Does AI train on your therapy notes? What ChatGPT, Gemini and Claude say

AI training means a provider reusing its users' conversations to improve its future models. For a psychologist, the question is simple. If you paste a session note into ChatGPT, Gemini or Claude, can that text be used to train the model? This article sets out what each provider's official pages say, as read on 15 October 2026, then explains why switching training off is not enough for patient notes. It is written for practitioners in France, Belgium and Switzerland.

Three questions not to confuse

When people talk about "data and AI", three issues often get mixed up. It helps to keep them apart, because one setting can answer one question and leave the other two untouched.

• Training. Does the provider reuse your conversations to improve its next models? This is what the "improve the model" option in chatbots controls.
• Retention. How long do your conversations stay on the provider's servers, and who can read them? A conversation can be excluded from training and still be stored for weeks.
• Data processing. Is there a contract stating what the provider may do with your patients' data, where it hosts it and how it protects it? With a personal account you accept general terms of use, not a contract of this kind.

A session note is health data. The French data protection authority, the CNIL, uses the GDPR definition. It covers data about a person's past, present or future physical or mental health that reveal information about their health status. This data falls under stricter rules than other personal data.

ChatGPT (OpenAI)

Personal Free, Plus and Pro accounts

According to OpenAI’s help centre, content from individual plans may be used to train its models. To prevent this, turn off "Improve the model for everyone" under Settings, then Data controls. According to OpenAI, once you opt out, your new conversations are no longer used for training. The page does not say what happens to past conversations.

OpenAI points out several limits. Giving a thumbs up or thumbs down can allow the whole conversation to be used for training, even after you have opted out. Temporary chats are not used for training, but OpenAI may keep a copy for up to 30 days for safety purposes. And a deleted chat is scheduled for permanent deletion within 30 days, unless OpenAI must keep it for security or legal reasons.

Business and Enterprise plans

ChatGPT Team has been called ChatGPT Business since 29 August 2025. OpenAI says this is a name change only. For Business, Enterprise, Edu and the API, OpenAI states on its enterprise privacy page that it does not train its models on your data by default. Workspace admins can control how long data is kept, and OpenAI can sign a Data Processing Addendum (DPA) for Business, Enterprise and the API. We found no mention of HDS certification on the pages we read.

Gemini (Google)

Personal Google account

According to the Gemini Apps Privacy Hub, when "Keep Activity" is on, Google uses your activity to provide, develop and improve its services, including training generative AI models. A subset of chats is read by human reviewers, including trained reviewers from Google's service providers. According to the page, these chats are disconnected from the account before being sent to the service providers. Reviewed chats are kept for up to three years.

If you turn the setting off, your future chats are no longer used for training, unless you send feedback. They are still kept with your account for 72 hours. Google itself asks you not to enter confidential information that you would not want a reviewer to see. A session note clearly falls into that category.

Google Workspace business account

The Google Workspace privacy hub states that customer data is not used to train models without the customer's permission, and that chats are not read by human reviewers without permission. It states that this content is customer data as defined by the Google Cloud Data Processing Addendum (CDPA). We found no mention of HDS certification on that page.

Claude (Anthropic)

Consumer Free, Pro and Max plans

According to Anthropic’s privacy centre, your chats may be used to improve its models if you allow it in the model improvement setting. You will find it under Settings, then Privacy, labelled "Help Improve our AI models". If the setting is on, Anthropic may keep your data in de-identified form for up to five years.

Turning it off excludes your chats from future training runs. Training runs already under way and models already trained are not affected. Giving a thumbs up or down keeps the conversation for up to five years, and it may be used for training. Chats flagged by safety systems may be kept for up to two years. Incognito chats are never used for training, and a deleted conversation is removed from back-end storage within 30 days.

Business plans

For its commercial products, such as Claude for Work and the API, Anthropic says it does not use your inputs or outputs to train its models by default. Here too, feedback you send can be an exception.

What about tools built for practitioners?

Note-taking tools designed for health professionals often rely on AI models supplied by large providers. So the real question is not only which model is used, but under what conditions. Each tool makes its own choices, and the word "secure" on a home page tells you nothing specific.

Before choosing, look for this information in the privacy policy or the contract.

• Who hosts the data, in which country, and whether that host is HDS-certified.
• Where the AI processing takes place.
• Whether the patient's name and identifying details are removed before the AI.
• What happens to the audio after the session.
• Whether the contract explicitly rules out training models on your data.

To compare the main tools, see our comparison of AI note-taking software for psychologists.

Why turning training off is not enough

Switching training off answers only one of the three questions. For patient notes, at least four problems remain.

Your conversations are still kept

Even without training, the provider keeps your conversations for a while. At OpenAI, a temporary chat may be kept for up to 30 days. At Google, chats stay with your account for 72 hours when activity is off. At Anthropic, a deleted conversation is removed within 30 days, and a conversation flagged by safety systems may be kept for up to two years. A note pasted into a chat is no longer under your control.

HDS-certified hosting is not stated

In France, article L.1111-8 of the Public Health Code requires anyone hosting health data collected during care, on behalf of the professional who collected it, to be approved or certified for that purpose. The French digital health agency, the ANS, quotes this text in its presentation of HDS certification. On the pages we read from all three providers, we found no mention of HDS certification. Our article on HDS-certified hosting explains how to check.

There is no data processing contract

When a provider processes patient data for you, the GDPR requires a contract or other legal act governing that processing (article 28). On a Free, Plus or Pro account, you accept the terms of a consumer service. Business plans offer a data processing agreement. That agreement and HDS certification are two separate obligations, and one does not replace the other.

Professional confidentiality is still at stake

Psychologists are bound by professional confidentiality. Handing the content of a session to a service whose retention period and reviewers you do not control raises a confidentiality issue, even with training switched off. We cover this in therapist confidentiality and AI.

HDS certification is a French rule. In Belgium and Switzerland, other laws govern health data and professional confidentiality.

Checklist before using AI for your notes

• I know whether the tool is a personal account or a business plan.
• I have read on the official page whether my conversations are used for training, and how to stop it.
• I know how long my conversations are kept, including after deletion.
• I know whether people can read my conversations.
• The data host is HDS-certified, and I have checked it.
• I have a data processing contract that covers health data.
• The patient's name and the details that could identify them are removed before the AI.
• I have informed my patients and obtained their agreement where required (see patient consent and AI note-taking).

If a single box is left unticked, do not put patient notes into the tool. A consumer chatbot is still useful for other practice tasks, as our article on ChatGPT for psychologists shows. The direct comparison is in Delta vs ChatGPT.

How Delta handles your notes

Delta is the clinical assistant for mental health and allied health professionals. You can take notes in two ways. During the session, Delta transcribes what is said. After the session, you can also dictate your observations. You can add written notes in both cases. Delta then generates the session report and updates the patient record, so you can find the whole course of care and how it has evolved in one place.

Data is hosted in France by an HDS-certified host, and AI processing takes place in France. The patient's name and identifying information are pseudonymised before the AI. No audio file is kept. The audio is used only during the session to produce the transcription. Data is never used to train a model, and it is encrypted in transit and at rest.

Delta works in the browser, with a Mac and Windows app recommended during sessions. The trial lasts 14 days, with no credit card. After that, the Pro plan costs €29 per month including VAT on an annual commitment, or €35 per month including VAT on a monthly plan. To find out more, visit the Delta page for psychologists.

Frequently asked questions

Sources

• OpenAI, How your data is used to improve model performance
• Google, Gemini Apps Privacy Hub
• Anthropic, Is my data used for model training?
• CNIL, Qu’est-ce qu’une donnée de santé ? (in French)
• Agence du numérique en santé, presentation of HDS certification, article L.1111-8 of the Public Health Code (in French)