Transparency is becoming a toggle
Google will now allow users to remove visible watermarks from its AI generations. Meta’s ‘open’ AI pitch and OpenAI’s health rollout add the same tension from other angles: AI companies want the trust benefits of openness and labelling, but they also want control over how those signals appear in the product.
TechCrunch
Google will now allow users to remove visible watermark from its AI generations
Google will now allow users to remove visible watermarks from its AI generations.
techcrunch.com

A watermark used to be the little confession stamped on the image: this was made by a machine. Google now wants that confession to be optional.
TechCrunch reported that Google will let users remove visible watermarks from AI-generated media, while keeping invisible watermarking and provenance metadata in place. That sounds like a reasonable product compromise. Creators get cleaner assets. Platforms keep some trace of origin. Everyone gets to pretend the trust problem has been handled.
I think the real story is sharper: transparency is becoming a product setting.
The old promise around AI labelling was moral clarity. Synthetic content would be marked. Users would know what they were seeing. Employers and teachers could set rules. Platforms could enforce policy. But the product reality is moving somewhere messier. Visibility is now split into layers: what the end user sees, what the platform can inspect, what the model provider can prove later, and what the customer would rather hide.
That split changes the politics of AI trust.
The label is no longer the system
Visible watermarks are socially useful because they are awkward. They interrupt the illusion. They tell the viewer, not some backend provenance service, that the media came from an AI system. Removing them may make sense for legitimate creative work, but it also moves the signal from public space into infrastructure controlled by the company.
That matters because invisible provenance is only useful to whoever has access to read it, verify it, or enforce it. A visible mark gives ordinary people a cue. An invisible mark gives institutions a tool. Those are different trust models.
This is why the anxiety around AI use in jobs and classes feels so combustible. People do not only worry that AI content exists. They worry about who gets to detect it, who gets punished for it, and whether the rules are clear before the detection happens. A watermark can be a consumer feature, a compliance mechanism, or an accusation machine depending on where it appears.
Google’s move is rational product management. It is also a sign that AI companies are learning to separate provenance from presentation. The former protects the platform. The latter affects adoption.
Openness has the same problem
Meta is running into a parallel tension from the other direction. TechCrunch’s Equity team framed the question bluntly: does Mark Zuckerberg really believe AI is “for everyone”? The word “open” does a lot of work here. It signals public benefit, developer freedom, and a challenge to closed labs. It also leaves room for strategic control.
OpenAI’s health push shows the same pattern in a more sensitive category. Making ChatGPT Health available to all US users expands access, but it also raises the product question every AI company now faces: where does user control end and provider responsibility begin?
That is the same pattern as the watermark story. The public signal says: trust us, this is open, labelled, or broadly available. The product architecture says: the important control points remain configurable.
TechCrunch also paired Meta’s open-AI positioning with a $250M deal dispute, which is a useful reminder that the rhetoric of access sits inside a very commercial machine. Open systems, APIs, provenance metadata, visible labels: these are not pure philosophical positions. They are market instruments.
There is a decent parallel in food labelling. “Organic”, “free range”, and nutrition panels were meant to inform consumers, but over time they became branding systems, regulatory artefacts, and pricing tools. The label still matters, but the fight moves to who defines it, who audits it, and how prominently it appears on the packet.
AI is heading there fast. Builders should assume users will want control over how AI involvement is displayed, while institutions will want reliable ways to detect and govern it. Those desires will conflict inside the same product.
The next trust war in AI will not be about whether systems disclose. It will be about who gets the toggle.
Read the original on TechCrunch
techcrunch.com