classified data as the next frontier
• $4.9 billion allocated for DoD AI R&D in 2025 alone, within a global military AI market valued at $9.31 billion in 2024
How Militaries Actually Use AI Today
Before debating the ethics, it is worth understanding what military AI looks like in practice. Most current applications are far from science fiction — they are operational tools for logistics, intelligence, and planning.
Decision Support and Planning
AI helps military staff build options faster by sorting data and surfacing patterns. AI-powered simulation models allow commanders to explore scenarios and evaluate courses of action before committing forces. Scale AI's Thunderforge program builds AI agents specifically for military planning and operations in the Indo-Pacific and Europe.
Intelligence and Surveillance
Project Maven — the program that triggered Google's withdrawal in 2018 — now uses AI to identify objects in surveillance imagery across the National Geospatial-Intelligence Agency. Over 20,000 active users rely on 35+ Maven-connected military tools.
Logistics and Maintenance
The Defense Logistics Agency's AI Centre of Excellence operates 55 AI models in production with 200+ use cases under development. Results include 15-20% logistics cost reductions, up to 30% reliability improvements, and predictive maintenance capabilities that could save the DoD up to $5 billion annually.
Edge Computing for the Battlefield
TurbineOne was awarded $98.9 million for on-device battlefield data processing, recognising that cloud connectivity cannot be guaranteed in combat environments.
The distinction matters: logistics optimisation and intelligence analysis are very different from autonomous weapons systems. Anthropic's refusal was not about AI in defence broadly — it was specifically about two use cases it considers incompatible with democratic values.
OpenAI's Military Contracts: The Pivot from 'No' to 'Yes'
OpenAI's journey into defence work has been rapid and controversial. In January 2024, the company quietly removed its blanket ban on military use from its usage policies, opening the door to defence work.
By February 28, 2026, OpenAI had announced a deal with the Pentagon for classified military use. The contract includes three stated prohibitions: no mass surveillance of Americans, no directing autonomous weapons where policy or law requires human control, and no "high-stakes autonomous decisions" requiring human approval.
However, critics have raised serious concerns about the contract's enforceability. Since OpenAI has not released the full contract text, The Intercept reported and Techdirt's Mike Masnick argued that the deal "absolutely does allow for domestic surveillance" because it references compliance with Executive Order 12333 — a Reagan-era intelligence directive with broad surveillance authorities.
The internal reaction was telling. CNN reported that many OpenAI employees "really respect" Anthropic's stance, at least one executive quit, and a "QuitGPT" campaign emerged urging subscription cancellations. Meanwhile, Claude was pushed to the top of the App Store.
MIT Technology Review characterised the situation bluntly: OpenAI's "compromise" with the Pentagon was exactly what Anthropic had feared — an arrangement where stated red lines could be quietly eroded by contractual ambiguity and government pressure.
Anthropic's Stand: What They Refused and Why
Anthropic's position is more nuanced than a blanket refusal of defence work. Before the standoff, the DoD had awarded Anthropic a $200 million transaction agreement in July 2025 for prototyping frontier AI capabilities. Claude was being used for intelligence analysis, modelling and simulation, operational planning, and cyber operations.
Amodei's public statement made the distinction explicit: "I believe deeply in the existential importance of using AI to defend the United States and other democracies." But he identified two areas where "AI can undermine, rather than defend, democratic values":
• Mass surveillance of civilians — AI systems capable of monitoring entire populations without meaningful oversight
• Fully autonomous weapons — lethal systems that select and engage targets without human decision-making
He also pointed out a contradiction in the Pentagon's position: "One labels us a security risk; the other labels Claude as essential to national security."
The government's retaliation was swift and severe. Hegseth designated Anthropic a supply-chain risk. Pentagon Undersecretary Emil Michael called Amodei "a liar" with "a God-complex." Federal agencies began phasing out Claude — OPM replaced it with Grok, Treasury engineers switched to Codex and Gemini, and the Secret Service dropped Claude entirely.
Anthropic filed suit in California federal court, arguing First Amendment violations and that the supply-chain designation exceeds presidential authority. A Lawfare analysis concluded the designation is unlikely to survive legal challenge. Microsoft filed a friend-of-the-court brief supporting Anthropic, and 37 engineers from OpenAI and Google, including Google chief scientist Jeff Dean, filed an amicus brief in their personal capacities. The federal hearing is scheduled for March 24, 2026.
The Defence AI Ecosystem: Palantir, Scale AI, and Anduril
While the Anthropic-OpenAI divide captures headlines, the defence AI market is dominated by purpose-built companies that have never questioned military work. Defence tech startups raised a record $49.1 billion in 2025, nearly double the prior year.
Palantir Technologies holds the largest footprint. Its $10 billion Army Enterprise Agreement (July 2025) consolidated 75 contracts into a decade-long deal. The Maven Smart System expanded to a $1.3 billion ceiling through 2029. NATO adopted Maven for military planning in April 2025. Palantir's U.S. government revenue reached $1.57 billion in 2024, with its stock up 340% in one year.
Anduril Industries, founded by Palmer Luckey, secured a $20 billion Army Enterprise Contract in March 2026, consolidating 120+ procurement actions. The company's valuation trajectory — from $14 billion in August 2024 to an estimated $60 billion in 2026 — reflects the sector's explosive growth.
Scale AI won the Thunderforge program, the DoD's flagship AI agent program for military planning, and was awarded a $250 million DoD contract. CEO Alexandr Wang called military AI a "moral imperative", arguing that China outspends the U.S. 10-to-1 on AI adoption.
For businesses evaluating AI vendors, this landscape reveals an important distinction: companies like Palantir and Anduril are built for defence from the ground up, while general-purpose AI labs like OpenAI and Anthropic face fundamental questions about where their technology should be applied.
The Google Project Maven Precedent
The current standoff has a direct historical precedent. In 2017, the Pentagon launched Project Maven — the "Algorithmic Warfare Cross-Functional Team" — to integrate AI for analysing drone surveillance footage in the fight against ISIS. Google won a roughly $9 million contract to provide AI capabilities.
When the arrangement became public in March 2018, over 3,100 Google employees signed an open letter to CEO Sundar Pichai: "We believe that Google should not be in the business of war." Nearly a dozen employees resigned, including researcher Meredith Whittaker.
Google withdrew from Maven in June 2018 and published its AI Principles restricting weapons development. The Pentagon pivoted to Palantir, Anduril, AWS, and Scale AI as successor integrators. Maven itself evolved into the $1.3 billion Maven Smart System now run by Palantir.
The lesson of Project Maven is twofold. First, employee resistance can force even the largest technology companies to change course. Second, the Pentagon adapted by building a diversified vendor ecosystem — meaning no single company's withdrawal halts military AI adoption. Anthropic's stand operates within this context: the Pentagon has alternatives, but Anthropic's refusal carries symbolic and legal weight that may define industry norms.
International Ethical Frameworks and the Race for Rules
The absence of binding international regulation on military AI makes the corporate ethical divide even more consequential. Governments are debating, but the results remain fragmented.
United Nations: The General Assembly adopted a resolution on Lethal Autonomous Weapons Systems (LAWS) in December 2024 with 166 votes in favour and only 3 opposed (Belarus, North Korea, Russia). By November 2025, 156 states supported a third consecutive resolution, with 42 calling for negotiations toward a binding instrument. However, no treaty exists — only a patchwork of voluntary commitments.
EU AI Act: In force since August 2024 with phased implementation through 2026, the Act is one of the world's most comprehensive AI regulatory frameworks. But it explicitly exempts military and national security systems, leaving defence AI regulation to individual member states.
U.S. Position: The DoD rejects autonomous weapons bans, favouring a governance framework based on its 2020 Ethical Principles for AI — responsible, traceable, governable. The 2025 NDAA requires periodic autonomous weapons development reports through 2029 but does not restrict development.
China: Pursues "intelligentization" of the PLA through civil-military fusion while publicly declaring support for an autonomous weapons ban — a contradiction that Carnegie Endowment analysts highlight as a significant governance challenge.
Transatlantic Divergence: The U.S. and EU are consolidating structurally different approaches. The EU mandates documentation, human oversight, and conformity assessments; the U.S. favours innovation-forward deregulation. For multinational businesses, this divergence creates a compliance landscape where the AI vendor you choose may determine which regulatory regime you can operate within.
Singapore's Approach: A Balanced Model
Singapore offers an instructive middle path. Rather than debating whether to use AI in defence, Singapore has focused on how — building institutional frameworks that integrate AI while maintaining governance guardrails.
In 2022, Singapore created a fourth armed service — the Digital and Intelligence Service (DIS) — on par with the Army, Navy, and Air Force. The SAF AI Centre was established in 2024 under the SAF C4 & Digitalisation Command to catalyse AI adoption.
The strategic imperative is demographic: SAF's National Service manpower supply is projected to decrease by one-third by 2030. Defence Minister Dr Ng Eng Hen stated that AI adoption "was not if, but how quickly, and how large a role."
Key developments include:
• Sovereign cloud infrastructure (March 2025): DSTA signed a contract with Oracle for an air-gapped, sovereign Oracle Cloud Isolated Region supporting MINDEF and the SAF for AI and data services
• AI Singapore partnership (May 2023): An MOU between AI Singapore and the DIS to deploy advanced AI techniques including LLMs and reinforcement learning
• Practical applications: DSO National Laboratories uses AI for tactical training simulations; MINDEF uses AI for cyber threat detection, analysis, and hunting
• International engagement: Singapore co-hosted the second REAIM Summit (2024), endorsed the Political Declaration on Responsible Military Use of AI, and committed S$680 million to national AI strategy
Singapore's approach demonstrates that meaningful military AI adoption and governance frameworks are not mutually exclusive — a point that both sides of the Anthropic debate would do well to acknowledge.
The Broader Debate: Public Opinion and the Chilling Effect
The Anthropic-Pentagon standoff has crystallised a debate that extends well beyond defence contracting.
Public opinion is clear on the fundamentals. An ITIF survey (February 2026) found:
• 75% of Americans say AI is not yet reliable enough for life-or-death military decisions without human oversight
• 54% say AI-powered mass surveillance is too dangerous
• 67% say private tech companies have a responsibility to set limits on how their products can be used, even if the government disagrees
The surveillance dimension is already real. ICE logged 25 new AI applications in 2025, including Palantir's ELITE tool using generative AI for deportation targeting. Facial recognition tools drawing on 1.2 billion face images were used 100,000+ times since May 2025 without completed privacy impact assessments.
The chilling effect on the tech industry is real. TechCrunch raised the question of whether the Pentagon's treatment of Anthropic will scare startups away from defence work entirely — not because they object to military contracts, but because contractual red lines could lead to government retaliation. The standoff signals the government's willingness to invoke Cold War-era authorities against domestic tech companies that resist state demands.
The EFF's analysis framed the core issue plainly: privacy protections should not depend on the decisions of a few powerful companies. Whether Anthropic wins or loses its legal battle, the precedent will shape how every AI company negotiates with every government for years to come.
What This Means for Businesses Choosing AI Vendors
For business leaders and policymakers, the military AI divide is not an abstract ethical debate — it has practical implications for vendor selection, risk management, and strategic planning.
Vendor values signal product philosophy. Anthropic's willingness to sacrifice a $200 million government contract — and face a federal ban — rather than remove safety guardrails tells you something important about how the company approaches its technology. When Anthropic says Claude has safety constraints, those constraints are tested under the most extreme pressure imaginable. For businesses in regulated industries — healthcare, finance, legal services — this matters. An AI vendor that maintains red lines against government coercion is more likely to maintain the data protection, accuracy, and ethical commitments it makes to commercial customers.
Regulatory exposure varies by vendor. The transatlantic regulatory divergence means that AI vendors with military entanglements may face different compliance burdens under the EU AI Act. Businesses operating across jurisdictions should assess whether their AI vendor's military relationships could create regulatory friction.
Key considerations for vendor evaluation:
• Transparency of use policies — Does the vendor publish clear, enforceable restrictions on how its technology can be used? Has it demonstrated willingness to enforce them?
• Track record under pressure — Has the vendor maintained its stated principles when challenged by powerful actors, or have its red lines shifted?
• Data governance — How does the vendor's relationship with military and intelligence agencies affect its data handling practices? Are there risks of classified or surveillance data mingling with commercial infrastructure?
• Long-term alignment — Defence contracts are lucrative but politically volatile. A vendor whose business model depends heavily on military revenue may prioritise government requirements over commercial customer needs during conflicts
The Anthropic case also establishes a new precedent: governments may retaliate against AI companies that refuse to comply with military demands. Businesses should factor this political risk into their vendor assessment — not just for Anthropic, but for any AI company that might face similar pressure in the future.
Whatever the outcome of Anthropic's March 24 hearing, the industry has permanently split on a fundamental question: should AI companies set limits on how their most powerful technology is used, or should that authority rest solely with governments? The answer each vendor gives to that question will shape the AI products businesses rely on for years to come.
PreviousClaude AI in Education: Transforming Teaching and LearningNext Claude AI for Admin & Office Productivity
Want to Apply This to Your Business?
We're a Singapore AI development and automation agency. Let's discuss how we can help solve your specific challenges.