infobro
Writing & Content53Image58Video28Audio & Voice33Chat & Assistants24Productivity46Coding & Development41Agents & Automation28Marketing & SEO21Sales & Support45Business & Finance96Operations21Data & Research23Education25Design22Health & Life38
Guide

What Your AI Tools Know About You, and the Settings to Change

What ChatGPT, Claude and Gemini keep about you, where each training switch sits, how long each holds your chats, and how to make a GDPR or CCPA request with the legal deadlines.

Stanislav VovchykMay 8, 202612 min readUpdated Nov 18, 2026
On this page · 10 sections
  1. 1. ChatGPT, Claude and Gemini side by side
  2. 2. Five changes in ten minutes
  3. 3. What AI tools collect, and how it can hurt you
  4. 4. What not to paste into a chatbot
  5. 5. Consumer accounts and work accounts get different handling
  6. 6. Memory, connected apps and built-in AI
  7. 7. What deleting a chat does and doesn't do
  8. 8. Who can read your chats, and who can ask for them
  9. 9. Who each step is wrong for
  10. 10. FAQ

In short

  • Every major chatbot stores your prompts, uploads and account details, and whether they train on them depends on the tool, the plan and your settings.
  • For sensitive questions, use Claude with training off: chats are kept 30 days, versus about two months for Gemini temporary chats and possible human review regardless of settings.
  • Switch off training in Claude under Settings > Privacy, in Gemini under Settings > History, and in ChatGPT under Settings > Data controls, then set Gemini auto-delete to 3 months.
  • Never paste passwords, API keys, health records or client data, and avoid free personal accounts for client work since only Enterprise, Business, Team and similar plans skip training by default.

Every major chatbot stores your prompts, uploads and account details. Whether a model gets trained on them depends on the tool and the plan. In Claude, one switch decides it. Gemini keeps your history by default, but training uses separate switches that start off. ChatGPT keeps its training choice on its Data controls page. No setting removes data held for safety review, and opting out doesn't pull your words out of models that have already been trained. Some things, like credentials and client files, should never go into a chat at all.

ChatGPT, Claude and Gemini side by side

Bar chart of chat retention. Claude: training off 30 days, flagged chats up to 2 years, training on up to 5 years, classification scores up to 7 years. Gemini: temporary and History-off chats about 2 months, default history 18 months, human-reviewed chats up to 3 years.

ToolConsumer chats used for trainingWhere to change itWhat's keptReview by the providerWork and API plans
ChatGPTYes, per its termsSettings > Data controls, which OpenAI's help center names as the place to choose how your data is usedHistory, memory and shared links, managed in SettingsHuman review possible, per Wirecutter's guide to AI chat privacyOpenAI says Enterprise, Business and Edu aren't trained on by default
ClaudeOnly if you allow it. Every consumer user had to choose (existing users by 8 October 2025)Settings > Privacy > Help Improve our AI modelsTraining on: up to 5 years. Training off: 30 days. Policy-flagged chats: up to 2 years, classification scores up to 7Chats flagged for safety review can be used for training whatever your settingTeam, Enterprise, Government, Education and API fall under Commercial Terms and aren't trained on by default
GeminiOff by default, through separate switches. History itself is on by default for users 18+Settings > History > Chat content including text & uploadsHistory auto-deletes at 18 months by default. Temporary and History-off chats: about two monthsHuman reviewers can read chats whatever your settings. Reviewed chats are kept up to 3 yearsFor Workspace accounts, the admin sets the terms

One thing the table can't show: turning Claude training off doesn't remove your data from training runs already under way or from models already trained.

For sensitive personal questions, I'd use Claude with training switched off. Anthropic keeps those chats for 30 days, and a chat you delete leaves its back-end storage within 30 days. Gemini holds even temporary and History-off chats for about two months, according to the Gemini Apps Privacy Hub, and its human reviewers can read chats regardless of your settings. ChatGPT's consumer terms allow training unless you change it. Claude's short window only holds if your chat isn't flagged for a policy violation, so it doesn't make it safe to type anything at all.

Microsoft Copilot comes in two forms: the free Copilot app, and Microsoft 365 Copilot on a work account, which runs under your organization's terms. A setting you find in one tells you nothing about the other. On a work account, your IT admin can tell you whether prompts are used for training and how long they're kept. For the free Copilot app and for Perplexity, ask the vendor which setting stops your conversations being used to improve its models before you type anything sensitive.

Five changes in ten minutes

The first three close the training switches. The last two clean up what's already stored.

  1. Claude: check the training switch. Go to Settings > Privacy and switch off "Help Improve our AI models" (the direct link is claude.ai/settings/data-privacy-controls). On mobile, tap your name > Settings > Privacy. Anthropic made every consumer user pick a position, so check which way yours is set. With it off, new chats fall under the 30-day retention.
  2. Gemini: close both training boxes and shorten history. History is on by default for anyone 18 or over, so Gemini stores your chats unless you change it. Under Settings > History, make sure "Chat content including text & uploads" is off. Then turn off the separate box for audio and Gemini Live recordings too, which matters most if you talk to Gemini: a clean recording of your voice is the raw material for a voice-cloning scam. While you're there, set auto-delete to 3 months.
  3. ChatGPT: switch off model improvement. Open Settings > Data controls, the page OpenAI's help center names for choosing how your data is used, and turn off the setting that lets your chats improve OpenAI's models.
  4. Clean-up: review ChatGPT's memory. OpenAI's memory announcement points to Settings > Personalization > Manage Memory. You can delete a single memory there or clear them all.
  5. Clean-up: clear old ChatGPT shared links. OpenAI's help center gives the path: Settings > Data controls > Shared links > Manage. Use "Delete shared link" on any link you no longer want people to open.

What AI tools collect, and how it can hurt you

A chatbot collects more than the text box suggests:

  • Prompts and outputs. Your questions and its answers, stored with your account.
  • Uploaded files. Contracts, spreadsheets and screenshots you attach.
  • Account and device details. Your email, your device and when you use the service.
  • Voice and images. Gemini gives audio and Gemini Live recordings their own training setting, apart from text.
  • Memory. Facts the assistant keeps about you from one chat to the next.

Your questions also reveal things you never typed. Three questions in one week about a medication's side effects, a clinic's opening hours and insurance coverage say a lot about your health without stating anything outright. "I never mentioned my name" protects you less than it sounds.

Stored chats can hurt you in several ways. Anyone who gets into your account can read them, and so can anyone holding a link you shared. A breach at the provider exposes them. Details about your job, your colleagues and your suppliers are what a convincing phishing email is built from. And once a chat has gone into training, Anthropic's own terms make clear you can't pull it back out of a model that's already trained.

What not to paste into a chatbot

  • Passwords, credentials and API keys. Anyone who can read the stored chat can use them.
  • Health records and financial details. These do the most damage if they leak, and you can usually describe them in general terms instead.
  • Client or employer data. Your contract with the client, not the chatbot's terms, decides whether you're allowed to.
  • Other people's personal information. They never agreed to it, and they can't opt out.

You can usually get the help you need by describing the situation. The model doesn't need your client's name to help you write a firm reply.

Consumer accounts and work accounts get different handling

The same prompt is handled differently depending on the plan. OpenAI's business-data page says it doesn't train on data from ChatGPT Enterprise, Business or Edu by default. Anthropic says its August 2025 consumer-terms change doesn't apply to Team, Enterprise, Government, Education or API use, and none of those are trained on by default. On Google Workspace accounts, the admin controls how data is handled.

Using a free personal account for client work is the worst setup. You get the consumer defaults and none of the work-plan protections. A paid work plan only helps if its contract covers what your client requires. Before any client material goes in, I'd get the data processing terms in writing. I wouldn't put anything confidential into a free tier at all, even with training off.

Memory, connected apps and built-in AI

Memory turns a series of separate chats into a profile of you. Connected apps go further. Once an assistant can read your email, calendar or drive, one question can pull in documents you never mentioned. What Is MCP? How ChatGPT and Claude Connect to Your Apps, and What Can Go Wrong explains how those connections work and where they leak. If you've connected apps, go through them the way you go through memory and remove the ones you don't use.

AI built into other apps works the same way. Notion AI works across your whole workspace, and its listed terms say it doesn't train on your content. Grammarly trains on your content unless you opt out. Automation tools add another vendor. If a workflow in Zapier or n8n sends a client email to an AI service for a summary, a second company now processes that email.

The same goes for posts and photos you published on LinkedIn, X, Meta's apps or Snapchat, and for pages you open in a browser with a built-in assistant: they can reach AI systems you never opened yourself. On each platform and on your phone, search the settings for "AI", "training" or "model improvement", switch off what you don't use, and take away the assistant's access to messages, photos and your screen if it doesn't need them.

Meeting recorders

Recorders sit in on your most sensitive conversations, and their terms vary a lot:

ToolTrains on your contentFree plan limitPaid from
Fireflies.aiNo400 mins storage per team$10/seat/mo billed annually
Otter.aiYes300 transcription minutes a month$8.33/user/mo billed annually
GranolaYes, opt-out availableMeeting history limited to 30 days$14/mo

If your team records client calls, the training column matters more than the price. Otter is the cheapest, and it trains on your content, so I wouldn't use it for client calls. Fireflies costs $10 a seat and doesn't train. Our Fireflies.ai review covers which teams should pay for it. If recordings shouldn't leave your computer at all, the open-source Otter alternatives run on your own machine. Whichever you use, tell people on the call that they're being recorded.

Screen recorders on your PC

Microsoft Recall keeps searchable snapshots of your screen history on Windows 11 Copilot+ PCs. Our directory lists it as not training on your content. A screen history still captures whatever was on screen, including passwords, banking pages and private messages. Screenpipe records locally, doesn't train on your content, and has a free plan for one device, with paid plans from $21 a month. If you want screen recall at all, I'd lean toward Screenpipe, because its code is open source and you or your IT team can see what it captures.

What deleting a chat does and doesn't do

"Delete" can mean three different things:

Three cards: 1 Not used for training, which doesn't reach already-trained models, flagged Claude chats or rating feedback. 2 Gone from your history, while Gemini temporary chats stay about two months. 3 Gone from servers: Claude within 30 days, flagged chats up to 2 years, Gemini reviewed chats 3 ye

  • Not used for training. A setting stops future use.
  • Gone from your history. You can no longer see it.
  • Gone from the provider's servers. This takes time and has exceptions.

Anthropic says deleted chats leave its back-end storage within 30 days and drop out of future training, but flagged content follows the longer windows in the table. Google keeps human-reviewed chats for their full 3 years even after you delete your history or change the setting. Deleting protects you from casual exposure, like someone opening your account. It doesn't mean the provider has erased everything.

Gemini's temporary chats keep the conversation out of your visible history and out of training, but the Gemini Apps Privacy Hub says they still stay with your account for about two months. A temporary chat disappears from your screen well before it disappears from Google.

Asking a provider for your data

In the EU, GDPR lets you ask a company what it holds about you and ask it to delete that data. In California, CCPA gives you the right to know what's collected and the right to have it deleted. For a quick look at your ChatGPT data without a formal request, OpenAI's help center points to Export data under Settings > Data controls. For a formal request:

  1. Find the provider's privacy page or privacy contact address.
  2. Name the law (GDPR or CCPA), and say whether you want access, deletion or both.
  3. Use the email address tied to your account so they can verify you.
  4. Note the deadline. GDPR gives the company one month to respond, which it can extend if the request is complex, as long as it tells you. CCPA gives 45 days, which the business can extend once if it notifies you.
  5. Save the confirmation and the date you sent the request, and follow up the day after the deadline passes.

A request doesn't override the retention rules in the table. Flagged and human-reviewed chats can outlast it. Before you ask for deletion, export anything you want to keep.

Who can read your chats, and who can ask for them

Does AI spy on you? The assistant itself doesn't watch you, but what you type is stored, and people can read it. Google says in its Gemini Apps Privacy Hub that it uses chats to protect its services and users, including through human reviewers, whatever your settings. Anthropic sends chats flagged for safety review into longer retention and can use them for training. Wirecutter's guide to AI chat privacy warns that ChatGPT and other chatbot conversations may be reviewed by people too.

Are your chats shared with law enforcement? The same Wirecutter guide says chats can be requested by law enforcement. A request can only reach what the provider still holds, so the retention column in the table sets your exposure. That's one more reason I'd keep sensitive questions in the tool with the shortest window.

To keep your text off a provider's servers completely, run a model on your own hardware. Ollama and LM Studio both say they don't train on your content, and models you run locally stay on your machine. If you'd rather stay with a cloud AI chatbot, the terms differ: DuckDuckGo, whose Duck.ai chat sits alongside its browser and search engine, says it doesn't train on your content, while Mistral trains unless you opt out. NotebookLM answers questions from the sources you upload, but it still runs in the cloud. Our NotebookLM review covers its limits.

Who each step is wrong for

  • Switching off training may cost you some personalization, and it won't undo data that's already been used.
  • Short auto-delete is a bad idea if you use old Gemini chats as a record. Export what you need first.
  • Clearing memory is a bad idea if you rely on the assistant remembering your style or projects. You'll have to explain them again.
  • Running a local model suits people with capable hardware and sensitive work. For casual use it's more trouble than it's worth.
  • A business plan is wasted on a solo user with no sensitive data. A free tier with training off is enough.

Claude and Gemini details last verified 18 November 2026.

FAQ

Does rating an answer with a thumbs up or down change anything?

In Claude, yes. Anthropic keeps thumbs-up and thumbs-down feedback for 5 years, and feedback you send is one of the exceptions it can use for training regardless of your model-improvement setting. If a chat is sensitive, skip the rating.

If I turn Claude training on later, does it reach my old chats?

Only the ones you touch again. Anthropic applies the 5-year training retention to new chats and to old chats you resume after switching the setting on. Leave an old sensitive thread closed rather than adding to it.

On a work Google account, who decides how long Gemini keeps my chats?

Your admin. Google's Workspace privacy hub says admins can set conversations to auto-delete after 3, 18 or 36 months, with 18 as the default. You can't shorten it yourself, so ask IT which setting your organization uses.

Is everything in Ollama local?

No. Ollama's free plan runs models on your own hardware without limit, but it also sells access to cloud models paid per token. Those requests go to Ollama's servers, so pick a model that runs on your machine if the point is to keep text at home.

Written by

Stanislav Vovchyk

Founder and editor of infobro.ai

I run infobro.ai: I decide what it covers and how tools are evaluated.

I research every tool from its documentation, pricing pages, product screenshots and public developer discussion, and say plainly what the evidence does and does not show.

More guides

Guide

Free AI Text to Speech With No Sign-Up: Where 'No Limit' Ends

Nov 16, 2026 · 12 min read