Last month, I clicked “I agree” on a software update for a tool I use in my clinical practice. I do this all the time. So do you. Most of us treat these agreements like airline safety briefings: just background noise before we get to work.
I don’t do that anymore. Here’s why.
Every AI tool coming into psychiatric practice, like the ambient scribe, documentation assistant, or the clinical decision support feature your EHR quietly added, comes with a terms of service agreement. We usually treat these documents as boilerplate, but they aren’t. They are the most honest risk disclosures in digital health, written by lawyers whose job is to map exactly where liability lands.
That liability lands on you.
If you read the disclaimers in any clinical AI product, you’ll notice a pattern. “Not intended to diagnose, treat, or prevent any disease.” “Not a substitute for professional clinical judgment.” “User assumes full responsibility for all clinical decisions.” These might sound like routine legal language, but each one is clear: if the tool makes a mistake, the clinician who used it is responsible for what happens.
This isn’t a mistake or something they missed. It’s built into the product on purpose. The vendor’s lawyers have already discussed what happens if the AI scribe leaves out a patient’s passive suicidal thoughts, if the decision support tool misses a lithium interaction, or if the documentation assistant creates a risk assessment that looks complete but synthesizes nothing. They decided the company shouldn’t be responsible for those risks, so they shifted that responsibility to you.
I spent two years on Texas Medical Board expert panels reviewing complex cases for standard of care. I can tell you what those reviews focus on: what the physician knew, what was documented, and what a reasonably prudent physician would have done. No one ever asks, “But the software’s terms of service said it wasn’t a medical device.” The disclaimer protects the vendor, not the clinician. In fact, it can make things worse for you, since you agreed in writing that the judgment was yours alone.
I know what real review involves because I’ve done it. Before this, I was a state-appointed independent medical examiner under Texas statute, using the AMA Guides to give independent opinions in legally defensible reports. Real review needs a framework, the records, enough time to compare them, and the authority to make your own decision. Just glancing at something and signing it isn’t enough. So when a tool gives you a finished note or a scored recommendation and asks you to take responsibility, you’re being given a reviewer’s liability without the conditions a real reviewer would have. That’s the main point: symbolic oversight just shifts liability to you with your signature.
Here’s an exercise I suggest to every psychiatrist using an AI tool, and it only takes ten minutes. Open the terms of service. Ignore the licensing parts. Look for the disclaimers and the limitation-of-liability section. Then flip every sentence around. “We are not responsible for clinical decisions” means you have to check everything the tool produces. “Output may contain errors” means mistakes are expected, and it’s your job to find them. “Not intended for use in emergencies” means if you use it during a crisis, you were warned.
What you end up with isn’t just a legal document. It’s really a clinical workflow guide, a list of the checks, documentation habits, and supervision steps the vendor expects you to follow. Most clinicians never pull out that list. The terms of service spelled out their job, but they just clicked past it.
For founders building these tools, it’s worth flipping the terms from your side too. Your disclaimer language isn’t just a legal formality, it’s a hidden barrier to adoption. The clinicians you want most as customers, the careful ones who will promote your product, are the ones who will eventually read what they agreed to. When they see that your terms put all the risk from your model’s mistakes onto their license, the trust your sales team built disappears. The companies that succeed in clinical markets will be the ones that share risk honestly: clear performance claims, defined intended use, and real indemnification. “Assume nothing, disclaim everything” worked when software just scheduled appointments. It doesn’t work when software writes a suicide risk assessment.
And for investors looking at these companies: add one question to your due diligence. Ask, “How do you divide liability between the platform and the prescribing clinician?” Founders who can answer this have thought about how their product might fail. Those who just point to the disclaimer haven’t.
The standard of care is changing for all of us. What a reasonably prudent psychiatrist does in 2026 increasingly includes how she oversees AI in her workflow. Vendors have already decided which parts are your responsibility. That map is in your inbox, attached to an agreement you probably clicked through.
Read it. Flip it. Then act on it.
ADVERTISEMENT
This essay is cited in the KevinMD record on artificial intelligence.
Erin J. Silvertooth is a dual board-certified psychiatrist and internist based in Austin, Texas, with more than 20 years of clinical practice. She directs TMS and Spravato/esketamine programs, provides neuropsychiatric consultation in brain injury care, and has consulted in sports psychiatry for Olympic, Paralympic, and collegiate athletes for over a decade.
She advises early-stage health technology companies and investors on AI safety and physician liability in digital health. She has served on Texas Medical Board panels evaluating standard of care and as a state-appointed independent medical examiner. She is Harvard-trained in health communication.
Her writing examines AI in psychiatric workflows, physician liability in digital health, interventional psychiatry, and what clinical practice looks like from inside the room. She shares updates on LinkedIn and Substack.
