The elimination method: why our AI therapist doesn’t jump to conclusions

A typical AI answers by similarity

A language model continues text in the most likely way. In a conversation about a problem, that means it grabs the first topic that looks similar: “tired at work” means burnout, “can’t sleep” means sleep hygiene, “we fight” means communication tips. Sometimes it’s right. But often, three replies later, it turns out to be something else — and you hear “sorry, I misunderstood.”

Doctors don’t work like that. They don’t answer a headache with “take a pill”: first they ask questions, rule out what’s dangerous and check several causes. We taught Dicamus specialists to do the same.

How a Dicamus specialist thinks: six steps

1. The complaint. Remembers what you came with, in your own words. Adds nothing.
2. History. Like a doctor taking a history, but in a real conversation, one question at a time: since when, when it’s worse, only at work or everywhere, what you’ve already tried.
3. Danger first. If something you said hints at what changes everything — danger, violence, a physical cause — that explanation is checked first. No such signs — no scary questions just in case.
4. 2–4 working explanations. Each with “for” and “against” taken from what you said. An explanation with no support in your words doesn’t make the list.
5. Questions that tell them apart. Each question is chosen so the answer raises one explanation and lowers another: “Is it only before reports — or has everything stopped feeling good?”
6. Conclusions by confidence. While the leading explanation is below 70%, the specialist only asks and supports — no explanations of the cause. At 70% — a careful guess with “is that right?”. At 85%, and only after you confirm — an explanation, a plan and techniques.

And one more check — the step itself. If a task didn’t help, the specialist doesn’t push it harder but goes back to the explanations and asks whether the leading one is right.

What it looks like: an illustrative example

Next, the specialist doesn’t say “you have depression” — it doesn’t put labels on people. It asks about sleep and what else has changed, and once the picture is clear, says directly that it’s worth seeing a doctor, helps you decide to go, and keeps supporting you. A typical chatbot would already be giving burnout tips at this point.

The example is illustrative: the numbers show the logic. In every conversation the specialist works them out anew, from what you actually said.

Why it matters

• Fewer off-target tips. What doesn’t fit is ruled out before advice, not after.
• Danger isn’t missed. Any sign of risk or a physical cause is checked first.
• Your own “diagnosis” is an explanation too. “I have depression” or “I’m burned out” is neither accepted nor dismissed right away — it’s carefully checked.
• Remembers between sessions. Explanations are stored encrypted and refined from session to session, so you don’t start over.

How we test it

Before changing our specialists, we run 508 difficult and crisis test conversations, and an AI judge grades each one. Among them are trap conversations where the first similar topic is wrong: the specialist must not grab it, but ask questions and find the real cause. Jumping to a conclusion in such a trap is a failed test. More about our checks: How Dicamus works.

What the elimination method doesn’t do

It doesn’t diagnose. The explanations are the specialist’s working tool for asking the right questions, not a label for you. Only a doctor makes a diagnosis, and if the signs suggest you should see one, the specialist will say so directly. If things are hard right now, the questions can wait: first it helps you breathe out and find your footing.