You flip one privacy toggle in ChatGPT, exhale, and assume you are now invisible. You are not. Yes, AI can train on you, and the default differs by vendor: ChatGPT and Gemini train on your chats unless you opt out, while Claude trains only if you opt in. Same question, opposite defaults. And the switch you just flipped fixes less than you think. Paid plans get no automatic exemption. Turning training off does not erase what already lives inside a trained model. That last part is the one thing no toggle can fix.
Most articles answer this with a single yes or no. The accurate answer needs four. Privacy panic starts when people collapse four separate things into one: they hear "AI trains on you," picture a single switch, flip it, and feel safe. That is not how any of these systems work. Memory, training, retention, and baked-in weights are four distinct mechanisms. Each has its own control and its own consequence. This post separates them, gives you the per-vendor defaults with primary sources, and names the one setting that reaches nothing.
The four things people confuse for one
These four mechanisms run independently. Switch one off and the other three keep going. First, the vocabulary.
1. Memory personalization
Memory is recall inside your own account. A chatbot remembers facts about you across sessions: your name, your preferences, an ongoing project. It makes replies feel personal. But recall inside your account is not the same as your data training the underlying model for other people. You can run memory on and training off at the same time.
2. Training opt-out
Training is when your chats become material that improves the model itself, often including review by human raters. This is the setting most people mean when they ask "does AI train on my data." It is a separate toggle from memory. On most platforms it also governs whether human reviewers can read sampled conversations.
3. The chat-retention window
Retention is how long the company keeps your conversation on its servers, regardless of training. Even after you delete a chat or opt out, providers hold the data for a window to screen for abuse, comply with the law, and keep the service running. Deleting from your view and deleting from their systems are two different events.
4. Data already baked into the weights
This is the one no setting reaches. Once a conversation has been folded into a finished training run, it is statistically absorbed into the model's parameters. Opting out is forward-looking only. It stops future use. It cannot pull your words back out of a model that already learned from them.
Rule of thumb: memory controls what the model remembers about you. Training controls what the model learns from you. Retention controls how long they keep the raw text. Weights are the part nobody can un-bake.
Which AI chatbots train on your data by default?
Only one of the three waits for your yes. ChatGPT and Gemini opt you in by default and make you opt out. Anthropic's Claude trains on consumer chats only if you opt in. The default is what matters, because most people never open settings. And none of the three exempts paid consumer tiers automatically. The exemptions cover business and API products, not Plus, Pro, or Max.
Anthropic Claude: opt-in to train
Anthropic updated its consumer terms so Claude can use your chats and coding sessions for model training, but only if you turn the setting on. "You're always in control of this setting," the company states. Existing users had until October 8, 2025 to make a choice. Decline, and your data stays on the 30-day retention period. Allow training, and retention extends to five years. This applies to Claude Free, Pro, and Max. It does not apply to Claude for Work, Government, Education, or API access.
Letting Claude train on your chats stretches retention from 30 days to five years. Decline and you stay on the 30-day window. The training choice and the retention window are the same lever.
OpenAI ChatGPT: opt-out, on by default
On personal ChatGPT accounts (Free, Plus, and Pro), OpenAI uses your conversations to improve its models by default. To stop it, open Settings, then Data Controls, then turn off "Improve the model for everyone." The change applies to new conversations going forward. OpenAI does not train on inputs or outputs from ChatGPT Team, Enterprise, or the API by default. So a paid Plus or Pro subscription does not opt you out. Only the business products do.
Google Gemini: opt-out, with human review
Gemini Apps Activity is on by default for eligible adults. With it on, humans, including trained service providers, review a subset of conversations to improve Google's services and the Gemini models. The default auto-delete window is 18 months, adjustable to 3 or 36 months. Google keeps any conversation a reviewer saw for up to three years, disconnected from your account, so you cannot later find or delete it. Turning off Keep Activity stops new chats from feeding review. Google makes no paid-tier exemption in this consumer policy.
| Mechanism | Claude (Free/Pro/Max) | ChatGPT (Free/Plus/Pro) | Gemini (consumer) |
|---|---|---|---|
| Trains on chats by default | No, opt-in only | Yes, opt-out to stop | Yes, opt-out to stop |
| Where to change it | Model training setting | Settings > Data Controls > Improve the model for everyone | Gemini Apps Activity |
| Human review of sampled chats | Tied to training opt-in | Limited reviewers for abuse and policy review | Subset reviewed by humans by default |
| Retention if you do not allow training | 30 days | ~30 days after deletion for abuse screening | Up to 72 hours when Activity is off |
| Retention if training is allowed | 5 years | Tied to your data controls and legal holds | 18 months default; up to 3 years if human-reviewed |
| Paid consumer tier auto-exempt | No | No (only Team/Enterprise/API) | No |
Default beats intention. The vendor's default decides what happens to the data of everyone who never opens settings, which is most people. If a platform is opt-out, assume your prior chats were eligible for training unless you changed the setting earlier.
"I opted out" and "my data is gone" are different sentences
Opting out of training does not delete your stored chats. Your text does not vanish. The common retention figure is around 30 days. Providers keep that window for abuse screening, security, and legal compliance. Delete a chat in ChatGPT Free, Plus, or Pro and it leaves your account immediately, then gets scheduled for permanent deletion within 30 days, unless retention is required for legal or security reasons. OpenAI keeps conversations during that window to screen for abuse and policy violations.
Anthropic mirrors this logic. Decline training and you sit on a 30-day retention period rather than the five-year window that training enables. Google's number looks shorter on paper, up to 72 hours when Keep Activity is off, but any chat a reviewer already touched is held on a separate three-year path you cannot reach. The pattern is consistent. Opt-out controls the training pipeline. It does not zero out the servers.
Legal holds can override everything. On May 13, 2025, a federal preservation order in The New York Times v. OpenAI required OpenAI to retain user conversations, including deleted ones, as potential evidence. That obligation was later lifted, but the lesson stands: a court order sits above any privacy toggle a product offers.
These companies are not acting in bad faith. The retention windows exist for real reasons: catching abuse, fixing outages, satisfying subpoenas. But "I opted out" and "my data is gone" are different sentences. Treat the gap between them as normal infrastructure, not a conspiracy.
Does deleting a chat or memory remove your data from the AI model?
No. Deleting a saved memory, or even your whole chat history, does not remove data already folded into a model's weights. This is the most misunderstood point in AI privacy. The memory feature and the training corpus are separate stores. Clearing one does nothing to the other.
Here is what most guides won't tell you: the opt-out is forward-looking by design, and that is a property of the math, not a loophole. Training does not file your sentence in a row you can later delete. It nudges billions of parameters by tiny amounts across the whole dataset. There is no "your data" object sitting in the model to pull out. Researchers call the attempt to reverse this "machine unlearning," and it stays an open, imperfect problem rather than a button vendors can ship.
You can opt out of tomorrow. You cannot opt out of yesterday. Anything you shared with an opt-out-by-default chatbot before you flipped the switch was eligible material, and no setting reaches it now.
Opting out today protects tomorrow's training runs. It does nothing about runs that already happened. If you pasted something sensitive into an opt-out-by-default chatbot before you flipped the switch, assume it was fair game. The fix going forward is prevention: do not type secrets into a system whose default is to learn from you and that you have not configured.
Audit, then decide per platform. Check the training toggle on every chatbot you use, note its default, and assume the pre-toggle history on opt-out platforms was fair game. For anything truly sensitive, prefer a tool that does not train on your data at all rather than one you have to police.
How to stop AI from training on your data: a 4-step checklist
Run this once per chatbot you use. Each step targets a different mechanism, so do not assume one switch covers the rest.
- Training: find the "improve the model" or "model training" setting on each chatbot and set it the way you want. On ChatGPT and Gemini that means opting out; on Claude it means leaving opt-in off.
- Memory: decide separately whether you want personalization. Turning memory off does not affect training, and vice versa.
- Retention: know the window. Deleting a chat starts a clock (around 30 days on ChatGPT) rather than an instant erase, and legal holds can pause that clock.
- Human review: on Gemini, the default samples chats for human reviewers; opting out of Activity is what limits that.
- Sensitive data: for anything you would never want in a training set, use a platform that does not train on user data by default, or do not type it at all.
If you want the deeper question of who actually owns and controls these memories, read our companion piece on who owns your AI memory at /blog/who-owns-your-ai-memory. For getting your data back out, see whether you can export your AI memory at /blog/can-you-export-your-ai-memory.
A private AI memory that does not train on your data
The cleanest way out of the toggle-juggling is a tool whose default is already the safe one. MemX (memx.app) ships that default out of the box. MemX is a consumer AI memory app, a personal second brain you capture into and ask in plain English. Save photos, PDFs, documents, voice notes, and WhatsApp messages, then retrieve them by asking a question. MemX answers from your own saved material and cites the source memory it pulled from. It reads text inside photos and scans with OCR, so the contents are searchable, not just the filenames. Say a reminder naturally and it writes the entry to your calendar, recurring included.
On the privacy axis this post is about, MemX defaults to the setting you would otherwise have to hunt for elsewhere: it encrypts your data, never trains AI on it, and keeps it out of reach of MemX staff. There is a free tier with no credit card on Android, iOS via TestFlight, and WhatsApp. It will not solve the un-baking problem for data you already gave to an opt-out-by-default chatbot, because nothing can. It does mean the memory you build here is not feeding a training run in the first place.
The honest version: AI can train on you, the default decides whether it does unless you intervene, retention outlives your delete button, and weights cannot be un-baked. Four mechanisms, four controls. Knowing which one you are touching is the entire difference between feeling private and being private.
Written by Aditya Kumar Jha, who works on AI memory at MemX.
01Does AI train on my data by default?
It depends on the vendor. ChatGPT and Gemini opt consumer accounts in by default, so you must opt out. Claude is opt-in, so it trains on your chats only if you turn the setting on. Check each platform's setting individually.
02If I turn off training, is my data deleted?
No. Opting out stops future training but does not erase stored chats. Providers keep a retention window, commonly around 30 days for abuse and security screening, and legal holds can extend it. Opt-out and erasure are separate actions.
03Does deleting a memory remove my data from the AI model?
No. Memory and the training corpus are separate. Deleting a saved memory or your chat history does not pull data out of a model that already trained on it. Training absorbs text into weights, which opt-outs cannot reverse.
04Are paid ChatGPT, Claude, or Gemini plans exempt from training?
No. Plus, Pro, and Max consumer tiers follow the same defaults as free accounts. Only business and developer products, such as ChatGPT Team, Enterprise, and the API, are not trained on by default. Paying does not opt you out.
05Can I make an AI forget data already in its weights?
Not reliably. Removing specific data from a trained model is called machine unlearning, and it remains an unsolved research problem. Vendors describe opt-outs as forward-looking only. The practical fix is to avoid sharing sensitive data with opt-out-by-default systems.
