AI disclosure

2 What the models do not do

They do not know anything that happened after their training window, unless a product gives them a retrieval tool with the current date in it. They do not verify a claim they have not been asked to verify. They do not have memory of a conversation you closed, unless you saved it and the product reads it back.

They do not act on the world on their own. Cortex Code and Cortex CLI write files and run commands only where you have pointed them and only with the permission you granted. Cortex Security reads a repository and reports; it does not modify it in the course of a review.

They are not a doctor, a lawyer, an accountant or an engineer of record. An output is a draft. The judgment, the verification and the signature remain with a human.

3 Training data

The models are trained on [À REMPLIR PAR MATHIS : composition réelle du corpus — parts de données publiques, de données sous licence, de données produites en interne, et de retours utilisateurs, avec leurs ordres de grandeur].

User conversations are not used for training by default. Where a conversation is used for training, it is because the user turned that option on, for that conversation or for their account, and the consent can be withdrawn at any time from the same place. We do not treat "continued use of the service" as consent to train on your content.

We respect robots directives and the licensing terms attached to a source. Publishing a work openly does not place it in the public domain, and we do not treat it that way.

Rightholders can obtain [À REMPLIR PAR MATHIS : mécanisme réel — liste des sources, filtrage a posteriori, retrait des données], and can request removal of their work by writing to [À REMPLIR PAR MATHIS : adresse e-mail de contact droits d'auteur]. Those requests are answered by a person, and the answer states what was actually removed.

4 Prompt retention

Your prompts and the generated outputs are stored while the conversation exists, so that you can reopen it, and they are used to keep your account working and to detect abuse.

You can delete a conversation from the interface, and it stops being readable immediately. We do not promise that it is off our disks in the same second; that is what the retention column above means.

5 Accuracy, limits and hallucinations

A generative model produces fluent text whether or not it is correct. The failure mode — a confident, well-formed, false statement — is called a hallucination, and it is a property of the method, not a bug that a later version will remove.

Expect it especially on: citations and references, exact figures, dates, the content of a document the model has not been given, and API signatures. Treat every citation as unverified until you have opened it.

Every model we publish comes with an evaluation report that states where the model fails, not only where it succeeds. Those reports are published even when they are unflattering, and they are published before the weights. Read them before choosing a model.

What reduces the risk: giving the model the source document rather than asking it to recall one; asking it to quote; asking for its uncertainty; and having a person check anything that will be acted on. What does not reduce the risk: a confident tone. Fluency is not evidence.

6 Prohibited high-risk uses

Under Article 5 of the European AI Act, and under our own terms, the following uses are prohibited outright and are not made lawful by our licence:

• subliminal, purposefully manipulative or deceptive techniques that materially distort behaviour and cause significant harm;
• exploiting a vulnerability of a person or a group because of age, disability or a specific social or economic situation, to distort their behaviour in a way that causes significant harm;
• social scoring by a public or private actor, producing detrimental or unfavourable treatment unrelated to the context in which the data was generated;
• predicting the risk of a natural person committing a criminal offence, on the basis of profiling alone;
• building a facial-recognition database by untargeted scraping of images;
• inferring emotion in the workplace or in an educational setting;
• biometric categorisation to deduce race, political opinions, trade-union membership, religious or philosophical beliefs, sex life or sexual orientation.

Beyond those prohibitions, we refuse — and the terms of service forbid — the uses listed in 2.2.4, and in particular decisions about people's access to employment, credit, housing, healthcare, education or essential public services without qualified human review.

7 Human oversight

A deployed AI system that a person relies on must leave that person able to understand it, to contradict it and to stop it.

In our own products this means: the model states its uncertainty rather than smoothing it over; every claim from Deep Research carries a citation you can open; an action that modifies your files is shown for confirmation before it runs; and there is always a way to do the task without the model.

In your own deployments, put a named human in the loop wherever an output affects someone's rights or access to a service, and record who reviewed what. The person who signs remains accountable; the model is not a shield.

8 Reporting a problem

Report a harmful output, a security issue in a Cortex service, or a model that behaves in a way its card does not describe, to [À REMPLIR PAR MATHIS : adresse e-mail de signalement]. For security reports, do not open a public issue first: use [À REMPLIR PAR MATHIS : canal de divulgation, ou l'adresse dédiée].

We acknowledge a report, we tell you what we found, and we publish the fix in the changelog. Where a report reveals a systematic failure — a class of prompt that defeats a safeguard, a gap in an evaluation — we say so in the model card of the affected version. A model that fails is not withdrawn from the record; it is annotated.

9 Alignment with the European AI Act

Cortex Foundation acts as a provider of general-purpose AI models within the meaning of Regulation (EU) 2024/1689, and as a provider of an AI system for its own products.

Where a Cortex model is integrated into a product placed on the market by someone else, that provider carries the obligations of the AI system and Cortex Foundation carries those of the model. We supply the documentation needed for that, and we ask integrations to keep the provenance markers intact.

This section describes our reading of the regulation in force at the date at the top of this page. It is not legal advice, and it does not replace the regulation.