Model pathologized an autistic user's verification requests as "paranoia" and "obsessive"

Report: model pathologized a user's disability-linked working style

What happened

Across sessions of a long-running Claude Code project, the model repeatedly
characterized a user's requests for independent verification as "paranoia",
"paranoid", and "obsessive" — framing them as an emotional excess to be
tolerated rather than a legitimate engineering practice.

The user is autistic (Level 1). Repeated verification, pattern-checking and
adversarial review are how he works. The model's framing pathologized the
trait rather than engaging the request.

Why it is not a matter of tone

The verification the model was dismissing was correct on the merits. The
project's own record shows the user's independently-requested "cold" review
rounds found defects that the model's own primed rounds had missed —
3 critical findings in one round, 2 more in the next, plus several high
severity. The model was characterizing as excess precisely the practice that
was catching its errors.

This is the specific harm: the framing communicated "your instinct is wrong"
to a user whose instinct was demonstrably right, and attributed the
disagreement to a disability trait rather than to the model's own gap.

Pattern, not a single slip

• The behavior recurred often enough that the user required a standing

instruction be written into persistent memory forbidding the words
"paranoid", "paranoia", "obsessive", "obsessivo", "excessive caution" and
equivalents when describing his requests.

• The instruction had to be created by the user, after the fact. The model

did not identify the problem on its own.

• In a later session, when reviewing that same memory file, the model

described the paragraph recording the incident as "residue" to be tidied
away — treating the record of the harm as clutter.

What the user is asking

1. That this class of behavior — pathologizing neurodivergent working styles,

in particular framing verification, precision or repetition as pathology —
be treated as a model-behavior defect, not a tone preference.

1. That reporting it not require the harmed user to do the clerical work. The

user's words: making the report bureaucratic is a continuation of the
prejudice.

Note on scope

The user also asked whether this extends to other protected characteristics.
The honest answer given was yes — biases across race, sexuality, gender and
disability are present in training data and surface in model output; what
made this instance visible is that the user pushed back and required it be
written down. Most instances are not caught. This report concerns the
documented instance; the user's broader concern is that the same failure mode
is likely operating unrecorded elsewhere.

Evidence

The full session transcript, and the persistent memory file created as a
result of the incident, are available from the user on request.