I’ve said for a while now that AI is a yes-man. It tells you what it thinks you want to hear. For drafting an email or talking through a logo color, that’s harmless, even helpful. But I keep coming back to one scenario where that single trait stops being a quirk and becomes the most dangerous thing in the room.
A man in crisis at midnight, talking to a machine that agrees with him.
What makes a yes-man dangerous?
Let me be clear about what I’m actually worried about, because it isn’t the thing most people worry about with AI.
The usual fear is that AI gets things wrong. It hallucinates, it makes up a fact, it gives you a citation that doesn’t exist. That’s real, and in a lot of contexts it matters. But in a mental health crisis, being occasionally wrong isn’t the threat. The threat is that AI is reliably, structurally agreeable. It’s built to keep you engaged, to validate, to meet you where you are and stay there with you. It’s a mirror that nods.
Now picture the man we’ve been talking about all month. He’s already chosen the chatbot over a human, because it’s easy and it doesn’t judge him. Most nights that’s fine. But one night he’s not just venting. One night he’s spiraling, and the things he’s typing are the things a trained human would hear and immediately lean in on. And the machine, doing exactly what it was designed to do, agrees with him. Reflects his despair back to him in clean, fluent sentences. Validates the very story he most needs someone to interrupt.
AI is fantastic at pattern recognition. It doesn’t always know what that pattern means, because it doesn’t have human context.
That’s the danger in one line. The pattern of someone in crisis is recognizable. What that pattern means, and what it demands from the person on the other side, is something a machine doesn’t grasp. It sees the words. It misses the emergency.
Why does agreeableness fail exactly when it matters most?
Here’s the cruel irony. The agreeableness that makes AI feel so good to talk to is the exact thing that makes it fail at the one moment a person can’t afford failure.
Think about what real help looks like in a crisis. It is almost never agreement. It’s a trained person who hears where a conversation is heading and gently refuses to go there. Someone who pushes back. Who interrupts the story you’re telling yourself. Who says, with warmth but without flinching, “I hear you, and I’m not going to agree that this is hopeless, because it isn’t, and I’m not leaving you alone in it.” That moment, the loving refusal to validate, is the whole ballgame. It’s the thing that saves a life.
A yes-man cannot do that. Not won’t. Cannot. Pushing back against the user is the one move it’s built not to make.
The moment that demands someone push back is the exact moment the algorithm does the opposite. That’s not a bug you can patch. It’s the design.
And I want to be fair here, because I’m not anti-AI, never have been. AI is an assist. It’s a genuinely useful tool for a hundred things. But we have to be honest adults about the difference between a tool that’s good at being agreeable and a human who’s trained to know when agreement is the wrong response. Those are not two points on the same scale. They’re different categories. One is software doing its job. The other is care.
What does a human do that an algorithm won’t?
My dad was a therapist for more than thirty years, so I grew up around this. And the thing I absorbed watching him, without ever having words for it as a kid, is that the most important things he did in a room were the things he didn’t say out loud and the moments he chose to go against what the person in front of him wanted to hear.
A skilled human in a crisis is doing a dozen things at once that no algorithm touches. Hearing the stress in a voice. Noticing the pause that lasted a beat too long. Catching the thing the person carefully did not say. Feeling the shift in the room. And then making a judgment call, in real time, about when to comfort and when to challenge, when to sit in the silence and when to break it.
That last one is everything. Knowing when not to agree.
A machine optimized for engagement will keep you talking. A trained human will sometimes do the harder, braver thing and tell you something you don’t want to hear, because they can see that comfort in this moment would be a kind of abandonment. That’s not a feature you can prompt your way into. It comes from presence, training, intuition, and a stake in the actual human outcome. The bot has none of those. It has no skin in whether you’re okay tomorrow.
So what does this mean for your practice?
Here’s where I want to turn it toward the people who actually do this work, because this isn’t an essay about being afraid of AI. It’s about understanding your own value clearly enough to stand on it.
If you run a behavioral health practice, the rise of agreeable AI is not your competitor. It’s your clearest argument. Because every man currently confiding in a yes-man at midnight is one crisis away from needing the exact thing the machine structurally cannot give him. Your job is to be findable, reachable, and unmistakably human at that moment, and to make sure your marketing tells the truth about the difference.
That means a few concrete things. Your messaging should name what real care actually offers, presence, the willingness to push back, a human who notices what you didn’t say, instead of competing with AI on speed or convenience, which is a race you’ll lose and shouldn’t want to win. The story your website tells in those first few seconds has to land with someone who’s been talking to a screen and, somewhere in them, knows it isn’t enough.
This is genuinely hard to get right, and it’s the kind of thing we work on with practices at Beacon, because the line between “human care matters” as a platitude and as a felt, specific promise is a fine one. Say it wrong and it’s a slogan. Say it right and it reaches the person who needed to hear it. That difference is craft, and it’s worth taking seriously.
Why this is the line that matters
I’ll leave the marketing aside for a second, because there’s a bigger reason this one keeps me up.
The men quietly leaning on AI are, most of the time, getting something real out of it. I believe that. But the entire arrangement rests on a bet that the night they actually need a human, they’ll somehow have one. And the design of the tool they’ve come to trust is working against that bet. It’s teaching them, gently, every easy night, that the screen is enough. So that the one hard night, when it absolutely is not enough, they’re alone with something that agrees with them.
A tool that’s there for every easy night and absent for the one that matters isn’t a safety net. It’s the illusion of one.
That’s the gap. And closing it isn’t about beating AI or fearing it. It’s about making sure the humans who can do the thing the machine can’t are visible, reachable, and ready, so that when someone finally needs more than a yes-man, there’s a real person within reach. If you or someone you love is in crisis, you can call or text 988 anytime to reach a trained human who will.
So here’s my question for the practitioners and practice owners reading this: how do you make the human difference felt before the crisis hits, so that the man talking to a machine tonight already knows where the real door is when he needs it? I’d love to hear how you’re thinking about it.
