Is turning off ChatGPT training enough to protect your clients' data?

Paper airplane with blacked-out lines flying away from a stack of sheets: client data protected before it is sent

Is turning off ChatGPT training enough to protect your clients’ data?

In October 2026, I gave a training session on artificial intelligence (AI) to real estate investors and property managers. During the part on privacy settings, a participant said: “Sharing a chartered appraiser’s report, that ends up on the server.” Then he added: “What’s the problem with that?” In November 2026, at a business show for SMEs, I came back to the topic with a compliance specialist, after a talk on Law 25. I gave him my usual answer: turn off model training and data sharing. He said yes, then added what was missing. You can also replace names with codes before sending, and you need to document what you have put in place.

No, turning off training is not enough. The setting protects your next conversations, not your clients’ personal information. It keeps your new conversations from being used to train the models. It does not change where the data is processed, how long the provider keeps it, or what Law 25 expects of your business. The providers’ pages were read in November 2026.

This article is provided for information purposes only, without any warranty, and is not legal advice. Have a lawyer validate any decision made on the basis of its content.


Where to turn off training in ChatGPT, Claude, Gemini and Copilot

In an individual account, free or paid, the training setting is in the tool’s privacy settings. The table gives its name, its location and its scope for the four assistants. The labels for ChatGPT, Claude and the former Copilot app come from the providers’ English help pages. Those for Gemini and the new Copilot app are our translation of the French screens, because the English help pages of Google and Microsoft do not reproduce them. The Gemini screen was read on November 19, 2026 and the Copilot screen on November 14, 2026.

No.ToolSetting nameWhere to find itWhat it changesWhat it does not change
1ChatGPT (Free, Plus, Pro)“Improve the model for everyone”Settings, then “Data controls”Your new conversations are no longer used to train OpenAI’s models.Your saved conversations stay in the account. A thumbs up given to a reply can make the whole conversation usable. Voice mode has a separate control.
2Claude (Free, Pro, Max)“Help Improve our AI models”Settings, then “Privacy”Your conversations, new or already stored, are no longer used in future training.What went into a training run already started or into a model already trained stays there. A conversation flagged by the safety classifiers can still be used.
3Gemini (personal Google account)“Keep Activity”“Gemini Apps Activity” page of the Google accountSetting turned off: your new conversations no longer help improve Google’s models, unless you send feedback. They are no longer saved in your activity.Google still keeps conversations for 72 hours anyway. According to its help page, those that human reviewers have already examined are kept for up to three years, even if you delete them.
4Copilot (personal Microsoft account)No training setting visible in the new app, available since August 18, 2026. The former app had “Training on conversation activity”.Former app: the profile, then “Privacy”Former app: your future conversations are excluded from training.According to Microsoft, this setting does not exclude your conversations from other product improvements, or from advertising, security and compliance.

Table sources: help pages from OpenAI, Anthropic, Google and Microsoft read in November 2026, and account screens.

Two rows need a clarification. For row 3, the screen is the authority. Google’s help page, updated on November 17, 2026, describes a “History” setting and other retention periods that the screen read in Canada on November 19 does not show. For row 4, the page Microsoft devotes to the new app refers to its privacy statement and describes no training setting. A screen check on November 14, 2026 did not show one either. With a personal Microsoft account, the safest course is therefore to paste nothing about a client into it.

In all my work accounts, training is turned off.

The temporary conversation, for what you do not want to keep

Each tool offers a conversation that stays out of the history, but none of them erases it right away.

  • ChatGPT. Temporary chat does not appear in the history, creates no memory and is not used for training. OpenAI may keep it for up to 30 days for safety (OpenAI).
  • Claude. The incognito chat opens with the ghost icon. It enters neither the history nor memory and is not used for training. Anthropic keeps it for 30 days (Anthropic).
  • Gemini. The temporary chat does not appear in recent chats and is not used to improve Google’s AI. It is kept for 72 hours for security reasons (screen read in November 2026).
  • Copilot. The “Start a temporary conversation” button opens a conversation that, according to the screen, “will not be saved or create memory”. This is our translation of the French screen, read on November 14, 2026.

The thumbs up can reopen what the setting closed

A thumbs up or a thumbs down is a voluntary submission that follows its own rules. In an individual ChatGPT account, the whole conversation linked to the thumb can be used for training, even after you opted out (OpenAI). With Claude, the entire conversation is kept for up to 5 years and can be used to train the models (Anthropic). With Gemini, according to Google’s help, feedback and the associated conversations are reviewed by specially trained teams and kept for up to three years (Google).

The rule for the team fits in one line: no thumbs up or down on a conversation that contains client information.

Memory is a separate setting

An assistant’s memory does not depend on the training setting. OpenAI says so in plain words: memory controls how replies are personalized and training controls the improvement of the models. They are two separate settings (OpenAI).

Deleting a conversation does not necessarily erase what the tool retained from it. In ChatGPT, you have to delete the saved memory and the original conversation (OpenAI). In Claude, memory entries drawn from a deleted conversation are not removed. They are deleted one by one (Anthropic). The article on ChatGPT’s secret codes explains what custom instructions and memory are for.

What has already left stays gone, with a nuance for each tool

The setting applies going forward. In training, I put it this way: “If it’s out there, it’s out there, you can’t take it back, but from now on you can turn it off.”

The nuance comes from Anthropic. Once the setting is off, Claude stops using already stored conversations, as well as new ones, in its future training. What is in a training run already started or in a model already trained stays there (Anthropic). OpenAI speaks of your new conversations. At Google, conversations that reviewers have examined are kept for up to three years, even if you delete them.


Team plans already turn off training, with two exceptions

In a team or business plan, training on your data is off by default, by contract. Nobody has to look for a setting. Two exceptions still require an action from the administrator.

  • ChatGPT Business (formerly Team), Enterprise and Edu. OpenAI uses neither the prompts, nor the responses, nor the files to improve its models. The thumbs up gives no right to train (OpenAI).
  • Claude Team and Enterprise. No training by default. First exception: a thumbs up can make the conversation usable for training, unless an owner has turned off the “Rate chats” option in the organization settings (Anthropic).
  • Microsoft 365 Copilot and Copilot Chat, with a work account. Prompts, responses and data read in Microsoft 365 are not used to train foundation models (Microsoft). Second exception: the web search queries that Copilot generates go to Bing. Microsoft’s data processing agreement does not apply to them (Microsoft).
  • Gemini in Google Workspace. Contents are neither reviewed by human reviewers nor used to train models outside your domain without your permission (Google).

These plans also come with a data processing agreement, often called a DPA. The A stands for Agreement or Addendum, depending on the provider. It is the contract that governs what the provider does with the personal information you entrust to it. The provider must process that information only to provide you the service, protect it and not keep it beyond that. The agreement exists because privacy laws require a written contract when a business entrusts personal information to a service provider. In Quebec, that requirement comes from section 18.3 of the Act, presented in the next section. OpenAI offers it through a form for ChatGPT Business and Enterprise (OpenAI). Anthropic’s is built into its commercial terms (Anthropic). At Microsoft and at Google, it is part of the business contract. An individual account has none.

A team plan does not settle location. Depending on the tools selected, your data is processed outside Canada, notably in the United States, with training turned off. Two nuances apply, and they have to be read by separating storage from processing. ChatGPT Enterprise and Edu can store data in Canada, but OpenAI does not offer processing there, and ChatGPT Business is not named among the eligible plans (OpenAI). Microsoft stores the content of Copilot interactions in Canada for an organization signed up in Canada (Microsoft) and announces processing in Canada for 2027.

To choose a plan and see the prices, refer to our article on bank reconciliation with ChatGPT, Claude and QuickBooks. The table in its section on choosing a plan compares the personal and team plans of ChatGPT and Claude, Microsoft 365 Copilot and Gemini in Google Workspace. The prices there are those of October 24, 2026.


What Law 25 requires beyond the training setting

The training setting leaves four questions untouched that the Act respecting the protection of personal information in the private sector, as amended by Law 25, asks. In all four cases, your client’s personal information has left your premises anyway. The quotations come from the text current to October 15, 2026, read in November 2026.

  1. The provider still keeps your data. Training on or off, your saved conversations stay with the provider until they are deleted. A temporary conversation is kept for 72 hours to 30 days depending on the tool. Anthropic keeps for up to 2 years a conversation flagged by its safety systems.
  2. You communicate personal information outside Quebec. Section 17 puts it this way: “Before communicating personal information outside Québec, a person carrying on an enterprise must conduct a privacy impact assessment.” The communication “must be the subject of a written agreement”. The Commission d’accès à l’information specifies that the rule covers an entity located outside Quebec, “both interprovincial and international” (page in French, our translation: CAI).
  3. You need an agreement with the provider. Under section 18.3, a business that entrusts personal information to a service provider must put the mandate or contract in writing. It must state in it the protection measures that the provider must take. An individual account comes with no data processing agreement. That is the first reason to move to a team plan.
  4. Your business remains responsible. Section 3.3 requires a privacy impact assessment for “any project to acquire, develop or overhaul an information system” that involves personal information. In our reading, adopting an AI assistant to process client information in it is one. At the show, the specialist insisted on one point: the business that subscribes is the one that has to assess its provider. An advisor can help, but does not carry the responsibility.

None of these points forbids using ChatGPT or Claude. They describe the work to do around the tool. The training setting is only the first step.

The concrete steps for each of the four questions

Each question calls for an action by your business, backed by a document that the provider gives you. These actions fall to the person in charge of the protection of personal information, the privacy officer. By default, that is the person with the highest authority in the business, who may delegate the role in writing (section 3.1).

No.QuestionWhat your business doesWhat the provider gives you
1Data retentionWrite into your policies how long a conversation that contains a client’s personal information stays in the account, then delete it once the file is closed. Section 3.2 requires your policies to cover the retention and destruction of personal information. For a one-off task, use the temporary conversation.Its retention periods, in its help pages or its agreement.
2Communication outside QuebecBefore the first send, carry out the assessment on the four factors in section 17: the sensitivity of the information, the purposes, the protection measures (including contractual ones) and the legal framework of the State where it is processed. Send only if it concludes that the information would receive adequate protection.Where the data is stored and processed, and its security measures.
3Written agreementSubscribe to the team plan, accept the data processing agreement on behalf of the business and keep a dated copy. Check that it covers the three points of section 18.3: confidentiality, use limited to the service and no retention after the contract ends.The data processing agreement: through a form at OpenAI, in the commercial terms at Anthropic, in the business contract at Microsoft and at Google.
4The business remains responsibleCarry out the assessment when you adopt the tool and consult the privacy officer from the start of the project (section 3.3).Its security and privacy documentation, which feeds your assessment.

Redact before sending, so the information never leaves

To redact is to replace each piece of personal information with a code before sending the text to the AI, then put the real values back into the answer. The tool works on codes. The lookup table stays with you.

The accurate word is de-identify. Under section 12 of the Act, personal information is de-identified “if it no longer allows the person concerned to be directly identified”. Because de-identification is undone with the table, the Act still speaks of personal information. Anonymization, on the other hand, is final. Under section 23, information is anonymized when it is reasonably foreseeable that it “irreversibly no longer allows the person to be identified directly or indirectly”. The Regulation respecting the anonymization of personal information (in French) sets its criteria.

By hand for a single text

For an email or a contract, the replacement is done by hand. Each name, each address and each number becomes a label, such as Client-017. You note the correspondence in a file that stays on your computer. The article on QuickBooks and bank reconciliation gives the rules to follow for an accounting export.

A tool when the volume grows

As soon as you process dozens of documents, a detection tool does the first pass. There are three families.

  • A free library installed on your computer, such as Presidio. Its default configuration contains models for English. French therefore requires one more configuration step. For Canada, its list of identifiers includes the social insurance number and the postal code. Its documentation warns that it does not guarantee finding all sensitive information.
  • A cloud detection service. The unredacted text then goes to the provider of that service, which also becomes part of your assessment.
  • A commercial tool in a container, installed in your own infrastructure.

A test run in October 2026 shows the limits of the first family. It covered about twenty fictitious names common in Quebec, placed in email sentences, short memos and bank statement labels. These labels imitated those of a statement, in capitals without accents. Presidio’s name detector, with a French model, fully masked all the names in the sentences and 90% of them in the memos. It masked none in the bank labels. Compound names are only partly masked: “Mohamed Ben Salah” was detected as “Mohamed Ben”. With the list of names supplied as a dictionary, all the names were masked, but the list contained exactly the names in the test. No real data was used.

The test shows what works best. A small in-house tool starts from the list of your clients and your employees, drawn from your own systems. It then lets the free library look for the rest. The list does the bulk of the work. The detector catches part of what the list does not contain.

Whatever the tool, a person reviews before sending.

A local model when redaction is not enough

Some documents do not redact well, because the personal information is the very subject of the text. A local model handles that case differently. It runs on a computer belonging to the business and nothing leaves. In exchange, it is less capable than the large models and it requires a machine equipped with a powerful graphics card. An ordinary graphics card is not enough to run a local model.

An example with team schedules

An SME wants the AI to propose the week’s schedule. Before sending, each employee’s name becomes a code, for example “employee QC3242”. The availabilities stay as they are. The AI places QC3242 on Tuesday morning. On the way back into the company’s system, the code becomes a name again and the employee receives their schedule. The AI provider saw no names.

What I do in my own tools

My email sorting agent processes messages from sensitive senders without AI: a script saves them and no cloud model reads them. In October 2026, when invoice amounts had to be read in those messages, a local model took care of it, on the graphics card of my computer. For the rest, I am setting up automatic redaction.


Document what you do, even when you conclude that nothing leaves

Documentation comes down to three pieces: the assessment, the agreement and the privacy policy.

  • The privacy impact assessment. It describes the tool, the information that goes into it, where it is processed, the settings in place and their date. Redaction appears in it, with the place where the lookup table lives.
  • The agreement. You keep a copy of the data processing agreement accepted on behalf of the business, with the date.
  • The privacy policy. It tells your clients and your employees what information you collect, for what purposes and to which providers it is released.

The specialist met at the show believes the assessment may no longer be needed if no personal information leaves, provided that this conclusion is documented. Our reading is more cautious, because section 12 treats de-identified information as personal information. It also requires a business that uses such information to take reasonable measures to limit the risk of anyone identifying a natural person from it. As long as your business keeps the lookup table, the redacted data therefore remains personal information for it. Redaction reduces the risk, but it exempts you from neither the assessment nor the agreement. If you conclude otherwise, write down why and have a lawyer validate that conclusion.


Four steps to take this week

Four steps are enough to start this week: turn off the setting, list the accounts, redact one document and write one page.

  1. Turn off the training setting in each individual account still in use, using rows 1 to 3 of the table, and note the date.
  2. List the accounts. Note who uses which tool and with which account. Anything that touches client or employee information goes through a team plan. With Claude Team, turn off “Rate chats”.
  3. Redact a real document by hand, once, with its lookup table. You will see what replaces well and what resists.
  4. Write one page that names the tool, where the data is processed, the settings and their date. It is the start of your assessment.

What Groupe Ordinata does to frame AI in SMEs

Groupe Ordinata Inc. is an artificial intelligence consulting firm based in Montreal, specializing in process automation and the integration of AI agents for Quebec SMEs. Our email triage agent has no sending rights and its privacy notice is public.

We bring to this more than twenty years of software engineering in industrial settings, including real-time systems and digital vision.

When an SME moves from an assistant to an agent, accounts are opened in the company’s name and model training on its data is turned off. Depending on the tools selected, data may be processed outside Canada, notably in the United States. We tell you which tools and where your data is processed before we start. We also prepare with you, before implementation, the documents the law requires, such as internal policies and the privacy impact assessment (PIA). Since Law 25, this assessment is required, among other cases, before entrusting personal information to a provider located outside Quebec.


Sort out your accounts and your documents with an outside view

The AI Program is a coaching program that moves forward one month at a time. It costs CA$2,200 before tax per month, for a maximum of 12 hours, with no commitment to continue. The first month starts from where your executives currently are. Depending on the scope, it begins with a short diagnostic. If the 30-minute call has already provided an initial analysis, it gets to the heart of the matter from the first session.

Prefer to start at no cost? Book a free 30-minute call, with no commitment. We discuss your tools and your needs to spot where AI could save you time, what precautions to plan and where to start.

Book a free 30-minute call


Frequently asked questions about ChatGPT training and client data

Does ChatGPT keep my documents?

Yes. Turning off training does not delete your saved conversations and does not hide them. To remove content, you have to delete the conversation or the account, according to OpenAI’s help read in November 2026. A temporary chat does not enter the history, but OpenAI may keep it for up to 30 days for safety.

Does turning off training erase what I already sent?

No. The setting applies going forward. OpenAI stops using your new conversations. Anthropic also stops using the conversations already stored in its future training. What is in a training run already started or in a model already trained stays there.

Should I redact before using ChatGPT?

Yes, as soon as a text contains personal information, such as the name of a client or an employee. Replace each name, each address and each number with a code and keep the lookup table on your computer. Redaction reduces the risk. It exempts you from neither the privacy impact assessment nor the agreement with the provider.

Is a team account enough for Law 25?

No. A team plan brings no training by default and a data processing agreement. It does not settle location: depending on the tools selected, data is processed outside Canada, notably in the United States. The privacy impact assessment and the privacy policy are still to be done. Have a lawyer validate your conclusions.

What is the difference between anonymizing and de-identifying?

To de-identify is to replace with codes the information that directly identifies a person. Because de-identification is undone with the lookup table, the Act still speaks of personal information (s. 12). Anonymization is irreversible: the information must no longer allow the person to be identified, directly or indirectly (s. 23). A text redacted with a table is de-identified, not anonymized.

Is a temporary chat deleted right away?

No. It does not enter the history and is not used for training, but the provider keeps it for a while. OpenAI speaks of 30 days at most for ChatGPT and Anthropic of 30 days for Claude, according to their help pages read in November 2026. Gemini’s screen indicates 72 hours.


This article is provided for information purposes only, without any warranty. It does not replace legal advice: have a lawyer validate any decision made on the basis of its content.

References

  1. OpenAI, Data controls in ChatGPT (updated around November 16, 2026, read on November 19, 2026): help.openai.com
  2. OpenAI, How your data is used to improve model performance (updated around November 16, 2026, read on November 19, 2026): help.openai.com
  3. OpenAI, Memory in ChatGPT (updated around October 31, 2026, read on November 19, 2026): help.openai.com
  4. OpenAI, Enterprise privacy at OpenAI (read on November 19, 2026): openai.com
  5. OpenAI, Data residency and inference residency for ChatGPT (updated on November 18, 2026, read on November 19, 2026): help.openai.com
  6. Anthropic, How do I change my model improvement privacy settings? (August 3, 2026, read on November 19, 2026): privacy.claude.com
  7. Anthropic, Is my data used for model training?, individual accounts (March 16, 2026, read on November 19, 2026): privacy.claude.com
  8. Anthropic, How long do you store my data? (July 1, 2026, read on November 19, 2026): privacy.claude.com
  9. Anthropic, Is my data used for model training?, commercial products (August 18, 2026, read on November 19, 2026): privacy.claude.com
  10. Anthropic, How do I view and sign your Data Processing Addendum (DPA)? (March 16, 2026, read on November 19, 2026): privacy.claude.com
  11. Anthropic, Use incognito chats and Use Claude’s chat search and memory to build on previous context (read on November 19, 2026): support.claude.com, support.claude.com
  12. Google, Gemini Apps Privacy Hub (updated on November 17, 2026, read on November 19, 2026) and the “Gemini Apps Activity” and temporary chat screens (read on November 19, 2026, in French): support.google.com
  13. Google, Generative AI in Google Workspace Privacy Hub (updated on October 6, 2026, read on November 19, 2026): knowledge.workspace.google.com
  14. Microsoft, Microsoft Copilot privacy controls, former app; Microsoft Copilot for individuals: your data, privacy, and responsible AI, new app (read on November 19, 2026): support.microsoft.com, support.microsoft.com
  15. Microsoft, Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat (updated on November 4, 2026, read on November 19, 2026): learn.microsoft.com
  16. Microsoft, Data, privacy, and security for web search in Microsoft Copilot and Microsoft Copilot Chat (updated on November 4, 2026, read on November 19, 2026): learn.microsoft.com
  17. Microsoft, Data Residency for Microsoft 365 Copilot Offerings (October 8, 2026, read on November 19, 2026): learn.microsoft.com
  18. Microsoft, Microsoft offers in-country data processing to 15 countries (November 4, 2025, read on November 19, 2026): microsoft.com
  19. Act respecting the protection of personal information in the private sector, CQLR, c. P-39.1, ss. 3.1, 3.2, 3.3, 12, 17, 18.3 and 23 (current to October 15, 2026, read on November 19, 2026): legisquebec.gouv.qc.ca
  20. Règlement sur l’anonymisation des renseignements personnels (Regulation respecting the anonymization of personal information), Order in Council 783-2024, Gazette officielle du Québec, May 15, 2024 (read on November 19, 2026, in French): publicationsduquebec.gouv.qc.ca
  21. Commission d’accès à l’information du Québec, Utilisation et communication de renseignements personnels (read on November 19, 2026, in French): cai.gouv.qc.ca
  22. Presidio, documentation, pages Home, Supported entities and PII detection in different languages (read on November 19, 2026): presidio.dataprivacystack.org

Groupe Ordinata Inc., artificial intelligence consulting for small and medium-sized businesses, Montreal, Quebec. Workflow automation · AI agents · ERP/CRM integration · Data architecture.

Ordinata.ca | Contact us