AI chatbots and agents: what you give away, and how to keep it
Every AI chatbot keeps something of what you type: your conversations, for a period, and sometimes for training or human review. The big services all offer controls, and their business plans keep more back than their free ones. But the surest way to keep a conversation private is to run the model yourself. Open-weight models and free software such as Ollama and Open WebUI now make that practical on an ordinary server. This article sets out what the main services keep, in their own words, and what running your own model involves.
What happens to what you type
When you send a question to a hosted chatbot, the text travels to the provider's servers, where the model answers it. What happens next depends on the service, the plan and your settings:
- Training. Whether your conversations help train future models.
- Retention. How long they are kept, including after you delete them.
- Human review. Whether people at the provider may read a sample.
For a business, the question is simple: is anything you type something you would be uncomfortable seeing outside the company? Client details, contracts, salaries and passwords usually are.
The big four, in their own words
As described on each provider's help pages on 19 October 2026:
| Service | Training on your chats | Retention | Notes |
|---|---|---|---|
| ChatGPT (personal plans) | Yes, unless you switch off "Improve the model for everyone" | Kept in history; deleted chats removed within 30 days | Temporary chats are not used for training. Business, Enterprise and Edu workspaces are not used for training by default. |
| Claude (Free, Pro, Max) | Your choice, set in privacy settings | Five years if you allow training; 30 days if not | Team, Enterprise and API use are excluded from training. |
| Gemini (consumer app) | Activity may be used to improve Google's services | 18 months by default, adjustable | People review a sample of chats; reviewed chats are kept up to three years, even after you delete them. |
| Copilot (personal accounts, app updated August 2026) | Microsoft says chats are not used to train its foundation models | Kept until you delete them | Memory of your preferences; adverts may be personalised using chat history and memories. |
Google's own advice is worth repeating: do not enter confidential information you would not want a reviewer to see. And for developers, OpenAI says data sent through its API is not used for training unless the customer opts in, with abuse-monitoring logs kept for up to 30 days by default.
These terms change. Check the current ones for the plan you actually use.
Agents hold more than words
AI agents go a step further than chatbots: they act for you. OpenAI's ChatGPT agent, for example, works in a remote browser, can sign in to websites and apps, and can then reach email, files and account settings. OpenAI says the agent's browsing history and screenshots stay in your conversation history until you delete them, that sign-in cookies persist between sessions, and that authorised staff and service providers may access agent content.
OpenAI's own advice is sensible: think about how sensitive a site is before you let the agent log in, and clear the remote browser's data after sensitive sessions. Better still, give an agent its own limited accounts rather than your main ones. Email access deserves the most care of all, because an agent that can read your mail can be steered by what the mail says.
The other route: a model on your own server
You can also run an AI model yourself. "Open-weight" models are published for anyone to download and run. When a model runs on your own server, your prompts and its answers go nowhere else: not to the model's maker, not to a reviewer, not into anyone's training data. You decide how long anything is kept.
It also keeps personal information under your control. Pasting client details into a hosted chatbot hands them to a third party. A model on a server you control keeps that decision, and where the data lives, in your hands.
The free software that makes it work
| Project | What it does | Licence |
|---|---|---|
| Ollama | Downloads and runs open models on your server, with a simple command line and API | MIT |
| llama.cpp | The engine underneath many tools; runs models efficiently on ordinary processors | MIT |
| LocalAI | Runs models with no graphics card required, behind an API compatible with the big providers' | MIT |
| vLLM | Serves models to many users at once; built for graphics cards | Apache 2.0 |
| Open WebUI | A chat interface, like the familiar ones, that runs entirely on your own server | Its own licence: free to use, with a requirement to keep its branding above 50 users |
A common set-up is Ollama to run the model and Open WebUI to give your team a familiar chat window, both on one server.
Read the model's licence
Open-weight does not always mean free of conditions. Some of the most capable models carry the plain Apache 2.0 licence: Google's Gemma 4, Mistral's open-weight models and Alibaba's Qwen3 family. Meta's Llama models use Meta's own licence, which asks you to show "Built with Llama", follow an acceptable use policy, and, for some multimodal versions, does not grant rights to companies based in the EU. Read the licence of the exact model you choose.
What it takes, honestly
Running your own model moves responsibility to you, and that is a security matter as much as a privacy one:
- Hardware. Large models need graphics cards. Small models, with a few billion parameters, run on an ordinary server's processor and memory, more slowly. They are good for drafting, summarising and answering questions about your own documents.
- Access control. Ollama listens only on its own machine by default and has no login of its own. Never put it on the internet without a login and encryption in front of it.
- Updates and backups, like any other software on a server.
New server owners can start with our guides to setting up an Ubuntu server, SSH key sign-in and a firewall.
Running it on Noiz
A Noiz Cloud server gives you a Linux server of your own, with full root access, Ubuntu or Debian, and a dedicated IP address, ready within one working day. Install Ollama and Open WebUI over SSH, following each project's own documentation. If you would rather we looked after the server, our Linux administration is available once, by the hour or monthly, and the managed Backup service keeps copies under terms agreed with you. If you are weighing up a set-up for your team, get in touch.
Sources
Checked 19 October 2026.
AI providers
- OpenAI, Data controls in ChatGPT
- OpenAI, Data controls in the OpenAI platform
- OpenAI, ChatGPT agent
- Anthropic, Updates to consumer terms and privacy policy (August 2025)
- Google, Gemini Apps Privacy Hub
- Microsoft, Copilot for individuals: your activity history and privacy controls
Self-hosted software
Model licences
Noiz
Note: while some of this content was produced with the help of AI, the final editing has been done by a real human being.
