Google shipped a prompt box into the satellite tool investigators are trained on

She runs verification at a newsroom or a human rights unit, and somewhere in her onboarding deck is a line that reads: get the coordinates, open Google Earth, compare, screenshot, file. Open-source investigators are formally trained on that product for conflict work in Syria, Yemen and Sudan, according to Rest of World ^[1]. The line was written for a tool that showed you what a sensor saw. On July 30 it became a tool that would also render what you typed ^[2][3].

Pulling the feature removed the feature, but it did not touch the exports ^[5], because a Google Earth screenshot carries no record of which build of Google Earth produced it, and the scenes users generated in the first hours were photorealistic and blended into real imagery by design ^[3][4]. Anyone reviewing a file next year gets a picture and a date, and now has to know what the product could do on that date.

The thing being pitched and the thing being done are two different products wearing one name. Google pitched a generative layer with stronger guardrails to follow ^[5]. What the field was doing with Google Earth was using it as an archive of record, the closest available stand-in for site access after the February strikes on Iran ^[6]. Same name, two products, and only one of them was ever load-bearing for a court exhibit.

The fallback that regulation leans on is machine detection, and the essay's author reports that colleagues tested detectors across 13 AI providers: seven had no public tool, and of those that did, all but one could be fooled ^[12]. Six providers offered something, five of those six failed, so one detector in thirteen held ^[13]. Meanwhile some EU AI Act transparency obligations only began on August 2, 2026, with the precise rules still a voluntary code ^[14], and California's law does not reach platforms until at least 2027, three years after passage ^[15][21].

Watermarks help only where they exist. Google's own watermarks caught Gemini output being passed off as satellite evidence ^[7], and Veo-marked missile footage was traceable until users learned to crop ^[17][18]. The Evin prison video had no mark, circulated within minutes, and was carried by Sky News and the BBC before any forensic result existed ^[9][10][20]. Citizen Lab later assessed it as part of an AI-enabled influence operation likely run by the Israeli government ^[11]. Chesney and Citron named that payoff the liar's dividend in 2019: doubt wins over belief every time an image can plausibly be fake ^[19].

So the forcing question is two questions per image in the file. Can this artifact be re-obtained from a second imagery holder, independent of Google? And will anyone outside the room rely on it later, in court, in a correction, in a government submission? If yes and no, the screenshot stands once it has a date on it. If yes and yes, the second copy is worth ordering while it can still be ordered, given that Washington has told satellite firms to restrict journalists' access in the region ^[8]. If it cannot be re-obtained and nothing external depends on it, mark it uncorroborated in-house. The cell with no second copy and real downstream reliance is where a platform screenshot stops being evidence and becomes a lead.

Nothing in the promised guardrails reaches the images already exported ^[5], and the cheapest fix for those is clerical: a date on every Google Earth artifact in the file, and a second holder for anything that has to survive somebody else's doubt.