Nina had a client complaint open beside ChatGPT when a news report stopped her hand over the paste button. She was a freelance designer, and the complaint needed a calm reply before an afternoon call. Her own draft sounded like someone arguing with a printer. The chatbot had helped with this sort of thing before. She paid for her personal account and had slowly started treating the empty message box as a room with the door closed.
The report concerned the February 10, 2026 shooting in Tumbler Ridge, British Columbia. OpenAI had banned an account associated with the shooter in June 2025. In its later letter to Canada's minister of artificial intelligence, the company described automated detection followed by human review. It said it had considered contacting police but had not found the credible, imminent planning needed under its rules at the time.
Nina read further. OpenAI contacted police after the attack, and Sam Altman later apologized for not alerting them about the banned account. Associated Press and Reuters reported the company's response. The event raised serious questions about how the warning had been handled. But another detail made Nina look back at the client document on her screen. Human review meant that the conversation could reach people working on the service.
She understood why a company would investigate threats. What bothered her was realizing how little she knew about the room she had imagined. She had pictured herself and a helpful machine, rather than a service with stored records and people who could review them. Her client's complaint included a private email address and details of a project that had not been announced. She returned to the client file without pasting it into the chat.
When her colleague Dan called about lunch, Nina asked whether paying for ChatGPT made a conversation private. He asked which account she meant. "It's my own paid account," she told him. Dan suggested opening the data settings instead. Nina opened them while he went to collect the sandwiches.
OpenAI's consumer data FAQ gave her a less cosy picture. A limited group of authorized staff and trusted service providers could access content when needed for tasks such as abuse investigations or support. Model improvement was another stated purpose, subject to the user's choice about training. OpenAI described access controls and logging of that access. She read the list again, paying attention to who was allowed access and why.
The same FAQ said content was stored on OpenAI systems and service providers' systems in the United States and elsewhere. Nina had been worrying about whether the model itself remembered her. She had barely considered the records kept by the service around it. A conversation did not have to become part of a model's training to exist on a server. She kept that page open while looking for the training controls.
She opened the model-improvement page next. For personal ChatGPT accounts, including paid Plus and Pro plans, eligible conversations could be used for training depending on the settings. She could turn off that use for new conversations. Paying for a personal plan had not automatically made that choice for her. Nina found Settings, then Data controls, and turned off Improve the model for everyone.
Her history stayed on the screen, which briefly annoyed her. Then she read the explanation. Turning off training did not delete the chats already saved in her account. Training use and keeping a chat history were separate choices, and opting out applied to new conversations. She had changed an important setting, but it was no reason to upload the original complaint now. The client had not given her permission to share it with this service.
Dan mentioned that business workspaces and the API had a different default. OpenAI said it did not train on their inputs and outputs by default, although organizations could agree to share data. Nina nearly decided that this settled everything. Then she noticed that the API documentation discussed retention separately. A service could promise no training and still keep data for another purpose. She kept the business page open for later and returned to her personal account.
Dan arrived with lunch while Nina was looking through her saved chats. She showed him an earlier request about a tense client email. The assistant had made her reply sound stiff, so she had asked for a warmer version with the same deadline. The next draft was much better. She had thought of this as getting her money's worth, and it was. It also looked like feedback a company could find useful when improving a writing assistant, if allowed to use it.
"Does it mark the answer right or wrong from what I say next?" Nina asked. Dan asked how the system would know that her own deadline was right. She could check the contract, but her message alone did not prove the date. She could confidently type the wrong one herself. And asking for warmer language mostly showed a preference about tone. It gave a different sort of feedback from checking a date against a document.
They searched for an explanation of how feedback was used and found OpenAI's published work on InstructGPT. It described people writing examples of desired answers and ranking model responses for later training. Nina could see how preferences could become useful training information through an organized process. Within her chat, a correction also gave the assistant context for its next reply. That did not show that its learned settings had changed. She still could not see which of her own corrections, if any, would be selected for later training.
Nina had sometimes pressed the feedback buttons after getting a useful draft. The current consumer policy said that submitting feedback could allow the associated conversation to be used for training even after opting out. She read that passage again. She had never thought to check the data rule before pressing a feedback button. She now treated it as a data-sharing choice and avoided sending feedback on private material.
Nina put down her sandwich. "So I'm paying them, and then helping train the thing I'm paying for?" she asked. Dan said that could be a fair description when her conversations were eligible for training. Her prompts supplied examples of real work, and her corrections could show where an answer disappointed her. The company could gain something useful alongside the subscription payment. She had budgeted for the fee without thinking about the possible reuse of her prompts and edited drafts.
She left the training switch off. If she wanted to contribute examples, she could choose material that was hers to share. She still wanted the service to help with her work. She also wanted any reuse of that work to be a decision she understood, rather than an assumption she had never checked.
She wondered whether deleting the older chat would undo everything. OpenAI's privacy policy described deletion within 30 days, with exceptions for matters such as safety or legal obligations. It also described an exception for content already separated from the account and stripped of identifying information. Nina requested deletion of the chat and saved the confirmation. She could not take that as a promise that earlier model training would be reversed.
She briefly considered keeping the whole drafting job on her computer and opened a guide to running AI locally and its practical limits. An offline model could draft text without sending it to a cloud service. A local app could still connect to a cloud model or an outside tool, though. She would need to check what actually ran on the machine and protect the local files. Nina had no tested local setup on that laptop, so she left the complaint out of this experiment.
For the reply still waiting on her desk, she considered Temporary Chat. The current help page said a chat was not used to improve models while it remained temporary. OpenAI could still keep a copy for up to 30 days for safety, and its privacy policy described cases requiring longer retention. Saving it as a regular chat would change which settings applied. Nina checked the current help page rather than relying on a tip from an old video.
That option helped with a particular kind of data use. It did not give her permission to send someone else's private file, or guarantee that nobody could review it for safety. Nina left the complaint on her laptop. She opened a blank document and wrote a made-up request for a reply about a missed design deadline. The assistant needed the tone and the basic problem. It could manage without the client's identity or the unreleased product name.
Even changing the name took some thought. Nina had once worked with a client whose unusual project description would have identified the company to anyone in that field. Replacing its name with Client A would hardly have hidden much. This time she removed those details too and described a general disagreement over timing. She read the request once as though she were the client, looking for anything she would be surprised to find in another company's system.
The assistant produced a useful draft from the made-up details. Nina put the actual deadline back in herself and checked it against the contract on her own computer. She could still get help with wording without handing over the whole argument. It took more care than pasting the original email. She compared the final wording with her first angry draft and kept the calmer version.
She sent the reply before the afternoon call. The assistant had helped with the wording, and her client had stayed out of its conversation. Nina still liked the tool. She just no longer mistook a subscription receipt for a locked door.