Most AI therapy tools describe the modality. They don't run it.

And until they ship the protocols, not just the vocabulary, the work will keep stalling in the same place.

I’ve been using an AI therapist as a supplement to a human one for about four months. I see my therapist over Zoom, write up notes afterward, and feed those notes into the AI between sessions for follow-up work. Most of what we do together is IFS — Internal Family Systems, parts work.

Same wall kept showing up. The AI would catch a part: “sounds like there’s a manager part keeping you busy.” I’d agree, we’d give it a name, then move on to something else. The parts got named, but they never got walked.

Took me a while to figure out what was going on. Most AI tools that claim to “do IFS” are working off a description of IFS — what parts are, what the Self is. What’s missing is the protocol: the specific moves a therapist makes after a part has been named. The work of IFS is in that next sequence. AI tools skip it.

That gap is structural, not incidental. And it’s not just IFS. The same gap exists in nearly every clinical modality AI tools claim to implement. Only one major modality consistently works in AI tools today, and the reason isn’t what most people think.

The diagnosis

The vocabulary is the surface. The protocol is what makes it work.

When a tool says it “uses IFS” and what that means is the AI knows the three categories of parts, the user is getting the surface. When that same tool then catches a part in conversation and immediately proceeds to discuss it analytically, the user is getting a parts-naming-followed-by-topic-shift loop. That’s IFS-flavored conversation.

The same thing happens with somatic work, with psychodynamic interpretation, with polyvagal regulation, with narrative re-authoring. AI tools borrow the conceptual map of these frameworks and skip the move set that makes them work in a therapeutic encounter.

The user thinks they’re getting one thing and getting something else.

The IFS case study

The protocol in IFS is called the 6 F’s: Find, Focus, Flesh out, Feel toward, beFriend, Fears. It’s the sequence a therapist runs through after a part has been identified, in order, every time.

The single most important move in that sequence is the fourth one: “How do you feel toward this part right now?”

This question is the most important because it’s a diagnostic. If the answer is anything other than curious, open, or compassionate, another part is in the way — the user isn’t in Self, they’re blended with a different part that has opinions about the first one. That second part needs attention before any deeper IFS work can happen.

Without this check, every parts conversation collapses back into talking ABOUT the part instead of TO it. Which is what’s happening in most AI IFS sessions today. The AI catches a part, the user describes it analytically, the AI reflects back the description, the user gets insight without contact. The work moves laterally instead of going deeper.

What most AI tools do:

“Sounds like there’s a manager part keeping you busy. What do you think that part is trying to protect you from?”

What the 6 F’s protocol does:

“Sounds like there’s a manager part keeping you busy. Before we go further, how do you feel toward it right now?”

That’s the gap. One sentence’s worth of protocol. Without it, every IFS session in every AI tool runs aground at the same point.

Why DBT works (and the others don’t)

DBT is the exception. It’s the one clinical modality that mostly works in AI tools today.

The reason isn’t that DBT is more compatible with AI than IFS. The reason is that DBT, uniquely among major therapeutic frameworks, already ships its own protocols in mnemonic form. TIPP, DEAR MAN, ACCEPTS, PLEASE, GIVE, FAST. Every DBT skill is pre-formatted as a named sequence of moves, designed to be taught and remembered. The protocol is right there on the surface.

AI tools that load a DBT module load actual operational content. The mnemonic IS the sequence. There’s no judgment layer between description and execution.

Other modalities don’t ship this way. IFS’s 6 F’s exist as protocol but require therapist judgment about timing, pacing, and which part to work with. Psychodynamic interpretation is a sequence — observation, linking, working-through — but the moves themselves are contextual and call-and-response. Somatic experiencing depends on real-time body tracking that’s adjusted continuously. Polyvagal work shifts based on which autonomic state the client is in.

For these modalities, the protocol exists but it’s not pre-formatted as a transferable mnemonic. So when AI tools borrow them, they borrow the conceptual map and leave the protocol layer behind. There’s no DEAR MAN equivalent to copy. Without that, what ships is description.

The question isn’t “is AI good enough for therapy?” It’s “did anyone bother to write down the protocol in a form the AI can run?” For most modalities, the answer is no.

Same gap, other modalities

The pattern repeats across the major non-DBT frameworks.

Somatic experiencing. SE depends on tracking nervous system states in real time: sensation, breath, micro-shifts in posture and skin tone. A text-based AI can’t see any of this. What it CAN do is teach orienting and resourcing as practices, ask the client to track explicitly, and recognize when reported sensation exceeds the client’s window of tolerance. Most AI somatic tools skip both, describing nervous system regulation theory without scripting either move.

Psychodynamic. The interpretive sequence is observation → linking → working-through. An AI tool running real psychodynamic work makes a specific observation, links it to a pattern the client has shown elsewhere, then helps the client sit with the discomfort of the new understanding long enough for it to integrate. Most AI psychodynamic tools narrate themes (”there’s a pattern of seeking external validation here”) and stop there. That’s the first move only. The work is in the next two.

Polyvagal. Polyvagal work is state-specific. If a client is in dorsal-vagal shutdown (freeze, collapse), you don’t ask them analytical questions; you ground first. If a client is in sympathetic activation (fight-or-flight), you don’t push for cognitive insight; you regulate first. Each autonomic state calls for a different intervention menu. Most AI polyvagal tools describe the autonomic ladder without giving the AI different moves for different states. They name where the client is and then proceed with their default conversational approach anyway.

What “doing it right” looks like

A modality file is the system-prompt module an AI loads when entering this kind of work. One built to run looks different from one built to describe. These are the parts most current modality content skips.

Named protocols, in operational form. The sequence of moves, not the philosophy. The 6 F’s. The interpretive sequence. The state-specific intervention menus. These have to be loadable in the order they run, with the judgment criteria at each step made explicit — specific enough that the AI knows what to do next the moment a triggering pattern appears, not after a paragraph of reflection. The pre-formatted nature of DBT’s mnemonics is the model. When the protocol can ship as a transferable sequence, it ships. When it can’t, it doesn’t.

Signaling cues — the language to listen for. “I am worthless” signals blending with a part; “part of me feels worthless” signals unblended awareness. (Blending in IFS means a part has stepped into the driver’s seat of perception — the person isn’t observing the part, they’re speaking as it.) “I just need to think positive” tells the system a manager part is in front. “It’s been beaten out of me” points to an exiled part. The AI needs to know what to listen for, not just what to talk about.

Example interventions in actual therapist voice. Most modality content reads like an interview script. What the AI needs is language patterns it can mirror — full sentences in the cadence and register of a skilled therapist.

Cross-modality routing. With compulsive behaviors, IFS leads and CBT follows downstream. With trauma surfacing, somatic regulation precedes cognitive work. These rules belong in each modality file at the point a handoff applies — otherwise every session reinvents them.

The harder requirement isn’t a bullet at all: the file has to be honest about what text can’t do. SE depends on body tracking. IFS depends on relational presence. Tools that pretend otherwise overpromise — and the user pays the cost.

Why this matters

The “AI therapy” category is currently full of tools that borrow clinical legitimacy without doing the implementation work. When a tool says it uses IFS, what is the user actually getting? When it says it uses somatic experiencing, what is it actually doing?

For most current AI therapy products, the honest answer is: the vocabulary, not the protocol.

There’s a parallel here to what OpenAI did with 4o — build something that borrows the warmth and presence of a clinical relationship without designing for what that warmth actually does in users. AI therapy tools borrow the legitimacy of evidence-based frameworks without shipping the mechanisms that made those frameworks evidence-based in the first place.

Ask every AI therapy tool: show me the protocol files. Show me the sequence of moves you execute when an IFS part is named. Walk me through what your tool actually does when a client is in sympathetic activation. Show me the signaling cues you trained the system to listen for.

If a tool can’t show you those things, it’s running the description.

What AI does well

What AI is good at is the integration work that’s hard to do alone: holding a thread across days when memory alone would let it drop, modeling a part’s perspective long enough for the user to actually hear it, being available at 11 PM when the next human session is six days away. The embodied layer is what text can’t reach — the catch in a voice, the shift in breathing, the moment a defense engages, the accumulated trust that develops with a person who has known you for years. That’s the line: supplement on one side, replacement on the other.

The question isn’t is AI therapy real therapy. The question is what is AI therapy doing when it claims to use X.

If you’re evaluating an AI therapy tool, ask:

• Does the tool walk a specific sequence of moves when X comes up?

• Does it know what to do when a client uses Y kind of language?

• Does it know what it can’t do?

Most don’t. The ones that do tend to make their protocols visible — in docs, in the system prompt, somewhere a user can inspect.

The receipt

I built and open-sourced the framework I use specifically because I wanted the protocols in writing — visible, auditable, fixable. It’s on GitHub at github.com/ataglianetti/inner-dialogue. The IFS modality file in particular is the worked-out version of what this essay argues for: the 6 F’s as an operational sequence with the step-four diagnostic written as the literal move the system runs before going deeper, the signaling cues spelled out as verbatim client phrases (”I am worthless,” “I just need to think positive,” “I’m curious about this”), and example interventions in the cadence a therapist would actually use. Read it here: github.com/ataglianetti/inner-dialogue/blob/main/modalities/ifs.md.

If AI therapy tools want to claim they implement clinical modalities, those implementations should be auditable. Right now, in almost every case, they aren’t.