On the night of 15 June 2026, as Russian drones struck Kyiv, Metropolitan Yevstratii Zoria was sent a photograph of Kyiv Pechersk Lavra. The image showed flames climbing between the gold domes of the Dormition Cathedral, one of Ukraine’s oldest religious sites.

Yevstratii did not take the photograph himself. He later told the Kyiv Independent that he had shared it after someone sent it to him but could no longer remember who. “At the time, the strikes and the fire at the Lavra understandably did not allow for a calm response,” he said.

He published the image as the attack was unfolding. Ukrainian media outlets quickly republished it as one of the first visual accounts of the strike. No photographer was identified. At the time, that absence seemed insignificant. The image appeared to speak for itself.

Then the image became the story.

Within days, posts containing the photograph were labelled as false information after an AFP (Agence France-Presse) fact-check connected it to a wider network of AI-generated images circulating alongside Russian propaganda. Pro-Kremlin accounts had already used fabricated photographs to claim journalists had been waiting outside the monastery before the strike, suggesting the attack had been staged. The burning cathedral image was drawn into the same atmosphere of suspicion.

A single disputed image triggered moderation measures: posts were labelled false, pushed lower in users’ feeds, excluded from recommendations, and in some cases demonetised. Authentic reporting on the attack was caught in the same system. Meta later acknowledged a technical error.

AFP’s fact-checkers were right to investigate synthetic images circulating around the attack. The photograph may well have been artificial. But the cathedral was still hit. And evidence of a real event was caught inside the uncertainty created by fake ones.

The doubt that surrounded the Lavra image is a reversal of everything open-source investigation was built to do. That work began with people like Eliot Higgins.

In 2012, Eliot Higgins began analysing weapons and videos from the Syrian conflict from his home in Leicester under the name “Brown Moses”. He had no military intelligence background and no access to classified material. Instead, he worked with what was publicly available: photographs, videos, satellite imagery, and the details hidden within them.

‘Open Source’ let ordinary people prove what states had done. A person with an internet connection could identify weapons, trace movements, and challenge official accounts. Bellingcat launched on 14 July 2014, just three days before Malaysia Airlines Flight MH17 was shot down over eastern Ukraine on 17 July 2014. In the weeks that followed, investigators began examining photographs and videos of the Buk missile system believed to have been used in the attack.

The same launcher could be traced across a single journey. It was photographed in Russia in the weeks before the attack, filmed again on the morning of 17 July as it moved through separatist-held territory toward the launch site, and recorded the following day heading back toward the Russian border. It was the same vehicle. Investigators proved it by comparing details most people would have ignored: the rubber skirts above the tracks, worn into a pattern unique to that launcher, like a fingerprint. At first, one kink appeared not to match. Investigators later concluded the discrepancy was likely an artefact of the image itself, an angled photograph that hadn’t been fully corrected for perspective when compared to the reference image.

The photographs mattered because they recorded a physical object with a history. A vehicle had travelled somewhere. Something had happened to it. The images preserved those traces.

Bellingcat says it provided its research to the Joint Investigation Team investigating MH17. On 17 November 2022, a Dutch court found that the aircraft was downed by a Buk missile fired from a field in eastern Ukraine, using a system that belonged to Russia’s 53rd Anti-Aircraft Missile Brigade and was transported into Ukraine and back afterwards.

The Syrian Archive, founded by Hadi Al Khatib, collected and preserved digital evidence of human rights abuses during the Syrian civil war. In 2020, the University of California, Berkeley’s Human Rights Center and the United Nations Human Rights Office published the Berkeley Protocol on Digital Open Source Investigations, establishing standards for collecting and verifying open-source evidence. What had once required a state’s intelligence agencies could now be done by anyone patient enough to look.

For a decade, that held.

The Buk launcher could be traced because there was a launcher to trace. It had travelled somewhere. It carried the marks of that journey from one photograph to the next. Artificial intelligence has made that work both easier and harder. Thousands of images, videos and social media posts can now be translated, transcribed, sorted and cross-referenced in minutes. For investigators working through vast archives of digital evidence, that is a genuine advantage.

But that archive is no longer made only by people. Alongside authentic photographs now sit synthetic images, fabricated videos, cloned voices and machine-written text, produced quickly and at little cost. They sit beside genuine evidence, making it harder to assume that an image records something that ever existed.

AI is now being used to analyse an information environment that AI is simultaneously contaminating.

That leaves verification at a permanent disadvantage. Every new image generator only has to produce material that escapes today’s detection systems. The detectors must identify convincing outputs from every generator already in use, as well as those still to come. The result is an arms race in which fabrication moves first and verification follows behind. The harder it becomes to establish whether an image has a real object behind it, the weaker the link between a photograph and the event it claims to record.

The danger is not only the fake image itself but also what happens around it. AFP was right to investigate synthetic images circulating alongside the strike on the Lavra. But the moderation did not remain confined to the disputed image. It reached authentic reporting of the attack. A real event became harder to document because fabricated images existed beside it. The facts had not changed. The authority of the photograph had.

When doubt becomes a weapon

Legal scholars Robert Chesney and Danielle Citron called this the liar’s dividend: the advantage created when people can dismiss genuine evidence as fabricated. The lie does not have to replace the truth. It only has to make the truth harder to establish. The danger was never simply that people would believe fake images. It is that they would stop believing real ones.

A government, military, or armed group does not need to produce a convincing alternative account of an event if it can cast doubt on the evidence documenting it. A photograph can be called AI-generated. A video can be dismissed as a deepfake. The person accused of wrongdoing does not have to prove that the evidence is false; they only have to make the accusation plausible enough for others to hesitate. The existence of convincing synthetic media gives denial a new source of credibility.

Gabon showed how early this could happen. President Ali Bongo had been largely absent from public life after suffering a stroke in 2018. When he appeared in a video address on 31 December, questions were raised about the video and Bongo’s condition. A week later, on 7 January 2019, Lieutenant Kelly Ondo Obiang and other soldiers attempted a coup. In his broadcast, Ondo Obiang said Bongo’s New Year’s address had reinforced doubts about the president’s ability to carry out the responsibilities of his office. The video was never established to be a deepfake. In 2019, a video did not need to be proven fake for doubt around it to have political force.

The technology was crude by today’s standards. That is what makes the episode important. The possibility of a deepfake was already enough to make uncertainty politically consequential before convincing deepfakes were commonplace. Today, synthetic images and videos are far more convincing and far easier to produce. The liar’s dividend once relied on a technology that barely worked. Now it can draw on fakes that do.

For open-source investigators, this changes the burden of proof. Evidence can survive physically and still fail politically if the people it implicates can persuade an audience to doubt it. Accountability depends not only on preserving the record but also on preserving the reason to believe it. An image can survive. The truth it records may not.

Who is accountable?

On 15 June, AFP fact-checked an image that appeared to show the Dormition Cathedral at Kyiv Pechersk Lavra on fire. The image may well have been artificial, and AFP was right to investigate it. Meta then applied its moderation systems to posts containing the image. Those systems did what they were designed to do – label and reduce the reach of content that had been identified as false. But the labels spread to authentic reporting of the strike because of a technical error. Real coverage of a real attack on a real cathedral was caught in the same system. No one had deliberately decided that the reporting was false, yet it was suppressed.

That raises a harder question than whether the image should have been flagged. At what point to whom does accountability attach? To the fact-checker who identified the false image? To the platform that acted on the fact-check? To the engineers who built the system? Or to the people who deliberately polluted the information environment in the first place? The same problem extends beyond Meta. An analyst may use AI systems to identify patterns across thousands of images, videos, and documents. Those systems may help prioritise information or assess credibility, but the material being analysed may itself contain fabricated content generated by other models. The resulting assessment can move upward into institutions where it informs whether sanctions are imposed, whether a strike is ordered, or whether a person’s claim for protection is believed.

Someone made that call. But who, exactly? Existing accountability frameworks were built for decisions that trace back to a person. They are not built for a conclusion that emerges from human judgement, machine assistance, and a pool already seeded with fakes.

The consequences are not abstract. A mistaken assessment can contribute to escalation. A credible account of violence can be ignored. A person fleeing persecution may be left with evidence no one is willing to believe.

On the night of 15 June 2026, flames climbed between the gold domes of the Dormition Cathedral. Yevstratii had been sent a photograph of them by someone whose name he could no longer remember. He shared it as the strike was unfolding. For a moment, the image seemed to be enough.

The cathedral was hit. The fire was real. The image may well have been artificial.

The photograph from nowhere is still there. And so is the uncertainty.