Last week,
Google introduced a feature in Google Earth that allowed users to generate artificial intelligence (AI) images directly from the platform’s satellite, aerial and 3D imagery. Within a day, it rolled back the feature after researchers demonstrated how it could create convincing but fictional images of real places, including military facilities, refugee movements and disaster zones.
Google said it had seen people sharing screenshots of AI-generated images that appeared to violate its policies and would reintroduce the feature only after implementing stronger safeguards.
At first glance, the reversal appeared to be another example of a technology company moving quickly and then stepping back after public criticism. However, the incident raises a much larger question.
“The real concern here is about trust, not any single law,” Akshayy S Nanda, Partner in the competition law practice at Saraf and Partners, told Business Standard. “The problem starts when a platform built to show the real world starts generating realistic scenes of events that never happened. That is not just about one fake image. It puts a question mark over everything else the platform shows.”
The Google Earth rollback may be the immediate trigger, but the questions it raises extend far beyond satellite imagery.
As AI is increasingly embedded in maps, legal documents, medical records, closed-circuit television analysis, digital archives and verification platforms, the debate is shifting from what AI can create to whether digital records themselves can continue to serve as trusted evidence.
Vijay Gopalakrishnan, Partner at Deloitte India, said organisations should evaluate AI deployments according to their business impact and implement appropriate safeguards, governance and human oversight, particularly where inaccurate outputs could affect business, compliance, privacy or public trust.
How Google Earth’s AI image feature worked
Google introduced the feature as an integration between Google Earth and its Nano Banana 2 image-generation model.
Users could zoom into a location on Google Earth’s website, select “Create image” and describe what they wanted to see. The AI would then generate a customised image using the platform’s satellite, aerial and 3D imagery as its starting point.
According to Google, the feature was intended to help users visualise ideas rather than document reality.
The company showcased examples such as recreating ancient Pompeii for classrooms, generating infographics about landmarks, previewing real estate developments and helping architects visualise projects before construction.
Within hours of the feature becoming available, open-source investigators began testing its limits.
As reported by Ars Technica, Dutch investigator Henk van Ess used the tool to generate fabricated images showing refugees near the Mexican border, a nuclear facility in Iran, a fatal road accident in Amsterdam and a hospital with a bomb crater in Gaza.
Investigative organisation Bellingcat also questioned why a platform frequently used to verify images and videos would allow users to alter satellite imagery using AI.
Why the controversy extends beyond Google Earth
The controversy has attracted attention because Google Earth occupies a unique position online.
Unlike image generators built primarily for creative expression, mapping platforms are widely treated as factual records.
Satellite imagery is routinely used in journalism, environmental monitoring, humanitarian investigations, insurance assessments, legal disputes and open-source intelligence investigations. In many cases, it serves as supporting evidence rather than illustrative content.
According to Nanda, that distinction fundamentally changes how generative AI should be deployed.
“The problem starts when a platform built to show the real world starts generating realistic scenes of events that never happened,” he said.
“Journalists, human rights investigators and lawyers who use satellite imagery to check things like environmental damage, new construction or conflict zones now have to ask whether any given image might be fake. That makes verification slower and harder, and it even makes it harder to trust the real images.”
He added that generative AI still has a place in mapping products, but only if it remains clearly separated from the authoritative map itself.
“A generative tool built for architects or urban planners, and clearly marked as different from the real map view, is far less risky than one that adds AI-generated content straight onto imagery that millions of people treat as an accurate picture of reality.”
The concern is not confined to satellite imagery.
Similar questions arise wherever AI is introduced into products whose value depends on maintaining an accurate record, including medical imaging, court documents, CCTV systems, digital archives and government databases.
Can AI watermarking preserve trust?
Google relied on its SynthID watermarking technology as one of the safeguards for the feature.
Unlike metadata, which may disappear when files are edited or shared, SynthID embeds an invisible watermark directly into the pixels of AI-generated images. Google says the technology is designed to remain detectable after common modifications such as compression and resizing.
However, researchers argue that watermarking addresses only part of the problem.
Online content is frequently reshared through screenshots, screen recordings or recompressed images. Provenance information may become harder to preserve as content moves beyond the platform where it was created.
Independent testing has suggested that SynthID can survive several common edits and screenshots, although sufficiently aggressive cropping can weaken or remove detection.
Google has also acknowledged that SynthID is intended to support transparency rather than provide a complete solution to AI-generated misinformation.
Nanda agreed that watermarking has value but should not be treated as a complete safeguard.
“Watermarking helps, but it is not a full legal fix on its own,” he said. “Watermarking shows good faith and can lower legal exposure, but real protection comes from a layered system that is actively maintained, not from any single technical fix.”
Regulation has not caught up with embedded AI
The Google Earth episode also highlights a regulatory challenge.
“The law is starting to draw a line between content that is artificially generated and content that is expected to represent reality, but regulation has not fully caught up with how these technologies are actually built into products today,” Nanda said.
He noted that most existing rules were designed with standalone AI-generated images, videos and audio in mind rather than generative features embedded in platforms that users already regard as authoritative.
“In these cases, the product itself lends credibility to the AI output, so users are more likely to mistake synthetic content for the real thing,” he said.
He also said current rules provide little guidance on how provenance information should remain attached after content is exported, shared or converted into screenshots.
Rather than introducing entirely new legislation, Nanda said existing transparency, labelling and due-diligence requirements could be strengthened for products that function as evidence or factual records.
Should high-trust systems face stricter AI rules?
The enterprise challenge extends beyond legal compliance.
AI capabilities have become a competitive differentiator across enterprise software, encouraging organisations to introduce generative features across a growing range of products.
Gopalakrishnan said AI adoption should not be driven merely by the availability of the technology.
“In most cases, thorough due diligence and thought is put into whether AI should be incorporated into a product,” he said.
Organisations should also evaluate the “technology stack, input data handling, guardrails around design, deployment and monitoring of AI solution’s performance and costs, and ensuring security and ethical AI”.
Where that process is missing, he said, organisations could face compliance, governance and business risks.
He added that enterprises should assess every AI deployment according to its potential impact rather than applying a common standard across all products.
“A use case by use case analysis in terms of impact, especially irrevocable ones, to business, brand and safety does need to be done,” Gopalakrishnan said.
He cited AI-generated social media content, commercial documents such as invoices and quotations, systems processing personal information, and applications affecting safety as examples where stronger governance and human oversight become essential.
Rather than treating every AI feature equally, enterprises may increasingly need to identify “high-trust systems” where preserving authenticity is as important as adding intelligence.
For such products, the Google Earth episode suggests that the central question may no longer be whether AI can generate convincing content, but whether users can continue to trust what they are seeing.
 "Google Earth's AI image rollback exposes trust crisis for digital evidence")
