How to tell if text was written by AI (what actually works)
By Claude Watermark Research · Updated
Only one class of signal is genuinely reliable: artifacts the tool physically left in the file, such as the HTML class names Claude's web interface carries onto the clipboard. Everything else — em dashes, tone, sentence rhythm, detector scores — is inference about style, and is wrong often enough that it should never be treated as proof.
Tier 1: artifacts, which are evidence
Copying rendered output from a chat interface carries HTML with it. Claude's paragraphs arrive wrapped in class names like `font-claude-response-body`. If that string is in the markup, something was pasted from Claude's web UI — a human would have to type it deliberately for that to be wrong.
The same tier includes zero-width characters, non-breaking and exotic spaces, and curly punctuation in contexts that never produce it. These are checkable facts about the bytes, not opinions about the prose. Our checker reports exactly which ones are present and where.
Tier 2: statistical detectors, which are estimates
Tools like Turnitin and GPTZero score how predictable the word choices are. Real signal, genuinely useful in aggregate, and unreliable for any individual document. They systematically misfire on non-native English speakers and on plain, uniform prose.
Treat any single score as a prompt to look closer, never as a verdict. None of these tools can show you the evidence behind the number, because there isn't any — only a probability.
Tier 3: vibes, which are not evidence
Em dashes, “delve”, tricolons, relentlessly even paragraph lengths, an upbeat closing summary. These correlate with unedited model output, and plenty of humans write this way — particularly people trained in professional or academic writing.
Accusations built on this tier are how innocent writers get flagged. The signals are also trivially removed by anyone actually trying to hide, which makes them weak in both directions.
What about watermarks
Anthropic announced a statistical watermark for Claude models launched on or after 2 August 2026, and Google already applies SynthID to Gemini text. Both are detectable only with the generating party's secret key.
So watermarks belong in no tier here for practical purposes: no third party can check them, and no shipped Claude model carries one yet.
Questions
- Is there a reliable free AI checker?
- For artifacts, yes — those are deterministic and ours is free and unlimited. For statistical detection, no tool is reliable enough to accuse someone on, free or paid.
- Does asking the AI whether it wrote something work?
- No. Models have no memory of what they generated and will confidently answer either way. This has caused real academic misconduct cases based on nothing.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Open the checker
