It estimates whether a sample looks closer to AI-generated, AI-assisted, or human-written writing, or whether there is not enough evidence. It does not identify the author.
Remote analysis · AI Text Detector
AI text detector
AI text detector?
Paste a draft, fetch a public article, or open a text file. The detector reports AI, human, mixed, and formulaic signals. It does not name a model, and it is not a byline. This is remote analysis. The text goes to the detection provider.
C2PA: Coalition for Content Provenance and Authenticity; C2PA specification 2.4
Last updated
Paste a draft, fetch a public article, or open a text file. The detector reports AI, human, mixed, and formulaic signals. It does not name a model, and it is not a byline.
This is remote analysis. The text goes to the detection provider. It is not an on-device model, and a high signal is not proof the words came from ChatGPT.
The interface is localized. You can paste writing in many languages; the sample is not translated first. Reliability can still vary by language, style, and length, and short samples are more likely to come back inconclusive.
Remote AI analysis
We do not keep the submitted text
Not proof of authorship
AI text detector
Your text is sent securely to our AI detection provider for analysis. We do not intentionally store submitted text. Anonymize text locally first →
The file is read locally. Extracted text is sent for AI analysis. Scanned PDFs are not OCR’d here — use Local OCR first.
Overall assessment
AI writing signals
These are separate signals, not a percentage of the text written by AI.
AI likelihood
—HumanAI
This is a model-estimated likelihood, not proof of authorship. It is not the percentage of the text written by AI.
Paragraph analysis
No paragraph-level signals were returned for this sample.
What this result means
AI text detection is probabilistic. Results are signals, not proof of authorship.
Supporting text stats
These counts are supporting measurements. They are not an AI detector.
How the AI text detector works
The text is analyzed remotely through CreatorPrivacyKit’s server using TypeSafe System One. One pass returns origin (human, AI, mixed, or uncertain) plus separate signals for AI writing, human writing, mixed authorship, formulaic structure, repetition, uniformity, and generic language. The detector can stay uncertain instead of guessing. Treat the output as signals, not a verdict for a disciplinary hearing.
AI detectors have limitations
Do not use a detector score as the sole basis for high-stakes decisions.
- Heavily edited AI text may look human.
- Formulaic human writing may look AI-generated.
- Short samples are difficult.
- Different languages and styles can behave differently. This model is described as covering English, Chinese, Japanese and Korean; other languages may be less reliable.
- New AI models can shift the distributions a detector was calibrated on.
Privacy
This is remote analysis, not a local-only tool. Submitted text must reach the detection provider (TypeSafe). CreatorPrivacyKit does not intentionally persist submitted text, does not put body content in analytics, and does not keep application logs of the writing itself. File mode only sends extracted text. You can clear the page state in the browser at any time. TypeSafe documents that Jev is not trained on customer requests; that statement is theirs, not an extra guarantee from us.
Related checks
- Private AI text assistant — rewrite and redact on this device. That page does not detect authorship.
- Content Credentials checker — for images, not paragraphs.
- MiniCPM5-2B in the browser — the local model behind the assistant, not this detector.
FAQ
It can flag writing that looks like AI-generated or AI-assisted text. It cannot prove the words came from ChatGPT or any other named product.
Yes. Edited AI text, formulaic human prose, short samples, and new models all create errors. An abstain is often more honest than a forced score.
The sample looks mixed: some passages sit closer to AI-generation signals, others look more human-written. That can be drafting, rewriting, or heavy editing.
The model did not have enough reliable evidence. Short text is a common reason. In that case it should not invent a score.
When the model returns paragraph-level localization, mixed documents can show which passages have stronger AI signals. Coverage depends on the sample.
Rewriting can weaken or hide signals. A human-leaning result after rewriting is not proof the first draft was human.
CreatorPrivacyKit does not intentionally keep submitted text. The writing has to be sent to the detection provider to run the check. See the privacy section on this page.
Not as the sole basis for a high-stakes decision. Use it as one signal among process, drafts, and conversation with the student.
You can submit writing in many languages. The sample is not translated first, because translation would change the style signals. Reliability can still vary by language, style, and length; short samples are harder.
No. The check is remote. Paste, URL extract, and text-file extract send the text for analysis. The browser does not receive the provider key. Submitted text is not intentionally stored.
No. You get separate signals for AI, human, mixed, formulaic, repetition, uniformity, and generic language. That is not an authorship percentage and not a named-model verdict.
Up to 80,000 characters. Short samples are more likely to come back inconclusive. The page language is not a limit on the language you paste, and reliability still varies by language, style, and length.
This page checks writing and sends the text. The image origin detector reads C2PA and metadata locally, with an optional OpenAI image check. They do not share a score.