ChatGPT is OpenAI's conversational assistant, released as a free research preview on November 30, 2022 and built on a language model tuned with human feedback. On consumer plans, conversations are saved to history and may be used to train OpenAI's models unless the user switches that off in Data Controls. Temporary Chats are not used for training and are not kept longer than 30 days in OpenAI's systems, according to its help center.
Key takeaways
- The first ChatGPT model was GPT-3.5, fine-tuned with supervised learning and reinforcement learning from human feedback.
- Training use of consumer chats is controlled by a single account-wide toggle under Settings, then Data Controls.
- Deleted chats leave the account view at once and are scheduled for permanent removal within 30 days, with exceptions for security and legal reasons.
- A U.S. court order required OpenAI to preserve output logs from May 2025 until September 26, 2025, with some accounts still covered afterward.
- On May 6, 2026, Canadian privacy regulators found early ChatGPT training did not meet their privacy laws and recorded OpenAI's commitments.
How the assistant was built
ChatGPT descends from a method OpenAI described in the paper "Training language models to follow instructions with human feedback", submitted March 4, 2022. Human labelers wrote example answers and ranked model outputs, and the model was then adjusted toward the answers people preferred. In the paper's tests, evaluators preferred outputs from a model with 1.3 billion parameters over those of the 175-billion-parameter GPT-3 it was tuned from. The authors also report better truthfulness and less toxic output, and they note that the models still make simple mistakes.
That last point carries over to the product. A chat assistant predicts plausible text, so it can state wrong facts in a confident tone. Important output should be checked against a reliable source. Medical, legal, and financial questions are the cases where a mistaken answer costs the most, and an assistant's reply is not advice from a licensed professional.
What happens to a conversation
The table summarizes OpenAI's help-center statements as found in October 2026. The help pages block automated page fetches, so the wording below comes from the text returned in search results for those pages. Rules differ by plan and may change, and the live help pages are the reference.
| Situation | What OpenAI states |
|---|---|
| Consumer plan, default settings | Chats are kept in history and can be used to improve models unless the training toggle is off |
| Training toggle off | Chats still appear in history but are not used to train; the setting covers the whole account on every device |
| Temporary Chat | Not saved to history, does not create memories, not used for training; OpenAI may keep a copy for up to 30 days; may be reviewed only to monitor abuse |
| Deleting a chat | Removed from the account view immediately and scheduled for permanent deletion within 30 days, unless de-identified or needed for security or legal reasons |
| ChatGPT agent chats | Deleted chats and screenshots removed within 90 days |
| Business, Enterprise, Edu, Healthcare, and Teachers workspaces | Not used for training by default unless the customer opts in |
The court-ordered exception to deletion
Deletion promises met a legal limit in 2025. In May 2025, a federal court in the New York Times copyright case required OpenAI to preserve and segregate output log data that would otherwise have been deleted. A later order, filed October 9, 2025, ended that obligation going forward from September 26, 2025, according to Engadget. Logs already saved stay accessible to the plaintiffs, and OpenAI must keep data tied to accounts the Times flagged. OpenAI stated at the time that the order did not affect Enterprise or Edu customers.
The practical lesson is narrow. A deletion setting describes the company's ordinary practice, and a court can override it for a period. Anyone typing sensitive material should treat the retention terms as conditional.
What Canadian regulators concluded
A joint investigation by the federal Privacy Commissioner and the Quebec, British Columbia, and Alberta authorities reported on May 6, 2026. They found that the first training of ChatGPT raised problems of overcollection of personal information, consent and transparency, inaccurate statements about individuals, limited access, correction and deletion rights, and accountability. OpenAI has since significantly limited the personal and sensitive information used to train new models, and it committed to better inform Canadians and to take further steps. The federal Commissioner treated the complaint as well-founded and conditionally resolved, while the provincial offices assessed compliance under their own laws.
A short settings checklist
- Open Settings, then Data Controls, and review the training toggle before entering anything personal.
- Use Temporary Chat for one-off questions that should not leave a history entry.
- Leave out names, health details, account numbers, and employer data that the task does not need.
- Review the Memory settings, which OpenAI documents separately from chat history.
- For work, use an organization's managed workspace instead of a personal account, and follow its policy.
Scale and plans
Search Engine Land reported in February 2026 that OpenAI announced more than 900 million weekly active users for ChatGPT. Later claims of one billion or more circulate in blogs, but no OpenAI statement for them could be opened, so they are not repeated. Plan names and prices, including a lower-cost tier with ads, are reported by third parties and vary by country and date. Current prices in Canadian and U.S. dollars should be read from OpenAI's pricing page, not from articles.
FAQ
Does ChatGPT use user conversations to train its models?
On consumer plans it can, unless the training setting under Data Controls is switched off. Business, Enterprise, and education workspaces are excluded from training by default, according to OpenAI.
How long does OpenAI keep a deleted chat?
OpenAI says deleted chats are scheduled for permanent removal within 30 days, with exceptions for security, legal obligations, or data already de-identified. Agent chats follow a 90-day window.
Is Temporary Chat private?
It keeps the conversation out of history and training, but OpenAI states copies are retained for up to 30 days and may be reviewed only to monitor abuse. It is a retention control, not anonymity.
Is ChatGPT allowed under Canadian privacy law?
Regulators judged the early training practices as not compliant, then recorded changes and commitments by OpenAI. The findings address the company's practices, not whether an individual user or employer may use the tool, which depends on the organization's own obligations.
Related reading
- An earlier look at the OpenAI chat bot
- How GPT-3 works as a language model
- Natural language generation in chat bots
- Using AI chatbots for conversations
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Ouyang et al., Training language models to follow instructions with human feedback, arXiv:2203.02155, March 4, 2022
- OpenAI, Help Center: Data Controls FAQ, Temporary Chat, Chat and file retention in ChatGPT, and ChatGPT agent, accessed October 2026 via search results
- OpenAI, Enterprise privacy page, accessed October 2026 via search results
- Office of the Privacy Commissioner of Canada, news release on the joint investigation of OpenAI's ChatGPT, May 6, 2026
- Engadget, OpenAI no longer has to preserve all of its ChatGPT data, with some exceptions, October 2025
- Search Engine Land, OpenAI: ChatGPT now has 900 million weekly active users, February 2026
- ChatGPT, entry consulted for the November 30, 2022 release date and GPT-3.5 origin