Every AI assistant you paste text into is a privacy decision, even when the assistant feels like a private conversation. The good news is that the controls exist, they are documented, and they work in your favor when you know which buttons to press. This guide walks through a practical threat model and the mitigations that actually change what leaves your device.
The threat model here is straightforward. Three things can leak: the prompts you send, the outputs you receive, and the metadata around them (timestamps, IP address, account identifiers). Each vendor handles these three differently, and each lets you tighten or relax that boundary through documented settings. Your job is to pick the right tightening for the data you actually care about.
1. Decide what you are protecting and from whom
Before touching any settings, write down two short answers:
- What kinds of text am I putting in here? Drafts, code, medical notes, business strategy, kids’ homework, fanfiction, and angry emails all deserve different treatment.
- Who is the realistic adversary? A future employer reading transcripts in a lawsuit discovery, a vendor training a model on your writing, an attacker scraping leaked logs, or a government issuing a subpoena all need different defenses.
This is a privacy tool. It is not an anonymity tool, and it will not hide you from a court order. If you need anonymity, route through Tor and avoid account-bound products entirely. For ordinary confidentiality (journalism, business, medical, family), the mitigations below are a reasonable baseline.
Our guide on reading an AI privacy policy before you sign up walks through the legal language itself, which is a useful companion read.
2. Audit what your current tools actually retain
Vendor policies differ more than their marketing suggests. Pull up the privacy or data-control page for each tool you use and look for three concrete facts: training opt-out status, retention window for prompts and outputs, and the deletion mechanism.
OpenAI’s API data controls documentation is unusually explicit. For API customers, the page states: “As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us).” The same page says abuse monitoring logs are generated by default and kept for up to 30 days, and that eligible customers can have their content excluded from those logs through Zero Data Retention or Modified Abuse Monitoring, subject to prior approval by OpenAI.
Anthropic’s consumer privacy policy, effective September 10, 2026, lets users opt out of having Inputs and Outputs used to train future models through account settings. The same policy says deleted individual conversations are removed from your history immediately and deleted from back-end systems within 30 days.
DuckDuckGo’s AI Chat privacy and terms, last updated October 1, 2026, takes a different route by sitting as a middleman. The page documents that metadata such as your IP is stripped before prompts reach the underlying model provider, and that providers must delete the data within 30 days at most, with some routes offering Zero Data Retention or Zero Provider Visibility.
Write down what you find. If a tool does not publish these three facts, that itself is an answer.
3. Tighten the settings that change behavior
Three classes of setting actually move the needle:
Training opt-out toggles. Both OpenAI (for consumer ChatGPT, not the API) and Anthropic expose a settings switch that says your conversations will not be used for training. Anthropic’s opt-out has documented exceptions for safety-flagged conversations and explicit feedback reports.
Retention windows. On the OpenAI API, abuse monitoring logs are kept for up to 30 days by default. Zero Data Retention is available to eligible customers but requires prior approval by OpenAI.
Account-level deletion. Vendors differ on what “delete” means. Our step-by-step guide to deleting your data walks through the exact buttons for the major assistants and what they keep on the server after you press delete.
If a vendor only offers opt-out through a settings page, take a screenshot of the setting showing it is off.
4. Move sensitive work to a local model when the tradeoffs fit
When the prompt is sensitive enough that no hosted retention window is acceptable, a local LLM removes the question entirely. Ollama’s privacy policy, last updated March 2026, makes this explicit for local operation: “Ollama runs on your local device. We do not collect, store, transmit, or have access to your prompts, responses, model interactions, or other content you process locally. Your data stays on your machine.” The same policy covers Ollama’s cloud-hosted models separately: content is processed transiently to provide the service, not stored beyond the time required to fulfill the request, and not used to train AI models.
The honest tradeoff: a 7B model on a mid-range machine is slower and weaker than GPT-5-class output, especially for long-context reasoning and code refactors. For tasks where quality matters more than confidentiality, keep using the hosted tool. For tasks where confidentiality matters more than quality, route them locally. Our beginner guide to local LLMs covers the install path.
Also remember: a local model only protects text that never goes near a cloud tool. A prompt typed into a local chat UI that calls a remote embedding API is no longer local. Read the integration layer before trusting the workflow.
5. Keep the human in the loop
Privacy controls drift. Vendors change defaults and update policies; two of the policies cited here were updated in September and October 2026. Set a calendar reminder to re-read each privacy policy you depend on once a year, and re-check the opt-out toggle after any major vendor announcement.
For deeper context on how privacy fits into the broader AI policy debate, the Electronic Frontier Foundation’s AI issues page frames the conversation in terms of existing civil liberties tools rather than a brand-new problem.
Quick checklist
- Write a two-line threat model before you change any setting.
- Confirm the training opt-out for every hosted AI account you use.
- Note the retention window in days for each vendor you trust.
- Take screenshots of opt-out toggles so you can show your settings later.
- Route genuinely sensitive prompts through a local model and verify the integration layer stays local.
- Re-audit every vendor’s policy once a year or after major announcements.
Privacy is not a setting you toggle once and forget. It is a habit of reading the actual controls, picking the right tool for each prompt, and confirming that what you expect is what the vendor does. The five steps above keep the most sensitive text out of hosted tools and their training pipelines.