Google pulls its AI satellite deepfake tool after a single day

The News

Google has switched off an artificial-intelligence feature in Google Earth barely a day after it went live, following demonstrations that the tool could fabricate convincing satellite imagery from a single line of text.

The feature, powered by Google's Nano Banana 2 image model, let users rewrite aerial, satellite and 3D views by typing prompts. It launched on Thursday and was rolled back within roughly a day once researchers showed how easily it manufactured scenes that never existed.

Henk van Ess of the investigations outlet Digital Digging produced images that inserted refugees near the Mexican border and a bomb crater beside a hospital in Gaza, using nothing more than text instructions layered over real-world imagery. Google initially defended the tool, saying it takes misinformation seriously and that every picture generated with Nano Banana in Google Earth carried a SynthID digital watermark. By the following day the company had removed the feature entirely.

Why It Matters

The episode is a compact lesson in how quickly generative tools can outrun the guardrails wrapped around them. A watermark such as SynthID may satisfy a compliance checklist, but it does little to stop a screenshot spreading across social platforms where provenance signals are stripped away in seconds.

Mapping carries a particular authority. People instinctively trust an overhead image as a record of the real world in a way they no longer trust a portrait or a voice clip. When that trust can be bent by a text box, the surface for manufactured evidence of conflict, migration or disaster widens sharply. The last comparable jolt came in 2024, when fabricated images of public figures forced platforms to scramble for labelling standards. Building the same capability into a map, the reference layer journalists and courts lean on, raises the stakes considerably.

That Google shipped and retracted the feature inside a day also signals how competitive pressure is compressing release cycles. Ship first and moderate later is becoming the default, and the correction is arriving in public rather than in testing.

Indian Angle

For India the story lands on unusually sensitive ground, because satellite imagery here is not just a misinformation problem but a regulated one. The Geospatial Data Guidelines of 2021 liberalised mapping for domestic firms yet kept fine-resolution imagery of border and defence zones tightly controlled. A tool that lets anyone fabricate an overhead view of a frontier area, real or invented, cuts directly against that regime.

It also feeds into MeitY's push on synthetic media. Draft amendments to the IT Rules circulated in 2025 would require platforms to label synthetically generated content prominently, and an easily faked map layer is exactly the kind of output those rules are meant to catch. A SynthID tag buried in metadata is unlikely to meet the visible-labelling bar Indian regulators have signalled they want.

There is a commercial thread too. India's listed mapping champion, MapmyIndia (CE Info Systems), has built its pitch partly on being a sovereign, auditable alternative to Google for a market wary of foreign control over geospatial data. Every stumble of this kind strengthens the argument for home-grown mapping infrastructure that Indian regulators can inspect directly.

FAQ

What exactly did the tool do?

It let Google Earth users edit satellite, aerial and 3D imagery by typing text prompts, using Google's Nano Banana 2 model. The output looked like genuine overhead photography but could depict events and scenes that never actually occurred anywhere on the map.

Why did Google remove it so fast?

Investigators demonstrated that the feature could generate believable fake imagery of sensitive events, such as border crossings and bomb damage near a hospital. Faced with that, Google withdrew the tool within roughly a day of launching it.

Does the SynthID watermark solve the problem?

Only partially. SynthID embeds a signal that can flag an image as AI-made, but watermarks are easily lost when pictures are screenshotted or re-shared, which sharply limits the protection they offer once an image is loose online.

What could this mean for Indian regulation?

It strengthens the case for MeitY's proposed synthetic-media labelling rules and highlights the sensitivity of India's geospatial guidelines, which restrict fine imagery of border and defence areas that a text-to-image tool could fabricate at will.

This story was reported by The Verge. Read the full original coverage at The Verge.