What is actually detectable, as of 11 August 2026
The useful question is not "can AI writing be detected". It is "can the specific person marking my work run a specific check today, and what would a hit prove". Those have different answers, and several of them changed in the last two weeks.
| The check | Can your marker run it today | What a hit would actually prove |
|---|---|---|
| A statistical AI detector GPTZero and the rest |
Yes. Anyone can paste text into a free one | Very little alone. These measure how predictable text is, so plain, careful writing scores high. A 2023 Stanford study found detectors misclassified more than half of essays by non-native English speakers |
| Turnitin's AI indicator | Only if the institution licenses it and left it switched on, which is not universal. Vanderbilt publicly disabled it in 2023 | Turnitin's own published position is a document-level false positive rate under 1%, caveated to documents where it reports over 20% AI writing, with an asterisk marking scores below that as less reliable. The company says the report should not by itself be used to penalise a student |
| Claude's text watermark | No, not yet. Anthropic started marking with EU launches on or after 2 August 2026 and says detection is covered in "forthcoming technical documentation" | When it arrives, that text "may have been processed by Claude". Which includes Claude having only edited writing that was entirely yours |
| Gemini's SynthID text watermark | No. The SynthID Detector portal is in limited testing behind an early-tester waitlist | The same. A watermark says a model touched the text, not who wrote the argument |
| C2PA credentials on an image | Partly. The public verify page reads them when they survived | Only that a signed record made it through. A screenshot destroys it, so absence proves nothing at all |
| Invisible characters and file metadata | Yes, trivially, and this is the one people forget | That a passage was pasted from somewhere. Not from where |
| Your document's version history | Yes, if you submitted a shared file or are asked for it | A drafting trail. This is the check most likely to help you, which is the whole argument for keeping one |
| Whether your sources exist | Yes. Free, no licence, about a minute per source, by anyone including your marker | Everything. A reference that resolves to nothing is not a probability score, it is a fact, and it is what actually ends up in front of a panel |
Read down the middle column. Every check that runs reliably today is either about your sources or about your drafting trail. Every check that is about your prose is either statistical and contested, or not yet available to the person grading you. That is not a loophole, it is a planning fact: the effort you spend worrying about the third column is misallocated, and the effort you spend on the last row is not.
What changed on 2 August 2026
Two things landed on the same day and they are connected. Article 50 of the EU AI Act, the transparency article, applies from 2 August 2026. It obliges providers to "add machine-readable marks to enable the detection of AI-generated or manipulated content", and obliges deployers to disclose deepfakes and AI-generated "text publications on matters of public interest without human review or editorial control".
Anthropic's response is visible in its own documentation: Claude models launched in the EU on or after that date "will support machine-readable marking at launch", using an imperceptible text watermark that "may persist through some editing" plus signed C2PA metadata on supported files.
Three things that law does not do, and it is worth being precise because a lot of coverage was not. It does not oblige you to disclose AI use in an essay; that is your university's rule, not Brussels'. It does not hand your marker a detector, and as of today none of the marking schemes above has a public one. And a machine-readable mark is not an authorship finding: Anthropic says plainly that a detected mark means the content "may have been processed by Claude", which is also true of a paragraph you wrote and asked it to tighten.
Where each guide fits
- The rule you are actually under. Is using AI cheating is the map of where institutions draw the line, which is the only line that binds you. Can I reuse my own paper covers the integrity rule that has nothing to do with AI and catches more students than any of the above.
- The detectors. How AI detection really works, what Turnitin can and cannot do, and what your AI score means, which is less than either side of the argument tends to claim. How accurate is GPTZero takes the free one students meet first and separates its benchmark number from your essay.
- What the repository keeps. Is my paper stored in Turnitin forever covers the setting your instructor chose and what removal actually involves.
- The watermarks. Does AI text carry one, how to inspect a document yourself, and what happens to an image credential when a figure is screenshotted into your paper.
- The shortcuts. Do humanizers work and what actually gives an essay away, which is the fabricated citations and the missing trail rather than the style. Is using QuillBot cheating draws the line for the tool most students reach for before any of these.
- If it has already happened. Accused of using AI covers what evidence helps, and why the right move depends on what really happened. I accidentally plagiarized is the same hour from the other direction, when you found it before anyone else did.
One thing this site will not help with, and says so on the front page: getting AI writing past a detector. Not because of the detectors, which are unreliable in both directions, but because the workflow that survives every check in the table above is the same workflow that produces work you can defend in a five-minute conversation. There is no second method.