The Pentagon revises its doctrine: AI could soon choose military targets

The Pentagon revises its doctrine: AI could soon choose military targets

🔎 An overlooked doctrinal shift

On June 25, 2026, Bloomberg revealed that the Pentagon had quietly revised its military targeting doctrine, JP 3-60, in April 2026. The change is radical: AI will no longer be a simple assistance tool, but a system capable of initiating target selection.

Why now? The revision comes amid a frantic race for the most powerful AI models. The June 2025 agentic ranking already placed OpenAI's GPT-5.5 at 98.2 points, Anthropic's Claude Opus 4.7 at 94.3, and Google's Gemini 3 Pro Deep Think at 95.4. In a single year, reasoning capabilities have exploded.

The Pentagon decided not to wait for the public debate. The Department of Defense's AI strategy from January 2026 already called to "fully integrate AI into military planning." Six months later, the doctrine follows.

The key points

• JP 3-60, the joint targeting doctrine, was revised in April 2026 to allow AI to initiate targeting actions, with humans serving only as supervisors.
• The shift from "human in the loop" to "human on the loop" represents an unprecedented paradigm shift in US military history.
• Anthropic and the Pentagon have been in open conflict since March 2026 over the guardrails that AI providers must impose on their models deployed in military contexts.
• Congress is trying to regain control, but the timeframe is considered "compressed" by several lawmakers.

Recommended tools

From the 2018 JP 3-60 to the April 2026 version: what really changes

The difference between the two versions is a calculated semantic shift. In the 2018 JP 3-60, the human is the initiator. AI provides data, the operator decides.

The April 2026 revision, which the May 2026 Air Force AFDP 3-60 version reflects the alignment of, reverses the sequence. AI proposes. AI identifies. AI initiates. The human validates — or not.

This is what the Straits Times publication directly quotes from the document: "Advances in AI will improve the targeting process and enable long-range precision engagements."

The shift from "human in the loop" to "human on the loop" is not a terminological detail. It is the difference between a driver who is driving and a driver who is monitoring the autopilot. Except here, the autopilot is designating human targets.

To understand the scope of this change, one must look at the official Congressional definition of LAWS (Lethal Autonomous Weapon Systems): systems that "select and engage targets without human intervention." The Pentagon denies that the new doctrine falls into this category. The nuance rests on the word "supervision."

This debate is not abstract. It determines who, in the decision-making chain, bears legal responsibility in the event of a targeting error. And with models like GPT-5.4 Pro (agentic score of 91.8) or Claude Opus 4.6 (84.7), the temptation to trust AI is structural.

The concrete role of AI in targeting: what models actually do

AI does not "shoot" at targets. Not yet. But it does all the upstream work, and that is where the doctrinal shift makes sense.

The military targeting process follows a codified sequence: find, fix, track, target, engage, assess (F2T2EA). In the old doctrine, each step required deliberate human action. AI provided "decision aids."

In the new version, AI can autonomously execute the "find," "fix," "track," and "target" phases, then present a set of recommended targets with a confidence level. The human operator steps in at "engage" and "assess."

Concretely, a model like Gemini 3.1 Pro (92 in general, 87.3 in agentic) can cross-reference satellite data, communication intercepts, drone feeds, and radar signature databases in real time. In a matter of seconds, it produces a prioritized list of targets with identification probabilities.

The AI Weekly summary points out that SOCOM (Special Operations Command) is already at the forefront of this type of deployment. Special operations, with their need for extreme reactivity, are the natural testing ground for this doctrine.

The question developers must ask themselves: does the boundary between fine-tuning, RAG and prompting change the nature of the recommendation? A model fine-tuned on classified military data is no longer a generic tool. It becomes a specialized system whose biases are invisible.

Anthropic versus the Pentagon: the battle of the guardrails

On March 28, 2026, Bloomberg revealed an open conflict between Anthropic and the Department of Defense. The issue: the mandatory guardrails that AI providers must impose on their models deployed in military contexts.

Anthropic wants technical limits hardcoded into the models. The Pentagon considers these safeguards to reduce operational utility and creates a dangerous precedent: one where a private provider dictates the terms of use for a system purchased by the State.

The Cloud Security Alliance analyzed this conflict by describing it as "vendor governance" — a new concept where tech companies become de facto regulators of military AI use.

This is a structural problem. When Claude Opus 4.7 (94.3 in agentic) is deployed in a targeting system, Anthropic wants to be technically able to prevent certain categories of recommendations. The DoD sees this as unacceptable interference.

The paradox is that Claude is precisely the model that some at the Pentagon prefer for targeting, thanks to its nuanced reasoning capabilities. The question of model choice is no longer just technical. It is geopolitical.

Meanwhile, Google signed an agreement in April 2026 to authorize Gemini in classified work, without the same public friction as Anthropic. The message to providers is clear: those who impose too many conditions lose contracts.

The timing: why April 2026, why in secret

The revision took place in April 2026. It was only made public by a leak to Bloomberg on June 25. Two months of silence on a doctrinal change touching on the ethics of war.

This timing is not insignificant. January 2026: the DoD AI strategy sets the political framework. March 2026: the Anthropic-Pentagon conflict erupts, revealing the pressure on suppliers. April 2026: the doctrine is revised. June 2026: the news breaks.

The senator who challenged the reform on June 15 — ten days before the Bloomberg revelation — was already speaking of a "compressed" timeframe and "serious questions." Congress was presented with a fait accompli.

The secret dimension is problematic. US military doctrines are normally public documents. the 2018 JP 3-60 was freely downloadable. The April 2026 version was not subject to an announcement, a public comment period, or a prior congressional hearing.

This modus operandi echoes the debates over self-replicating models: when technical capability outpaces the legal framework, public and private actors tend to act first, and justify later.

The implications for AI developers: you are no longer neutral

This is the part that most articles on the subject ignore. The revision of JP 3-60 is not just about generals. It directly concerns the engineers building these systems.

When you fine-tune a model for ISR (Intelligence, Surveillance, Reconnaissance) data analysis, you are participating in the targeting chain. When you optimize a RAG pipeline to query classified military databases, you are building a link in the system. When you reduce inference latency so that a model like GPT-5.4 (87.6 in agentic) responds in real-time on a theater of operations, you are making targeting faster.

The distinction between "tool" and "weapon" becomes blurred. The Congress document on LAWS defines lethal autonomous weapons as those that "select and engage targets." But what about a system that selects the target and presents it to a human who presses a button in 2 seconds?

Developers working at OpenAI, Google, Anthropic, or in defense must understand that the choice between fine-tuning, RAG and prompting is no longer an architectural choice. It is a choice that determines the level of autonomy of the targeting system. A RAG model with retrieval on dynamic databases will be more "autonomous" in its recommendations than a prompted model with fixed instructions.

This reality is new. There is no industry ethical framework preparing engineers for this responsibility. Model safety guidelines (RLHF, constitutional AI) were not designed for lethal contexts.

Congress Tries to Catch Up

The legislative response was swift but potentially insufficient. On June 15, 2026, a senator publicly questioned the Pentagon regarding the revision of the policy on autonomous weapons, denouncing a "compressed" timeline.

Then, Air & Space Forces reported that Congress wanted to extend oversight to AI systems used in operational planning — not just to autonomous weapons in the strict sense.

The legislative stakes are high. The definition of LAWS by the Congressional CRS is clear: no human intervention. But the new JP 3-60 doctrine maintains formal human supervision. Congress must decide whether "on-the-loop" human supervision is enough to exclude a system from the LAWS regime.

This is a legal debate, not a technical one. And it is taking place after the doctrine has already been modified.

The risk is that Congress ends up legislating on systems that are already deployed. Automation tools like those compared in Make, Zapier, n8n show how quickly AI pipelines can be assembled and put into production. In a military context, this speed is amplified by budgets and operational urgency.

What the revision says about the real state of AI in 2026

Beyond the military topic, this revision says a lot about the maturity the Pentagon attributes to current AI models. In June 2025, benchmarks already placed GPT-5.5 at 98.2, Claude Opus 4.7 at 94.3, Gemini 3 Pro Deep Think at 95.4 in agentic. The Pentagon would not revise its doctrine for systems it deems incapable.

The fact that the revision comes just one year after these scores suggests that the 2026 models — whose public scores are not yet available — have crossed a reliability threshold deemed sufficient for targeting. That is information in itself.

The DoD does not publish its own evaluations. But the sequence is revealing: AI strategy in January, conflict over guardrails in March, doctrinal revision in April, Google agreement for classified in April, revelation in June. Each step assumes that the technology is ready.

For civilian developers, the signal is clear: if the Pentagon trusts these models for targeting, the LLM reasoning capabilities in 2026 are well beyond what public benchmarks suggest. Either the scores underestimate the models, or military evaluations measure dimensions that mainstream benchmarks do not capture.

❌ Common mistakes

Mistake 1: Confusing targeting and engagement

AI does not fire. It identifies, prioritizes, and recommends. Engagement remains formally human. But in a context where AI initiates the sequence and human validation takes a few seconds, the distinction is more theoretical than practical. The solution: understand the full F2T2EA chain before judging.

Mistake 2: Thinking that LAWS do not exist yet

The CRS Congress document defines LAWS as systems without human intervention. The new doctrine maintains supervision. But the boundary between "supervision" and "absence of intervention" is a matter of legal interpretation, not technical.

Mistake 3: Believing that the debate is purely American

The revision of JP 3-60 will fuel the global autonomous arms race. China, Russia, and Israel are developing similar capabilities. The Pentagon's silence on its own doctrine encourages other actors to do the same, without any binding international framework.

❓ Frequently asked questions

Can AI already kill without human intervention?

No according to official doctrine. The revision of JP 3-60 maintains human supervision. But AI now initiates the targeting process, which reduces human intervention to a formal validation.

Which AI model does the Pentagon use for targeting?

The DoD does not publish its technical choices. But Google signed a deal for classified in April 2026 with Gemini, and Anthropic is in open conflict over Claude's guardrails. The landscape of available models suggests that several providers are involved.

Can Congress block this revision?

Theoretically yes, through the power of the purse. Congress controls the Pentagon's budget and can condition funds to restrictions on the use of AI for targeting. But oversight mechanisms are slow in the face of rapid technical deployments.

How does this concern civilian developers?

The architectures (RAG, fine-tuning, agents) deployed in the civilian sector are the same as those adapted for the military. The difference between these approaches determines the level of autonomy of the system. Developers who build these pipelines are participating, often unknowingly, in the technology ecosystem that makes autonomous targeting possible.

Is Anthropic refusing to work with the Pentagon?

No. Anthropic is disputing the conditions imposed by the Pentagon on guardrails, not the principle of collaboration. The conflict is about governance: who decides the ethical limits of a militarily deployed model, the provider or the end user?

✅ Conclusion

The secret revision of JP 3-60 is the moment when targeting AI moved from theory to operational doctrine, without prior public debate or oversight. For developers, the question is no longer whether AI will be used for targeting — that's done — but to understand how much every line of code contributes to this decision chain.