How to Detect AI-Generated Text in 2026: Check the Watermark First

Here is the one fact to hold on to before anything else: no detector proves a piece of text was written by AI. A detector score is a probability signal, and a flag is the start of a conversation, not a verdict. That matters more in November 2026 than it did a month ago, because the biggest providers have started marking their own text — so the first check is no longer a detector website at all. This guide explains how to detect AI-generated text in 2026, in order: check for a watermark first, run a free detector second, and read the score honestly third.

Step 1 — Check for a watermark first

A watermark is an invisible mark woven into the words a model chooses. It adds no hidden characters and no extra text, and it carries no information about who wrote the prompt. It can only be read by a detector that knows the secret key — usually the provider’s own. That makes it the strongest signal available, and also the hardest for an ordinary reader to check, because most providers only let approved organisations run the check.

The picture across providers, as of 10 November 2026:

Provider	What is marked	Since	Who can check it
Grok API	Text from seven Grok 4 models	7 November 2026	Developers, via the x-watermark response header; detection API by application
Claude	Model text, SynthID-Text-based	14 August 2026, extended 3 November	Detection API in private preview for eligible organisations
ChatGPT (textGrain)	Text for eligible EU users	Rolling out from 5 October 2026	Approved researchers and expert organisations only
Gemini	Google says SynthID-Text	—	Contested — treat as unconfirmed

Grok’s marking covers the API only. xAI’s docs state the watermark switched on on 7 November for seven models — grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning and grok-4.20-multi-agent — across the Responses and Chat Completions APIs. A response header reading x-watermark: true means the output was watermarked; any other value, or no header, means it was not. Team admins can opt out per model, but must confirm a legal exception applies, and every opt-in and opt-out is recorded in the team audit log. The docs say nothing about the Grok app, grok.com or Grok on X, and xAI has announced no watermarking date for those — so do not assume consumer Grok text is marked.

Claude’s watermarking has been running globally since 14 August 2026, built on Google DeepMind’s SynthID-Text method. Anthropic extended it on 3 November to Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6, with one wrinkle: on Sonnet 4.6, text from a request that sets temperature, top_p or top_k to a non-default value might not carry the mark. The rollout on Amazon Bedrock was due to complete by 10 November. Note that Opus 4.6 is not in the 3 November extension — do not count it as watermarked.

OpenAI’s version, called textGrain, was announced on 5 October 2026. Eligible ChatGPT and Codex users in the EU get it by default over the following weeks; everywhere else it stays off by default, though API customers can switch it on for selected models. There is no public detector for ChatGPT text — access is limited to approved researchers and expert organisations, and checking student work sits outside the programme. OpenAI’s own tests put detection at roughly 80% for 200-token passages and 95% for 400-token passages, at a 1% false-positive target. Paraphrasing collapses it: swapping 10% of words for synonyms cut detection on 400-token passages from about 92% to 66%, and swapping 25% cut it to 17%.

Gemini needs one careful sentence. Google says Gemini text carries a SynthID watermark, but at least one tracker grades that claim as contested for the app and the API, so treat Gemini watermarking as unconfirmed rather than settled.

The rule that governs all of them: a detected mark means the model was “likely involved” — Anthropic’s own wording — and nothing more. No mark proves nothing at all. The text might come from an older model, a provider that does not mark, or a passage edited hard enough to strip the signal. OpenAI puts it the same way: “the absence of a detected watermark does not prove human authorship.”

Why the law is pushing this

The watermark wave is not a voluntary trend. Article 50 of the EU AI Act requires providers of AI systems that generate synthetic text, audio, images or video to mark their outputs in a machine-readable format, so they can be detected as artificially generated. Those transparency duties have applied since 2 August 2026.

Systems already on the market before that date get a transition period — but only until 2 December 2026, which is 22 days from today. That deadline comes from Regulation (EU) 2026/1744, the Digital Omnibus on AI, and it covers the marking requirement only. You will see 2 December 2027 quoted in some commentary; that date belongs to a different set of duties for high-risk systems, not to marking. The December deadline is why the providers moved in a rush through October and November, and why more marking will appear before the year ends.

Step 2 — How to detect AI-generated text with a free detector

If no watermark check is available to you — and for most readers, none is — a free detector is the next step. Every figure in this table was re-checked on the vendor’s own site on 10 November 2026, and anything that could not be confirmed has been left out.

Detector	Free allowance (re-checked 10 Nov 2026)
GPTZero	Scans up to 10,000 characters without an account; longer scans need a free account
Copyleaks	25,000 characters per scan, no login needed
Originality.ai	100 free credits per account; a free scan needs no account
Pangram	20 free checks a day after signing up for the free plan
ZeroGPT	15,000 characters per check
Scribbr	Unlimited 1,200-word checks, no sign-up needed

A few practical notes. Copyleaks offers the largest no-login scan, which makes it the easiest first stop for a long document. Scribbr’s unlimited 1,200-word checks suit short passages, and it needs no account at all. Turnitin has no individual free check — it is sold to schools and institutions, so do not go looking for a free Turnitin scan. One more thing to know before you paste anything in: free scans on some tools may be used for training, so avoid pasting unpublished or sensitive work.

Step 3 — Read the score honestly

Vendor accuracy figures and independent measurements disagree by a wide margin, and that gap is the most useful thing to know about detectors. No single accuracy number belongs to a tool on its own — it depends on the test set, the model that wrote the text, and the threshold the tester chose. Treat every figure below as a claim with a named source, not as a property of the tool.

Detector	What the vendor claims	What independent testing found
GPTZero	99.9% accuracy (vendor’s own testing, per gptzero.me, checked 10 Nov 2026)	RAID benchmark: 95.7% of AI texts detected; false-positive rate disputed — GPTZero reports 1% on RAID, one independent comparison measured 6–8%; Ryne.ai measured about 18% false positives in the real world; Leap AI scored it at F1 0.94
Originality.ai	Up to 97–99% accurate	RAID: 85% average, and 96.7% on paraphrased text; a GPTZero-run benchmark (competitor-run, so read it as such) gave 83.0% overall and 31.7% on GPT-5 text; independent false-positive rate 2–5.7%
Copyleaks	99.1% accuracy, 0.03% false-positive rate	F1 score 0.87; real-world accuracy estimated at 76–91%; about 7% false-positive rate in one independent 2026 test
Turnitin	92–100% accuracy in various studies	Turnitin’s own chief product officer admitted the platform intentionally skips about 15% of AI content to cut false positives, putting real detection near 85%

The strongest independent evidence comes from a June 2026 study in the International Journal for Educational Integrity, which tested four detectors on 160 papers. Pangram came out best. Turnitin, GPTZero and Copyleaks scored roughly 0–14% on fully AI-written papers — a stark result — while all four tools correctly identified the human-written papers. The RAID benchmark, published at ACL 2024 and built on more than six million generations across 11 models and 11 adversarial attacks, tells a similar story: detectors do far worse on paraphrased or “humanised” text than on raw AI output, and every tool degrades when the text has been edited.

False positives, and who they hurt

A false positive is human writing flagged as AI-generated, and it does not fall evenly. A Stanford study published in Patterns in 2023 found GPTZero incorrectly classified over 60% of TOEFL essays by non-native English speakers as AI-generated. That study is three years old and GPTZero says it has since trained its model to cut false positives for second-language writers — but the pattern it exposed, that fluent-but-formulaic writing looks machine-like to a detector, has not gone away.

Small samples mislead in both directions. One comparison found zero false flags across 42 human texts, which sounds reassuring until you run the numbers: a clean run that small still allows a true false-positive rate of about 7%. Short texts are the weakest case of all — there is simply less signal for any tool to work with — and edited text and newer models both weaken detection further. Much of the detector evidence circulating today dates from 2023, while vendors have replaced their models since: Turnitin in July 2026, Originality.ai in September 2026.

What to do when text is flagged

Treat the flag as a starting point, not a verdict. Check for a watermark first, if the text came through a channel where marking is switched on. Ask the writer for drafts or process evidence — version history, notes, earlier drafts — which is far stronger proof than any score. Run the text through a second detector and compare the two results. And never accuse someone on the strength of one score alone, especially a student or a non-native English speaker, where the evidence says false flags are most likely.

Start with the watermark check on anything that came through an API or a provider with marking switched on, then run the text through two free detectors and compare. If both flag it and the writer cannot show drafts, you have a conversation to have — not a verdict to hand down.