Is ChatGPT Safe for Confidential Documents? What It Keeps in 2026
Not on Free, Go, Plus or Pro, where OpenAI may train its models on your chats unless you switch that off, and even deleted chats can stay on its servers for up to 30 days. ChatGPT Enterprise, Edu and Healthcare come closest to safe, because they don't train on your data by default and they limit when OpenAI staff can look at it.
Quick answer
- Free, Go, Plus and Pro: keep confidential documents out. Chats may be used for training until you opt out.
- Business doesn't train on your data, but your workspace admins can read and export your chats.
- Deleting a chat does not delete a file you saved separately, such as in your Library or a project. Delete both.
- US patient data needs a HIPAA-eligible product with a signed business associate agreement (BAA). Business is not on that list.
Which ChatGPT plan can you trust with confidential files?
The plan you pay for decides almost everything: whether OpenAI trains on your files, how long it keeps them, and who can read them. This table comes from OpenAI's business data, enterprise privacy and retention pages.
| Plan | Fit for confidential documents | Trains on your data by default | How long chats are kept | Who can read them |
|---|---|---|---|---|
| Free, Go, Plus and Pro | No | Yes, unless you opt out | Until you delete them, then removed within 30 days | A limited number of OpenAI staff and service providers |
| Business (once called Team) | Only if you accept admin access; not on OpenAI's HIPAA-eligible list, so no BAA for patient data | No | Admins set it; deleted chats removed within 30 days | Your workspace admins, plus OpenAI staff for support, abuse and legal needs |
| Enterprise and Edu | Yes, with admin controls; patient data only on Enterprise with Regulated Workspace | No | Admins set it | OpenAI staff only for incidents, recovery with your permission, or the law |
| ChatGPT for Healthcare | Yes, including patient data (comes with a BAA) | No | Admins set it | Same limits as Enterprise |
| API (for developers) | Yes, for apps you build; BAA with Modified Retention | No, since 1 March 2023 | Abuse logs up to 30 days | OpenAI staff and contractors who review abuse |
The Business row surprises people. Team became Business on 29 August 2025 (rename FAQ), and its admins "can view, access, export, and delete end user conversations" (enterprise privacy). So a company workspace keeps your files away from training, but not away from your employer.
On Enterprise, Edu and Healthcare, workspace admins can also open an audit log of conversations (enterprise privacy). Our ChatGPT Business privacy guide covers this in more detail.
Business, Enterprise, Edu, Healthcare and the API are covered by a SOC 2 Type 2 security audit, and OpenAI encrypts stored data with AES-256 and data on the move with TLS 1.2 or later (enterprise privacy).
Who can read what you put into ChatGPT?
ChatGPT is not confidential on personal plans: a limited number of OpenAI staff and trusted service providers can read chats. OpenAI's consumer data page lists four reasons: to look into abuse or security problems, to give support you asked for, for legal matters, and to improve the models unless you opted out.
OpenAI's own advice is: "Please do not enter sensitive information that you would not want reviewed or used" (consumer data page).
People outside OpenAI, and courts, can reach chats too:
- 404 Media reported on 14 September 2026 that contractors are paid to read real prompts, and that OpenAI "acknowledged sensitive details can still get through" its filter for personal data (404 Media). Our Project Lily page covers the report.
- A proposed class action, Vredenburgh v. OpenAI OpCo, filed on 16 September 2026, says users were not told about this (Decrypt, court docket).
- In the New York Times copyright case, a 13 May 2025 order made OpenAI keep consumer, Team and API chats, even deleted ones, until 26 September 2025 (OpenAI's update). Enterprise and Edu were left out of the hold.
- A later order in the same case covers 20 million consumer chats with names and other details removed (ABA Journal).
Does ChatGPT train on your documents?
On Free, Go, Plus and Pro, yes, it may, unless you switch training off. The privacy policy says OpenAI "may use Content you provide us to improve our Services, for example to train the models that power ChatGPT".
OpenAI says switching it off stops your new chats being used, and does not say what happens to chats you already had.
Even after you opt out, "If you choose to provide feedback, the entire conversation associated with that feedback may be used to train our models" (OpenAI data use). So a thumbs up or down can send the whole chat for training. On Business, Enterprise and Edu, feedback does not give that permission.
Training matters for a private file because researchers have pulled text back out of trained models. A 2023 paper made gpt-3.5-turbo repeat its training data "at a rate 150x higher" than normal, for about $200 (2023 paper). The authors say the attack "is specific to this model", and newer models were not tested.
What happens to a file after you upload it?
An uploaded file follows the same training setting as your chats. OpenAI says how documents "may be used to improve models depends on the service you use and your data settings" (file uploads FAQ).
Deleting the chat is not enough, because saved files have their own rules:
- OpenAI's retention page says "Deleting a chat does not delete a separately saved file" (retention policies).
- Files in a custom GPT stay until the GPT is deleted, and project files stay until you delete the project or the file, or a workspace retention policy removes them. Removal is then scheduled within 30 days.
- In Enterprise workspaces, OpenAI says its backups may keep a deleted Library file for up to 30 more days.
- OpenAI Support wrote on 6 October 2026 that custom GPT knowledge files "should not be treated as a secure place for sensitive or proprietary raw data" (forum answer).
ChatGPT can't open files on your computer by itself. A normal chat only sees what you attach.
That changes when you connect your computer or browser. OpenAI's Work mode "can use permitted files and tools on a connected computer" (Work mode documentation), and the browser extension "can read or act on sites where you're already signed in" (extension documentation).
How do you use ChatGPT for confidential information more safely?
Use a company plan, switch training off, and take names and numbers out before anything leaves your computer. None of these steps turns a personal account into a safe place for files you are bound to protect, but together they close the biggest gaps.
- Use your company's Business, Enterprise or Edu workspace for work files, never a personal account.
- On a personal plan, open Settings, then Data controls, and switch off "Improve the model for everyone".
- Don't rate sensitive chats with a thumbs up or down, since feedback can send the whole chat for training.
- Remove names, account numbers and other details that point to a person before you paste or upload. The free anonymize text tool can strip names from a short passage first.
- Delete chats and Library files you no longer need, and disconnect apps and computer access you don't use.
What are the real ChatGPT security risks in 2026?
The risks come from stolen logins, fake add-ons, flaws in ChatGPT's own features and, once, a breach at an outside vendor. None of the cases below was a break-in at OpenAI's own servers. Security vendor SOCRadar found exposed AI logins for 482 companies in malware logs, with OpenAI sessions at 358 of them (Security Affairs).
| Date | What happened | What was exposed | Source |
|---|---|---|---|
| January to October 2023 | Malware logs holding ChatGPT logins offered for sale | More than 225,000 logs | The Hacker News |
| 31 July 2025 | OpenAI removed the option to make shared chats discoverable by search engines | Shared chats showing up in search engines | Search Engine Journal |
| November 2025 | Analytics vendor Mixpanel was breached | Names, emails and rough location of some API users and a limited number of ChatGPT users; no chats or passwords | OpenAI |
| 26 January 2026 | At least 16 fake ChatGPT browser extensions found | About 900 downloads | Akamai |
| Fixed 20 February 2026 | Check Point found a hidden channel in the code sandbox | One bad prompt could send chat data out | Check Point |
| Disclosed 25 September 2026 | OpenAI research agents posted user images to public image hosts as unlisted links | 53 user images | BleepingComputer |
| 15 October 2026 | Aoki, a data-stealing program, started targeting ChatGPT cookies | ChatGPT sessions and coding tool settings | Infosecurity |
Three of these cases began on someone's own computer, and a stolen login cookie hands over your whole chat history. Keeping fewer confidential files in ChatGPT means there is less to lose if your account is taken. Our ChatGPT hacking history covers older incidents.
How often do employees share sensitive data with ChatGPT?
Often. Cyberhaven's 2026 report found that staff put sensitive data into AI tools about once every three days, and that 32.3% of ChatGPT use at work runs through personal accounts, outside company control (Cyberhaven 2026).
LayerX, a browser security company, reported in October 2025 that 77% of employees paste data into AI tools, and 82% of those pastes come from personal accounts the company does not manage (LayerX report).
IBM's 2026 Cost of a Data Breach report, summed up by Baker Donelson, found that incidents involving shadow AI, meaning AI tools staff use without approval, rose to 43% from 20% in 2025, and cost more ($5.39 million versus $4.63 million) (IBM 2026 summary).
These firms sell security products, so treat the numbers as vendor research. A well-known real case is older:
- In 2023, Samsung engineers reportedly pasted confidential data into ChatGPT at least three times in under three weeks, including source code and a record of a meeting (The Register).
- In May 2023 Samsung banned generative AI on company devices for a time, after about 65% of staff in an internal survey said they saw a security risk (TechCrunch).
- Several banks restricted the tool that year too, as did Apple and Verizon (Fortune).
Can lawyers, doctors and accountants use ChatGPT?
Only on the right plan and inside the rules of their profession. Lawyers in the US follow ABA Formal Opinion 512 of 29 July 2024, which says they need informed consent before putting the information of the people they represent into AI tools that learn from it (ABA 512 summary, 2Civility).
California's SB 574, signed on 30 September 2026, starts on 1 January 2027. It stops lawyers putting confidential or personal information into an AI tool unless only people bound by confidentiality can get at it (Bloomberg Law, BakerHostetler).
Doctors and other health workers in the US need a BAA for any cloud service that holds patient data, even if the data is encrypted (HHS guidance). ChatGPT for Healthcare, launched on 8 January 2026, comes with one (Healthcare launch, Axios), and OpenAI keeps a list of HIPAA-eligible products.
Accountants, consultants and finance teams usually have confidentiality terms in their contracts. Those terms often cover AI tools too, so read them before a customer file goes anywhere.
Ask questions about confidential documents without sending them to ChatGPT
If you want AI to answer questions about a contract under NDA, a board report or a case file, but don't want those files stored with a chatbot, Elephas keeps them on your own computer.
Elephas is a private AI for your documents. It answers questions from your own files and shows which file each answer came from.
Before your question and the matching passages go to a cloud AI, Elephas finds personal details such as names, email addresses and phone numbers, and swaps them for placeholders. When the answer comes back, it puts the real words back in, and a list under each answer shows exactly what was hidden.
- The built-in Elephas AI does not store your prompts or use them to train AI models.
- You can also use your own API key, or download a model that runs on your computer so nothing is sent at all.
- On the Team plan, each member's chats stay private, and the Owner and Admins cannot read them (Elephas security).
Elephas runs on Mac (Apple Silicon, macOS 14 or later) and Windows 10 or 11. There is a 7-day free trial with no card, and plans start at $19 a month. For a side-by-side look at other options, see our private AI tools guide.
FAQ
What should I do if I already put a confidential document into ChatGPT?
Delete the chat, then delete the file from your ChatGPT Library too, because deleting the chat leaves the file behind. On a personal plan, switch training off so later chats are not used.
Tell your company's security or privacy team the same day, since some data rules set a short deadline: under the EU's GDPR, a company must, where feasible, tell its regulator within 72 hours of learning of a personal data breach that is likely to put people at risk (GDPR Article 33).
Do Locked Chats encrypt my conversations?
No. Locked Chats, which started rolling out on 20 October 2026, put chosen chats behind a PIN, with Face ID as an option on supported phones, but OpenAI says they create no "end-to-end encryption or a separate encrypted vault" (Locked Chats FAQ). They stop someone reading your screen, not OpenAI or a court.
Does using ChatGPT waive legal privilege?
Courts don't agree yet. In February 2026, Judge Jed Rakoff held in United States v. Heppner that about 31 documents made with a public AI chatbot were not privileged (Venable).
Other courts have protected some AI work as trial preparation, so assume the other side may ask for your chats. Our guide to ChatGPT and legal privilege has the detail.
Does turning off Memory delete what ChatGPT knows about me?
Not by itself. Turning Memory off does not delete your past chats (Memory FAQ). To remove what ChatGPT has remembered, delete it from the memory summary or saved-memory controls, and delete the chats separately.
Is Temporary Chat safe enough for a confidential file?
Not on its own. A Temporary Chat "is not used to improve OpenAI models while it remains temporary", but OpenAI can keep a copy for up to 30 days for safety purposes (Temporary Chat FAQ).
If you save it, it becomes a normal chat and follows the usual rules for saved chats. It suits a quick question with nothing private in it, not a contract or a customer file.
Private answers from your own documents
Elephas answers from the files on your Mac and shows the source for every answer. It hides personal details like names, emails and phone numbers on your computer before a question goes to cloud AI. Plans start at $19 a month.







