What Happens to What You Type Into ChatGPT, Claude and Gemini (and the Settings to Change)

When you type something into ChatGPT, Claude or Gemini, it can end up in up to four places: your chat history, the app’s memory of you, the company’s training data, and, if you use your voice, a recording. Each one has its own switch. Most advice about AI privacy settings covers only the training switch, and that leaves three of the four doors open.

There are also a few things no setting stops: the thumbs-up button, safety review, and someone stealing your login.

You don’t have to turn everything off. I leave training on in Claude on purpose. It’s a choice I’m comfortable with, and I’d rather make it knowingly than find out about it later. Below: what each app keeps and for how long, where every switch lives, and a five-minute checkup.

Verified as of October 27, 2026, against OpenAI’s Data controls, Chat and file retention and Locked chats help pages; Anthropic’s Privacy Center articles on retention, model training, location and dictation; Google’s Gemini Apps Privacy Hub and activity help pages; and the Privacy page in Claude’s Mac app. These settings change often, so check the app before relying on a detail.

The quick answer

• ChatGPT trains on your chats by default on Free, Plus and Pro. Turn it off under Settings > Data controls > Improve the model for everyone.
• Gemini saves your activity by default, keeps it for 18 months and uses it to improve Google’s models. Turn it off under Settings & help > Activity.
• Claude has a training switch called Help improve our AI models, under Settings > Privacy. Anthropic doesn’t publish which way it starts for new accounts, so check yours.
• Memory is a separate switch in ChatGPT and Claude. Turning off training doesn’t stop the app remembering things about you.
• Voice can have its own switch. ChatGPT’s voice mode has a separate training control, and Claude has started asking voice users for permission to train on recordings.
• Private-chat modes still keep a copy for a while. ChatGPT keeps temporary chats up to 30 days, and Gemini keeps them 72 hours. Claude’s incognito chats are never used for training.
• Pressing thumbs up or down can undo your opt-out for that chat. OpenAI and Google say so outright, and Anthropic keeps the whole rated conversation for five years.
• Flagged chats can reach a human, whatever your settings.

What each app keeps, side by side

Two caveats sit behind the neat rows. Work and school accounts (ChatGPT Business, Claude Team, Google Workspace) follow their organization’s rules, and companies keep data longer where the law or a security investigation requires it.

Switch 1: training

Training means the company can use your conversations to improve future models. Turning it off is forward-only. Anthropic states this plainly: chats already used in a training run, or in a model that’s already trained, stay there.

ChatGPT. Open your account menu > Settings > Data controls, and turn off Improve the model for everyone. On a phone, open the sidebar, tap your profile icon, then Data controls. It applies to your whole account across devices, and your chats stay in your history. If you use Codex on a personal plan, the same switch covers your Codex tasks.

Claude. Open Settings > Privacy and switch Help improve our AI models on or off. The description reads “Allow the use of your chats and coding sessions to train and improve Anthropic AI models,” so it covers Claude Code too. With it on, Anthropic may keep de-identified chats for up to five years in its training pipeline. With it off, deleted chats leave Anthropic’s servers within 30 days.

Gemini. Go to gemini.google.com > Settings & help > Activity, click On near the top, and choose Turn off or Turn off and delete activity. With Keep Activity off, Google still holds new chats for 72 hours so Gemini can respond, and it won’t use them for training unless you send feedback.

What surprised me looking these up was how firmly ChatGPT leans toward training by default. I’m still fine leaving training on. I’m only here for a short time, and if something I’ve worked through helps a model help someone else, that’s a small piece of legacy I’m happy to leave.

Training on is a fair choice. What matters is making it yourself.

Switch 2: memory

Memory is what the app remembers about you between chats: your job, your preferences, the fact that you’re training for a race. OpenAI’s help page says outright that memory and training are separate settings, and the same holds in Claude. Turning off training doesn’t erase what the app has learned about you, and turning off memory doesn’t stop training.

ChatGPT. Memory has its own controls, separate from Data controls. Turning memory off doesn’t disable OpenAI’s safety features, which can still use limited context “in rare, high-risk situations.”

Claude. Memory has its own tab in Settings and is on by default on Free, Pro and Max. Claude stores what it learns as topics you can read, edit or delete. A few details matter:

• Sensitive topics are off by default. Health, religion, politics and similar subjects aren’t stored unless you switch that setting on.
• Memory is shared between chat and Claude Cowork when Cowork runs in the cloud. While memory is on, there’s no way to keep the two apart.
• Incognito chats are never saved to memory.

Gemini. Gemini can personalize answers from your past chats, but only while Keep Activity is on, so that switch is your main control. Google has renamed the menus more than once, so look for memory or past-chat settings in your own app.

I keep a Claude project as a personal coach for money, workouts and the sort of talking-it-through you’d do with a therapist. That’s exactly the material worth checking memory for. Every few weeks, open the memory list and delete anything you wouldn’t want resurfacing in an unrelated chat.

Memory is the setting that follows you into the next conversation.

Switch 3: voice

Voice is newer, and it’s the switch people miss.

ChatGPT. OpenAI’s Data controls page says voice mode “has a separate control for sharing data for model training.” Opting out of text training doesn’t automatically cover your voice.

Claude. In early October 2026, Anthropic began asking voice users whether it can use their recordings and voice-chat data to improve its models. The option is off unless you agree. It’s separate from the text-training switch, so you can allow one and refuse the other, and you can switch it off later and delete stored voice data (source: Mayank Parmar — Anthropic asks Claude users to share voice data for AI model training (bleepingcomputer.com), 2026-10-04). Claude shows voice users an Allow or Not now prompt. It hadn’t reached the Privacy page of my Claude Mac app by October 27, so it may not have reached yours either. Anthropic hasn’t said how long it keeps recordings from people who opt in.

Dictation is different. Anthropic says audio from dictation in the Claude mobile apps is deleted once it’s turned into text and isn’t used for training. The text then sits in your chat like anything you typed.

Talking to an AI and dictating to it can follow different rules.

Switch 4: history and the private-chat modes

Each app has a mode for conversations you don’t want kept. They’re useful, as long as you know what “private” covers in each one.

ChatGPT temporary chat. It stays out of your history, doesn’t create or update memories, and isn’t used for training. OpenAI may keep a copy for up to 30 days for safety. Choose before your first message whether it’s personalized. A personalized temporary chat still reads your existing memories and custom instructions, so pick unpersonalized when you want a clean slate. If you save a temporary chat, it becomes a normal chat under your normal settings.

ChatGPT locked chats. These are new. Locked chats sit under Projects > Locked chats behind a four-digit PIN or Face ID, out of your sidebar and search. OpenAI is clear about the limits: locking isn’t encryption, and it doesn’t change how long the chat is kept or whether it’s used for training. Resetting a forgotten PIN permanently deletes every locked chat, and OpenAI can’t recover them.

Claude incognito chats. They aren’t used for training even if your training switch is on, and they’re never saved to memory.

Gemini temporary chats. Google keeps them for 72 hours. They aren’t used for training and are left out of the human review Google uses to improve its products.

Deleting chats. ChatGPT removes a deleted chat from view immediately and schedules it for permanent deletion within 30 days. A file saved to ChatGPT’s Library stays there until you delete it separately. Claude removes deleted chats from its servers within 30 days. Gemini deletes activity whenever you choose, plus on its auto-delete schedule.

Hidden, kept out of training and deleted are three different things. A locked chat is hidden. A temporary chat is kept out of training. Only deleting removes it, and even that takes up to 30 days.

What no setting stops

The thumbs-up button

The rating buttons under each reply carry more than a rating.

• ChatGPT: if you rate a response after opting out of training, “the entire conversation associated with that feedback may be used to train OpenAI models.”
• Gemini: with Keep Activity off, Google won’t train on your chats “unless you choose to send Google feedback.”
• Claude: rating a reply stores “the entire related conversation” for up to five years. Anthropic removes your account details first and lists training among the uses. You can review what you’ve sent under Settings > Privacy > Your feedback.

Rate the reply to a recipe question as often as you like. Skip the thumbs on the conversation about your health or your money.

Safety review

Automated systems scan conversations for serious harm, and flagged chats can reach human reviewers, whatever your training setting. In October 2026, a Florida woman was arrested after Claude flagged a conversation in which she allegedly threatened a sheriff’s office. Reviewers read the messages and reported them to police (source: Emma Roth — Florida woman arrested for allegedly making threats in an AI chat. (theverge.com), 2026-10-05). She has been charged, and the case hasn’t gone to trial.

Anthropic’s privacy policy allows sharing with law enforcement to prevent serious harm, and it keeps flagged chats for up to two years. Google has a subset of Gemini chats reviewed by people. Those chats are disconnected from your account first, kept for up to three years, and not removed when you delete your activity.

Write in a chat window the way you’d write anything that could be read later.

A stolen login

In September 2026, Anthropic warned some paying subscribers that malware on their own computers had stolen their logged-in Claude sessions, letting someone else run up their usage (source: Julie Bort — Hackers are stealing Claude tokens from subscribers (techcrunch.com), 2026-09-08). A stolen session skips your password and your two-factor code entirely.

The defence is ordinary device hygiene: no pirated software, no unknown browser extensions, sign out of sessions you don’t use, and check your usage page for jumps you can’t explain.

When you connect accounts or let AI log in

Connecting Gmail, a calendar or a bank adds a fifth place your data goes. Two posts cover those in detail: Is it safe to connect AI to Gmail and your calendar? and Is it safe to connect ChatGPT to your bank account?. Two privacy details belong here.

• Gemini makes you choose. Connected apps such as Gmail and Google Drive stop working when Keep Activity is off. If you want Gemini in your inbox, your activity gets saved.
• ChatGPT can stay signed in to websites for you. When ChatGPT’s cloud browser logs in to a site, you type the password into a secure form the model can’t see, but the logged-in session persists until it expires so later tasks can reuse it. Clear it under Settings > Cloud browser > Browser data. The same caution applies to any AI browser: see Are AI browsers safe?.

One more for Claude users: the Privacy page has a Location metadata switch, which lets Claude use your approximate city or region, from your IP address, for features like web search. It was on in my account. Turn it off if you’d rather Claude didn’t know roughly where you are. Anthropic still uses rough location to enforce regional rules either way.

The five-minute privacy checkup

1. Export first if you might delete anything: ChatGPT, Claude, Gemini.
2. Decide on training in each app you use, and set it on purpose.
3. Open memory in ChatGPT and Claude, read what’s there, and delete anything you wouldn’t want resurfacing.
4. Check voice if you use voice mode: ChatGPT’s voice control, and Claude’s opt-in if you’ve been asked.
5. Look at Claude’s Location metadata switch and Gemini’s auto-delete period (3, 18 or 36 months).
6. Pick your private mode for sensitive chats: ChatGPT temporary (unpersonalized), Claude incognito or Gemini temporary.
7. Leave the thumbs alone on anything personal.
8. Sign out of old sessions and glance at your usage page.

Frequently asked questions

Does turning off training delete my chats? No. Your history stays. Delete chats separately, and allow up to 30 days for them to leave the servers.

Are temporary chats really private? They’re kept out of history and training. ChatGPT may keep one for up to 30 days and Gemini for 72 hours. A personalized ChatGPT temporary chat still reads your memories.

Can a human read my AI chats? Yes, in some cases. Claude’s flagged chats can go to human reviewers, OpenAI keeps even temporary chats for up to 30 days for safety, and Google sends a subset of Gemini chats to reviewers after disconnecting them from your account.

Does pressing thumbs up share my conversation? Yes. In ChatGPT and Gemini, rating a reply can send that conversation for training even if you’ve opted out. Claude keeps the whole rated conversation for up to five years.

Does Claude train on my chats? Only if Help improve our AI models is on in Settings > Privacy. Incognito chats are never used.

Do locked chats in ChatGPT keep my data safer? They keep chats out of sight on your screen. They don’t change retention or training, and resetting your PIN deletes them permanently.

Key takeaways

• Your words can land in four places: history, memory, training and voice. Each has its own switch.
• ChatGPT and Gemini train by default. Claude’s default isn’t published, so check your own.
• Memory and training are separate. Turning off one leaves the other running.
• Hidden, not trained on and deleted are three different things.
• The thumbs buttons, safety review and a stolen login sit outside every setting.
• Leaving training on is a reasonable choice. Make it on purpose.

A chat window is not a diary. Set the switches once, and you’ll know exactly what you’re writing in.