AI & Future
AI tools are helpful, but they can quietly collect what you feed them. Here is practical, jargon-free advice on guarding your personal data without giving up the tools.
The single highest-impact move is to open each AI tool's settings and switch off "use my data to train the model," then adopt one hard rule: never paste a password, a card number, or someone else's private records into a prompt. Everything else here refines those two actions.
When you type into ChatGPT, Gemini, Claude, or Copilot, your text leaves your device, travels to the provider's servers, gets processed, and — depending on the tier and your settings — may be logged, retained, sampled by human reviewers, and folded into the training data for a future model. The details differ sharply by tier, and that distinction is the whole game.
Consumer tiers (the free apps and low-cost personal plans) are the ones that historically lean toward using your conversations to improve the product. Business and developer tiers usually flip that assumption. OpenAI's API and ChatGPT Enterprise/Team, Microsoft 365 Copilot, and Anthropic's commercial API all state they do not train their foundation models on your inputs by default — Microsoft brands this "Commercial Data Protection." So the first question is not "is this tool safe?" but "which tier am I on?" The same brand can behave very differently depending on whether you signed in with a personal or a work account.
One reality check: your own deletion settings are not absolute. In 2025, a court order in the New York Times v. OpenAI copyright case required OpenAI to preserve ChatGPT logs — including conversations users had deleted — for standard consumer and API tiers. A legal hold can override the delete button. That is not a reason to panic; it is a reason to keep out of a prompt anything you would not want surfaced in a lawsuit.
This is the highest-leverage setting, and it is only off-by-default in your favor on the paid business tiers. Menus move, so treat these as "look here":
Go to Settings > Data Controls and turn off "Improve the model for everyone." For anything sensitive, use Temporary Chat (the toggle near the model name at the top of a new chat) — those sessions are not saved to history, do not use memory, and are kept only briefly for safety. Separately, check Settings > Personalization > Memory, because ChatGPT can quietly remember facts across chats; clear it or switch it off if you would rather it did not.
Open Gemini's Settings and find "Gemini Apps Activity," or manage it at myactivity.google.com. Google is unusually blunt here: when this is on, human reviewers can read conversations, and its own guidance says do not enter anything you would not want reviewed. Turning it off stops future saving, but conversations already selected for human review are kept separately for up to three years and are not deleted when you clear your activity. Set auto-delete (3, 18, or 36 months) as a backstop.
In Settings, look under Privacy for the control over whether your chats help improve Claude, and opt out if you prefer. Anthropic's consumer policy changed in 2025 so that users choose this explicitly; if you opt in, retention of your data can extend for years, so the choice matters. Claude used through the API or commercial plans is not trained on by default.
In consumer Copilot, open your account privacy settings and turn off the model-training toggle for your conversations; the work version, Microsoft 365 Copilot, already carries Commercial Data Protection. In Perplexity, go to Settings > Account and disable "AI Data Retention."
Some inputs should never leave your device through a general chatbot, whatever your settings say:
.env files, or emails to "get help" without noticing the secrets inside. Once sent, treat that credential as compromised and rotate it.When you genuinely need help with sensitive material, sanitize it first:
You almost always get the same quality of answer, because the model needs the structure of your problem, not the identifying details.
Modern assistants do more than answer — they remember, and they reach into your accounts, and each capability widens what you expose.
Persistent memory (ChatGPT's Memory, Gemini's saved info) means one stray sentence can be stored and reused for months. Review and prune it periodically.
Connectors and plugins that link Gmail, Google Drive, Notion, or your calendar grant standing access to whole accounts, not just the item you asked about. Connect only what you actively use and disconnect the rest. On phones, scrutinize permission prompts: an assistant asking for Contacts, Microphone, or Location — on iPhone (Settings > Privacy & Security) or Android (Settings > Apps > [app] > Permissions) — should have an obvious reason. Deny anything that does not match a feature you use, and grant it later if a feature genuinely needs it.
Browser extensions that bolt AI onto every page are the easy-to-miss risk: some stream the content of pages you visit — including logged-in dashboards — to a third-party server. Install only reputable ones, and read what data they capture.
For genuinely confidential work, the strongest guarantee is data that never leaves your machine. Local runners like Ollama and LM Studio let you run open-weight models (Llama, Mistral, Qwen) offline, so prompts stay on your hardware. Apple Intelligence handles many requests on-device and, when it needs more power, uses Private Cloud Compute, engineered so that even Apple cannot retain or see the data. These are not as capable as the largest cloud models, but for sensitive drafts they trade a little quality for a lot of privacy.
The biggest is assuming "I deleted it" equals "it is gone" — history deletion, human-review copies, and legal holds are three different things. The second is trusting the free tier for a work task; if it touches company or customer data, use the sanctioned, contract-backed tool instead. The third is ignoring the tier distinction and pasting into a personal account what should only go through a business account with data protection. And the quiet one: leaving memory and connectors switched on and forgetting they are steadily building a profile of you.
No — it stops your conversations from feeding future model versions, which is the biggest single win, but the provider may still log and retain chats for a limited period for abuse monitoring and legal reasons. Combine it with deleting history and using temporary or incognito chat modes for anything sensitive.
Usually, and for a specific reason: business and developer tiers (ChatGPT Enterprise/Team and the API, Microsoft 365 Copilot, Anthropic's commercial API) typically exclude your data from training by default and add contractual protections. A personal paid plan is not automatically a business plan, so confirm which one you are on.
Only through a tool your employer has approved with a data-processing agreement. Consumer apps can retain and, on some tiers, human-review inputs, which may breach confidentiality or data-protection law. Ask IT whether there is a sanctioned enterprise version before pasting anything.
Run a local model with Ollama or LM Studio so nothing leaves your device, or use a provider tier that offers zero data retention. For everyday sensitivity, redacting identifiers before pasting into a cloud tool is usually enough.
Theo writes about online safety the way a good friend would — clearly, calmly, and without trying to scare you. He's interested in the simple habits that stop most problems, and he thinks staying private online is a skill anyone can learn.
Keep reading
AI news moves fast and most of it is noise. Here is a calm, jargon-free system for staying informed about what matters without burning out on every headline.
Priya Nadar
Recommendation systems shape what you watch, buy, and read every day. Here is a clear, jargon-free look at how they work and how to stay in control.
Priya Nadar