Since Israel and the U.S. launched strikes against Iran in February, plunging the Gulf region into conflict, the countries have also been waging an information war using tools powered by artificial intelligence. To separate fact from fiction, authorities, analysts, and journalists alike turned to commercial satellite imagery, including Google Earth, the closest thing to “ground truth,” with limited access to sites.
Google Earth has long been a trusted infrastructure for the verification field, and is the platform on which open-source investigators are formally trained for conflict work in Syria, Yemen, and Sudan. Yet on July 30, Google launched a new feature that built image generation directly into Google Earth, the product that for two decades was synonymous with satellite image truth.
It meant that users could zoom in to real coordinates, type a prompt, and get an AI-generated photorealistic scene blended into actual satellite and aerial imagery. Within hours, users had conjured a nuclear plant in Iran, a bomb crater beside a Gaza hospital, and floods lapping the U.S. Capitol.
Faking satellite imagery is not new. Google’s own watermarks had exposed its Gemini AI model being used to fake satellite evidence, a recurring feature of Iran’s state-affiliated propaganda. But the Google Earth feature provided greater ease and accessibility inside the very tool that was known for trusted information.
Within a day, following an outcry from investigative experts and the media, Google pulled the feature, saying it would return with “stronger guardrails.” That is our information environment in a nutshell: New ways to fabricate launch overnight, a storm of criticism ensues, and guardrails arrive only after the harm.
Meanwhile, the U.S. government ordered satellite companies to restrict journalists’ access to imagery in the region, citing wartime security, narrowing the supply of real information even as the counterfeits spread.
Fabricated footage #
This was not the first generative AI product from a big tech company to be pulled or shelved recently. OpenAI shut down its Sora video app in March following sustained public demands for its withdrawal. But the recall came only months after a flood of abusive content of women and problematic recreations of historical figures. The pattern is similar: The harm arrives, the company withdraws the product and says it will work on the guardrails it could have built — or been required to — before any harm.
The recalls are the exceptions. The rule looks more like tools such as Veo 3, the video generator Google launched in May 2025, weeks before Israel and Iran went to war. Within days of the first strikes, AI-generated videos of missile barrages on Tel Aviv were circulating with the Veo watermark still visible. Later, users became savvy enough to crop out the visible watermark.
Watermarks only apply to products that carry them, though. When fabricated footage of a blast at Iran’s Evin prison hit social media within minutes of the strike and was shared by outlets including Sky News and the BBC, there was no watermark with which to check. Citizen Lab later assessed that the video belonged to an AI-enabled influence operation likely run by the Israeli government.
But detection and attribution run on different clocks, and both lag the lie. Forensic experts determined the Evin video to be AI-generated within days, but in an environment of millions of pieces of content, these efforts are not enough. The facts arrive long after the doubt has done its work.
Liar’s dividend #
Regulation has not kept pace with the rapid development of AI models, and new laws don’t go far enough. Some of the transparency obligations required by the EU AI Act, enacted in 2024, began on August 2, 2026. Despite the long timeline, implementation is shaky, with precise rules remaining a voluntary code rather than enforceable legislation. California’s AI transparency law, also passed in 2024, took effect in August for companies that make generative AI tools, but won’t reach the platforms until at least 2027. And the detection tools these laws require are not ready: When my colleagues tested them across 13 AI providers, seven had no public detection tool, and of those that did, all but one could be fooled.
This environment has allowed the theory of the liar’s dividend to become the reality of modern conflict. When legal scholars Bobby Chesney and Danielle Citron coined the term in 2019, it was a warning: The mere existence of convincing fakes would let bad actors dismiss authentic evidence as fabricated. That term is now simply how contemporary conflicts work. The fake no longer has to survive scrutiny; it only needs to exist long enough for the truth to arrive damaged.
The Google Earth episode demonstrated in 24 hours what the last three years of AI content distorting conflicts and crises have taught more slowly. Guardrails for these tools need to be tested where the stakes run highest — in wars, atrocities, and crises — before release, not retrofitted after an outcry. Some capabilities will fail that test outright: Generative image creation has no legitimate place inside verification infrastructure such as Google Earth.
AI transparency experts have outlined what is needed: provenance that is not optional, traveling with an image to wherever a viewer sees it; robust and transparent efforts to detect AI content and communicate how the content is made, including uncertainties and complexities outside of “AI versus real” binaries. And a burden of proof that rests with the companies deploying these systems, not the people documenting their own catastrophes.
Google knows how to build for trust. A year ago, the Google Pixel 10 became the first smartphone to build the open C2PA provenance standard directly into its camera, signing photos at the moment of capture. In Apple’s newest iPhone 18 Pro models, unveiled this week, a set of features enables users to produce cryptographically signed reference images and identify images generated or edited with AI. “When one of the largest companies in the world decides this problem is worth solving at the hardware level, it is an acknowledgment that our collective loss of trust in the visual record is no longer a niche concern for forensic researchers,” noted Hany Farid, the world’s leading deepfake expert. “It is a mainstream problem, and it demands mainstream solutions.”
Google recalled its feature back in a day, but no one can recall the doubt that now attaches to every image out of a war zone. The counterfeit satellite frames from this war are still being screenshotted into new claims, and the authentic record now carries an asterisk it does not deserve. Somewhere in that gap, the photojournalists proving their own pictures, the families weighing which warnings to believe, and the fact-checkers running behind the lie keep doing the slow work of restoring what a product launch made deniable in an instant.
Google said the feature will return once its “stronger guardrails” are ready. One hopes the led to the realization that no benefit of integrating generative AI into Google Earth is worth the harm and the doubt it creates.