“Guarded,
poor alliance,
flat affect” —
the clinical
shorthand that
is now training
the machines to
misunderstand
autism

The prior you can’t see — what “good rapport” is actually measuring

Every clinical encounter runs on assumptions nobody wrote down. A new conceptual analysis in Frontiers in Psychiatry by Eik Niederlohmann, a physician psychotherapist working in German psychosomatic and public mental health settings, gives those assumptions a name: relational-proximity priors. These are the background expectations about how much closeness, warmth, eye contact, quick reciprocity and visible gratitude should be present before an interaction counts as empathic, collaborative or therapeutic.

The word prior is doing deliberate work. These expectations operate before explicit reasoning begins — they shape what a clinician notices, which patients feel easy to help, and what gets rewarded as progress. More closeness is presumed to mean better alliance. Warmth is presumed to be empathic. Accepting help in a visibly grateful way is presumed to signal readiness. And when a patient sits at a distance, answers briefly, asks for questions in writing, or shows little facial affect, those same priors quietly reclassify regulation as resistance.

The analysis is autism-first by its own account, because autism research holds the richest evidence on relational mismatch — masking, healthcare access barriers, interpersonal distance, and Damian Milton’s double empathy problem, which the Directory has covered through the study that measured mixed-neurotype conversation from the inside. Notably, the paper resists overextension: it uses double empathy as a stress-test logic rather than a universal theory of neurodivergent relationality, and cites Livingston, Hargitai and Shah’s 2025 Psychological Review caution against turning the concept into a total explanation of all social mismatch. That restraint deserves marking, because this corpus has leaned on double empathy too, and the honest position is that the construct explains breakdowns of mutual intelligibility — not everything.

A prior becomes a problem, on Niederlohmann’s account, when three things happen together: a behaviour is interpreted through majority-coded norms of closeness before alternatives are considered; plausible explanations like sensory processing load, interoceptive or alexithymic differences, autonomy protection or previous invalidation are never seriously weighed; and the resulting inference changes care in an escalatory or exclusionary direction. Distance read as poor alliance. Reduced eye contact read as disengagement. A request for written communication read as rejecting help. Meanwhile the patient who becomes socially smoother over time — through masking, appeasement, or exhaustion — is recorded as improving. The service is rewarding neurotypical performance and calling it recovery, a dynamic this corpus has documented in the engagement scales that measure the mask rather than the student.

From the therapy room to the file — how an interpretation becomes a fact

The paper’s sharpest move is tracing what happens after the encounter. A clinician’s in-the-moment reading — hedged, contextual, revisable — gets compressed into record language: guarded; limited insight; noncompliant. Niederlohmann calls these theory-laden compression devices, and the description is exact. Each phrase contains an entire interpretive framework while presenting itself as observation. Written without the context that produced it — who was in the room, how bright and loud it was, what medium the patient asked for, what adaptation was tried — the phrase hardens a local hypothesis into institutional truth. The next clinician inherits the label before they meet the person.

Institutions sediment the same priors at scale. Telephone-only triage for people who communicate better in writing. Intake interviews that reward rapid autobiographical coherence. Bright waiting rooms, group formats equated with progress, documentation forms with many ways to mark disengagement and none to record sensory overload or successful low-intensity contact. None of this is intentionally exclusionary. All of it is an institutionalised proximity preference, and the burden of adaptation falls on the person the service exists for — the same structural inversion the Directory has traced in why cognitive distance drives mass exclusion.

There is also an uncomfortable mirror held up to professionals. Drawing on Jon Frederickson’s interpersonal account of empathy and Kohut’s language of selfobject needs, the paper argues that helping roles are never psychologically empty: clinicians feel effective when patients respond with warmth and gratitude, and a patient who stays literal, distant and minimally expressive can frustrate that wish. The escalations that follow — pressing harder, offering more emotional intensity than the person can bear, moralising gratitude — are countertransference wearing the costume of clinical concern. Frederickson’s formulation, which the paper builds on, is that many clinicians mistake empathy for warmth, when the empathic question is how this patient will experience the intervention. It is a theory of mind problem pointed in the direction the field rarely points it — and the research on the imaginary conversations neurotypical people conduct when predicting how an exchange will go suggests the modelling error is not one-directional. Warmth can soothe one person and invade another. Working side-by-side, reducing eye contact, putting the agenda in writing — these are not anti-relational manoeuvres. They are relationally precise ones.

From the file to the model — scaling a misreading

Here is where the argument reaches the machines, and where it extends what this corpus began in the automation of the neurotypical norm at work. Mental health AI does not learn from behaviour in a vacuum. It learns from documentation, annotation, alliance ratings, and decisions about what counted as cooperation or risk. If those inputs are already overfitted to one narrow picture of good relationship, the model does not correct the bias. It stabilises it, legitimises it, and scales it.

One correction to make before going further, because I have made this error in print myself before and the paper flirts with it too. It is tempting to say these systems were built for neurotypical people. They were not. Nothing in a health service was ever built for a kind of person. The intake interview was built for throughput; the telephone triage for administrative reach; the alliance rating for auditable evidence of therapeutic progress; and all of it was calibrated, as every instrument since Quetelet has been calibrated, to a statistical middle that describes no actual human being. The person who passes through without friction is not a satisfied customer of a system designed around them. They are simply someone whose output happens to clear a bar normed to nobody — and whose own strain, being invisible to the instruments, never reaches a form. Saying the records were “built for neurotypicals” flatters the machinery with a beneficiary it never had, and converts a universal misfit into a minority grievance. The accurate version is worse and more useful: the record was built to be legible to institutions, and legibility has never had anything to do with accuracy about a person.

The paper’s scenarios are hypothetical and flagged as such, but each is recognisable. A triage tool treating sparse affective language or delayed response as elevated risk, with no way to know the person is autistic, overloaded, or simply safer in writing. A note-writing assistant reproducing the older record phrases in which flat affect was implicitly pathological. A conversational agent trained on generic empathy scripts, assuming faster reciprocity and more affiliative warmth are always improvements — an assumption anyone who has read the Directory’s coverage of ChatGPT use by autistic adults or the market for AI co-regulation will recognise as already in production. This is the epistemic injustice literature — Fricker’s term for being wronged as a knower — arriving in machine learning, and it converges with what this corpus has called shadow epistemology: the community knew masking and burnout were real long before the instruments could see them, and would have been coded “misleading” for saying so.

The proposed safeguards are unglamorous and correct: annotation schemas that separate observation from interpretation; lived-experience input into label design; governance that tests whether systems disproportionately penalise literal style, asynchronous communication and low-affect expression. Interestingly, the field has occasionally moved the other way — AI researchers adopting monotropism theory to describe attention allocation — which shows the traffic between autistic self-knowledge and machine architecture can run in the validating direction when someone bothers to steer it.

But the deepest version of the risk is older than any of these systems, and Norbert Wiener named it at the founding of the cybernetic age: a system comes to decide on the basis of decisions it has already made. That is the whole mechanism, stated seventy years early. A clinician’s hedged reading becomes a record; the record becomes a label; the label becomes training data; the model learns the label and returns it as a prediction; the prediction shapes the next clinician’s reading, which produces the next record. The loop feeds on its own outputs, and every turn of it is downstream of the interpretation that started it. Nothing in that circuit has to be malicious, or even wrong at any single step, for the whole to become self-confirming.

And notice what such a system actually needs in order to steer someone. Not accuracy. A description does not have to be complete, or fair, or true to the person, to govern what happens to them. It only has to be better at predicting them than they are at surprising it — and a patient whose file already says guarded is being predicted before they open their mouth. We are living in an epoch that runs on descriptions of us, and the clinical record is one of the oldest and least examined of them: a description built from your behaviour, which then steers your treatment, which then updates the description.

What this paper is, and what it isn’t

Provenance, plainly. This is a single-author Conceptual Analysis. It presents no original empirical data, no systematic review, and no validated construct — the author says all of this explicitly, and the audit tables are offered as heuristics to slow interpretation, not instruments. It appears in Frontiers in Psychiatry, a venue this corpus treats with more caution than the flagship journals. And the declaration section states that generative AI (ChatGPT) was used in manuscript preparation and language editing, with the author reviewing and taking responsibility for all content. In a paper about AI systems inheriting the assumptions of their source material, that declaration is at minimum an irony worth naming — and at best an argument for exactly the transparency it models, since most manuscripts now touched by these tools declare nothing.

Why cover it anyway? Because the argument stands on checkable primary material — Milton, Fricker, Crompton’s peer-to-peer transfer studies, Nicolaidis on healthcare communication — and because the construct organises a pattern this corpus has documented piecemeal for two years into a single testable claim. The paper is honest about needing empirical validation and participatory refinement. That is a research programme, not a finding, and it should be read as one.

Auditing the audit

The deepest point survives every caveat, and it is not really about AI at all. What research and clinical practice can see is constrained by what their measures were built to see — and alliance ratings, engagement scores and record shorthand are measures, whether or not anyone validates them. A care system that only asks how a patient can move toward the institution’s picture of good relationship has already cast its verdict; the case notes merely transcribe it, and the models merely learn it. Normalising the autistic brain is logically impossible. Normalising the record that describes it has been standard practice for decades. The machines are simply the first readers to take the record entirely at its word, because that’s what they’re built to do.

But there is something structurally unusual in what this paper proposes, and it deserves naming, because almost nothing in two centuries of this history resembles it. Every instrument in the apparatus — the curve, the checklist, the threshold, the engagement score — has pointed the same way: at individuals. The person is assessed, scored, filed, and read against the bar. What Niederlohmann proposes is an audit, and the audited party is the clinical encounter, the service design, the documentation standard and the model. The room sits the assessment. The demand structure answers the criteria.

Do not mistake that for rescue, and I offer none. An audit is also a market, and the certification economy that grows around one is sold by the same people who sold the training that changed nothing; no patient’s file closes because a service passed an inspection, and there will simply be two ledgers where there was one. But a direction has changed. The measuring eye, which has looked at nothing but persons since it first opened, has begun — in a Frontiers paper, tentatively, with heuristic tables and no validation — to look at the other end of the arrow. That is the more interesting thing here than any claim about AI. What happens when the eye keeps turning is a question nobody in this literature has yet asked.

Citations

Niederlohmann, E. (2026) — Neurodiversity as an epistemic stress test for psychotherapy and health care: a neuroaffirmative conceptual analysis of relational-proximity priors in clinical, institutional, and AI-supported models — Frontiers in Psychiatry

Milton, D. E. M. (2012) — On the ontological status of autism: the “double empathy problem” — Disability & Society

Livingston, L. A., Hargitai, L. D. & Shah, P. (2025) — The double empathy problem: a derivation chain analysis and cautionary note — Psychological Review