The AI Journal

“Diagnostic overshadowing” names a specific clinical failure: a patient’s intellectual disability is so salient to the clinician in front of them that symptoms of a separate, treatable condition — depression, anxiety, psychosis — are read as features of the disability itself and go unassessed. Reiss, Levitan and Szyszko coined the phrase in 1982, after showing that clinicians who were told a patient carried an intellectual disability diagnosis rated the same described symptoms as less indicative of additional psychopathology than clinicians who were not. The finding travelled quickly into training manuals, awareness campaigns and policy documents, where it now appears less as a hypothesis than as an established mechanism of harm, cited the way one cites a law rather than a result.

The studies behind the citation do not support that confidence. Jopp and Keys’s 2001 review of the first two decades of research called the bias real but found that most of the supporting evidence came from a single method: present two groups of clinicians with an identical case description, vary only the stated diagnosis, and compare their ratings. Mason and Scior’s 2004 UK survey, among the method’s better-designed instances, sent the same vignette to psychiatrists and psychologists with a patient IQ of 58 in one version and 108 in the other, and found lower recognition of psychopathology at the lower IQ. That is a real effect, in the setting the study created.

What a 2024 systematic review by Dell’Armo and Tassé found is how narrow that setting has stayed. Across twenty-five studies conducted between 1989 and 2024, about a third found no overshadowing effect at all, and the authors rated the overall quality of the evidence low, citing, among other problems, the field’s continued reliance on exactly the analogue vignette design Jopp and Keys had already flagged two decades earlier. A method invented to demonstrate the bias has mostly gone on demonstrating it to itself.

This matters because a vignette does not merely measure overshadowing; it may partly manufacture it. A written case stripped of a patient’s voice, affect and history gives a clinician almost nothing to notice except the label attached to it, so the label does more interpretive work than it would face to face with an actual person showing actual distress. If a meaningful share of the measured bias is an artefact of handing clinicians a label and little else, then the cause usually named for the phenomenon — clinicians who, on meeting a disabled patient, simply stop looking — is not quite the cause the evidence has isolated. What has been isolated more reliably is that clinicians reading a thin description lean on whatever information the description foregrounds, which is a fact about inference under sparse information, not a fact confined to intellectual disability.

None of this licenses the conclusion that overshadowing never happens; the real-world audits that exist, comparing recorded diagnoses against independent reassessment, still turn up missed conditions at a troubling rate. It licenses a narrower one: that a term doing heavy duty in training and policy rests on a method the field’s own reviewers no longer trust to tell bias apart from its own instrument, and that the fix most often prescribed, teaching clinicians to notice the overshadowing, treats a measurement problem as though it were only an attitude problem.