AI Policy

Google Marked Every Image and Pulled It in a Day

Google added Nano Banana image generation to Google Earth on 30 July and withdrew it on 31 July. Every image carried a SynthID mark. It did not matter.

AI PolicyBy Servola Tech Desk2026-07-315 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Google Marked Every Image and Pulled It in a Day

Key takeaways

Google watermarked every image its Google Earth feature produced and withdrew the feature in under 24 hours anyway; from 2 August the EU AI Act makes that same machine-readable marking the legal test under Article 50, so a system can satisfy the regulation and still be too damaging to keep running.
Article 50 applies from 2 August 2026 and infringement is priced at up to 15 million euros or 3 percent of worldwide annual turnover; generative systems already on the market on that date have until 2 December 2026 to meet the machine-readable marking duty.
If you buy a generative tool you are almost certainly a deployer rather than a provider, so the machine-readable mark is your supplier's obligation while yours is a disclosure a person can actually see, covering deepfakes and published text on matters of public interest.
Google named the trigger for the rollback as people sharing screenshots, which is the one distribution path where a machine-readable mark has no reader; test every synthetic asset you publish by assuming it will be screenshotted and reposted with nothing left but the pixels.

A Thursday launch, withdrawn by Friday

Joseph Cox of 404 Media typed a request into Google Earth and produced a blast crater in Los Angeles. Another prompt put a crowd of protesters outside Google's own campus in Mountain View. A writer at PC Gamer turned their hometown into a warzone. None of it had happened, and all of it sat on top of real satellite imagery of real places.

The feature was one day old. Google announced on 30 July 2026 that Nano Banana, its image generation model, could now create imagery grounded in Google Earth's satellite, aerial and 3D views. By 31 July the company had rolled it back and appended an update to its own announcement post.

The statement explaining the reversal reads: "We've seen geospatial professionals using this feature for a range of useful purposes, however we've also seen people sharing screenshots of generated imagery that appear to violate our policies. We're rolling back this feature in Google Earth while we work on implementing stronger guardrails." The trigger named in that sentence is worth holding on to. It was screenshots.

Every image carried the mark

This was not an unlabelled system. Every image the feature produced carried SynthID, Google's machine-readable watermark. Google's own update states that the generated images were watermarked as AI generated and that they did not appear in the main Google Earth experience where other people would come across them.

When first challenged, Google pointed at precisely that. Anyone could check an image using Gemini or Google Lens, and the company said it prevents image creation on harmful topics and updates its protections continually. On the technical claim Google was correct. The images were marked, the marking was detectable, and a route to verification existed for anyone who wanted one.

Cox's response was that this misreads how misinformation actually moves. People do not run the images they scroll past through a verification tool, and a caption travels faster than a provenance check. The mark was present in every file and consulted in almost none of them.

Brussels makes that mark the test on Sunday

Article 50 of the EU AI Act applies from 2 August 2026, which this year falls on a Sunday. It requires providers of AI systems, including general purpose systems, that generate synthetic audio, image, video or text to ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The duty is not restricted to high-risk systems, which is exactly why it reaches ordinary businesses.

The European Commission's guidance describes the required marks as effective, reliable, robust and interoperable, and sets no technical specification beyond that language. Exemptions cover assistive editing, source code, short sequences of symbols, machine-to-machine output and artistic, satirical or fictional work. Infringement is priced at up to 15 million euros or 3 percent of worldwide annual turnover.

Generative systems already placed on the market on 2 August have until 2 December 2026 to meet the machine-readable marking duty in Article 50(2), a transition agreed in the AI Omnibus text of May 2026. Content generated before 2 August does not have to be labelled retrospectively. The United Kingdom has no equivalent statutory marking duty, so a British business is reached by this rule through what it publishes into the European market rather than through anything at home.

Compliance and safety turned out to be different questions

Run the two facts together and the result is uncomfortable. Google marked every output. Measured against the standard that binds providers from Sunday, the system passed. Google then withdrew it in under a day, because correctly marked images in circulation were doing the damage regardless.

A machine-readable mark is read by a machine that has been asked to look. In a screenshot pasted into a group chat, a social post or a slide deck, nothing asks. The mark survives as a property of a file that nobody interrogates, which makes it an excellent audit artefact and a weak defence. Article 50 sets the auditable version as the requirement, and Google has just demonstrated at global scale that meeting it is not the same as being safe to ship.

This is not an argument against the rule. Marking is a floor, and a floor is worth having, because without it there is no way to establish provenance even when someone does go looking. It is an argument against treating an Article 50 project as the end of the work. The largest company to deploy this technique in a consumer product satisfied the floor and still reversed itself inside 24 hours.

Three things to settle before Sunday

First, work out which side of the rule you are on. If you buy a generative tool, you are almost certainly a deployer, and the machine-readable marking duty sits with the provider. Your own obligation is different and more visible: disclose deepfakes, and disclose AI-generated or manipulated text published on matters of public interest where no person has reviewed it. That is a disclosure a reader sees, not a signal buried in a file.

Second, list every synthetic asset that leaves your organisation. Product renders, marketing imagery, training material, generated voice on a phone line, drafted copy on a public site. For each one, name who applies the mark, and get that in writing from the supplier rather than assuming the model handles it.

Third, run Google's test on your own output. Assume the asset is screenshotted, stripped of its metadata and reposted with a caption you did not write. If your answer to that scenario is that the file contains a watermark, you are holding the position Google held on Thursday and abandoned on Friday.

Frequently asked questions

Does the EU AI Act ban a feature like this?

No. Article 50 is a transparency rule rather than a prohibition. It requires providers to mark synthetic output in a machine-readable form and requires deployers to disclose deepfakes and certain published text. The Google Earth feature would have satisfied the marking requirement. Google withdrew it on its own policy judgement, which is the part worth carrying: the legal floor and the decision to ship are separate calls.

We only use AI images for marketing. Does any of this reach us?

The marking duty falls on the provider of the system you use, so in practice it reaches you through your supplier contract rather than directly. The deployer duty in Article 50(4) is narrower than most marketing teams assume, covering deepfakes and text published on matters of public interest without human review. Ordinary product imagery generally sits outside it. If a supplier fails the marking duty, what you inherit is the commercial problem rather than the fine.

What is the deadline that actually applies to us?

2 August 2026 for the obligations generally. For generative systems already on the market before that date, the machine-readable marking duty under Article 50(2) runs to 2 December 2026 under the AI Omnibus transition agreed in May 2026. Content generated before 2 August does not need retrospective labelling.

Google did the marked, documented, verifiable thing, and it took one day for that to be beside the point. The rule that starts on Sunday asks for exactly what Google had already done. Treat it as the beginning of the question rather than the answer to it.

Google EarthEU AI ActArticle 50SynthIDSynthetic mediaNano Banana

1,134 Lab Staff Say The Brake Does Not Exist

Chief scientists at OpenAI, Anthropic, Meta and Google signed one statement asking Washington to build pacing tools. Their argument is that no such tool exists.

6 min read

AI Policy

December 2 Binds Whoever Runs The Model

Regulation (EU) 2026/1744 took force on 27 July. An audit published the next day found output moderation on 3 percent of Hugging Face Spaces. The duty reaches deployers.

6 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles