AI for Life
Which AI assistant for what: comparing what the publishers actually document
Every comparison of AI assistants ends with a winner. This one does not, for a simple reason: we have not tested these tools yet, and a ranking without a test is an opinion with a table around it. What we can do today is read what each publisher documents, and on the five questions below the answers differ far more than the marketing does.
What this page compares, and what it does not
It does not compare quality, speed, price or how the answers feel. It compares what the publisher writes on its own documentation pages about your data: what gets used for training, how long it is kept, who can read it, and what happens when you delete. Since autumn 2026 a fifth question applies to the agents that act for you: what they can do without asking you.
One rule throughout: where a page does not state something, we say so instead of filling the gap. In particular, none of the pages we read states its own default setting plainly enough for us to publish it as a fact, so this guide describes the documented conditions and controls rather than what is switched on when you open an account. Check that in your own settings, today, because it is the kind of thing that changes.
1. What is used to train the models
- ChatGPT. One setting governs it, called Improve the model for everyone. When it is off, OpenAI documents that new conversations are not used to train its models. Separately, it states that by default it does not use content from Business, Enterprise, Edu or ChatGPT for Healthcare workspaces to train models.
- Claude. Anthropic documents three cases, and only three, in which consumer chats and coding sessions are used to improve its models: you allow it in your privacy settings, your conversations are flagged for safety review, or you explicitly opt in, for example through its Trusted Tester Programme. It also documents what is excluded: raw content from connectors such as Google Drive, including remote and local MCP servers, unless you copy it into the conversation. Incognito chats are never used, even with model improvement enabled.
- Gemini. Google states that its use of your data, as described in its privacy policy, extends to the generative AI models and other machine learning technologies that power its services.
- Grok. SpaceXSI documents that it may use your content and interactions, along with Grok's responses, to train its models, and that you control this through Settings, Data, Improve the Model. Private Chat keeps content out of training. Two documented exceptions matter: feedback you choose to send after opting out may still be used for training, and if you use Grok without logging in, in some regions outside the EU and the UK, there is no opt-out at all.
2. Whether a human can read your conversation
This is the question people never think to ask, and the one with the bluntest answer.
- Gemini. Google states plainly that human reviewers, including specialised reviewers from its service providers, read some of the data it collects, and asks you not to enter confidential information you would not want a reviewer to see or Google to use to improve its services.
- Claude. Human involvement appears through the safety route: conversations flagged for safety review may be used or analysed to improve detection and enforcement of its Usage Policy.
- ChatGPT. The data controls page documents that temporary chats may be retained for up to 30 days for safety purposes, and that safety features may use limited, safety-relevant context in rare, high-risk situations.
- Grok. SpaceXSI documents that a limited number of authorised personnel may review your conversations for specific business purposes, including improving model performance, investigating security incidents and misuse, and complying with legal obligations.
Whatever the tool, the practical rule does not change: anything you would not want a stranger to read does not go in the box. Our guide on protecting your personal data with an AI assistant covers the settings one by one.
3. How long your data stays
- Claude publishes the most precise figures we found. Delete a conversation and it leaves your history immediately and back-end storage within 30 days. If you allow training, data may be kept de-identified for up to 5 years in model training pipelines, and only for chats that are new or resumed after you enabled the setting. If a chat is flagged by automated systems as violating the Usage Policy, inputs and outputs are kept up to 2 years and classification scores up to 7 years. Feedback sent with the thumbs buttons is kept 5 years, de-linked from your user ID.
- Gemini runs on an auto-delete setting in your activity: 18 months by default, changeable to 3 months, 36 months, or no automatic deletion, with manual deletion available at any time.
- ChatGPT documents the 30-day safety retention for temporary chats, and self-service export and account deletion for signed-in Free, Go, Plus and Pro users. On managed workspaces, export and deletion go through the workspace owner.
- Grok publishes no fixed retention period in its privacy policy, only an ongoing legitimate business need. Its FAQ is more precise on deletion: what you delete, and Private Chat conversations, leave its systems within 30 days, unless de-identified and disassociated from your account or kept for safety, security or legal reasons. Unregistered users have no data retained for their use after the session.
4. What deleting actually deletes
If you read one section of this page, read this one.
- Gemini. Google documents that conversations reviewed by humans are not deleted when you delete your activity: they are kept, with associated data such as language, device type, location information and your feedback, for up to three years. It also states that deleting your Gemini activity does not delete data saved by the other Google services you used it with.
- Claude. Turning off the model improvement setting stops future training on your previous and new chats, but data already included in training runs under way, or in models already trained, stays there. Deleting a chat means it will not be used to train future models.
- ChatGPT. Turning off Improve the model for everyone does not delete or hide saved chats, and archiving a chat does not change whether it can be used for training.
- Grok. The 30-day window applies to what you actually delete. The catch is elsewhere: grok.com, the mobile apps and Grok inside X do not share the same settings, so opting out in one place settles nothing in the others, and the X route is handled through the X Help Center.
In every case, the lesson is the same: a privacy setting is about the future, not the past.
5. And the agents that act for you
Three products launched between August and October 2026 move the question. An assistant answers; an agent sends the email, fills in the form and pays. Muse from Meta, announced on 8 September, Grok Bot from SpaceXSI, announced on 11 August, and Claude Duet from Anthropic, announced on 21 October, share the same architecture: a computer of their own in the cloud, a browser of their own, and work that carries on when your phone is off.
Two things before anything else. Neither consumer agent is available in Europe: Meta's help centre states that the Muse app and website are not yet available everywhere, and the announcements we read name the US, Canada, then Singapore, never a European country; Claude Duet is a research preview, by waitlist, in the United States only. And all three ask for approval before acting, which changes the question to ask: no longer only what it does with your data, but what it can do without asking you.
- What needs your approval. Anthropic documents that Claude Duet shows the action, the destination and the content, every recipient included, then waits, before it sends a message, makes a payment, signs a document or deletes anything. Meta documents that Muse checks with you before a sensitive action such as sending an email or making a purchase, and that you can allow once, always allow, or deny.
- What stays closed even with your approval. Anthropic puts it in writing: trading securities, moving money between accounts, reaching into someone else's account. The Muse pages we read state no equivalent list.
- What never leaves. Both publishers keep credentials away from the model: Meta files them in a secure credential store and documents three checks before one is used, Anthropic has you type them into a form encrypted with a single-use key. Claude Duet adds a classifier that never shares a sign-in credential, a full card or bank account number or a government ID number, in its own words even if you ask.
- What disconnecting does not do. The most useful sentence in either documentation. Disconnecting an app stops further access, but Anthropic states that what Claude Duet already learned stays in memory until you reset it; Meta offers to let you ask Muse to forget a subject, or a reset it describes as permanent, chat history, files and active tasks included.
- Training. For Muse, Meta documents a setting that is on the first time you use it, changeable at any time, and whose change also applies to your previous interactions. For Claude Duet, the pages we read do not say whether its conversations follow the same rules as the rest of a Claude account, and its page says exactly that.
An agent also changes the security question, because it reads web pages and emails written by other people. Both publishers say so: a page or a message can carry instructions meant to hijack the agent, and their answer is the same, move the check outside the model. At Meta, a second agent isolated at system level, the Sentinel, approves anything that goes out to the internet. At Anthropic, the check that compares the action with your request runs outside the model, so the text on a page cannot change its outcome.
And for a specific task?
Once the data question is settled, the choice becomes a question of task. Each of these pages reports what the publisher documents, and says what we have not verified:
- Writing and everyday questions: ChatGPT, Claude, Gemini, Grok.
- Code: Cursor and GitHub Copilot. Read our guide on using a code assistant without shipping vulnerabilities first, because the study it cites matters more than the choice of tool.
- Images: Midjourney and Canva. Ownership differs sharply between the two, and our guide on publishing AI images explains why that matters before you publish.
- Video: Runway.
- Audio and meetings: ElevenLabs, Descript, Fireflies.
- Automation: Zapier and n8n, whose free routes are documented very differently.
- Acting for you: Muse and Claude Duet, neither of them available from Europe at the date of this update.
The three questions to ask of any tool
- What does it do with what I put in? Find the training setting before the first serious file, not after.
- What happens when I delete? Ask about the back end, not the interface. The two are rarely the same.
- Who else can read it? Human review, service providers, connected apps. The answer is in the documentation, and it is usually one sentence long.
If your team uses these tools for work, these three questions belong in writing: our guide on a one-page AI charter turns them into rules.
What we have not measured
No quality test, no speed test, no price, no ranking, no star rating. The hands-on tests are still to come, and when they happen each tool page will say so with its date. Until then, this page is worth exactly what its sources are worth, which is why they are listed below with the date we read them.
Tools in this guide
- Vendor information
ChatGPT
General-purpose assistant for drafting and research.
WritePrice: to verify
- Vendor information
Claude
Assistant for long documents and careful writing.
WritePrice: to verify
- Vendor information
Gemini
Google's assistant, linked to Workspace.
WritePrice: to verify
- Vendor information
Grok
Conversational assistant that can search public X posts and the live web.
WritePrice: to verify
- Vendor information
Muse
Meta's personal agent, which takes actions in the apps you connect.
AutomatePrice: to verify
- Vendor information
Claude Duet
Anthropic's proactive agent: it starts the conversation and asks before acting.
AutomatePrice: to verify
Questions
Which assistant is the best?
We cannot tell you, and anyone who does without a published method is guessing. We have not run hands-on tests yet, so this guide ranks nothing. It compares what each publisher documents, which is verifiable today.
Which one is the safest for confidential work?
Read the deletion section below rather than any badge. Claude documents deletion from back-end storage within 30 days; Google documents that conversations already reviewed by humans are kept up to three years and are not removed when you delete your activity.
Can an AI agent act without asking me?
The two consumer agents launched in autumn 2026 document the opposite: Claude Duet shows the action, the destination and the content, every recipient included, before it sends, pays, signs or deletes, and Muse checks with you before a sensitive action. What is worth reading is the list of what stays closed even with your approval, and what happens to its memory when you disconnect an app.
Do these settings change what my employer sees?
No. These are the publishers' rules. Your employer's rules are separate, and on managed workspaces the administrator controls export, retention and deletion.
Sources
- OpenAI Help Center, Data controls in ChatGPT
- Anthropic Privacy Center, Is my data used for model training?
- Anthropic Privacy Center, How long do you store my data?
- Google, Guide sur la confidentialité dans les applications Gemini
- SpaceXSI, Consumer FAQs
- SpaceXSI, Privacy Policy
- SpaceXSI, Introducing Grok Bot
- Meta Newsroom, Introducing Muse: The World's First Personal AI Agent Built for Everyone
- Meta Help Centre, How to manage your Muse data
- Meta Help Centre, How Muse handles your privacy, safety and security
- Anthropic, Introducing Claude Duet
- Anthropic, Claude Duet Safety, Privacy, and Security
One AI email a week. Only what changes something for you.