Forum

When Image Editing Becomes Conversation, Watermarks Stop Being Locks

Raindrops covering a window pane, with a gray sky and blurred rooftops behind the glass

On 12 March, Google opened native image generation in Gemini 2.0 Flash to developers through an experimental release in AI Studio and the Gemini API. The headline feature is not a new style or a higher resolution. It is a way of working: you show the model an image and ask for changes in plain language, then keep talking until it looks right.

Remove the chair. Make the light later in the afternoon. Keep the character, change the jacket. Each instruction builds on the last, and the model carries the context forward. Google’s own announcement emphasizes multi-turn editing and keeping characters and settings consistent across a sequence of images.

It took a few days for users to find the uncomfortable use case. As TechCrunch reported this week, people on social media showed the model removing watermarks from photographs, including images from stock agencies such as Getty Images, and filling in the gaps convincingly. The feature is labeled experimental and sits in developer tools, not in the consumer app, and it struggles with some semi-transparent or very large marks. But the point was made.

The lock that was never a lock

Visible watermarks have always been a deterrent rather than a barrier. Skilled retouchers could remove them before generative AI existed; content-aware fill made it faster years ago. What changes now is the effort required. When the instruction is a single sentence, the deterrent mostly disappears.

That does not change the law. Removing a copyright owner’s watermark without permission remains unlawful in the United States outside narrow exceptions, and using the result without a license is still infringement. What it changes is the practical calculation for photographers, illustrators and agencies who relied on a visible mark to keep casual copying in check.

Where protection is moving

If visible marks are weak, protection has to live somewhere less visible. Three layers are worth understanding.

  • Invisible watermarks. Google says images made with this model carry SynthID, its imperceptible watermark for AI-generated content. That marks what the machine made; it does not protect what a photographer made.
  • Provenance metadata. Content Credentials, based on the C2PA standard, record how a file was created and edited. Metadata can be stripped, but when it is present it gives clients and platforms something verifiable.
  • Contracts and licensing records. The least glamorous layer is the most reliable. Clear licenses, delivery records and registration where it is available remain the tools that hold up when a dispute starts.

None of these is complete on its own. Together they move protection from the surface of the image to the record around it.

What conversational editing means for craft

The watermark story overshadows a more interesting shift. Conversational editing changes who can do sophisticated retouching and how that work is described.

For years, image editing was a craft of tools: layers, masks, brushes, curves. Skill was visible in the handling. When editing becomes conversation, the craft moves to judgment: knowing what to ask for, noticing what the model got subtly wrong, deciding when an image has stopped being a photograph and become an illustration.

That last point matters most. A photograph edited through ten rounds of natural-language instructions may be accurate in mood and false in fact. Photographers, editors and designers need clear language for this, both for themselves and for the people who see their work.

A practical note for creators

If you publish images online, assume that a visible watermark slows down only the least motivated copier. Keep original files and licensing records organized. Consider attaching Content Credentials where your tools support them. And when you use conversational editing yourself, note what you changed. The same tool that removes someone else’s mark can quietly remove the evidence of your own process. Competitions increasingly ask for that evidence: the entry rules of the International AI Award, for example, ask entrants to explain how AI was used in their work.

The bigger picture

Every capability in image AI arrives with a misuse attached. The response that ages best is not panic, and not dismissal, but a clearer account of what an image is and how it was made. Conversational editing will spread into consumer apps soon enough. The habits we set while it is still experimental will decide whether it makes images easier to trust or harder.