Can ChatGPT Remove Watermarks? Here’s What Actually Works

It’s one of the most common things people type into ChatGPT: “remove the watermark from this image.” The short answer is no — not reliably, and for the one watermark that matters most (Google’s invisible SynthID) it can’t do anything at all. Here’s why, and what actually works.

Why ChatGPT can’t remove watermarks

• It often refuses. Removing watermarks brushes against usage policies, so ChatGPT will frequently decline or hedge.
• It doesn’t edit your pixels precisely. When it does “edit” an image it re-generates a new one — so you get a different picture, not your original with the mark cleanly lifted.
• It can’t touch metadata. The C2PA “Made with AI” tags live in the file’s metadata, which a chat model doesn’t read or rewrite.
• It can’t remove SynthID. Google’s invisible SynthID watermark is woven into the pixels and survives ordinary editing — a language model has no way to strip it.

Ironically, ChatGPT’s own images are watermarked

Images you generate with ChatGPT/DALL·E carry C2PA metadata that flags them as AI-made — see cleaning ChatGPT image metadata. So the model that can’t remove watermarks is busy adding them.

What actually works

• Visible logos (like the Gemini sparkle) — lifted with a purpose-built remover, not a chat model.
• Metadata / “AI” tags — stripped losslessly from the file.
• Invisible SynthID — only removed by regenerating the image on a GPU (a look-alike), which is a dedicated process, not a prompt.