What Sparked the Sudden Rollback?
Last week, Google introduced a feature integrating its Nano Banana 2 image-generation model with Google Earth, allowing users to create photorealistic images based on the platform's satellite data. The idea was to let users visualise creative projects,
like new gardens or historical scenes. However, within a day, the company pulled the feature. The reason? Users immediately began using the tool to generate convincing but fake images of disasters, military activity, and other sensitive scenarios, from explosions in Paris to false terror attacks. Despite the images being watermarked, researchers and journalists expressed alarm, arguing the tool could erode public trust and be weaponised to create and spread disinformation with unprecedented ease.
A Predictable Problem
For many observers, the speed with which the tool was misused was not a surprise. The incident highlights a core dilemma for tech companies: how to build guardrails strong enough to prevent malicious use without stifling innovation. Google stated it was rolling back the feature to implement stronger safeguards, noting that people uniquely trust Google Earth as a reliable view of the world. But critics pointed out that the potential for misuse seemed obvious, questioning how the feature was released without more robust protections in the first place. The event underscores a recurring pattern in the AI race, where powerful tools are often released before the ethical and safety frameworks can fully catch up, placing the burden of identifying misuse on the public.
The Great Labelling Debate
This rollback puts the broader challenge of identifying and labelling AI-generated visuals into sharp focus. The two main approaches are visible watermarks and invisible metadata. Visible watermarks are easy for humans to see but can be cropped or edited out. Invisible methods, like Google's SynthID watermark and the industry-wide C2PA standard, embed a cryptographic signature into the file's data. This 'Content Credential' acts like a nutrition label, showing who made the image and what tools were used. Major players like Adobe, Microsoft, and OpenAI are all part of the C2PA coalition, aiming to create a traceable chain of authenticity for digital content.
Why Can't We Just Spot the Fakes?
If standards exist, why is this still so hard? The first problem is that these systems are opt-in. A company or creator has to choose to add the C2PA label. Secondly, metadata can be stripped, either intentionally or accidentally, when an image is screenshotted, compressed, or uploaded to a platform that doesn't preserve the data. This has led to a constant cat-and-mouse game. Detection tools are getting better, with some claiming up to 99% accuracy in ideal conditions, but AI generation models are also evolving rapidly. As soon as a detector learns to spot the tell-tale signs of one AI model, a new model is released with fewer of those giveaways. This makes 100% reliable detection a constantly moving target.
The Human Element
Ultimately, technology alone cannot solve the problem. Studies have shown that humans are not very good at distinguishing AI images from real photos, with one experiment showing an accuracy rate of just 62%—only slightly better than chance. People were particularly bad at identifying fake landscapes, the very type of image the Google Earth tool was designed to create. This suggests that even with labels, building societal resilience to disinformation requires a massive uplift in digital literacy. We need to shift from blindly trusting what we see to questioning the origin of content by default. The Google Earth incident was a warning shot, demonstrating that the integrity of our digital information ecosystem depends on more than just code.
