AI Policy
Probator.ai is built on artificial intelligence. This policy explains which AI systems we use, what they are for, how reliable they are, how we protect your data, and the rules for using AI-based results responsibly. It forms part of our Terms of Service and is written with the EU Artificial Intelligence Act (Regulation (EU) 2024/1689) in mind.
1. The AI systems in Probator.ai
| Feature | AI used | What it does |
|---|---|---|
| AI detection | Probator detection model, our own classifier developed by Spawncore, which scores sentences using multilingual text embeddings (bge-m3 and EmbeddingGemma) computed on Cloudflare Workers AI | Estimates, for each sentence and for the whole text, how likely it is to have been generated by an AI system |
| AI detection (additional signals) | Third-party general-purpose language models. The expert reading uses OpenAI GPT-6 Luna, reached through OpenRouter and served by OpenAI, Microsoft Azure or Amazon Bedrock; it runs on deep scans and on standard checks in languages our detection model was not trained on, and the deep scan lets it reason at greater length. The rewrite test uses the open model OpenAI gpt-oss-20b, run on Cloudflare Workers AI, on deep scans | A "rewrite test" (how much a model changes the text when asked to polish it) and an "expert reading" (a sentence-by-sentence forensic review with quoted reasons); the deep scan is a more thorough expert reading |
| Grammar & style | Open language model Mistral Small 3.2, reached through OpenRouter and served by one of several inference providers (currently DeepInfra, Parasail, Venice or Mistral AI) | Suggests corrections and explains each one; you decide whether to accept it |
| Plagiarism | No generative AI: text fingerprinting against our own index, web search and academic databases | Finds passages that match other sources |
| Hidden marks | No AI: rule-based character and file analysis | Finds invisible characters, look-alike letters and concealed text |
The final AI-detection likelihood combines these signals with forensic and stylistic evidence. The report shows each signal separately so you can see how the result was reached.
Workers AI runs the embeddings for our detection model and the rewrite test in Cloudflare's network, in our own Cloudflare account; Cloudflare does not guarantee that this inference takes place inside the EU. The expert reading and grammar suggestions are sent through OpenRouter to the providers named above, mostly in the United States (Mistral AI is in France). Our database and file storage (Cloudflare D1 and R2) are in the EU. See our Privacy Policy.
2. Intended purpose
Probator.ai is an assistive tool that gives writers, teachers, editors, publishers and organisations information about a text, to support their own judgement. It is designed to:
- help authors improve their writing and check their own work before submitting it;
- give reviewers indications that a text may need a closer look, together with the reasons.
It is not designed or intended to make or replace decisions about people, and in particular not to be the basis for automated decisions on access to education, assessment of learning outcomes, monitoring of students during tests, recruitment, or employment. Results are not proof of authorship or misconduct.
3. Accuracy and limitations
- No detector is fully reliable. False positives happen: human text can be flagged, especially formal, technical, formulaic, translated, heavily edited, or non-native writing. False negatives happen too: AI-generated text that has been paraphrased or edited can go undetected.
- Short texts (under about 150 words) give much less reliable results; the report warns you when this applies.
- Languages. Detection works across many languages, but accuracy varies by language and genre. Our stylistic lexicons cover English, Portuguese, Spanish, French, German and Italian.
- Mixed authorship. Texts written partly with AI assistance can produce intermediate scores; the sentence view helps identify which passages drive the result.
- Grammar suggestions can be wrong or change meaning. Review each one.
- Watermarks embedded by AI vendors (such as Google SynthID) can only be verified with the vendor's keys; Probator does not claim to verify them. We report provenance marks we can read (see section 8).
We evaluate our detection model on held-out test sets of human and AI-written text, including AI-polished and "humanized" text, in each language, and publish the figures at probator.ai/accuracy before a model is put into use. Each language's decision threshold is set so that at most 1 in 100 human documents in our validation data is flagged. The engine never reports a text as AI-generated from a single weak signal, and very short texts are reported as inconclusive unless there is hard evidence. Performance on your texts may differ.
4. Human oversight: rules for using results
If you use Probator.ai to review other people's work, you must:
- never use a Probator result as the only basis for a decision that affects a person (for example a grade, a sanction, an accusation, or a hiring decision);
- review the full report, including the reasons and the confidence level, and consider other evidence such as drafts, version history, and the person's other work;
- give the person the opportunity to explain and to respond before any adverse decision;
- comply with your institution's policies and with the law, including data-protection law, when you submit other people's texts;
- if you are an organisation, inform the people whose texts you check that you use an AI-assisted tool, where required.
Using Probator.ai as an automated decision system in high-risk areas (such as education or employment decisions under Annex III of the AI Act) is outside its intended purpose and prohibited by our Terms.
5. Admission, assessment and recruitment
Probator.ai is not intended to be the basis for decisions on admission to education or training, on grading or other assessment of learning outcomes, or on the recruitment and selection of people (for example filtering applications or evaluating candidates).
The EU AI Act treats AI systems intended for these uses as high-risk (Annex III, points 3 and 4); those rules apply from 2 December 2027. A school, university, employer or other organisation that uses Probator.ai for one of these purposes is using it outside its intended purpose and may itself take on obligations under the AI Act, as the deployer or, where it puts the system to that use under its own responsibility, as the provider of a high-risk AI system. In any case it must ensure that a person reviews each case with the full context, can disregard the result, and gives the person concerned the chance to respond (see section 4).
6. Your data and AI
- No training on your content. We do not use the texts and documents you submit to train or improve our detection model or any other AI system. Our model is trained on public and licensed datasets and on text we generate ourselves for that purpose.
- Where the models run. The embeddings for our detection model and the rewrite test run on Cloudflare Workers AI, which processes the text for us as a sub-processor. The expert reading and grammar suggestions are sent through OpenRouter, which we instruct to route the text only to providers that do not collect it under OpenRouter's provider data policies, so it is not used to train models. See our Privacy Policy and sub-processor list.
- No profiling. We do not build profiles of you or of the authors of the texts you check.
7. Transparency about AI-generated content
Certificates of analysis record what our systems concluded about a text at a point in time. They carry the same limitations as the results themselves (section 3) and state on their face that they are probabilities, not proof, and that they do not certify authorship.
Grammar and style suggestions are produced by AI and are always shown as suggestions that you accept or dismiss; they are not inserted into your text without your action. Explanations in the expert reading are generated by AI and labelled as such, together with the model used. Probator.ai does not generate full texts for you.
8. The EU AI Act and Probator.ai
Article 50 of the AI Act sets transparency obligations that apply from 2 August 2026. This is how Probator.ai stands with them, and how it helps you with yours.
- Our grammar and style corrector is an assistive editing tool: it proposes corrections to text you wrote and does not substantially alter it. Every suggestion is shown as a suggestion, labelled as AI-generated where it comes from a language model, and applied only when you accept it. We therefore consider it to fall within the exemption in Article 50(2) for AI systems performing an assistive function for standard editing.
- We do not generate texts for you. Probator.ai does not produce new content that would need to be marked as AI-generated. If this changes, we will mark that output in a machine-readable way as the Act requires.
- We never remove marks. Probator has no feature that removes hidden characters, watermarks or metadata: it shows them and explains what they may be, and grammar suggestions are never offered for passages that contain invisible characters. File metadata such as Content Credentials is never altered. Using the Service to remove such marks is prohibited by our Terms.
- Provenance detection. Every check reports machine-readable AI marks we find: C2PA Content Credentials and IPTC "digital source type" values in document metadata and embedded images, explicit AI labels in document properties and HTML meta tags, and invisible character patterns in the text that may encode a watermark. We report Content Credentials as found but do not validate their cryptographic signature, and we cannot read vendor watermarks that require the vendor's key (such as SynthID). Marks are frequently lost when text is copied or converted, so finding none does not mean a text was written by a person.
- Human review records and disclosure labels. To help deployers who publish text on matters of public interest, certificates can record a human review (see our Terms) and we offer disclosure labels in six languages. Whether you must disclose, and whether the editorial-responsibility exception in Article 50(4) applies to you, depends on your circumstances; Probator.ai provides information and records, not legal advice.
9. Reporting problems
If you believe a result was wrong or harmful, or you have concerns about how Probator.ai is being used, write to ai@probator.ai. Include the report if you can (you can download it as JSON from the editor). We review reports to improve our models and this policy.