---
title: "Private AI chatbots: chat training, retention and local options | Ethical Stack"
url: https://ethicalstack.ai/guides/private-ai/
content_signal: search=yes, ai-input=yes, ai-train=no
---

# What does “private AI” mean for a chatbot?

A hosted chatbot has to process your prompt on a provider’s infrastructure. “Private” depends on what it retains, who can access it, whether it is used for training, and which model host receives it. A setting or product name alone does not answer all four questions.

## Are my chats used to train the model?

Check the product, account type, and setting. OpenAI says personal ChatGPT users can turn off “Improve the model for everyone”; chats remain in history but are not used for training. This is separate from business and API terms. [OpenAI’s Data Controls FAQ ↗](https://help.openai.com/en/articles/7730893-data-controls-faq).

Duck.ai says its chats are not used for training by DuckDuckGo or its model providers, with limited provider-specific retention exceptions. Proton says Lumo does not train on chats, with an exception for anonymized feedback submitted by the user. [Duck.ai policy ↗](https://duckduckgo.com/duckduckgo-help-pages/duckai/ai-chat-privacy) · [Lumo privacy policy ↗](https://proton.me/lumo/privacy-policy).

## How long are chats kept, and who can read them?

Look for chat history, temporary-chat behavior, deletion windows, abuse review, feedback controls, and file handling. A provider may process text briefly to generate a reply even when it does not retain chat history. When an assistant can route to several models, check the upstream provider too. The catalog profiles for [Duck.ai](https://ethicalstack.ai/products/duck-ai/), [Proton Lumo](https://ethicalstack.ai/products/proton-lumo/), [ChatGPT](https://ethicalstack.ai/products/chatgpt/), [Gemini](https://ethicalstack.ai/products/gemini/), [Brave Leo](https://ethicalstack.ai/products/brave-leo/), and [Microsoft Copilot](https://ethicalstack.ai/products/copilot/) state the specific product and route limits reviewed.

## Local options

When a model runs on your computer without calling a cloud service, the prompt can stay on that device. Ollama says it cannot see prompts run locally; its cloud-hosted models are processed on its servers. [Ollama privacy policy ↗](https://www.ollama.com/privacy).

Local setup needs downloaded model files and enough memory and compute. It changes where inference runs, not what was in the model’s training data. Check that the selected model is local rather than cloud-hosted, and review the app and model separately. See [local model options](https://ethicalstack.ai/alternatives/chatgpt/#which-alternatives-can-i-try) or [compare providers](https://ethicalstack.ai/compare/providers/).

## What should I avoid sharing?

Do not paste credentials, private keys, confidential work, or another person’s sensitive information unless your organization has approved that exact tool and route. For a public answer, remove names, account numbers, and identifying details first. For legal, medical, or financial decisions, verify with the original source or a qualified professional.