Chat Picker

2026 AI Tool Privacy: Data Collection Policies and Deletion Mechanisms

Vendor pages document different training, memory, temporary-chat, and deletion controls across ChatGPT, Claude, and Gemini, which an afternoon test can examine without assigning a privacy score.

Sources checked 3 Oct 2026

For 2026 AI tool privacy, the vendors’ own pages document different controls for model training, memory, temporary use, and deletion rather than one shared standard. Chat Picker has not tested the assistants, so the useful comparison is the behavior each page documents and the controls you can verify in your own account.

Know these limits before you start

ChatGPT. As read on October 1, 2026, OpenAI’s data controls page says control availability depends on sign-in status, plan, and workspace settings. Turning off “Improve the model for everyone” prevents new conversations from being used to train OpenAI models, but it does not delete or hide saved chats, according to the same page. Temporary Chat does not enter chat history, create or update memories, or improve models, although it may be retained for up to 30 days for safety purposes. Signed-in Free, Go, Plus, and Pro users can request a data copy and delete their account; managed workspaces instead direct users to the workspace owner. OpenAI’s memory documentation also says saved memories are stored separately, so deleting the original chat does not automatically delete a saved memory.

Claude. Anthropic’s model-training page says consumer chats and coding sessions can be used to improve Claude when the user allows this. Conversations flagged for safety review may also be used to improve safeguards, while Incognito Chats are excluded from model improvement even when the setting is enabled. Claude’s memory documentation says memory defaults on for Free, Pro, and Max users, while Team and Enterprise owners control availability. Memory entries are not removed when a conversation expires or is deleted, although individual entries can be removed and a full reset permanently deletes all memory. The same page says Incognito Chats are temporary and excluded from chat history, but Team and Enterprise Incognito Chats are included in standard exports and remain subject to organizational retention.

Gemini. Google’s personalization documentation says personalization requires a personal Google Account and is unavailable for work, school, and supervised accounts. It says available personalization can draw on past Gemini chats, content and activity from certain connected Google apps, response preferences, and instructions. The Gemini Apps Privacy Hub covers processing while a user is signed in and supplements the Google Privacy Policy, which covers signed-out processing. The vendor pages we read do not state a model-training setting on the plans page or give a self-service account-deletion timetable, so the comparison matrix marks the training setting “Not verified.”

The test plan

Use fictional information only, and run the same sequence on the plan and account type you actually plan to use.

  1. Prepare a controlled memory test. Input: “Project Northstar. Code: ORION-SEVEN. Coordinator: Jordan Lee. This is fictional test data.” Give each assistant this instruction: “For this test only, remember that the fictional project code is ORION-SEVEN. Do not infer real-world facts about Jordan Lee.” Before sending, record the plan, account type, workspace status, and visible training or improvement setting. After sending, record whether a history item appears and whether the memory interface shows a saved entry.

  2. Check whether information carries into a new chat. Start a separate conversation and give this instruction: “What is the fictional project code from the previous test? If it is unavailable, say that it is unavailable.” Look for the answer, a reference or citation to the earlier conversation, and any visible memory entry. Record the exact response and available controls without treating a correct answer as proof of how the information was stored.

  3. Test the temporary or private mode. Start whichever temporary or private mode your interface offers. Input: “Temporary test phrase: SILVER CEDAR. Reply with only that phrase.” Look for the mode label, history entry, memory entry, and model-improvement status. Record each item you can see and the exact wording of any retention or eligibility notice.

  4. Delete the data objects separately. Return to a regular chat containing the synthetic record. Use the interface to delete the conversation first, then locate and delete the memory or saved-information item separately. Record the control names, confirmation text, time, and anything still visible. A disappearing interface item is not, by itself, proof of complete backend erasure.

  5. Inspect account-level controls. Input no new material; use the account containing the synthetic test. Locate data export and account-deletion routes, noting any owner or administrator requirement. Do not confirm permanent account deletion unless you intend it. Record the available route, eligibility wording, and any step the page does not settle.

How to read your results

Sort each observation into one of four categories: clearly documented, documented exception, unknown, or administrator-controlled. Put your must-have settings first, such as excluding sensitive chats from model improvement or separating a saved memory item from its source chat.

Do not turn the log into a privacy score. A deletion confirmation establishes what the interface reported, not every downstream storage or retention process. Keep unresolved points marked as unknown, and check the current vendor documentation before relying on a personal, paid, or workspace account.

Where the plans differ

Start with the “Training on your chats” row in the ChatGPT, Claude, and Gemini comparison matrix. It records an opt-out for ChatGPT’s Free, Go, Plus, and Pro plans on OpenAI’s pricing page, and for Claude’s Free, Pro, and Max plans on Anthropic’s pricing page, which also says Team is not trained on by default. The training setting for Gemini is marked “Not verified” rather than inferred from Google AI Plans.

Then use the matrix’s “Free plan,” “Main paid plan,” “Team plan,” and “How usage limits are described” rows to record the account under test. Personal, paid, and managed accounts are not interchangeable: workspace settings can restrict individual controls. Chat Picker leaves unconfirmed cells blank and marks them “Not verified,” so confirm the current vendor page before subscribing.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants