Google Earth has been the backbone of open-source conflict investigation for two decades. When reporters, human rights researchers, and forensic analysts needed to verify what was happening on the ground in Syria, Yemen, or Sudan, they turned to Google Earth as the closest thing to ground truth. That trust lasted exactly 24 hours after Google decided to add generative AI to the product.
On July 30, Google rolled out a feature that let users zoom into real satellite coordinates, type a text prompt, and get an AI-generated photorealistic scene blended directly into actual imagery. Within hours, users had conjured a nuclear facility in Iran, a bomb crater beside a hospital in Gaza, and floodwaters lapping at the steps of the U.S. Capitol. All of it looked real. All of it was fake. All of it was served inside a platform people trusted to show them the world as it actually is.
Google pulled the feature by July 31, saying it would return with stronger guardrails. But the damage to trust was already done, and no software update can undo that.
Why Google Earth matters for conflict verification
Google Earth has occupied a unique position in the information ecosystem. It is the platform where open-source investigators are formally trained. Organizations teaching people to do conflict documentation in active war zones use Google Earth as their primary tool. When a strike hits a building, analysts cross-reference the location against Google Earth's imagery to confirm or deny what happened.
This workflow depends entirely on one assumption: the imagery in Google Earth is real. Not curated, not interpreted, not generated. Real. The moment Google added AI image generation to that tool, it broke the contract with every investigator, journalist, and fact-checker who relied on it.
Faking satellite images is not new. Iran has used Google's own Gemini model to fabricate satellite evidence, which the company has exposed through watermarks. But those fakes lived on social media and propaganda channels. The Google Earth feature put the fabrication capability inside the verification tool itself. It is the difference between a counterfeiter operating in a back alley and one setting up shop inside the Federal Reserve.
The information war context
The timing made this worse. Since Israel and the United States launched strikes against Iran in February 2026, the region has been locked in both a military and an information conflict. Journalists and analysts need reliable imagery more than ever. At the same time, the U.S. government ordered satellite companies to restrict journalists' access to imagery of the region, citing wartime security. Real information was getting harder to obtain while fabricated content was getting easier to produce.
The Evin prison strike illustrates the problem. When footage of a blast at Iran's Evin prison appeared on social media minutes after the strike, major outlets including Sky News and the BBC shared it. There was no watermark. Citizen Lab later assessed that the video was likely part of an AI-enabled influence operation run by the Israeli government. But forensic experts only confirmed it was AI-generated days later. By then, the doubt had already done its work.
This is the liar's dividend in action. The term, coined by legal scholars Bobby Chesney and Danielle Citron in 2019, described a future where convincing fakes would let bad actors dismiss real evidence as fabricated. That future arrived. The fake does not need to survive scrutiny. It only needs to exist long enough for the truth to arrive damaged.
Detection tools are not ready
Laws have tried to address this gap. The EU AI Act, enacted in 2024, began requiring certain transparency obligations on August 2, 2026. But implementation is shaky, with key rules remaining voluntary rather than enforceable. California's AI transparency law, passed the same year, took effect in August 2026 for companies making generative AI tools but will not reach platforms until at least 2027.
Even where detection tools are required, they do not work. When researchers tested detection tools across 13 AI providers, seven had no public detection tool at all. Of the remaining six, all but one could be fooled. The laws assume a detection infrastructure that does not exist.
The pattern repeats across the industry. OpenAI shut down its Sora video app in March 2026 after sustained public pressure, but only after months of abusive content targeting women and problematic recreations of historical figures. Google's own Veo 3 video generator, launched in May 2025 just weeks before the Iran conflict began, was immediately used to create fake missile attack videos with the watermark still visible. Users eventually learned to crop the watermark out.
What provenance actually requires
Hardware is starting to move. Google's Pixel 10 became the first smartphone to build the open C2PA provenance standard directly into its camera, signing photos at the moment of capture. Apple's iPhone 18 Pro, announced this week, adds features for creating cryptographically signed reference images and identifying AI-generated or edited content. Hany Farid, the leading deepfake detection expert, called this a shift from niche forensic concern to mainstream problem requiring mainstream solutions.
But hardware provenance only helps if every link in the chain supports it. An image signed at capture means nothing if the platform hosting it strips the metadata or if the viewer's device cannot verify the signature. Provenance needs to travel with the image to wherever it is seen, and it needs to be mandatory, not optional.
Google's response to this crisis was to remove the feature in a day. That speed is notable, but it raises the question of why the feature shipped without testing it against the highest-stakes use case first. The people who rely on Google Earth are not casual users browsing their hometown. They are investigators documenting airstrikes and human rights abuses. If your guardrails are not ready for that scenario, they are not ready for launch.
The counterfeit satellite images from this conflict are still being screenshotted and shared as evidence in new claims. The authentic record now carries an asterisk it does not deserve. Somewhere in that gap between fabrication and verification, photojournalists are proving their own pictures are real, families are weighing which evacuation warnings to trust, and fact-checkers are running behind a lie that a product launch made deniable in an instant. Google says the feature will return with stronger guardrails. The better realization would be that some capabilities should not return at all.