Does ChatGPT keep your data? An Irish guide to OpenAI
· Updated · Written and maintained by Joaquín Trapero, Nonimo
Yes, and the full answer is longer than the settings screen suggests. ChatGPT stores your conversations in your account until you delete them, and OpenAI says deleted data leaves its systems within 30 days, with named exceptions for legal process, abuse and financial records. Turning off model training does not change any of that. It removes one use of the text.
For a firm in Ireland the second question matters more than the first. The company that answers for all of this under the GDPR is OpenAI Ireland Limited, in Dublin 1, and the regulator that decides is the Data Protection Commission, three kilometres away on Pembroke Row. This guide works through what OpenAI’s own documents say, and what Irish law asks of you rather than of OpenAI.
Is ChatGPT confidential, and in what sense?
Not in the sense a client means by the word, and not in the same way on every plan. As of 25 September 2026, OpenAI’s help pages and its enterprise privacy page draw the line by account.
On Free, Go, Plus and Pro, OpenAI may train on your conversations by default. Turning off “Improve the model for everyone” keeps new conversations out of training, but they can still appear in your chat history, and a thumbs up or thumbs down on a reply means the entire conversation may be used to train the models even with the setting off. A Temporary Chat is not used for training while it stays temporary, but OpenAI may keep a copy for up to 30 days for safety purposes.
Business, Enterprise, Edu and the API are not trained on by default. Their content may still be run through automated classifiers, and on Business and the API a group of specialised contractors can review it for abuse, as the sections below set out.
So when people call ChatGPT confidential, they are describing OpenAI’s promises about its own staff and systems. The duty of confidentiality to your client is a different thing: it is yours, and no supplier’s privacy page discharges it. The practical step is to keep out what identifies the client: names, PPS numbers, file references. Nonimo, a Mac and Windows app, replaces names and identifiers with labels before the text reaches the chat, so the originals stay on your computer.
Does ChatGPT keep your data, and for how long?
OpenAI’s Europe privacy policy, updated on 24 August 2026, splits retention into three groups rather than giving one number. It is worth reading them in that shape, because firms tend to remember the first and act as if the other two do not exist.
| What OpenAI says | The period | The catch |
|---|---|---|
| Data you delete | Removed from systems within 30 days | Unless it was already de-identified for model improvement |
| Temporary Chats | Deleted automatically within 30 days | Unless kept for safety or legal reasons |
| Data kept for legal or safety reasons | For as long as the obligation lasts | Lawful subpoenas, banned accounts, financial records |
Source: OpenAI, Europe privacy policy, 24 August 2026, section 4.
The 30 day clock, and the three things that stop it
The policy names the exceptions rather than gesturing at them. Content or accounts banned for usage policy violations may be retained. Data covered by a lawful subpoena is kept for the duration of the obligation. Payment and transaction records are kept for accounting and regulatory purposes. There is a fourth, quieter one: when you ask for erasure, OpenAI keeps the audit record of the request so it can show it complied.
None of those exceptions is unusual. What matters is the shape. Your deletion request is a request to a third party who has already read the document, and it runs against a clock that other people can stop. How long Claude, Gemini and Copilot keep a conversation is set out in our other guides.
Temporary Chat, memory and what “delete” reaches
Temporary Chats are the one place where deletion is the default. OpenAI’s help pages, read on 25 September 2026, say that while a chat stays temporary it does not appear in history, is not used to improve the models and does not create or update memories, and that OpenAI may keep a copy for up to 30 days for safety purposes.
That last point is the one to remember: a Temporary Chat stays out of your history, not out of OpenAI. A personalised one can also draw on the memories and custom instructions you already have.
Memory is separate again. If it is on, ChatGPT keeps details across conversations, and those memories are deleted individually, not by clearing a chat. A firm that tells staff to “delete the chat afterwards” has not reached the memory, the export, or the copy an administrator can pull.
Training and passing through your data are not the same thing
This is the confusion that costs Irish firms the most, and it survives every settings change. There are two distinct events when someone puts a client letter into ChatGPT, and only one of them has a switch.
Call the first event disclosure. The text leaves the office and arrives at a company outside it. That happens the instant the request is sent, and no setting, plan or server location reverses it. The second event is use for model training, which is a decision OpenAI makes afterwards about text it already holds.
What turning off “Improve the model for everyone” removes
It removes one purpose. OpenAI states that for consumer services it may use your content to train its models, and that by default it does not train on inputs or outputs from business users, including ChatGPT Business, ChatGPT Enterprise and the API. The consumer switch lives under Settings and then Data Controls, applies to the whole account across devices, and can be changed at any time with no restrictions.
That is a real control and it is worth using. It is also the only thing on this page that the switch touches.
What it does not remove
It does not delete the conversation, shorten retention, stop an administrator on a business plan from reading it, stop a safety classifier from running over it, or put the disclosure back in the box.
If the document held a client’s PPS number, a bank account and a child’s initials, all of that has been sent to a company in another jurisdiction. Whether that is a reportable breach is decided by Irish and EU law rather than by a toggle, and our guide to that question goes through the test and the first hour.
Who can read what you put into ChatGPT
OpenAI’s enterprise privacy page, updated 8 January 2026, is unusually specific about this, and it differs by plan. On ChatGPT Enterprise, Edu and Healthcare, authorised OpenAI employees access conversations only to resolve incidents, to recover conversations with your explicit permission, or where required by law, and workspace admins get an audit log through the Compliance API. A plan named for healthcare still leaves an Irish practice to decide what a referral letter may carry into it.
ChatGPT Business is described differently, and the difference is the part to read twice.
Business accounts: your administrator, and a named third party
On ChatGPT Business, workspace admins can view, access, export and delete the conversations of anyone in the workspace. OpenAI’s access is limited to authorised employees who need it for engineering support, abuse investigation and legal compliance, and to “specialized third-party contractors who are bound by confidentiality and security obligations, solely to review for abuse and misuse.”
For a practice where the managing partner is also the workspace admin, that first sentence is a feature. For an associate drafting a personal matter on the firm’s account, it is a surprise. Both facts belong in the single page of rules that our AI policy template is written to produce.
Free, Go, Plus, Business, Enterprise: what actually changes
The plan decides the contract, and the contract decides the answer. This is the single largest lever a small Irish firm has, and it costs nothing to pull.
| Plan | Trained on by default | Who else can see it |
|---|---|---|
| Free, Go, Plus, Pro | Yes, unless you turn it off | OpenAI staff and safety systems |
| ChatGPT Business | No | Your workspace admin, plus abuse review contractors |
| Enterprise, Edu, Healthcare | No | Authorised OpenAI staff for incidents or law |
| API Platform | No, since 1 March 2023 | Authorised staff and abuse review contractors |
Source: OpenAI, Enterprise privacy at OpenAI, 8 January 2026.
OpenAI will execute a Data Processing Addendum for ChatGPT Business, ChatGPT Enterprise and the API. That is the document your insurer and your clients will ask for, and it is the same document the cyber questionnaire expects you to name.
Ads arrived in ChatGPT, and they can read your chats
This is new since the August 2026 policy and most Irish coverage has not caught up with it. OpenAI now collects ads history and interests for Free and Go users, and where ads personalisation is available and switched on, it uses past chats to make the ads more relevant. The legal basis it names for that is consent.
Consent in the GDPR sense is a high bar, and it is withdrawable. The practical point for a firm is narrower and sharper: the free tier is now an advertising product, and the content of a conversation is one of its inputs.
A client matter is not an input you want in an advertising pipeline, whatever the consent screen said. It is also the clearest illustration of the distinction this guide is built on. Nothing about advertising is model training, the training switch does nothing to it, and it is a second use of text that was already sent. Paid plans are outside it, which is one more reason the account decision comes first.
Connectors, apps and web search: the second door out
Most of this guide is about the front door, the box you type into. There is a second door, and it opens outwards. ChatGPT can send and retrieve information from connected services, and the policy is explicit that what you share with partners outside OpenAI is governed by their own terms and privacy policies.
Why a connector is harder to govern than a chat box
A connector works out of sight: nobody sees it at the moment it is used. A connector to a document store hands ChatGPT a permissioned view of your files. An app or a web search step passes a fragment of the request to a provider you never contracted with.
OpenAI says it does not train on data accessed from apps by default on business plans. That is a real commitment, and it is a different promise from the data staying inside the perimeter.
| Route | Who receives the text | What governs it |
|---|---|---|
| The chat box | OpenAI | Your plan and, on business plans, the DPA |
| Connected apps and connectors | The other company’s service | That provider’s own terms |
| Web search steps | The search provider | Outside the residency region |
| Business account export | Your workspace admin | Your own internal rules |
Source: OpenAI, Europe privacy policy, 24 August 2026, section 3; Help Center, Data residency, consulted 19 September 2026.
The practical rule is short. Every connector an administrator enables is a new processor in your chain, and it should be named in the register you already keep. Our partners page exists because in most Irish firms of this size the person who enables them is an external provider.
Where OpenAI keeps Irish data, and what European residency does not cover
OpenAI offers data residency in Europe, covering the EEA and Switzerland, and it is included at no extra cost on the plans that qualify. It is also narrower than the phrase suggests, and the help centre says so plainly.
New workspaces only, and Enterprise or Edu only
Apart from those, only eligible API customers qualify. A Plus subscription does not have it. A ChatGPT Business workspace is not on the list. An existing Enterprise workspace that was not provisioned with residency does not acquire it by asking nicely.
The categories that stay outside the region
Even with residency on, OpenAI names data that may sit elsewhere: anything handled by external integrations such as apps, connectors or web search, transient processing steps, workspace metadata, the workspace name, billing information and user logins. Inference residency, which keeps GPU execution in region, is a separate setting available in Europe, the United States and the United Arab Emirates, and OpenAI states that it does not confine all processing: authentication, routing and analytics may still happen outside.
| What residency covers | What it does not |
|---|---|
| Conversations, files, images, memory, at rest | Web search, apps and connectors |
| Backups and replicas in region | Workspace metadata, billing, logins |
| With inference residency, GPU execution | Authentication, routing, analytics |
Source: OpenAI Help Center, Data residency and inference residency for ChatGPT, consulted 19 September 2026.
The line worth carrying into a client meeting is the one this guide keeps returning to. A server in Frankfurt does not unsend a document. It changes where a copy rests, which is a genuine improvement, and it changes nothing about the fact that the text left your office.
That distinction is also what an insurer is testing when it asks where your data is processed, and a residency region is a better answer than a shrug. The cyber questionnaire guide sets out how to answer it without overclaiming, and the Irish cover guide covers what the policies here actually say about AI.
The regulator is in Dublin, and that is not a technicality
Here is where Ireland stops being a footnote. OpenAI’s Europe privacy policy names OpenAI Ireland Limited, at 1st Floor, The Liffey Trust Centre, 117 to 126 Sheriff Street Upper, Dublin 1, as the controller responsible for personal data of people in the EEA and Switzerland. The same policy tells EEA residents that they can reach out to “the Irish Data Protection Commission as our lead supervisory authority.”
What the DPC can actually do
The DPC opens statutory inquiries under section 110 of the Data Protection Act 2018. It did exactly that to Google Ireland Limited on 12 September 2024 over whether a Data Protection Impact Assessment was required before developing the PaLM 2 model, and that inquiry was announced as a cross-border inquiry under the one-stop-shop mechanism. The same machinery is available against any controller whose main establishment is here.
The opinion that Dublin asked for
In September 2024 the DPC used Article 64(2) of the GDPR to ask the European Data Protection Board for an opinion on AI models, and the Board delivered it on 17 December 2024. The DPC’s four questions included when an AI model may be considered anonymous, and how a controller demonstrates that legitimate interests is the right basis for developing and for deploying a model.
That is worth sitting with. The European rules that govern what OpenAI, Google, Anthropic and Microsoft may do with training data were shaped by a request from the Irish regulator. If you practise in Ireland, the authority you would write to is the one that asked the question.
It cuts both ways, and firms should know the second edge. A regulator that is lead authority for most of the industry carries most of the industry’s caseload, so an Irish complaint joins a queue that is European in scale rather than national.
In 2025 the DPC received 16,160 new cases from individuals, a rise of 45 per cent, took in 6,521 valid breach notifications and concluded 208 valid cross-border complaints, up 43 per cent on the year before. That is a great deal of work and not a great many complaints. The EU AI Act guide covers the separate regime that now sits on top of this one.
What Irish law asks of you, not of OpenAI
Your obligations do not move because OpenAI has a Dublin entity. You remain the controller for your client’s file, and the GDPR and the Data Protection Act 2018 apply to what you decided to send. A local authority carries those same controller duties with two more regimes on the same file, and what a council can put in a prompt narrows accordingly.
The Law Society line that closes the argument
On 12 November 2025 the Law Society of Ireland published guidance for the profession on generative AI, written against the Solicitors’ Guide to Professional Conduct. It names the four systems by name, Copilot, ChatGPT, Claude and Gemini, and it says this about the versions most people use: “By default, free and paid consumer versions of GenAI systems are not suitable for securely handling personal data or client confidential data.”
It does not stop at enterprise either. The guidance warns that solicitors “should never assume that using an enterprise-grade tool guarantees compliance”, and asks for due diligence and a detailed review of the terms before use. That is a workable standard for a firm of twelve people, and it is the one an Irish regulator would read back to you. What replaces that assumption is a criterion by criterion look at each assistant.
Privilege, and the document that left the firm
The same guidance makes a point that has nothing to do with data protection and everything to do with litigation. Intentionally providing privileged communications to a free or paid model without appropriate safeguards may lose the benefit of privilege, because it is an intentional release of a privileged document to a third party outside the firm. The High Court now says the same in the practice direction on AI in court documents.
Among the safeguards it lists is a contractual zero data retention period. OpenAI offers zero data retention on eligible API endpoints for qualifying use cases, on request. It is not a checkbox in ChatGPT, and an external IT provider is usually the one who knows whether your tenant has it.
How to turn training off, and delete what is already there
-
Decide the account first. Which login staff use decides which contract governs the text, and it is the only step on this list that changes the legal position rather than the settings.
-
Turn off “Improve the model for everyone”. Account menu, then Settings, then Data Controls. It applies to the whole account across web and mobile, and Codex tasks on a personal plan follow it.
-
Export before you delete. The account export carries dates. A screenshot does not, and dates are what a breach assessment runs on.
-
Delete the conversation, then the memories. They are two different stores, and clearing one does not clear the other.
-
Ask your provider about the workspace region. Owners and administrators can see the Data Residency Region under Workspace settings, General. If the field is missing, it was never configured.
-
Write the rule down and collect the acknowledgements. An unsigned rule is hard to enforce, which is why our template ends with a signature block.
-
Diary the review. Every document quoted in this guide carries a date, and three of them changed in 2026 alone. A setting you checked last winter is not a setting you can describe to a client today.
Is ChatGPT safe for client data in Ireland?
Safe is not a property software has, and no Irish regulator has said the word about any of these products. The DPC’s own AI guidance of 18 July 2024 does not name a tool either.
It asks a controller to understand, before it starts, what personal data the system uses, where that data goes when a third party is involved, whether the provider retains or reuses it, and how the product lets you meet your obligations. Those four questions are answerable for ChatGPT from the documents quoted above, which is more than can be said for a good deal of software.
What personal data the system uses.
Where that data goes when a third party is involved.
Whether the provider retains or reuses it.
How the product lets you meet them.
On a consumer plan, the Law Society has already told solicitors it is not suitable for client confidential data by default, and the same reasoning reaches accountants, brokers and anyone else carrying a professional duty. On a business or enterprise plan with a signed DPA, it becomes an ordinary processor question, answered with due diligence rather than with a toggle. Chartered accountants face a stricter CCAB-I line, so whether accountants can use ChatGPT turns on the data itself.
Neither answer is changed by the training setting, and that is the whole point of this guide. The same question for the other three products is in our Claude guide, our Gemini guide and our Copilot guide.
What a tool can do here, and what no tool can do, including ours
Software of this category masks identifiers before the text is sent, and it does not make an organisation compliant. What ours does is pseudonymisation: the mapping back to the person is kept, encrypted, on your own machine, so it is reversible by design. For you the result is still personal data under the GDPR, and for OpenAI it turns on what else it could match the text against.
In practice that is a Mac or Windows app: select the text, press one key, and the identifiers it recognises become labels before the paste, with the originals restored in the reply. A payment card number is cleaned silently; a PPS number is shown to you to confirm, because the app cannot tell whether it belongs in the document, and who is allowed to use a PPS number is set out in a statute of its own.
What the app keeps is on our security page: the map stays encrypted on your computer.
If your answer is to change the account and write one page of rules, that is a good answer and it costs nothing. Our organisations page is for the firms that get there and want the control as well.
Sources
- OpenAI, Europe privacy policy, updated 24 August 2026. The controller for the EEA and Switzerland is OpenAI Ireland Limited; the Irish Data Protection Commission is named as lead supervisory authority; the 30 day deletion window and its exceptions; Temporary Chats deleted within 30 days; ads for Free and Go users on the basis of consent. Source
- OpenAI, Enterprise privacy at OpenAI, updated 8 January 2026. No training on business data by default; who can view conversations on Enterprise, Edu, Healthcare and Business; the abuse review contractors on Business; DPA availability; API retention of up to 30 days and zero data retention on eligible endpoints. Read again 25 September 2026: business data may be run through automated content classifiers and safety tools. Source
- OpenAI Help Center, Data residency and inference residency for ChatGPT, consulted 19 September 2026. Eligibility limited to eligible API customers and new Enterprise or Edu workspaces; Europe covers the EEA and Switzerland; the categories that may sit outside the region; inference residency does not confine all processing. Source
- OpenAI Help Center, Data controls in ChatGPT, read 25 September 2026. With “Improve the model for everyone” off, new conversations are not used for training but “can still appear in your chat history”; Temporary Chats may be retained for up to 30 days for safety purposes. Source
- OpenAI Help Center, How your data is used to improve model performance, read 25 September 2026. Consumer content may be used for training by default; opting out in Settings or the Privacy Portal; feedback with a thumbs up or thumbs down may put the entire conversation into training even after opting out. Source
- OpenAI Help Center, Temporary chat in ChatGPT, read 25 September 2026. Not used to improve models while it remains temporary; OpenAI may keep a copy for up to 30 days for safety purposes. Source
- Data Protection Commission, Data Protection Commission launches inquiry into Google AI model, 12 September 2024. A cross-border statutory inquiry under section 110 of the Data Protection Act 2018, into whether a DPIA was required before developing PaLM 2. Source
- Data Protection Commission, Irish Data Protection Commission welcomes EDPB opinion, 18 December 2024. The opinion was sought by the DPC in September 2024 under Article 64(2) GDPR; its four questions, including when an AI model may be considered anonymous. Source
- Data Protection Commission, Annual Report 2025, published 30 June 2026. 16,160 new cases from individuals, up 45 per cent; 6,521 valid breach notifications; 208 cross-border complaints concluded as lead supervisory authority. Source
- Law Society of Ireland, guidance on the use of generative artificial intelligence, announced 12 November 2025, PDF edition dated December 2025. Names Copilot, ChatGPT, Claude and Gemini; by default, free and paid consumer versions are not suitable for securely handling personal data or client confidential data; never assume an enterprise tool guarantees compliance; privilege may be lost by intentional release to a third party; zero data retention listed among appropriate safeguards. Source
Nonimo is the software that does this on your own computer: it masks client names and IDs before your text reaches ChatGPT . No account, and your client's details never leave your machine.
Common questions
Does ChatGPT keep your data?
Yes. Conversations are stored in your account until you delete them, and OpenAI says deleted data is removed from its systems within 30 days unless it must be kept for legal, safety or accounting reasons. Temporary Chats are deleted automatically within 30 days.
Is ChatGPT confidential?
Not in the sense a client means, and it depends on the plan. On Free, Go, Plus and Pro, OpenAI may train on your chats unless you turn that off, and Temporary Chats may be kept for up to 30 days for safety purposes. Business, Enterprise, Edu and the API are not trained on by default, but automated classifiers can still run over them. The duty of confidentiality to your client stays with you.
Does turning off training in ChatGPT delete my chats?
No. OpenAI's own answer is that new conversations you keep out of training can still appear in your chat history. The setting removes one use of the text. It does not remove the text, end retention, or undo the disclosure to a third party.
Who is the data controller for ChatGPT in Ireland?
OpenAI Ireland Limited, registered at the Liffey Trust Centre, Sheriff Street Upper, Dublin 1. OpenAI's Europe privacy policy of 24 August 2026 names it as the controller for the EEA and Switzerland, and names the Irish Data Protection Commission as its lead supervisory authority.
Can I complain to the Data Protection Commission about ChatGPT?
Yes. OpenAI's own policy points EEA residents to the DPC as its lead supervisory authority, or to their local authority. The DPC handled 16,160 new cases from individuals in 2025, a 45 per cent rise on the previous year.
Does ChatGPT store Irish data in Europe?
Only for some customers. OpenAI offers data residency in Europe for eligible API customers and new ChatGPT Enterprise and Edu workspaces. Consumer plans are not on that list, and workspace metadata, billing and logins sit outside the region in any case.
Does OpenAI use my chats to show me ads?
For Free and Go users it can, with consent. The Europe privacy policy of 24 August 2026 says past chats may be used to make ads more relevant where ads personalisation is available and switched on, and that this is based on consent.
Is ChatGPT allowed for client files in an Irish solicitor's office?
The Law Society of Ireland's guidance, announced on 12 November 2025, says that by default, free and paid consumer versions of GenAI systems are not suitable for securely handling personal data or client confidential data. It points firms to enterprise versions of these tools and to due diligence on the terms.
Does deleting a ChatGPT conversation remove it from the training data?
Not from models already trained. OpenAI says deleted content is removed within 30 days unless it has already been de-identified and disassociated from your account when you allowed it to be used to improve the models.
What should an Irish firm do first about ChatGPT?
Decide which account staff sign in with, because that decides which contract governs the text. Then write one page saying which documents may never be put into it, and record who acknowledged the rule.