Frequently asked questions

What is a paraphrase pass?

A single rewriting sweep over a document in which every sentence is reworded while the meaning is kept. Writers use it to improve a draft; researchers use it as the standard attack on AI text watermarks and detectors, because changing most of the words removes the statistical patterns both rely on.

Which AI models watermark their text?

As of October 31, 2026: Google's Gemini app has marked text with SynthID since May 2024. Anthropic announced on August 14, 2026 that future Claude models carry a SynthID-based mark worldwide. OpenAI announced textGrain on October 5, 2026, on by default only for ChatGPT and Codex users in the EU and opt-in for API customers everywhere. Open-weight models carry no provider mark unless whoever deploys them adds one.

Can I check whether a text is watermarked?

Not with a public tool. Each provider's detector needs the provider's secret key, and Google, Anthropic and OpenAI all limit detector access to approved organisations such as regulators, researchers, media and educators. The open-source SynthID library only detects marks made with your own key.

Does paraphrasing remove an AI watermark?

A full rewrite does. Anthropic says light editing probably leaves the mark and a complete rewrite removes it. OpenAI's own test found that replacing a quarter of the words with synonyms cut detection on 400-token passages to 17 percent. Research papers report the same pattern: the mark weakens in proportion to how much text changed.

Does paraphrasing make text undetectable?

It removes watermarks and lowers classifier scores, but detectors now train on humanizer output, Turnitin flags rewritten text as a category of its own, and process evidence such as document history is unaffected by rewriting. Nothing on this site tests or ranks evasion tools.

How accurate are AI detectors?

Every vendor publishes its own numbers and independent evidence is thin. On a shared benchmark such as RAID, accuracy varies widely by model, domain and attack. The one peer-reviewed fairness result, from Stanford in 2023, found a 61.3 percent average false positive rate on essays by non-native English speakers.

Why do AI detectors flag non-native English writers?

Classifiers treat predictable word choice as machine-like. Writers with a smaller working vocabulary produce more predictable text, so their writing scores as AI more often. The 2023 Stanford study showed that enriching the vocabulary flipped the results.

What should I do if I am falsely accused of using AI?

Gather process evidence: version history in Google Docs or Word, drafts, notes, research trail, and any writing replay. Ask which tool produced the score and read that tool's own guide, which usually says the score is an indicator, not proof. Cite the false positive research, and ask for the institution's written policy on how detector scores may be used.

What does a Turnitin AI score mean?

The percentage of a submission that Turnitin's classifier judges likely AI-written, including text it believes was run through a humanizer. Turnitin's own guide describes it as an indicator that should start a conversation, warns that short submissions are unreliable, and limits bypasser detection to English.

Is AI detection reliable enough to fail a student?

Detector vendors themselves say no. Turnitin calls its score an indicator, not a verdict. Several universities have switched the AI indicator off. Decisions about a person should rest on process evidence and conversation, not a classifier percentage.

Is it illegal to remove an AI watermark?

Not for an individual in the jurisdictions covered here as of October 2026. The EU AI Act puts marking duties on providers and deployers. California's AI Transparency Act bars large online platforms from knowingly stripping provenance data from January 1, 2027. The model providers' usage policies restrict building tools that pass generated text off as human.

What is the EU AI Act Article 50?

The transparency article. From August 2, 2026, providers of generative systems must mark output in a machine-readable, detectable way, and deployers must disclose deepfakes and AI-generated text published on matters of public interest. The Code of Practice, final on June 10, 2026, interprets marking as secured metadata plus a watermark.

What is California's AI Transparency Act?

SB 942 as amended by AB 853 (signed October 13, 2025). Covered providers owe disclosures and a free detection tool from August 2, 2026. Large online platforms and generative AI hosting platforms are covered from January 1, 2027, including a ban on knowingly stripping provenance data. Capture devices follow on January 1, 2028.

What is SynthID?

Google DeepMind's watermarking family. For text, it adjusts the model's word choices with a keyed pattern during generation and detects the pattern later. It has run in Gemini since May 2024, was published in Nature in October 2024, and is open-sourced as a library. Anthropic's Claude watermark is a version of it.

What is textGrain?

OpenAI's text watermark, announced October 5, 2026. It adds a statistical signal to word choices. OpenAI reports about 95 percent detection on 400-token passages at a 1 percent false positive rate, dropping sharply after synonym replacement. It is on in the EU for ChatGPT and Codex, opt-in in the API, and the detector is gated.

What is C2PA and does it work for text?

An open standard for signed provenance manifests attached to media files. It is widely used for images, audio and video and is one of the two marking layers in the EU Code of Practice. It rarely helps with text, because text is usually pasted without a file to carry the manifest.

What is an AI humanizer?

A rewriting tool sold on the promise that its output will pass AI detectors. Technically it is a paraphraser with a detector check attached. Detectors now train on humanizer output, platform usage policies restrict building them on frontier APIs, and Google will not run ads for them.

Is QuillBot an AI humanizer?

QuillBot is a general paraphraser with modes for fluency, formality and length. It is not marketed as a detector bypass, but heavy paraphrasing of generated text is treated by most integrity policies the same as submitting the generated text.

What is the best AI detector?

There is no reliable ranking. Vendor accuracy claims are not comparable. The fairest comparison is a shared benchmark such as RAID, where results depend on the model that wrote the text, the domain and the attack. Use more than one tool, treat any score as a signal, and prefer process evidence.

How does this site decide what to list?

An entry gets in if it is a primary source, a widely used tool, a published paper, a law, or a feed that reports changes first. Every entry says what it is, who runs it, roughly what it costs and any known limit, and carries the date a person last checked the link. No affiliate links, no sponsored entries.