Was This Image or Paragraph Made by AI? In 2026 the Watermark Is Already Inside — the Checking Tool Is What You Cannot Get

“Is this image AI-generated?” “Did a chatbot write this essay?” Almost everyone has had to ask by now.

The good news: AI companies have been marking their output for a while. The bad news: the tools that read those marks are mostly not available to the public yet.

Here is the honest state of play as of September 2026.

1. This is not the logo in the corner

When people hear “watermark” they picture a translucent logo on a photo. That is not what this is.

Google’s SynthID hides a marker inside the pixel or audio data itself — invisible and inaudible, but detectable with the right tool. Text works differently again; see section 3.

Google states SynthID currently covers four content types:

2. Does it survive cropping and compression?

The practical question. Google’s stated position is that the watermark is designed to withstand cropping, filters, frame-rate changes and lossy compression. For audio, it says the mark survives added noise, MP3 compression and speed changes.

Note the wording: “designed to withstand” is not a guarantee of survival. This article will not upgrade that into “always detectable” — no such guarantee has been published.

3. How do you watermark text? Anthropic explains it best

Text has no pixels to hide data in, so what is the trick? Anthropic published Claude’s approach on 2026-08-14 (the page was updated on 2026-09-01 with detection API details). In plain terms:

When a model writes, many choices are genuinely interchangeable — after “the weather was cold and…”, either “grey” or “overcast” reads fine. The watermark uses a secret key to steer those low-stakes choices. The text reads perfectly naturally, but a party holding the key can verify that Claude produced it.

Anthropic is also specific about where the effect weakens, which matters:

• Code: because code demands precision, watermarking has limited room; comments are where it may appear.
• Light editing of your own writing: the effect is minimal when Claude only revises text you wrote.
• Translation: every word is chosen by Claude, so translations do carry it.
• Image files: a different mechanism — a C2PA content credential in the file’s metadata, not embedded in the picture itself.

Can rewriting defeat it? The stated position: light edits cannot fully remove it; the watermark disappears only when every word is replaced.

4. The gap that matters: you cannot check it yourself

This is the part to remember.

• On Google’s side: there is a SynthID Detector portal for uploading images, video or audio. But Google states it is currently being tested with journalists and media professionals, with an early-tester waitlist to apply to. The route that is open now: upload the file in the Gemini app and ask whether it was generated or edited by Google AI.
• On Anthropic’s side: the detection API is in private beta, open to “qualified organisations” under EU rules — the stated list includes regulators, law enforcement, media, fact-checking organisations, independent researchers, educational institutions and EU civil-society groups. Enterprises with compliance needs may also apply. Anthropic says access will widen over time.

Plainly: if you suspect a colleague’s report was AI-written, there is no official button for you to press today.

5. So what can you actually do?

1. For images, video and audio: try the Gemini app upload-and-ask route. It is the method Google names, and it needs no application.
2. For text: stop trusting the percentage scores from consumer “AI detectors.” Those tools have nothing to do with either official watermark, and false accusations are common. No product is recommended here.
3. If you are the one producing content: the watermark works in your favour too — it evidences both “this was AI-assisted” and “this was not forged in your name.” Claude’s C2PA credentials on image files follow that logic.
4. For anything that matters, fix it with process, not detection: ask for drafts, version history, and a verbal walkthrough. All more reliable than any detector.

6. What this article does not claim

• When SynthID Detector opens to the general public. Google states only that testing is with journalists and media professionals; no timeline is published — unverified.
• Accuracy or false-positive rates for either watermark. Neither official page publishes numbers, and no third-party test figures are cited here.
• What other AI companies (OpenAI and others) currently do. Not checked this round, so not folded into the conclusion.
• The conditions under which a watermark fails. Official pages describe what the design withstands, not where it breaks — unverified.

Official links

• Google DeepMind SynthID (source for coverage and verification routes): https://deepmind.google/science/synthid/
• Anthropic, how Claude’s text watermark works (source for the text mechanism and detection access list): https://www.anthropic.com/news/claude-text-watermark

Related reading

• Gemini tool page (plans and free-tier verification log)
• Claude tool page (plans and free-tier verification log)
• Which AI video “price cuts” are not real cuts