How AI Is Used in Modern Warfare—and Where Oversight Can Fail

Artificial intelligence now supports intelligence analysis, operational planning, logistics, cyber operations, navigation, and information campaigns in armed conflict. It can process data faster than a human team, but it does not remove uncertainty or legal responsibility. The central risk is “decision compression”: a system produces recommendations so quickly that human review becomes a formality rather than a meaningful check.

What is known about Claude and Operation Epic Fury

U.S. Central Command began Operation Epic Fury against Iran in February 2026, alongside Israel's Operation Roaring Lion. Multiple news organizations reported that U.S. forces used Anthropic's Claude through military systems during the campaign. Reuters reported that a source confirmed the use of Anthropic services but that the precise role was not disclosed.

That distinction matters. Public reporting supports the claim that Claude was used, but it does not establish every task attributed to it. Statements that the model personally “selected,” legally approved, or executed a particular strike should not be presented as fact without an official record describing the workflow.

Anthropic has separately confirmed that its systems supported U.S. defense work in areas including intelligence analysis, modeling and simulation, operational planning, and cyber operations. In a March 2026 statement about its dispute with the Pentagon, the company said operational decisions belong to the military, not to a private AI provider.

Where AI enters a military workflow

Project Maven and machine-assisted analysis

Project Maven began as a U.S. Department of Defense effort to apply computer vision to large volumes of aerial imagery. Its broader ecosystem has evolved into software that combines data from sensors and other sources to support analysts and commanders. Anthropic announced in 2025 that Claude was integrated with Palantir systems on classified networks for defense and intelligence customers.

“AI-assisted” does not reveal how much authority the software had. A system may only retrieve documents, it may assign confidence scores to imagery, or it may recommend a prioritized list. A responsible account should ask who reviewed the output, which evidence was visible, whether dissenting assessments were recorded, and who had authority to act.

Autonomous navigation is not autonomous targeting

Computer vision can help a drone follow terrain or maintain a track when GPS is jammed. A loitering munition may also have automated functions for searching a defined area. Those capabilities are not identical to a weapon independently deciding that a person or object is a lawful target.

The distinction can be separated into three questions:

1. Mobility: can the system navigate and avoid obstacles without continuous control?
2. Perception: can it detect or track an object from sensor data?
3. Use of force: who determines that the object is a valid target, selects the weapon, and authorizes an attack?

Anthropic's public position in its 2026 Pentagon dispute was that it would not remove safeguards concerning fully autonomous weapons and mass domestic surveillance. The company also said it offered to support research on the reliability of military AI. Those are the company's stated limits; independent oversight is still needed to determine how systems are deployed in practice.

Why faster targeting can increase risk

Speed can be useful when tracking an imminent threat, but a shorter decision cycle can also remove opportunities to discover an error. Large models and classifiers can produce plausible output from incomplete or contradictory inputs. A confident recommendation may encourage automation bias—the tendency to accept a machine result even when other evidence should prompt doubt.

Human involvement is meaningful only when the reviewer has enough time, information, authority, and technical understanding to reject the recommendation. A person clicking “approve” after seeing a summary is not equivalent to an analyst examining the underlying sources.

Minimum controls for high-consequence use include:

• clear separation between intelligence hypotheses and verified facts;
• source provenance, timestamps, confidence, and contradictory evidence shown with every recommendation;
• independent human review and a documented chain of authority;
• legal assessment that is not reduced to an unverified model output;
• testing against spoofing, data poisoning, sensor failure, and adversarial inputs;
• audit logs that preserve model version, prompts, inputs, outputs, edits, and final decisions;
• fail-safe behavior when communications or confidence fall below defined thresholds;
• after-action investigation and reporting when civilian harm or system error is alleged.

AI-generated misinformation during conflict

Conflict creates ideal conditions for false media: events move quickly, original footage is scarce, audiences are emotional, and old clips can be relabeled before verification. Generative AI adds cheap synthetic images and voices, but not every false post is a deepfake. Video-game footage, old war footage, altered captions, and unrelated photographs remain common.

Reports during the 2026 Iran conflict described fabricated imagery of damage to U.S. military assets as well as recycled videos from earlier wars. The safest approach is to verify the specific file and claim rather than infer authenticity from how realistic it looks.

Verification checklist for readers

1. Find the earliest available upload and identify who originally posted it.
2. Look for confirmation from multiple independent reporters or official sources close to the event.
3. Extract key frames and perform reverse-image searches to find older uses.
4. Compare landmarks, weather, shadows, uniforms, insignia, aircraft, and damage with known reference material.
5. Check whether the audio is continuous and whether lip movement, reflections, and compression artifacts change abruptly.
6. Inspect metadata when the original file is available, but remember that metadata can be removed or altered.
7. Do not treat an AI detector's score as proof; detectors can produce false positives and false negatives.

TipsMake's guide to AI-video detection tools is a starting point, while the site's metadata and file-inspection tools can support manual checks. Neither replaces source verification.

The policy question is accountability

The debate is not simply whether militaries will use AI; they already do. The more important questions are what the system is permitted to do, how its limits are tested, whether commanders can understand and challenge its output, and who is accountable when it fails.

AI may help analysts find relevant information in a flood of data. It may also amplify mistakes at operational speed. Responsible coverage should distinguish confirmed deployment from anonymous reporting, decision support from decision authority, and generated propaganda from independently verified evidence. Those distinctions are essential when lives and escalation decisions are at stake.