AI War Footage Provenance Claims

Introduction

Since the outbreak of large-scale conflict in Ukraine in February 2022 and the escalation of fighting in Gaza beginning in October 2023, the provenance of war footage has become one of the most contested domains in contemporary information warfare. Researchers, journalists, and governments face simultaneous challenges: the proliferation of genuine documentation from ubiquitous smartphones, the deliberate publication of misleading or decontextualized footage by state and non-state actors, and the increasing difficulty of distinguishing authentic footage from AI-generated or digitally manipulated material.

Ongoing investigations by Bellingcat, the BBC's Verify unit, ProPublica's Tracking Ukraine project, the New York Times Visual Investigations team, and academic researchers have advanced the state of the art in video forensics — but even specialist teams routinely encounter footage whose provenance cannot be definitively established in real time.

The Documented Challenge

Volume overwhelms verification capacity. In the first months of the Russia-Ukraine conflict, researchers estimated that tens of thousands of new pieces of conflict footage were being posted daily across Telegram, Twitter, TikTok, and other platforms. Existing open-source intelligence (OSINT) teams — Bellingcat, the Centre for Information Resilience, and others — are staffed in the dozens, not thousands. Verification is inherently incomplete.

Miscontextualization is the dominant form of disinformation. The majority of provenance-disputed footage examined by fact-checkers involves real footage attributed to the wrong event, location, or actor — not AI-generated fabrications. Reverse-video search tools, geolocation through visible landmarks, and shadow analysis have allowed researchers to resolve hundreds of misattributions. The BBC, AP, and Reuters have published detailed methodological guides.

AI-generated footage is emerging but remains a small fraction. While AI-generated synthetic video (deepfakes) has been used in conflict information operations — notably a poorly executed 2022 deepfake of Ukrainian President Zelensky urging surrender, quickly identified by metadata and visual artifacts — the technical quality required to produce convincing combat footage at scale has not yet made this the dominant form of manipulation. Independent researchers including those at the Citizen Lab and MIT Media Lab note that authentic footage remains far easier and faster to manipulate through selective editing and miscontextualization than to fabricate wholesale.

State actors deliberately exploit provenance ambiguity. Russia's Ministry of Defense and affiliated Telegram channels have systematically published footage with false attributions. The Ukrainian government has done the same in specific documented cases, according to independent assessments by the UN Commission on Human Rights in Ukraine and independent researchers. Both sides' information operations exploit the genuine difficulty of real-time verification.

Ongoing Investigations

Several major provenance questions in the Ukraine and Gaza conflicts remain under active investigation as of 2025–2026:

• Al-Ahli Hospital explosion (October 2023, Gaza): Within hours of the explosion, attribution was disputed between Palestinian officials blaming Israel and Israeli and U.S. intelligence asserting a misfired Palestinian rocket. Multiple independent investigations — by the NYT Visual Investigations team, AP, BBC, and Human Rights Watch — produced assessments ranging from "consistent with a misfired rocket" to "inconclusive." The event illustrates that even well-resourced forensic teams can reach different provisional conclusions from the same visual evidence under time pressure.
• Various Bucha imagery (March–April 2022): Russian government officials claimed Bucha footage was staged. Satellite imagery from Maxar Technologies, timeline analysis from the NYT, and independent OSINT teams confirmed bodies had been present on Bucha streets for days before Russian withdrawal — a strong rebuttal of the staging claim. However, independent satellite imagery access is not available for all conflict claims.

The Epistemic Problem

The genuine difficulty of war footage provenance creates a predictable exploitation pattern: actors with incentives to dispute authentic atrocity documentation can introduce uncertainty in public discourse by pointing to acknowledged verification limitations, even when specific footage has been verified. The existence of some genuinely unresolved cases does not invalidate footage whose provenance has been established.

Conversely, official attribution of footage to specific actors by governments with strategic interests in the outcome should be treated as one input among several in provenance assessment, not definitive confirmation.

What Responsible Verification Looks Like

Established open-source investigation methodology includes:

• Geolocation: Matching visible landmarks, terrain, and infrastructure to satellite imagery using tools including Google Earth, Planet Labs, and Maxar.
• Shadow and lighting analysis: Using solar position calculations to verify claimed time and location of outdoor footage.
• Metadata analysis: Examining EXIF data, upload timestamp sequences, and encoding artifacts.
• Cross-referencing: Matching footage against independently published accounts from multiple sources.
• Reverse-video search: Identifying prior publication of identical or manipulated footage.

Bellingcat, the NYT Visual Investigations team, BBC Verify, and the EU DisinfoLab have published detailed guides to these methods.

Verdict

War footage provenance is a genuine and important area of ongoing investigation, not a settled domain. The claim that all war footage is AI-generated or fabricated is unsupported and contradicted by the substantial volume of footage whose provenance has been established through rigorous open-source methods. The valid concern — that a meaningful fraction of circulating footage is misattributed, selectively edited, or decontextualized, and that real-time verification is impossible at scale — is well-documented. Investigations by established OSINT organizations continue, and provisional conclusions in specific cases should be understood as subject to revision as evidence accumulates.

The 2025 Israel-Iran Conflict: A Documented Escalation

The June 2025 Israel-Iran war produced what several researchers have described as the first conflict in which generative AI played a central role in shaping public perception in near real time. The European Digital Media Observatory (EDMO) documented a cluster of specific, independently verified cases. A video captioned "Doomsday in Tel Aviv," posted by the Tehran Times, was traced to an Arabic-language TikTok account that regularly produces AI-generated "fantasy destruction" content. Iranian outlets including Fars News, Nour News, and Shargh circulated a "Tel Aviv before and after" compilation that the Persian-language fact-checking outlet Factnameh found was almost entirely AI-generated. And an image purporting to show downed Israeli F-35 wreckage — bearing telltale flaws including a misaligned Star of David insignia and undisturbed grass beneath supposed debris — was assessed by an AI-detection tool as 98 percent likely synthetic. Iranian state television did not retract the F-35 claim until July 10, 2025, weeks after independent researchers had flagged it, illustrating that correction of AI-driven falsehoods lags across the propaganda spectrum, not only in adversarial reporting.

Euronews's Euroverify team separately identified a widely shared TikTok clip falsely depicting Iranian strike damage, distinguished from authentic footage by unnatural light rays emanating from a cooling tower and emergency workers who appeared to vanish mid-frame — artifacts consistent with synthetic video generation. The same team traced viral claims of Ben Gurion Airport damage to accounts with a documented history of spreading disinformation. Separately, PolitiFact debunked a set of images claiming to show Egyptian youths floating food-filled bottles toward Gaza, published July 24, 2025; several bore a visible watermark from Grok, X's AI image generator, along with distorted hands and physically implausible buoyancy.

The Sanitization Risk

Not every documented instance of AI-generated conflict imagery was created to deceive. In May 2024, an AI-generated graphic reading "All Eyes on Rafah" — depicting an aerial view of orderly tent rows arranged to spell the phrase — was reshared an estimated 47 million times on Instagram, becoming, by some measures, the most widely circulated AI-generated image on record. Its creator was not attempting war-crime denial; the graphic was explicit advocacy art, not documentary footage. But researchers who studied its spread noted a distinct risk: because platform content-moderation systems are tuned to filter graphic real photography, a sanitized AI depiction can travel farther and faster than the authentic images it stands in for. That inverts the usual "fake displaces the truth" framing into "the sanitized fake outcompetes the truth for attention" — a media-literacy hazard distinct from, and not evidence for, claims that the underlying events are fabricated.

Verification Infrastructure and Its Limits

Newsroom-side responses have matured alongside the threat. The Coalition for Content Provenance and Authenticity (C2PA) — whose steering committee includes the BBC, Adobe, Sony, Microsoft, Google, and Truepic — has developed the Content Credentials standard: a cryptographically signed manifest attached to a photo or video at the moment of capture, recording subsequent edits. Camera manufacturers have begun building it in. Leica's M11-P, released in 2023, was the first consumer camera to sign images via a hardware security chip at capture; Sony and Canon have since added support aimed partly at photojournalists, and Reuters has worked directly with Canon on testing. The BBC has the most mature newsroom implementation to date, with the New York Times and AFP focusing on photojournalism use cases.

Content Credentials only establish provenance for footage captured on compliant hardware — the standard cannot retroactively verify the vast body of video already circulating without it, and it says nothing about whether a compliant camera was pointed at a genuine or staged scene. Columbia Journalism Review's Tow Center for Digital Journalism has separately cautioned that automated deepfake-detection tools remain an unreliable substitute for open-source verification: accuracy drops sharply outside a tool's training data, confidence scores such as "70% human, 30% artificial" often fail to specify what was altered or whether the substantive content changed, and journalists risk overconfidence when a tool's verdict happens to match their prior assumption. The Tow Center's guidance is that detection software should function as one input alongside reverse-image search, metadata review, and cross-referencing — never as a standalone verdict on its own.