Stellar Reviewsby flylabs.ai

En breve

  • La IA escribe los borradores, tú decides qué publicar.
  • Stellar Reviews nunca publica nada por su cuenta.
  • Las reseñas y tus respuestas no se usan para entrenar la IA.
  • Quien lee tu respuesta solo ve el texto. Si alguien la comprueba en /verifica sabe que viene de Stellar Reviews, nunca de qué establecimiento.
  • Puedes corregir cada borrador o pedir otro.

Los detalles

El modelo de IA de cada plan

PlanModelo para los borradores
FreeIA básica (Claude Haiku)
StandardIA estándar (Claude Sonnet)
PremiumIA avanzada (Claude Opus) para las reseñas difíciles, IA estándar para las demás
BusinessIA avanzada (Claude Opus) para las reseñas difíciles, IA estándar para las demás

Difícil significa con nota de 3 sobre 5 o menos, o de más de 600 caracteres. La prueba usa el modelo de Premium. El análisis del estilo de tus respuestas usa siempre la IA avanzada. Los modelos son de Anthropic y usamos siempre la última versión.

Este documento existe en italiano y en inglés: aquí tienes el texto en inglés.

Versión 2026-10-10

Artificial intelligence in Stellar Reviews

Version 2026-10-10. How we use AI, what whoever publishes the replies has to do, and how a text generated by Stellar Reviews can be recognised, also for the purposes of Articles 4 and 50 of Regulation (EU) 2024/1689 (AI Act). This is a translation: if it differs from the Italian version, the Italian version prevails.

What the AI does

  • Writes reply drafts to reviews, in the reviewer's language and in the property's tone, derived from the replies the property has already published and from the context the property provides.
  • Translates the review and the draft for whoever manages the property, when they are in another language.
  • Flags sensitive reviews (low ratings, concrete complaints, contact requests) and cases where the language of the draft should be checked.

The model is Claude by Anthropic, used through its API: the data sent is not used to train Anthropic's models and is kept by Anthropic for 30 days. It receives only the review and the property context needed for the draft; for the tone analysis of published replies it also receives excerpts of at most 30 already published replies (800 characters each) and of the reviews they reply to (300 characters each).

A person stays in control

Drafts are proposals. Stellar Reviews publishes nothing: the property reads, corrects where needed and publishes on the platform. For critical reviews the draft is a structured outline (opening and points to clarify) to be completed. Drafts never promise refunds, discounts or compensation.

Known limitations

  • It may misread irony, sarcasm, dialects or very short reviews.
  • It uses only the facts in the review and in the property context: if the context is incomplete or outdated, the draft may be generic or inaccurate.
  • It may get names, timing and references wrong: always check them.
  • Language detection may be wrong: when that happens, the app says so.

Instructions for use

  1. Always read the whole review before the draft.
  2. Check names, dates, services mentioned and every fact about the property.
  3. For critical reviews, complete the outline with what you actually know about that stay.
  4. Keep the property context up to date: that is where the AI takes its facts from.
  5. If the draft is not right, fix it with "Edit" or ask for another with "Rewrite". "Suggest a change" tells us what is wrong: it helps improve the next drafts.

How generated texts can be recognised

  • In the app every draft carries the label "AI-generated: review it before publishing".
  • Machine-readable marking: in the app pages, and in the formatted version of a draft copied to the clipboard, the draft carries the data-ai-generated and data-digital-source-type markers with the IPTC standard value http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia, the same vocabulary used by C2PA.
  • Register and fingerprints: for every generated draft, previews included, we keep, for 5 years, a non-reversible fingerprint computed from its sentences, together with the model and the date. We do not keep the text for this purpose.
  • Public verification: at /verifica anyone can paste a text and find out whether it contains sentences from a draft generated with Stellar Reviews, even if partly edited. The verification does not reveal which property the draft belongs to.

Why there is no marking inside the published text: platforms accept only plain text, which has no room for metadata; invisible characters would be stripped and would alter length limits. Detection therefore relies on fingerprints. Translations for the property are not marked: the law treats them as standard editing.

Obligations of whoever publishes

Article 50(4) of the AI Act requires disclosing AI-generated text only when it is published to inform the public on matters of public interest, and not when a person reviews it and takes responsibility for it. Replies to reviews, reviewed by whoever manages the property, normally fall outside it. This is information, not legal advice on a specific case.

Contact

For questions or reports about the use of AI: info@flylabs.ai.