What happens to what you type into an AI service

Often yes to the first, and sometimes yes to the second, depending on the product and your settings. Consumer apps for individuals often let the provider train its models on your chats unless you opt out, while business, school, and developer versions of the same services usually leave your content out of training by default. Turning training off is a separate step from deleting chats, and a deleted chat can take up to 30 days to leave a provider's systems. Temporary or incognito chats are kept out of training, but providers still keep a copy for 72 hours to 30 days in the consumer examples below. People working for a provider may read some chats, and flagged chats or feedback can be kept longer. Policies change, so check the date on the page you read; this page is general information, not legal advice.

Consumer apps and business products have different terms

A consumer app is one you sign up for yourself. A business, school, or enterprise plan is bought by an organization, which signs a contract and gets administrator controls. An API (application programming interface) lets a developer's software send text to a model directly; see hosted API. Each can carry different terms. Anthropic's privacy policy (external site: anthropic.com) says it does not apply to content processed for customers of its business offerings, which customer agreements govern. Google's Gemini Apps Privacy Hub (external site: support.google.com) says a work or school account may have different data handling terms. So first ask which product and account you are using.

What to look for in a policy

Model training means using examples, such as past conversations, to adjust a model so later versions behave differently. Retention means how long the provider keeps a copy. Look for answers to these questions:

• Training. May your chats be used, and which setting stops it? Does it cover past chats or only new ones?
• Feedback. What is sent when you rate a response?
• Retention and deletion. How long are chats kept, and how long does deletion take?
• Human review. Can people read conversations, and are reviewed chats kept longer?
• Temporary chats. Are they excluded from training, and how long is a copy kept?
• Files and voice. Do uploads and audio follow the same rules?

Four U.S. providers, as published on October 8, 2026

OpenAI (ChatGPT). OpenAI publishes ChatGPT's training settings, temporary chats, and retention periods in its Help Center (external site: help.openai.com). USASI could not re-check those pages on October 8, 2026, so this page does not summarize them; read them directly. What OpenAI's API documentation says is summarized under workplaces and schools below.

Anthropic (Claude). Its privacy policy, effective September 10, 2026, says it may use your inputs and outputs to train its models unless you opt out, and that even after an opt-out it uses chats flagged for safety review, or material you explicitly report (for example through its feedback mechanisms), for model improvement. Its Privacy Center says data from users who allow training may be kept de-identified for up to five years, deleted chats leave back-end storage within 30 days, and opting out later does not remove data from training runs under way or models already trained (retention (external site: privacy.claude.com)). Incognito chats (external site: support.claude.com) are not used for training and are kept for 30 days by default. Anthropic says it does not train on inputs or outputs from commercial products such as Claude for Work and the API by default (commercial (external site: privacy.claude.com)).

Google (Gemini Apps). The privacy hub (external site: support.google.com), last updated September 24, 2026, says that with Keep Activity on, Google uses your activity to improve its services, including training generative AI models, and human reviewers read a subset of chats. Activity auto-deletes after 18 months by default, but reviewed chats are disconnected from your account and kept for up to three years even if you delete your activity. The hub asks you not to enter confidential information you would not want a reviewer to see. With Keep Activity off (unless you send feedback), or in a temporary chat, chats are not used for training but are kept for 72 hours. Audio, Gemini Live video, and screen shares are not used to improve Google services by default.

Microsoft (Copilot). Microsoft's pages differ by app version. For the updated app, available since August 18, 2026, with a Microsoft account rather than a work or school account, it says prompts, responses, and file contents are not used to train foundation models (activity history (external site: support.microsoft.com)); the page does not define that term. Its older FAQ (external site: support.microsoft.com), for the older app only, describes training on conversations, voice, and uploaded files unless a user opts out or is excluded, 18-month default storage, and no opt-out from human review when a Code of Conduct violation is suspected.

Policies change, so check the date

Anthropic's policy shows an effective date and links to its previous version. Google's pages show "Last updated." Microsoft keeps separate live pages for its older and updated Copilot apps, which describe different training practices. Note the date, product, and plan a page covers, and keep a copy.

FTC staff have written that companies offering AI models as a service may be liable under laws the FTC enforces if they break privacy commitments, including promises not to use customer data for secret purposes such as training or updating their models (January 2024 (external site: ftc.gov)), and that starting to use consumers' data for AI training while disclosing it only through a surreptitious, retroactive change to terms may be unfair or deceptive (February 2024 (external site: ftc.gov)). Both are staff posts on the FTC's Technology Blog.

What a workplace or school should check

Read the business terms that match your plan, not the consumer pages, and check:

• Training controls. OpenAI's API documentation says data sent to its API is not used to train or improve its models unless you explicitly opt in to share it. Anthropic says feedback may be used for training and Team and Enterprise owners can turn its feedback buttons off.
• Provider logs. OpenAI's API documentation (external site: developers.openai.com) says abuse-monitoring logs, which may include prompts and responses, are kept for up to 30 days by default, with exceptions such as legal requirements, and approved customers can use Zero Data Retention.
• Who inside can see chats. Microsoft says prompts and responses in Copilot Chat for work or school (external site: support.microsoft.com) are logged and IT admins can view them. Anthropic says Team and Enterprise incognito chats appear in owners' data exports.
• Contract wording. Google's Workspace privacy hub (external site: knowledge.workspace.google.com) says Workspace does not use customer data for training models without the customer's prior permission or instruction. Some statements add "outside of your domain"; if that matters, ask Google or your administrator what it covers.
• Guidance. Tell people that personal accounts bring consumer terms.

Keeping inputs on your own machine

Local inference means running a model's weights on your own computer, so your text need not go to a provider. Ollama's privacy policy (external site: ollama.com) (March 2026) says it does not collect, store, transmit, or access content you process locally, though it may collect limited device and usage metadata. A shared or network-exposed machine is not private by default; Choosing hosted access or local inference covers those checks.

Worked example: a teacher with two accounts

Rae is a fictional reader invented for this page. Rae teaches high school English, has a personal Google account and a school Workspace account, and wants help drafting comments on student essays.

1. What goes in? Students' names and work. Rae checks the school's student-data rules first.
2. Which terms? The Gemini hub says school accounts may differ, so Rae reads the Workspace hub and notes its date, October 6, 2026.
3. The personal account. With Keep Activity on, chats may train models and be read by reviewers; even with it off, Google says chats can be reviewed to protect users and the public. Rae rules it out for student work.
4. School settings. Rae asks the IT office about retention and who can see logs.
5. Feedback. On a personal account with Keep Activity off, Google says feedback shares context including the last 24 hours of chats, so Rae skips rating buttons near student work.
6. Share less. Rae replaces names with numbers.

Outcome: use the school account if approved, with names removed, and keep the personal account for lesson ideas. A teacher without an approved account might try a local model or leave student work out.

What you can do next

• Read Choosing hosted access or local inference and browse runtimes such as Ollama and llama.cpp.
• Look up hosted API and self-hosting / local inference in the glossary.
• Visit the enterprise AI hub and the records for OpenAI, Anthropic, Google, and Microsoft.
• Read Finding AI policy and standards sources.

Sources

All read on October 8, 2026.

• OpenAI: Data controls in the OpenAI platform (external site: developers.openai.com) (API documentation). The ChatGPT Help Center is linked above but was not re-checked.
• Anthropic: Privacy Policy (external site: anthropic.com) (effective September 10, 2026); Privacy Center articles How long do you store my data? (external site: privacy.claude.com) and Is my data used for model training? (commercial) (external site: privacy.claude.com); Help Center: Use incognito chats (external site: support.claude.com)
• Google: Gemini Apps Privacy Hub (external site: support.google.com) (last updated September 24, 2026); Generative AI in Google Workspace Privacy Hub (external site: knowledge.workspace.google.com) (last updated October 6, 2026)
• Microsoft: Microsoft Copilot for individuals: your activity history (external site: support.microsoft.com), Privacy FAQ for Microsoft Copilot (external site: support.microsoft.com) (older app), Data protection when using Microsoft Copilot Chat for work or school (external site: support.microsoft.com)
• Federal Trade Commission, Technology Blog: AI Companies: Uphold Your Privacy and Confidentiality Commitments (external site: ftc.gov) (January 9, 2024); AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive (external site: ftc.gov) (February 13, 2024)
• Ollama: Privacy Policy (external site: ollama.com) (last updated March 2026)