desenmascara.me

Use of artificial intelligence

What AI does on desenmascara.me, what it does not, and who is accountable. Transparency under Article 50 of the EU AI Act.

Last updated: 18 October 2026

• Reports are generated automatically. Some are written by an AI.
• They are published without prior human review.
• Every report says how it was decided: by AI, by rules, by an external source or by a person.
• A person reviews any report someone disputes, free of charge, and no automated process ever overwrites that correction.

1. Which AI we use

• OpenAI language models (currently GPT-5.6 Luna, with reasoning mode off) read the content of the analysed website and write the reasoning behind the verdict.
• A statistical model of our own, which generates no text, only decides whether an AI call is worth spending. It never issues a verdict.
• Screenshots are real captures of the website, not generated images.

• Many verdicts are not decided by AI: security lists (Google Safe Browsing, PhishDestroy, Cloudflare), financial regulators’ warnings, reputation (Tranco), verified trust seals and heuristic rules. In those cases the text is a template of ours.
• When the AI decides, it may call a website “fraudulent” only if it quotes, verbatim from the page itself, the mechanism of the deception. Without that quote, “suspicious” at most.
• Rules sit on top of the AI and limit its conclusions: for example, it never accuses an official body or a very well-known domain on its reasoning alone, and it does not accuse a website with a verified legal identity just because of what it did not see.
• Every report carries a label with its method: “AI analysis”, “Heuristic analysis”, “Blocklist”, “Regulator notice”, “Human review”, and so on.

• Automated reports are published without a person reading them first. That is why the editorial exemption in Article 50(4) does not apply, and we say so on every AI-generated text.
• Oversight is reactive and real: any website owner can dispute their report free of charge; a person reviews it and, if it is wrong, corrects it in the report itself. No automated process overwrites that correction.
• Before sending a verdict to VirusTotal we ask for human approval, except in unmistakable cases (confirmation by a third party, tricks no legitimate site uses, or newly created investment sites).
• A person is accountable for everything published: the owner of the project. Contact: [email protected].

• An AI can be wrong, and a website changes after we analyse it. Every report is a dated assessment, not a statement of fact about anyone.
• We only read what the website shows us that day. What we did not see is not proof that it does not exist.
• Check for yourself before paying or making an important decision.

• The API, the MCP server and the Agent Trust API return the assessed_by field in every response, stating what decided the verdict (ai_reasoning when it was the AI).
• Every report links to this page in its structured data (publishingPrinciples).

• No chatbots posing as people and no fictitious authors: reports are signed by desenmascara.me.
• We do not generate images, audio or video.
• We do not use AI to make decisions about individuals: we analyse websites and domains.