Technically responsible answers about AI text watermarks and this detector.
An AI text watermark is a machine-readable signal intentionally embedded during text generation. Modern approaches usually work by subtly biasing the model’s token selection according to a secret pattern or key, producing a statistical signature that a matching detector can later test for. It is not the same as visible hidden characters.
The language model produces a probability distribution over possible next tokens. A watermarking algorithm slightly adjusts those probabilities. The generated text remains fluent, but a detector that knows the same scheme can measure whether the observed tokens are more consistent with the watermarked distribution. Results are typically expressed as confidence or a statistical test, not absolute certainty.
OpenAI has researched statistical watermarking but does not currently provide a public text watermark detector for ChatGPT that third parties can use. This product therefore marks OpenAI-related detection as “Not currently supported.” We do not invent or reverse-engineer proprietary algorithms.
OpenAI has published research on watermarking. Whether any given ChatGPT response carries a production watermark is determined by OpenAI’s current product configuration, which is not publicly guaranteed for all users or all outputs. There is no public third-party detector for it at this time.
Anthropic has not released a public text watermarking scheme or corresponding detector that third parties can integrate. Detection availability depends on publicly supported technical capabilities. This detector will be enabled when a reliable public method is available. We do not fabricate Claude’s algorithms.
Google has developed SynthID, including research and product work related to text. Public third-party detection of SynthID Text is limited; this product lists it as “Coming Soon” pending a reliable public detection path. We do not claim an official partnership with Google.
SynthID is a family of watermarking technologies developed by Google DeepMind. For text, it refers to statistical techniques that embed a detectable signal into model outputs. Detection typically requires Google’s tools or an equivalent published method. Third-party access is constrained.
Often only partially, and sometimes not at all. Statistical watermarks depend on the specific token sequence. Substantial paraphrasing, rewriting, or summarization can destroy or severely weaken the signal. Robustness varies by scheme and by how aggressive the edit is.
Yes. Translating text into another language generally produces a new token sequence that no longer carries the original statistical watermark. Cross-lingual watermarks are an active research area but are not widely deployed in public detectors.
Yes. Detectors return confidence scores or hypothesis-test results, not perfect knowledge. Short text, noise, editing, and mismatches between the generator’s scheme and the detector’s assumptions all increase error rates. False positives and false negatives are possible.
No. A missing watermark does not prove that text was written by a human. Many AI systems do not apply watermarks, watermarks can be removed by editing, and public detectors for commercial schemes are limited. Absence of evidence is not evidence of absence of AI involvement.
Longer samples are better. Very short passages (under ~200 characters) provide little statistical evidence. For more reliable analysis, use several hundred words when possible. Extremely long documents may be truncated or rate-limited depending on the plan.
Basic analysis in the current implementation runs in your browser. Submitted text is not sent to a remote server for persistent storage. We do not keep the content of your scans by default. See the Privacy page for details.
It is primarily a watermark and provenance detector. It also provides a separate statistical AI-likelihood estimate. It is not a generic “this is AI / this is human” classifier that claims high accuracy across all models. Those two functions are kept distinct on purpose.
Watermark detection looks for an intentionally inserted machine-readable signal. AI-likelihood (or “AI detection”) estimates how similar the text is to typical model output using statistical features. They answer different questions. A watermark can be strong evidence of a specific generation process; statistical AI scores cannot prove authorship.
You may use the tool as one input among others. Because public watermark detectors are limited and statistical estimates are imperfect, results should not be treated as definitive proof for academic integrity decisions. Always combine with human judgment and institutional policy.