On February 28, 2026, U.S. strikes hit the Shajareh Tayyebeh Elementary School in the southern port city of Minab, Iran. Iranian authorities reported that around 170 people were killed, at least 110 of them schoolchildren. Days later, when mourners buried their children, Iran’s foreign minister posted a photograph of rows of small, freshly dug graves. Some Iranian diaspora opposition accounts immediately began labeling the photo as AI-generated. Yet subsequent verification work confirmed it was indeed authentic. While the Iranian state has long dismissed evidence of its violence as foreign-produced fabrication, this time, the tables were turned: Iranian officials found their evidence of war crimes dismissed as fabrication. The burden then fell on the victims of the strike, who were forced to prove that their deaths and grief were real.

AI in war manufactures what shirin anlen from WITNESS has termed “ambient uncertainty” about everything. In Iran, we see the worst-case consequence of this uncertainty: A license is given to every party to dismiss documented evidence as fake, while the effects are keenly felt by communities whose deaths or experiences are denied and by families who are told their grief is propaganda.

As a result of this uncertainty, in June 2025, many Iranians did not know what to believe during the intensive attacks on Iran. After decades of Iranian state lies and denial of human rights abuses, portions of the country’s population were not convinced that U.S. and Israeli strikes were dangerous and really killing civilians. Instead, they assumed that the regime was exaggerating or fabricating the death toll until they saw the consequences themselves.

With no trusted actors, Iranian citizens treated Israeli warnings about strikes in civilian areas as suspect by default, heightening the risks of incoming bombardments. Shaped by Israel’s own record of brutal violence in the region and by lies about its state and military actions, even the minimal “evacuation” notices Israel provided to residents in Tehran created further panic and opened a new canvas for AI-generated saturation of reality. The messages, asking millions to evacuate with little to no time, which Amnesty International described as “sinister and fear-inducing,” either came with inaccurate or vague descriptions of the city or were issued in the middle of the night when residents were asleep and under an internet and communications blackout.

This vacuum of reliable information created confusion and panic among Iranians: People were left trying to determine whether a warning was manipulated or a real sign of an incoming bombardment. Cascading into that fog, AI-generated emergency alerts and spoofed official messages circulated alongside state ones, capitalizing on public fear, distrust, and the information vacuum and adding synthetic fuel to a preexisting propaganda crisis. In wartime, AI accelerates doubt: States with a record of official deception are more readily dismissed in moments of real danger, while AI gives all sides new tools to exploit that skepticism and obfuscate their abuses.

Weaponizing Doubt

Inside Iran, the regime has perpetuated conditions for this disbelief for decades, but now, all of the warring parties are exploiting them. In February 2026, as the United States and Israel began bombing Iran, their denials of civilian casualties fed an AI-saturated information environment where war crime documentation was preemptively dismissed as fabricated. When every side lies, AI turns uncertainty into a weapon; AI tools become a means to fabricate evidence, dismiss real documentation, and make truth arrive too damaged to protect anyone.

This is hardly the first time this dynamic has unfolded. Take Gaza, for instance, where Israel has been responsible for more than 75,000 deaths and where the UN Independent International Commission of Inquiry concluded that Israel committed genocide against the Palestinian population. The case of Saeed Ismail is particularly illustrative. Saeed was displaced from northern Gaza at age twenty-two and went online to raise money to feed his family. Soon, however, accounts on Bluesky accused him of being an AI creation. The accusations arose in response to the garbled embroidered text on his blanket. It was the kind of misspelled English often found on mass-produced textiles in markets where English is used as a decorative design element rather than a carefully proofread message, but online it was treated as the telltale fingerprint of AI. After surviving the massacres in Gaza, Saeed spent hours defending the fact that he is a real person. Such is the cruelty of AI dismissal.

Saeed is not the only one who has experienced this ordeal. In April 2024, a photograph of an injured man at Al Shifa Hospital was manipulated with AI by unknown actors and made to look like he had a third leg. The photo was then circulated as supposed proof that images of Palestinian victims are AI-generated. Open-source investigator Tal Hagin traced the image back to its original form, but the doctored version remained in circulation. While AI does not invent the machinery of denial, it furnishes new language for long-standing efforts to dismiss Palestinian documentation as staged, manipulated, or terrorist propaganda. This dynamic has a name: the “liar’s dividend,” coined in 2018. Researchers predicted that as deepfakes became more convincing, bad actors would benefit not only from passing fakes off as real but also from passing real evidence off as fake. What was framed as a legal theory has hardened into a standard feature of modern conflict, refined in each subsequent war.

As early as 2021, Sam Gregory, the head of WITNESS and a synthetic media expert, warned that authoritarian regimes could turn the very existence of deepfakes into a weapon to dismiss genuine smartphone footage of state violence, including in Myanmar, as fabricated. In a few short years, the scale of use has exploded. After the onset of Sudan’s civil war in 2023, AI-manipulated propaganda began circulating, and the same dynamic occurred during Israel’s 2023 campaign in Gaza, with instances numbering in the hundreds. By the 2025 Iran-Israel AI war, the tactics had become routine. And by 2026, propagandists were generating manipulated information at an industrial level, with journalists and fact-checkers conceding that the volume of AI content tied to the Iran war was unprecedented and overwhelming.

Denial is an old strategy, but generative AI has democratized it. Previously, the most effective way to cast doubt on war crimes was to make documentation itself dangerous or impossible: control the networks, seize the footage, and cut the connection. During the 2020–2022 Tigray War in Ethiopia, which resulted in hundreds of thousands killed, the Ethiopian state imposed communications blackouts, likely to hide atrocities and contest the truth.

But that level of centralized control is no longer necessary. Anyone with a generative AI model and a social media account can now amplify the same doubt, simply by circulating something synthetic or calling real evidence fake. Almaz, a prominent Tigrayan human rights defender whose name we have changed for her safety, experienced both forms of denial. During the blackouts, those able to document abuses—or amplify evidence from the diaspora—faced reputational attacks and efforts to discredit their documentation. “They would edit pictures and videos and try to fabricate evidence to back whatever narrative they were trying to push against me,” she told the De|Center, a digital human rights organization, in 2024, as the atrocities continued.

Moreover, when witnesses cannot safely record what they see, and advocates are made into targets, state denial efforts get a head start. Verification may eventually arrive, but it arrives late, after the public understanding of events has already become anchored. Now, the vacuum is being filled not by state denial and information blackouts but by AI-generated images, creating an industrial-level environment of unknowability.

The problem is not limited to fake images being used to discredit real ones. Sometimes the footage itself is authentic, while the political framing around it is deceptive. In May 2025, video footage from the U.S.- and Israeli-backed Gaza Humanitarian Foundation (GHF) aid site in Rafah went viral, showing lines of people waiting—and cheering—for food. The video was propaganda meant to suggest that Palestinians welcomed the GHF. On social media it was accused of being AI-generated and untrustworthy. But Hany Farid, a digital forensics authority and co-founder of GetReal, found no signs of AI generation. The people who called it fake were not wrong to distrust the GHF. As human rights organizations have noted, Palestinians were being subjected to starvation as a weapon of war, and the GHF was itself a militarized aid operation outside the UN, documented to have committed war crimes by killing Palestinians who came for food. The footage was real; the framing was the lie. In an information environment shaped by violence and official deception, the two had become almost impossible to separate.

These examples reveal the epistemic damage caused by AI. The accusation of fakery has no fixed political target; it is no longer just a claim about whether an image is real, but also a line trotted out when it suits a particular party. Weaponization of doubt can escape the control of those who manufacture it.

Insights from Iran

This dynamic is now most visible and most advanced in Iran. For decades, the state’s internet shutdowns, surveillance infrastructure, and suppression of independent media produced an information vacuum that authorities adeptly exploited. But the onset of generative AI has progressed the denial strategy even further. The January 2026 regime crackdown against protesters offers insights.

People across Iran reported streams of civilian casualties from the brutal state response to protests. As reported in a Tech Policy Press article, contacts inside Iran sent WITNESS several photographs of blood pooled on a Shiraz sidewalk, each frame holding nothing but the ground, because raising a phone any higher risked being seen. A woman whose name was changed for her safety, Shahla, was in Tehran when she watched cleaning crews wash protesters’ blood from her own neighborhood streets. She had not taken pictures because it was too much of a risk with the security presence. At checkpoints, guards went through phones and detained anyone whose camera roll held protest footage. Whatever record made it out of Iran that January was fragmentary, dangerous to gather, and often destroyed before it could travel.

Those inside Iran struggled to share evidence of the regime’s atrocities, allowing others to exploit the vacuum. In April 2026, Eyal Yakoby, a pro-Israel commentator with a large social media following, posted an AI-generated image on X of blood-soaked Tehran streets, claiming it depicted the January massacres against protesters months earlier. His goal did not seem to be to document Iranian suffering, but rather to reinforce the political case that the regime’s brutality justified the ongoing war on Iran. BBC Verify journalist Shayan Sardarizadeh then debunked the image, alongside other fact-checkers at Full Fact.^[1] The regime already had what it needed though: a discredited AI fake to reinforce the argument that all documentation of the January massacres was suspect and being used to advance foreign agendas.

AI-generated lies do not have to survive verification to succeed. They only have to create an interval of doubt.

Those with power understand that AI-generated lies do not have to survive verification to succeed. They only have to create an interval of doubt. A fake image can be debunked and an accusation can be disproved, but by then, the first wave of disbelief has already circulated. The seed has been planted—perhaps the massacre was staged, perhaps the grieving mother was an actor, perhaps the witness was never real. Denial does not need to win the argument; it only needs to make the truth arrive to its audience damaged.

The mechanism is structural, but the harm is individual and cumulative, and the cost is concentrated on specific people: Saeed in Gaza; Almaz in Tigray; Shahla in Tehran; the parents of the children buried in Minab; and the Iranian families who could not tell regime propaganda from genuine warnings.

Countering AI Propaganda Is Not Futile

Despite these challenges, the people targeted by these lies keep doing the work of making truth believable. Saeed Ismail is still raising money to feed his family, and Palestinians continue to show the real face of Israeli atrocities in Gaza. Tigrayans are still preserving testimony from a war in which documentation was dangerous and denial had a head start. The Human Rights Activists News Agency in Iran has compiled a 1,350-page named record of the January crackdown, despite a regime determined to erase what happened. The parents in Minab continue to share their children’s stories, keeping their memories alive while caught between regime propaganda and opposition denial. 

As war continues across Iran, Lebanon, and Gaza today, the struggle over whether violence can be seen, verified, and named has become inseparable from whether it can ever be punished. Even the institutions meant to hold perpetrators accused of war crimes to account are now under attack, as the United States moves to dismantle the International Criminal Court and sanction its officials and allied human rights defenders. In this environment, we have a duty to resist the denial of truth and the technological systems and tactics that manufacture ambient uncertainty around war.

The struggle over whether violence can be seen, verified, and named has become inseparable from whether it can ever be punished.

There is no silver bullet to end wartime denialism, but the space for it can be narrowed if the problem is treated as both a content one and a witnessing one. When treated as a content problem, the task is to spot and label individual fakes once they are already circulating, an endless game of catch-up that the dismissal mechanism is built to win. Treated as a preemptive witnessing problem, the task is to protect the people who record atrocities and the processes that let their evidence travel intact, so that an authentic record can survive the accusation of fakery before it is ever made.

Therefore, the first task should be ensuring that witnessing can happen safely, which requires tools designed to address the risks to people documenting violence. Apps like Proofmode or Tella, for example, allow witnesses to record, encrypt, and control the identifying data attached to evidence. Resilient channels that keep working when a state cuts the network are also vital.

Device-to-device messaging apps like Briar or Delta Chat can help. Distributed archives with dispersed copies—like the Syrian Archive of war crimes or the Palestinian Museum’s “unlootable” heritage archive—keep records from vanishing when a server is attacked, platforms removed, or a connection drops.

Safe witnessing also requires harm reduction and privacy protections that recognize the physical dangers witnesses face, such as device searches, confiscation, destruction of evidence, and mass surveillance. It also means pushing back on policies that destroy safety online and defending online anonymity so those documenting violence, where being identified can be risky, have layers of safety.

The second task should then be creating a layered set of trust signals that can protect documentation from dismissal. There must be provenance and authenticity infrastructure embedded at the point of content creation so that the documentary record carries its origin with it. Content lifecycle tracking needs to dispel the binary of “AI equals fake,” since AI editing now appears in real photos. And content platforms must communicate AI content transparently, without overpromising on detection that does not reliably work. These efforts require building provenance and content credentials into capture and editing by default. They also require preserving that provenance while content moves through platform systems instead of stripping it on upload.

Neither of the tasks are sufficient alone. But together, these protections can make denial harder to weaponize. Tech companies have the resources, obligation, and reach to slow the acceleration of epistemic collapse. They cannot end wartime denialism, but they can protect the conditions under which evidence is created, preserved, and understood. The alternative is accepting ambient uncertainty as the permanent condition of modern war documentation.

^[1] See Shayan Sardarizadeh’s reply and post on X. Note that age-restricted content may only be available to users signed in to X.