Lock Down the Privacy Settings on AI Chatbots

Most AI chatbots use your conversations to train their models unless you opt out. Here’s how to find the off switches, cut back on stored history and clean up what’s already there.

This guide is general information, not technical or legal advice. If you’ve lost money or personal information, contact your bank, IDCARE on 1800 595 160, or report it through ReportCyber at cyber.gov.au.

Quick summary

1. A Stanford study of frontier AI developers found six leading US companies feed user conversations into their models, and some offer no way to opt out.
2. Turning off training usually only affects future chats, so deleting old conversations is a separate job with its own settings.
3. Opting out on social platforms often only stops future use, and some platforms enable AI training by default, including LinkedIn.
4. Never paste passwords, financial details, health information or confidential work material into a chatbot, including in uploaded files.
5. A short quarterly review of training toggles, chat history and connected apps keeps your settings working as you intended.

Most AI chatbots keep your conversations on their servers and, unless you change the settings, may use them to train future models. A Stanford study of leading developers found six major US companies feed user conversations back into their models, and some offer no way to opt out at all. This routine walks through what to check before you even create an account, then the settings to change, the history to clean up and the habits that keep things under control.

Step 1: Read the Privacy Policy Before You Sign Up

Before creating an account, look for three things in the privacy policy: whether conversations are used for training by default, how long the company keeps them, and whether staff or contractors review chats. The Stanford researchers flagged long retention periods, training on children’s data and a general lack of transparency as recurring problems across the industry. If a policy is vague on any of these, treat that as information in itself.

In Australia, the Office of the Australian Information Commissioner publishes guidance on privacy and generative AI for both companies and users, and it is worth a skim if you want the regulator’s view. Australian privacy law also sets some ground rules. Organisations generally need your express consent to handle sensitive information, and if a company relies on implied consent through an opt-out, that option must be clearly presented, easy to use and low cost. You can opt out later and, as far as practical, be in the same position as if you had opted out earlier. If a service buries the opt-out five menus deep, that sits badly with those principles.

If children will use the chatbot, check what the policy says about under-age accounts. Australia’s Children’s Online Privacy Code is being developed by the OAIC specifically to tighten rules for apps and services children use.

Step 2: Turn Off Training on Your Conversations

Once you have an account, find the training toggle. The location varies by service:

1. ChatGPT. Click your profile icon, then switch off the “Improve the model for everyone” toggle. This setting applies to Free, Plus and Pro accounts. Temporary chats stay out of training on their own, but conversations you have while logged out are collected no matter what, with no setting to stop it.
2. Gemini. Two places to visit. In your Google account, turn off Web & App Activity, then open the Gemini settings and switch the activity setting to off. Google has renamed that option more than once, so look for Gemini Apps Activity or something similar.
3. Claude. In the settings, switch off “Help improve our AI models”, and while that screen is open, switch off location metadata too.

One thing catches almost everyone: these toggles only affect future conversations. Anything you typed before switching them off has already been collected. Switching training off is job one, but it does not undo the past, which is what the next step handles.

Step 3: Clean Up and Control Your Chat History

Go through your chat list and delete anything you would not want sitting on a server. Delete conversations by selecting them and using the delete option, and check whether the service offers a bulk delete or a clear-all for a date range.

A few specifics worth knowing. Turning off Web & App Activity in your Google account stops new searches and other data being saved, but anything already stored stays put. Use the delete button on the same page and pick a range such as all time to clear it. The same applies to Gemini’s own activity. On Claude, Anthropic’s privacy policy states that chats you delete can linger in its systems for up to 30 days before they are fully removed.

If you use Siri on an Apple device, open the Siri settings and look for the Keep Conversations option, which lets you set retention anywhere from forever down to 30 days. Individual chats can be removed with a swipe, and the select menu in the top corner clears several at once. Apple says this data stays on the device, is not used to train its models unless you explicitly agree, and is deleted from cloud processing as soon as your device receives the response. Turning on Face ID or Touch ID for the app stops anyone else reading your chat history.

Step 4: Know What Never to Paste

The safest privacy setting is the one you apply before typing. Keep out of chatbots entirely:

• Passwords, or anything resembling one. Protect your accounts with a proper passphrase rather than testing one in a chat window.
• Financial details such as account numbers, tax file numbers or invoices with personal information.
• Health information about you or your family.
• Confidential work material, client data or source code.
• Anything identifying your children.

This is not paranoia. In March 2023, engineers in Samsung’s chip division pasted confidential source code and meeting notes into ChatGPT for help with work tasks. Reporting at the time said that material then entered the model’s training data, and Samsung responded by barring staff from the tool. In the same month, a flaw in an open-source component behind ChatGPT let some users see other people’s chat titles and payment details, which OpenAI confirmed affected about 1.2 per cent of ChatGPT Plus subscribers. Uploaded files count too: the Stanford researchers note that sensitive information in a file you attach during a conversation may be collected just like typed text.

Step 5: Check Connected Apps and Other Training Settings

Chatbot accounts are often signed into through Google, Apple or another provider, and each of those connections is another place where settings live. Open your account page, review the list of connected apps and remove anything you no longer use. If your chatbot can hand conversations to another service, such as the ChatGPT extension available on Apple devices, turn it off unless you deliberately want that data flowing to a third party.

Then widen the net to your social accounts, because several of them train AI models on your activity. On LinkedIn, open Settings and Privacy, head to Data Privacy and switch off “Data for Generative AI Improvement”. Note that this setting is enabled by default, covers details like your name, photos, job history, education and location, and opting out only prevents future use. On X, open Settings and privacy, then “Grok & Third Party Collaborators” under data sharing. Meta says it trains on public posts, photos and comments but not private messages unless you share them with Meta AI, and posts restricted to friends or private accounts are excluded, though a comment left on a public post can still be picked up.

While you are in account settings, check that your logins are protected. Adding multi-factor authentication means a stranger who learns your password still cannot read your chat history.

Step 6: Set a Routine for Reviewing Your Data

AI settings drift. Services rename options, add features and change defaults, and an update can quietly switch something back on. Put a recurring reminder in your calendar, once a quarter is plenty, and run through a short checklist:

• Confirm the training toggles are still off.
• Delete conversations you no longer need.
• Review connected apps and remove stale ones.
• Check the privacy policy for updates about training or retention.

That last point matters because policies do change. Anthropic, for example, moved to using customer conversations for training by default unless users opt out, which is exactly the kind of shift a quarterly check catches. If a service you use suffers a breach, our guide on what to do after a data breach covers the first steps.

Open the settings of the chatbot you use most today, switch off the training toggle and delete the conversations you no longer need.

Your questions answered

Q: Does turning off model training delete my old conversations?

A: No. On most services the training toggle only applies to conversations from that point forward. Existing chats usually need to be deleted separately, and some services keep deleted chats in their systems for a period, such as up to 30 days for Claude.

Q: Do chatbots collect my data if I use them without an account?

A: They can. OpenAI’s documentation says conversations had while logged out of ChatGPT are collected no matter what, and no setting stops it. Temporary chat modes generally stay out of training.

Q: Can I stop my social media accounts being used to train AI models?

A: Often, yes, though the setting is not always on by default. LinkedIn enables generative AI training unless you disable it under Data Privacy, while Meta says it uses public posts and photos but not private messages unless you share them with Meta AI.

Q: Is my chatbot conversation data ever exposed to other people?

A: It has happened. In March 2023, a flaw in ChatGPT let some users see other people’s chat titles and payment details, which OpenAI confirmed affected about 1.2 per cent of ChatGPT Plus subscribers. Samsung barred staff from using ChatGPT after engineers pasted in confidential source code.