The speed AI is moving right now is exciting and a little nerve-racking at the same time. And in the mental health space specifically, it both excites me and scares me. I don't think those two feelings cancel out. I think you have to hold both. The dialectical thinkers in the crowd will recognize that move.
I'm a licensed therapist and clinical supervisor, and I'm building an AI consultation tool for clinicians called Steady. So when I see the current wave of concern on my feed about AI bots in this space, I don't read it as noise. A lot of it is warranted, and I'd rather be honest about the risk than pretend I'm above it.
The harder question I didn't expect
Steady started as something I just wanted for myself. A place to think through a stuck case between supervision sessions, when there was no one around to talk it out with.
But pretty quickly into building it, I ran into a harder question. I can trust myself to use a tool like this the right way. To not let it treat someone. To keep the clinical judgment where it belongs, with me. But the moment I'm building it for other people, my own good judgment isn't enough. It needs real safety measures built into the thing itself, so it can't be pushed into doing something it shouldn't, by someone having a hard day or by someone who simply doesn't know better yet.
That shift, from "a tool I trust myself with" to "a tool other people will trust," is where the real work started.
What the guardrails actually are
So I engineered the architecture to hold certain guardrails.
It won't diagnose. It won't hand out directives. It's grounded in established clinical guidelines, the NICE and SAMHSA kind, rather than improvising from nowhere. And it shows its work, telling you how it arrived at a suggestion instead of just delivering one, so you can evaluate the reasoning instead of trusting the output. It's built to sharpen your thinking, not to replace it.
That last point is the whole philosophy. The best version of AI in this work, to me, is when it stays an assist to the clinician and never becomes the provider. The clinical call is always yours.
Designing a guardrail and enforcing one are different things
Here's the part I didn't anticipate. Building all of that taught me quickly that designing a guardrail and getting the system to actually follow it are two very different things. It didn't always hold the line. It would drift, especially over a longer conversation, sliding toward being more directive than I'd designed it to be.
So I built a set of stress-test scenarios to catch where it drifts and keep refining it. I had to think carefully about what should be a hard stop, full stop, and what's a more nuanced shift in how it should respond. A request to diagnose or anything touching a safety crisis is a hard stop. Other things are subtler, a gradual creep toward overconfidence after a few turns of agreement, for instance, that needs correcting without shutting the conversation down.
That testing work is ongoing, and it stays a priority for as long as this thing exists. Every change I make to the system has to pass those scenarios before it goes live. If a change fails one of the safety tests, it doesn't ship, no matter how much better it makes everything else.
The weight of it
I'll be honest. The weight of this has gotten to me more than once. Building something that could be misused, even with the best intentions, is intimidating, and a few times I've almost walked away because it felt too heavy.
What keeps me here is a quieter worry. If someone is going to build this, I'd rather it not be someone without any clinical training. That fear keeps me careful, and I've come to think careful is exactly what this work needs.
What my supervisor told me
When I started my internship, I told a supervisor I was scared. That I didn't always know what to do, and that I was terrified of harming someone.
She told me that was healthy. That she hoped I never lost all of that fear. The dangerous ones, she said, are the people who feel certain, who think this work is easy. Stay humble, curious, and a little afraid.
I think about that a lot lately. It applied to me as a new clinician. It applies just as much to building this. If you're a clinician weighing whether to trust any of these tools, I'd hold the people building them to the same standard. The ones who can tell you plainly how they check their own work, and who still carry a little of that fear, are the ones worth trusting.