← Back to Blog

How to Tell If Something Was Written by AI

By Pic-Wiki Content Team ·

The announcement that started this

Since August 2, 2026, AI companies serving the European market have been required to mark AI-generated text in some detectable way. This comes from the EU AI Act's Code of Practice on Transparency of AI-Generated Content, which Anthropic, Google, and roughly 190 other signatories signed onto in July 2026.

Anthropic's response was to start embedding an invisible watermark directly into everything Claude writes. Not hidden characters, not extra tokens, nothing you can see or copy-paste around. A statistical pattern, woven into the words themselves, that only shows up if you know how to look for it. Anthropic published the mechanics, including a Monopoly-and-pi analogy for how the watermark actually works, in How Claude's text watermark works — worth reading directly.

The rollout covers every Claude model launched on or after August 2, 2026, applied globally rather than just in Europe. Older models are being added over the following months.

Where this idea actually came from

Anthropic didn't invent this from scratch. The method, called SynthID-Text, comes from a paper published by Google DeepMind researchers (Dathathri et al.) in Nature in October 2024: "Scalable watermarking for identifying large language model outputs."

The paper's real contribution was making watermarking production-ready for the first time. Earlier watermarking ideas existed, but none had been deployed at the scale of a live product. DeepMind's version changes only how the model samples the next word, not how it was trained, so it adds almost no extra computing cost or delay. In their testing, watermarked text scored no worse than unwatermarked text on quality. DeepMind also has a shorter, plainer writeup on their SynthID announcement post.

The paper was also upfront about the limits, ones that still apply today: the signal weakens after paraphrasing or translation, and it barely shows up on very short or highly factual text.

Key tips for judging AI-generated content

Read for the tells before reaching for a tool.

  • The "it's not X, it's Y" construction, used more than once or twice
  • Stock transitions: "moreover," "furthermore," "in today's fast-paced world"
  • Suspiciously even sentence rhythm across a whole piece
  • Specific-sounding but unsourced claims ("studies show...")
  • No real anecdote — no exact quote, odd detail, or specific date that couldn't have been guessed
  • Heavy em-dash use as a substitute for a comma or period

Read it out loud. Human writing has quirks that sound natural spoken but strange on paper. AI writing tends to read too smoothly — optimized for skimming, not speaking.

Treat watermark and detector results as evidence, not a verdict. A match tells you the AI could have generated the words in front of you. It doesn't tell you who wrote the final version, and there's no public tool yet to check every provider's watermark on demand — Anthropic has said a detection API is coming but hasn't shipped one.

Check images and files separately. Many AI tools attach Content Credentials, a signed metadata note based on the C2PA standard, to generated images and documents. Verify any file at contentcredentials.org/verify — though a missing credential proves nothing, since screenshots and re-uploads strip metadata automatically.

When it actually matters, ask for process, not a score. Drafts, edit history, or someone's ability to explain their own reasoning on the spot are stronger evidence than any single tool's output.

Useful links