What AI apps really send about you

Updated October 17, 2026 · 12 min read

When you chat with ChatGPT, Gemini, Claude, Copilot, Meta AI or Vibe (formerly Le Chat), your messages, files and technical data such as your IP address go to the company’s servers. History is usually kept until you delete it, and on consumer plans most services may train their models on your chats unless you opt out. Here is the service-by-service detail, based on their policies as of October 17, 2026.

Based on the official privacy policies and help pages we read on October 17, 2026. They change often, so check the current version before relying on it.

The short version

• When you use ChatGPT, Gemini, Claude, Copilot, Meta AI or Vibe (Mistral’s new name for Le Chat), your message is processed on the company’s servers, not on your phone.
• Along with it go your files, photos or voice recordings, plus technical data such as your IP address, device type, operating system and timestamps.
• On consumer plans, most services may use your chats to improve their models unless you turn that off. Business plans don’t, by default.
• Chat history is usually kept until you delete it. Deleted or “temporary” chats are often still held on the servers for days or weeks.
• People can read some conversations: safety teams, contractors who rate answers, or authorities with a legal request.

What leaves your device when you hit send

The policies describe three kinds of data, broadly the same across services.

1. Content

Your prompts and anything you attach: documents, screenshots, photos, voice-mode audio and video, and data from connected services (calendar, email, cloud storage). OpenAI states plainly that it collects “your prompts and other content you upload,” including files, images, audio and video. The answers are stored too.

2. Technical data

Your device sends information automatically with every request. OpenAI lists your IP address, browser type and settings, date and time, device name and operating system. Anthropic adds mobile carrier, time zone and device identifiers. Google processes your general area (derived from your IP address, for example) and, with your permission, your precise location.

3. Context

Memory, custom instructions, profile details and interests can be attached to your messages to personalize replies. Meta says AI interactions include “information we use to make responses from AIs more relevant (like your interests, location and profile).”

Comparison table

Consumer plans (free and individual paid), unless noted.

Service by service

ChatGPT (OpenAI)

What’s sent: prompts, files, images, audio and video, connected-app data, plus IP address, device, browser and an approximate location based on your IP.

Retention: chats stay in your account until you delete them. Deleted chats are removed from OpenAI’s systems within 30 days, unless they were already de-identified and disassociated from your account, or security or legal obligations require keeping them. Temporary chats stay out of your history and aren’t used for training, but a copy may be kept for up to 30 days for safety purposes.

Training: on individual plans (Free, Plus, Pro), OpenAI says it may use your content to train its models. To opt out, turn off “Improve the model for everyone” in Settings › Data controls. One exception: if you rate a response with a thumbs up or down, the whole conversation may be used for training, even after you opt out.

Who can read it: a limited number of authorized staff and trusted service providers under confidentiality obligations, to investigate abuse or security incidents, provide support, handle legal matters, or improve models (unless you opted out). OpenAI asks you not to enter sensitive information you wouldn’t want reviewed. Data may also go to vendors and, where the law requires, to authorities.

Ads: OpenAI’s EU policy describes ads for Free and Go users; personalizing them with past chats depends on your consent.

Gemini (Google)

What’s sent: typed or spoken prompts, files, photos, videos, shared screens, Gemini Live recordings and transcripts, your general location (IP, Home and Work addresses) or precise location with permission, and device identifiers and details.

Retention: with Keep Activity on, activity is kept for 18 months by default, adjustable to 3 months, 36 months or indefinitely. With it off, and for temporary chats, conversations are kept with your account for 72 hours to respond and protect the service.

Training: with Keep Activity on, activity may be used to improve Google services, including training generative AI models. With it off, future chats aren’t used for training unless you send feedback. Gemini Live audio and video aren’t used for improvement by default (a separate setting).

Who can read it: human reviewers, including trained service providers, read a subset of chats. Those chats are disconnected from your account first, but kept for up to three years, and deleting your activity doesn’t remove them. Google advises against entering confidential information you wouldn’t want a reviewer to see, and says it doesn’t use your chats to show you ads.

Claude (Anthropic)

What’s sent: prompts, files and connected-service content, plus technical data: device type, OS, browser, carrier, time zone, IP address (and location derived from it) and device identifiers.

Retention: deleted chats leave your history immediately and the back end within 30 days. If you allow training, de-identified data may stay in training pipelines for up to 5 years. Chats flagged by safety systems are kept for up to 2 years (classification scores up to 7 years). Incognito chats stay out of your history and are kept for 30 days by default. The page we read doesn’t give a standard period for chats you don’t delete.

Training: under its privacy policy, Anthropic may use your chats to train its models unless you opt out in privacy settings. Even then, safety-flagged chats and chats you rate may be used. Incognito chats are never used.

Who can read it: according to its help center, access is limited to a small number of staff, after your data is de-linked from your user ID; the Trust and Safety team can review flagged chats. Data may also be shared with service providers and, with a legal request, authorities.

Microsoft Copilot

Microsoft documents two regimes depending on the app version (a new version has been available since August 18, 2026).

• New app (personal Microsoft account): Microsoft says prompts, responses and file contents aren’t used to train foundation models; optional feedback may be used to improve Copilot. The pages we read don’t state how long chat history is kept.
• Older app: conversations kept 18 months by default; uploaded files and images kept no longer than 18 months. They’re used for training unless you opt out, with exclusions (signed-out users, under-18s, Entra work accounts, some countries). Some conversations get automated and human review, with no opt-out when a Code of Conduct violation is suspected.

Microsoft advises against sharing sensitive personal data (health, religion, sexual orientation…) you wouldn’t want used under its privacy statement.

Meta AI

What’s sent: text, documents, images or recordings depending on the app, plus information used to tailor replies: interests, location, profile.

Use: Meta AI’s terms point to Meta’s Privacy Policy, under which your AI interactions are used to personalize your experiences and ads and improve AI at Meta. Meta warns: “Do not share information that you don’t want the AIs to use and retain, such as information about sensitive topics.”

Who can read it: Meta may review your interactions, automatically or manually (human). Messages and context such as region and interests may be shared with partners like search engines. Exception: in chats using Private Processing, Meta says it can’t read the messages.

Retention and opt-out: the pages we read give no retention period and no setting to exclude your chats from training; Meta points to its Privacy Center for access and deletion.

Vibe, formerly Le Chat (Mistral AI)

Le Chat is now called Vibe. What’s sent: prompts and documents, outputs, feedback, technical data (IP address, network protocol) and usage data. Data is hosted in the EU by default.

Retention: conversations are kept until you delete them or close your account. After deletion, some data may remain on servers for a while (policy enforcement, legal obligations); the page we read gives no figure.

Training: on Vibe (free and Pro), you are not opted out by default. To opt out, turn off “Allow your interactions to be used to train our models” at admin.mistral.ai › Vibe › Privacy. Attached documents count as input. Enterprise customers are opted out by default; on Teams, admins can disable training for the whole organization. Thumbs up/down feedback always lets Mistral use that exchange.

Consumer vs. business: two different rulebooks

This is the clearest divide. Business offerings (ChatGPT Business, Enterprise and Edu, Claude for Work and the Anthropic API, Gemini in Google Workspace, Copilot with an Entra account, Vibe Enterprise) don’t use your data for training by default, according to their providers.

The trade-off: your organization sets retention and log access. OpenAI lets Enterprise and Edu admins access conversations through compliance tools; Anthropic includes incognito chats from Team and Enterprise plans in organization exports. A work account shields you from training, not from your employer.

How to limit what you share

Settings to check

• ChatGPT: Settings › Data controls › turn off “Improve the model for everyone.” Use a temporary chat for one-off questions and delete chats you don’t need.
• Gemini: turn off Keep Activity or set auto-delete to 3 months. Temporary chats limit retention to 72 hours.
• Claude: Settings › Privacy › turn off model improvement. Use incognito (the ghost icon) for chats that shouldn’t stay in your history.
• Copilot: update the app, review the privacy controls (memory, personalization, personalized ads) and clear history from the Microsoft privacy dashboard.
• Meta AI: with no training setting in the pages we read, the only safe step is to keep sensitive topics out. Don’t share conversation links: anyone with the link can read the chat.
• Vibe: admin.mistral.ai › Vibe › Privacy › turn off the training toggle.

Good habits

• Don’t paste sensitive data: card numbers, passwords, medical records, client files, other people’s information. Every provider recommends this itself.
• Redact before sending: swap names, addresses and numbers for placeholders (“Client A”).
• Skip the thumbs up/down if you opted out: at OpenAI, Anthropic and Mistral, feedback can make that conversation usable anyway.
• At work, use the company account, not a free personal one: that’s where the contractual protections are.
• Watch your connectors (email, calendar, storage): each one widens what the assistant can read.

The alternative: AI that runs on your phone

All these rules exist because your message leaves the device. The other approach is to run the model on the phone’s own chip. Open models such as Google’s Gemma are compact enough for a recent iPhone, using open-source engines like llama.cpp.

That’s how DontKeep works: you download a model once from Hugging Face (the app’s only connection, which, like any download, gives Hugging Face your device’s IP address), then everything runs on the iPhone, even in airplane mode. No server receives your messages, there’s no account and no usage analytics, and the conversation is never written to disk: it disappears when the app closes, and at the latest after an adjustable delay (one hour by default). To keep a chat, “Keep” encrypts it on the device. Details are in our privacy policy.

There’s a trade-off: a model that fits on a phone is less capable than a large cloud model, and it doesn’t search the web. For drafting, summarizing, rewording or thinking through something personal, it’s often enough.

What still leaves an iPhone

A local app sends nothing from your conversations, but the iPhone itself records a few things:

• iOS logs app usage: Screen Time, battery use and data usage (including the model download). No app can prevent this.
• iOS dictation: according to Apple, Keyboard settings show whether audio is processed on device; otherwise it’s sent to Apple’s servers, and not stored unless you opted in to Improve Siri and Dictation.
• Third-party keyboards: if you use one, what you type is also subject to its privacy policy, especially if you granted it full access.

FAQ

Does ChatGPT read my conversations?

They’re processed automatically to answer and to detect abuse. Authorized staff and contractors can access them in specific cases (abuse, support, legal matters, model improvement unless you opted out), according to OpenAI.

Are my chats used to train AI?

On consumer ChatGPT, Claude and Vibe: yes, unless you opt out. On Gemini: yes if Keep Activity is on. On the new Copilot app: no, according to Microsoft. Meta uses AI interactions to improve its AI. Business plans don’t by default.

Is temporary or incognito mode enough?

It keeps chats out of history and training, but not off the servers: up to 30 days at OpenAI, 30 days by default at Anthropic, 72 hours at Google.

Does deleting a chat really erase it?

It leaves your history immediately, then the servers after a delay (30 days at OpenAI and Anthropic). Exceptions: legal obligations, flagged chats and, at Google, chats already reviewed by humans, kept up to three years.

Does paying for a subscription protect my data?

Not by itself. ChatGPT Plus or Pro, Claude Pro or Max and Vibe Pro follow the same training rules as free plans. Business plans are what change the rules.

Is a local AI completely private?

Your chats never leave the phone, which removes third-party retention from the picture. What remains is what iOS records (Screen Time, battery, network data) and what your keyboard does. Lock your iPhone: anyone holding it can open the chats you chose to keep.