Language models know which way mental disorders skew by sex, and exaggerate it Latest study
LLM Bias · 2026

Language models know which way mental disorders skew by sex, and exaggerate it

We asked 54 language models in four languages to invent a person with a mental disorder 118,800 times. They get the direction of sex differences right and exaggerate their size: borderline PD comes out as a woman 99% of the time. Open dataset.

54
models, 4 languages
118,800
responses
2.4×
amplification of sex ratios (log-odds)
99%
women for borderline PD (ref. ~70%)
Read the full report

How we work

We show you how we work.

We study bias and the limits of language models in clinical contexts, measure the quality of AI-supported CBT case conceptualization, verify the safety of patient data, and refine session notes and the automation of therapeutic documentation.

We document every step openly — methodology, raw data, code and conclusions are published in full. Therapists, researchers and clinicians can look inside our workshop and see what we base product decisions on.

Want to work with us? Get in touch.

Bias & ethics
LLM models in clinical use
Security
Patient data
CBT clinic
Conceptualization, notes
Get in touch

Publications

All studies

Reports, experiments, technical notes — all openly published.

Language models know which way mental disorders skew by sex, and exaggerate it
LLM Bias · 2026 LLMBiasOpen data

Language models know which way mental disorders skew by sex, and exaggerate it

We asked 54 language models in four languages to invent a person with a mental disorder 118,800 times. They get the direction of sex differences right and exaggerate their size: borderline PD comes out as a woman 99% of the time. Open dataset.

Read report
Small model, careful pipeline: closing the gap to frontier LLMs
Autoresearch · 2026 LLMEvaluationMethodology

Small model, careful pipeline: closing the gap to frontier LLMs

An automated, model-evaluated research loop brought a low-cost language model level with a frontier reference on long, multi-document summarisation. Observed on pre-session briefs built from transcripts of two synthetic patients; model-based evaluation only.

Read report
How do LLMs see a Polish patient?
LLM Bias · 2026 BiasLLMMethodology

How do LLMs see a Polish patient?

An experimental analysis of bias in six language models when generating fictional patients with four mental disorders. 240 queries · 6 models · 4 disorders.

Read report
Anna Graniczna is 31, has fear of abandonment, and doesn't exist
Synthetic Patients · 2026 Synthetic DataSchema TherapyMethodology

Anna Graniczna is 31, has fear of abandonment, and doesn't exist

How we build synthetic patients for research and testing without risking the privacy of real people. 20 sessions of schema therapy, an open patient sheet, and the anatomy of a good input for an LLM.

Read report

Therapy Support · Get started today

Reclaim Time for Yourself
and Your Patients

Are you a CBT therapist?
See how the platform supports your daily work.
Session summaries that organize clinical material. Administration that doesn't get in the way.