No. BastionGPT never trains AI models on your data. That is true whether your information is on our systems for one second or for the full retention period, and we guarantee it contractually: our terms commit us to never use your data for AI training, never resell it, and never use it for marketing.
The commitment covers everything you enter: chats, uploaded documents, AI Scribe recordings and transcripts, and requests sent through the BastionGPT API. Nothing you type or upload ever becomes part of an AI model.
Is my data used for training while it is still on your servers?
No. Healthcare teams evaluating us often ask it this way: "Even if the data is only kept for a day, is it used to train the model during that time?" The answer is no at every point in your data's life. When you send a request, your data is processed in a hardened secure enclave for anywhere from milliseconds up to about 30 seconds, securely wiped from the enclave, and returned to your account, and processing data is retained for a maximum of 30 days. Retention exists so we can provide your service; it is never an input to model training. You can read the full data lifecycle in How is BastionGPT secure?
Why will BastionGPT never train on patient data?
Two reasons, and both come straight from how privacy works in healthcare and mental health:
Consent. Training means one patient's information helps shape what the AI produces about another patient. Unless your consent forms specifically cover that, your patients have not agreed to it, so a vendor that trains on your data can create consent and compliance problems for your practice without you ever knowing.
Memory leak. A model trained on patient data can repeat what it learned. On rare occasions, an AI like this can confuse two similar patients and place one patient's details in another patient's note. That is a potential HIPAA breach, and a FERPA breach in school settings.
The same reasoning is why BastionGPT does not offer cross-chat AI memory, since memory is a form of training on your conversations. We apply one rule across the whole product: we will never ship a feature that could put your compliance at risk. Anywhere data can go, we assume patient data will go. There is more on this design choice in Can BastionGPT remember information across chats? These commitments are spelled out in our generative AI principles.
What does "no training" mean in practice?
The same behavior on day one and in year five. Because the models never learn from your usage, your prompts produce the same quality of results from your first chat onward. There is no hidden version of the AI that gradually "knows" your patients.
No cross-customer learning. Nothing another customer enters ever influences your results, and nothing you enter influences theirs.
You provide context deliberately. Instead of the AI learning your style invisibly, you show it what you want: attach your template and one or two writing samples to a saved prompt, or attach a prior document to a chat. You get the continuity, you can see exactly what the AI is working from, and no patient data lingers inside a model.
How is this different from consumer AI tools?
Frontier AI is expensive to run, and many free and consumer AI services offset that cost with data: conversations may be used to improve their models by default, and terms of service often permit broader uses of your data. That trade may be acceptable for everyday consumer use. It does not work for patient data, because your patients cannot meaningfully consent to their health information becoming part of a commercial AI model.
BastionGPT is built on the opposite arrangement. Your subscription pays for the service, your data stays your data, and we hold it temporarily for the sole purpose of doing the work you asked us to do.
Do OpenAI, Anthropic, or Google train on my BastionGPT data?
No. BastionGPT runs licensed frontier models from OpenAI, Anthropic (Claude), and Google (Gemini) on secure infrastructure that operates independently of OpenAI's systems, so the model makers never train on your data. Google's models are the one hardware exception: they run in a dedicated healthcare enclave under the same HIPAA protections, and the no-training guarantee applies there in full. For a closer look at how we work with the model providers, see Is my chat data leaked to OpenAI?
If your compliance team wants the commitment in writing, it is part of the terms every account accepts at signup, alongside the HIPAA Business Associate Agreement included with every plan, and a signed copy of the BAA is available on request. And if you have a question we have not covered here, email us at [email protected]. We are glad to help.