You usually can’t tell whether AI-generated text has a watermark by looking at it. Text watermarks are designed to be subtle statistical patterns in a model’s token choices, not visible marks or hidden characters. To check for one, use a detector that supports the specific provider and watermark scheme—and treat its result as a limited signal, not proof of who wrote the text.
What a text watermark looks like—and what it doesn’t
A text watermark is generally not a special symbol, invisible space, or formatting artifact that survives copy and paste. Google says SynthID Text influences token choices during generation; OpenAI says textGrain subtly adjusts random word choices rather than inserting hidden characters, invisible spaces, or unusual punctuation. Google’s SynthID documentation and OpenAI’s provenance documentation describe these provider-specific approaches.
As an Amazon Associate I earn from qualifying purchases.
In broad terms, a watermarking system subtly alters a property of generated content—such as the statistical prevalence of words in context—so a detector can look for the corresponding pattern. NIST describes this general approach in its AI 100-4 report. The method is intended to avoid noticeably changing the text, but detection depends on the scheme and its configuration.
Free tools Windows power users keep installed
One-click scans. No signup required.
How to check a passage for a watermark
- Identify the likely source. If you know which model or service generated the text, start there. Different providers use different schemes, and a detector for one is not a general-purpose checker.
- Find the provider’s documented verification route. Check that it accepts text and covers the relevant product, model, and generation path. Availability and supported inputs can change.
- Submit only content the tool supports. Read its result in the tool’s own terms, such as “watermarked,” “not watermarked,” or “uncertain.” Do not translate that result into a broader claim about all AI-generated text.
- Keep the result qualified. A negative result means the detector did not identify its supported signal; it does not establish that a person wrote the passage.
Google says SynthID detection can use thresholds to manage false positives and false negatives, and describes detection as probabilistic. A detector’s output is therefore an estimate about a particular signal, not a definitive authorship verdict. Google AI for Developers: SynthID
#1 Best Overall
Which provider tools can check text?
Google SynthID
Google’s SynthID Text documentation describes a detector for supported SynthID watermarks that can report whether text is watermarked, not watermarked, or uncertain. Google says the open-source implementation is available in Hugging Face Transformers v4.46.0 and later. The Google DeepMind SynthID Text GitHub repository identifies its code as a reference implementation for research and reproducibility, not production use, and points to Transformers for production deployments.
On May 20, 2025, Google announced a SynthID Detector portal for submitted text and other media created with Google AI tools. The announcement said it was initially rolling out to early testers and that the portal could highlight portions likely to contain a watermark. That announcement does not establish current access or the portal’s present supported inputs; check Google’s announcement for current details. Google also reported that more than 10 billion pieces of content had been watermarked with SynthID as of that announcement; this is Google’s cumulative figure from May 2025, not an independently measured or current count.
OpenAI textGrain
OpenAI says ChatGPT-generated text includes textGrain watermarks in the EU, while API customers globally can enable watermarking for supported models. Coverage can differ by product, model, export path, file type, and generation date; OpenAI says it is extending coverage. Its text-detector access is limited to qualifying organizations on a case-by-case basis. Check OpenAI’s provenance documentation for current coverage and access information.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →OpenAI distinguishes checking for an embedded provenance signal from using an AI-writing classifier, which estimates AI authorship from the text. A classifier is not a watermark detector, and a watermark detector does not test every provider’s output.
Rank #3
Why a detector can miss a watermark
- There may be too little text. Short passages can lack enough token choices to embed or detect a reliable signal.
- The content may constrain word choices. Code and precise factual text offer fewer plausible next-token choices, making watermarking harder.
- Editing can weaken the pattern. Extensive paraphrasing or translation may reduce a signal’s detectability.
- Language matters. OpenAI reports that in a test of 500 synthetic English prompts translated into 23 other official EU languages, detection at a 1% false-positive rate was 69.0% for Spanish and 42.2% for Romanian. Those are results for OpenAI’s stated test setup, not a general guarantee for other text, languages, or detectors. OpenAI’s provenance documentation
- The detector may not support the signal. A tool built for SynthID, for example, cannot establish whether text carries a different provider’s watermark.
What a watermark result can—and cannot—show
A positive result can be evidence that a supported model likely generated or processed the content. OpenAI cautions that a watermark alone does not establish who authored or owns the text, who is legally responsible for it, or how much a person contributed. It also cannot show how much a person edited the output. A missing signal is not proof of human authorship.
For that reason, do not use a detector result by itself to accuse someone of misconduct or make a high-stakes authorship claim. Report what the tool actually found, name the supported scheme, and consider other evidence separately. No universal checker is established by the provider documentation cited here; these tools address specific signals and coverage.
Quick Recap
Rank #4
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
