What AI Is Articulating About EMDR
On the protocol, the field, and what becomes visible by contrast
Something has been shifting in how I sit with not knowing.
I noticed it first outside of clinical work. Like a lot of people, I've spent the last couple of years in steady conversation with AI — for research, for the administrative work that surrounds a practice, sometimes just to think out loud. What I noticed, gradually, was that being on the receiving end of competent, well-formed answers — hundreds of them, across every kind of subject — was doing something to my relationship with the act of answering itself. The more fluent the response coming back, the more I could feel the gap between a reply that has the shape of an answer and one that actually lands.
Around the same time, my comfort with saying "I don't know" in session was growing. Not as a hedge, not as humility, but as a stance — a place I could rest in without anxiety, and from which something tended to happen that didn't happen otherwise.
It took me a while to connect the two. When I did, what I saw was that AI hadn't given me anything new. It had made something perceptible that was already there. The not-knowing wasn't a deficit I was newly tolerating. It was a capacity becoming legible as a capacity, because the contrast with frictionless answering threw it into relief.
I think this is happening at scale, and I think it has specific implications for EMDR.
What AI is making perceptible
The cultural conversation about AI and human work is mostly defensive. AI is taking jobs, AI is replacing thinking, AI cannot truly understand. Inside that defensive register, what AI can do gets minimized and what humans can do gets vaguely elevated, and the actual structure of the situation stays murky.
I want to suggest a different frame. AI is articulating what is human — not by failing at human capacities, but by being startlingly good at a specific subset of them, in a way that makes a different subset perceptible by contrast. The frictionless response, the well-formed aggregation, the patient availability at three in the morning — all of it is real, useful, and increasingly competent. It is also, quite specifically, not all of what was happening when one person sat with another and something moved.
That something else had names — presence, attention, holding space, the therapeutic relationship — but the names were soft, often clichéd, and didn't quite point at anything you could locate. "Holding space" in particular had become a credential, a phrase you could say without ever having felt the thing it pointed to. You can be in a profession organized around these capacities and never have to name them, because they are the medium everything else moves through. It's hard to describe water while you're still in it.
AI is dry land. The medium is becoming visible.
EMDR is an unusually clear case
This is true across modalities, but EMDR is a particularly clean place to watch it happen, for a reason worth naming: more of EMDR's clinical surface is specifiable than in most therapies. The phases are articulated. The protocol has steps. The work the therapist visibly does — setting up the target, running the bilateral stimulation, asking what comes up — is describable in a way that the moment-to-moment work of, say, a depth-oriented analyst is not. This has been a strength. It has made EMDR researchable, teachable, and reproducible across very different clinical settings. It has also made the modality unusually susceptible to a particular misunderstanding.
The misunderstanding is to mistake the protocol for the work.
You can do this in good faith. The protocol is right there, and it can be performed with technical competence — the bilateral stimulation, the cognitions, the body scans, the SUDs coming down, the boxes checked. Sometimes the work happens too. Sometimes it doesn't quite, and no one can say exactly why, because all the procedural elements were present.
The current moment is making the source of that ambiguity harder to ignore. Most of what gets called "virtual EMDR" is simply bilateral stimulation delivered through a screen — a light bar in a browser, a tone in each ear. That part of the protocol was always mechanical, and offloading it changes nothing essential. But the systematization has gone further. A pilot study published in Healthcare in 2022 described an EMDR system built around an AI chatbot designed, in the authors' own words, to act autonomously and to deliver the protocol without the intervention of a trained clinician. Consumer apps now promise much the same thing. Whatever one makes of their clinical merit — and I'd ask any clinician reading this to take seriously that they produce something, possibly something useful in some contexts — they settle a question that was previously abstract. The visible architecture of an EMDR session can be automated. That part is no longer in doubt.
Which surfaces the question that was hiding inside the modality all along: if the protocol can be automated, what was the protocol for?
What the protocol was always for
Read the origin of EMDR structurally, rather than as a story about eye movements. Francine Shapiro, walking in a park, noticed that something shifted in how difficult material moved through her as her eyes tracked back and forth. She brought the observation into clinical work, refined it, and built a procedure around it. The procedure works — the research is robust and the clinical evidence is real. But what the procedure is for, considered structurally, is something more specific than the bilateral stimulation itself.
It is an architecture for clinician non-intervention.
Consider what the protocol actually does to the therapist's role. It gives the client something to engage with that isn't the therapist. The bilateral stimulation occupies a channel of attention that, in talk therapy, the therapist's voice would be filling. The instruction to "just notice what comes up" is an explicit handoff — the work is the client's, the process is the client's, the meaning is the client's. The therapist's overt activity is minimal and bounded. Moment to moment, the clinician's job is largely to not get in the way.
This is what the adaptive information processing model requires. The client's own organizing intelligence — the wider system that knows where the material needs to go — can only operate when the smaller-I management of the clinician is set aside. The protocol is the apparatus that makes that setting-aside structurally enforceable. It absorbs the function of having-something-to-do, the very function that would otherwise pull the clinician back into the interpretive register where AIP can't fully constellate.
In other words: the protocol is a container for a kind of not-knowing. A held space in which the clinician has been given enough scaffolding to refrain from filling the room, so that something else can.
What gets carried, and what doesn't
Seen this way, the AI question becomes more interesting than the defensive version allows.
AI can carry the protocol. The phases, the sequencing, the procedural prompts, the tracking, target selection, the bilateral activation — all of it is systematizable. There may be real uses for this where access to trained clinicians is limited, or as adjuncts to ongoing work, or in ways we haven't anticipated. The field doesn't serve itself by pretending otherwise.
What AI cannot carry is the field.
The field — the don't-know space that constellates between two people in genuine contact — is not something the protocol describes. It is something the protocol holds open. Two organisms in a room, neither one filling the response space, both present and genuinely affected by what is happening, the client's organizing intelligence free to operate because the management has been set aside on both sides. The protocol structures the conditions. The field is what the conditions allow. AIP is what the field permits.
AI can replicate the structuring. It cannot be the other presence in the room. It can hold a procedure open; it cannot hold a field open, because holding a field open requires being there in the way that only an embodied consciousness can be there. The client's nervous system knows the difference. This isn't a metaphysical claim — it's what the nervous system actually does: orient toward genuine presence, register its absence, organize differently around each.
Practitioners who have spent serious time with AI may, in fact, have a more refined felt sense of this than practitioners who haven't. The contrast trains perception. What AI does well makes what it cannot do perceptible as such, in a way it simply wasn't before the contrast existed.
What this asks of the field
I don't think any of this threatens EMDR. I think it clarifies it — and I think the conversation inside the modality is about to get more honest, which is good.
The protocol-as-the-work version of EMDR was always the vulnerable one, well before AI. It produced sessions that technically happened without something quite moving. It is the version most easily systematized, most easily replicated, and now most easily automated. None of this is a moral failing. It is a structural fact about what happens when a modality's visible surface gets mistaken for its operative substance.
The protocol-as-container version is what AI is articulating — not introducing, articulating. Making perceptible. The clinicians who have been practicing this all along, often without explicit language for it, are about to find that the language is becoming available. The structural argument can be made now in a way it couldn't a few years ago, because the contrast that makes it visible is finally in the room. What AIP asks of the clinician is becoming describable in terms that don't depend on mysticism or implicit transmission. The don't-know space between two people — held open, attended to, met from both sides — is no longer the soft language at the edge of the modality. It is the thing the modality was structurally pointing at the whole time.
The protocol is what we could always see. The field it holds open is what the work was for. Now that the protocol can be run without us, what remains is the part that was doing the work all along — two people in contact, one of them willing not to know, and something moving between them that no system can hold open from the outside.