"Will ChatGPT train on what I paste into it?" It is the most-typed version of this question, and it has the wrong shape. Privacy in AI chat is a function of which tier and which contract you are on, not which logo is in the browser tab. The same company will happily train on your consumer chats and contractually promise never to touch your API traffic.
So the short answer: if you want an AI assistant that does not train on your data, you do not necessarily need a different product — you need a different plan. Business, team, enterprise, and API tiers across the major providers generally exclude your inputs from model training by default. Consumer free tiers historically have not. And if your threat model goes beyond training — if the data genuinely cannot leave the building — then the only real answer is running an open-weight model locally, and accepting that it will be worse.
First, get the threat model right
Most "AI privacy" advice collapses four separate fears into one word. They have different answers.
- Training. Will my text be used to improve a future model, so that some fragment of it could theoretically surface elsewhere? This is the fear everyone names, and it is the easiest one to solve — it is a settings toggle or a plan change.
- Retention. Is my text stored on someone's server, and for how long? Providers typically retain conversations to run the product and to monitor abuse, even when they are not training on them. Different problem, different fix.
- Legal exposure. Could that stored text be subpoenaed, breached, or handed to a government? Retention creates this risk regardless of training policy.
- Internal exposure. On an enterprise plan, your own admins may be able to see conversations. For the employee pasting a private grievance into the company chatbot, this is the actual risk, and no vendor policy protects them from it.
Notice that only the first one is about the model. If your worry is (3) or (4), switching from one hosted chatbot to another accomplishes precisely nothing.
The tier is the setting, not the brand
Across the major assistants, the pattern is consistent enough to state plainly, though every provider revises its terms and you should read the current policy page before betting a company on it.
- ChatGPT. The consumer app has historically used conversations to improve models unless you turn that off in the data controls, or use temporary chat. The API, and the Team/Enterprise/Edu tiers, are stated not to be used for training by default. The single highest-value action for most people is either flipping the data-controls toggle or moving the team onto a business plan.
- Claude. Anthropic has positioned commercial and API traffic as excluded from training by default, and its consumer terms have changed more than once — which is exactly why you check the policy rather than the reputation. If you are weighing the two assistants on more than privacy, our ChatGPT vs Claude comparison covers the rest.
- Gemini. The consumer product and the Workspace/Cloud products are governed by different terms, and the enterprise side is the one with the no-training commitments. Do not assume the free app inherits your organisation's Google contract — it usually does not.
The counter-intuitive consequence: the free tier is the leaky one. People reach for a free consumer chatbot precisely when they want to be discreet, which is the exact inversion of what the terms support.
What "zero data retention" actually means
Zero data retention (ZDR) is a contractual arrangement, usually available on API or enterprise agreements, in which the provider agrees not to persist your inputs and outputs after serving the request. It is the strongest promise a hosted provider can make, and it is worth understanding precisely what it does and does not cover.
- It is negotiated, not toggled. ZDR generally has to be enabled on your account, often after a conversation with sales, sometimes only above a spend threshold.
- Default retention exists for abuse monitoring. Without ZDR, providers commonly hold inputs for a limited window so they can investigate misuse. ZDR is the thing that removes that window. Get the specifics in writing.
- It says nothing about your own logging. If your application writes prompts to your own database, or your gateway keeps a trace, ZDR at the provider is irrelevant — you built the retention yourself.
- It is a promise, not a physical property. You are trading verification for capability. That is a rational trade for most businesses, and an unacceptable one for a few.
If a vendor's privacy page says "we take your privacy seriously" and does not use the words training, retention, or sub-processors, you have learned nothing. Our glossary unpacks the rest of the vocabulary these pages hide behind.
The fully local option
If the data cannot leave the machine — client-privileged material, patient notes, unannounced financials, source code under a strict NDA — then no policy is good enough and you run the model yourself.
Ollama is the standard entry point: install it, pull an open-weight model (Llama, Mistral, Qwen, Gemma and DeepSeek variants are all commonly used), and talk to it with no network connection at all. LM Studio and Jan wrap the same idea in a desktop UI for people who would rather not touch a terminal. Pair any of them with a local UI and you have a private chatbot with a genuinely different guarantee: the packets never leave.
The trade-off is real and you should not be sold past it. A model small enough to run on a laptop is meaningfully weaker than a frontier hosted model at reasoning, long-context work, and code. It has no web access unless you build it. Quality scales with your RAM and your patience. For summarising a sensitive document, drafting, and Q&A over your own files, a good local model is entirely sufficient. For the hardest analytical work, you will feel the gap immediately.
Also worth naming: routing an API key through a self-hosted UI is not local privacy. Open WebUI pointed at a hosted API is a self-hosted interface in front of someone else's model. The data still leaves. This confusion is extremely common and it defeats the entire exercise. If you want tools that genuinely function without a network, our guide to AI tools that work offline is the companion piece to this one.
| Option | Protects you from | Does not protect you from | Best for |
|---|---|---|---|
| Consumer free tier, training toggled off | Training on your chats | Retention, breach, subpoena | Casual, non-sensitive use |
| Team / Enterprise plan | Training; adds admin controls, SSO | Your own admins reading chats | Most companies, most of the time |
| API with zero data retention | Training and retention, contractually | Trust — it's still someone else's server | Regulated products, sensitive workloads |
| Local model (Ollama, LM Studio, Jan) | Everything — the data never leaves | Weaker output; your own device security | Legal, clinical, air-gapped, offline |
Swipe the table sideways to see every column →
The mistake most companies make
They ban the tools. Then the work does not stop, so the sales rep pastes the contract into a personal ChatGPT account at home — the consumer tier, on a personal login, entirely outside any policy the company could enforce. The ban did not remove the risk; it moved it somewhere invisible and made it worse.
The organisations handling this well do the opposite: they sanction a good tool loudly, put it on a business tier, and make it the path of least resistance. Shadow AI is a UX problem wearing a compliance costume. This is one of the first governance decisions worth getting right when you are building out a startup's tooling, and it is far cheaper to make now than after an incident.
A six-point checklist before you paste anything sensitive
- Which tier am I on — consumer, team, or API? Assume the consumer default is the permissive one.
- Is the training toggle off? Check it directly; do not rely on what you remember setting.
- Is retention specified in writing, with an actual window, or just implied?
- Who at my own company can read this? Admin consoles exist.
- Would I be comfortable if this text appeared in a breach disclosure? If not, it belongs in a local model.
- Does the vendor name its sub-processors? A chatbot built on someone else's API inherits that provider's terms too — and rarely says so on the homepage.
For anything privileged, the honest answer is often "do not use the chatbot for this part." Redact the names, ask the general question, keep the specifics in-house. Cheaper than any tool.
FAQ
Which AI chatbot does not train on your data?
Any of them, on the right plan. The API and business/enterprise tiers of the major providers — including ChatGPT, Claude, and Gemini — are stated not to use your inputs for model training by default, while consumer free tiers historically have unless you opt out. The choice that matters is the tier, not the brand. Verify on the provider's current policy page, because these terms genuinely do change.
Is there a truly private ChatGPT alternative?
Only a local one. Running an open-weight model through Ollama, LM Studio, or Jan means your prompts never touch a network, which is a categorically stronger guarantee than any policy. The cost is capability: local models that fit on a laptop are noticeably weaker than frontier hosted models, and you supply the hardware.
What does zero data retention actually mean?
It is a contractual commitment, usually on API or enterprise agreements, that the provider will not store your inputs and outputs after serving the request. Without it, providers typically keep data for a limited period to monitor abuse. It has to be arranged with the vendor rather than switched on in settings, and it does not stop your own application from logging prompts.
Does turning off chat history make ChatGPT private?
It helps with training, not with everything else. Disabling history or using temporary chat generally removes the conversation from model-improvement pipelines, but the data still transits the provider's servers and may be retained briefly for safety monitoring. It is a meaningful step and not a privacy guarantee.
Are self-hosted AI chat interfaces private?
Only if the model is also local. A self-hosted UI like Open WebUI pointed at a hosted API sends your data to that API exactly as the official app would — you have self-hosted the front end, not the intelligence. Privacy comes from where the model runs, not where the interface runs.
Pick the tier deliberately, read the retention terms once, and keep a local model around for the handful of things that genuinely cannot leave your machine. That combination covers almost everyone. Compare the assistants in the AI chatbot tools category, browse ChatGPT alternatives if you want to move off it entirely, or see which tools you can use without an account at all.