Bias in Virtual Therapy Clients: Who Gets Left Out of AI Training?

Imagine a training platform in which almost every virtual client has a similar age, educational background, family structure or way of speaking. Trainees might complete many sessions while encountering a surprisingly narrow range of experiences.

Diversity is not just about appearance. It shapes expectations about how distress is expressed, what counts as “cooperative” behaviour and whose experiences feel familiar.

What did a 2025 study find?

In a study published in Advances in Simulation, researchers prompted two language models to generate 500 fictional UK patient profiles. The resulting cohorts differed significantly from demographic reference data in age, sex and ethnic representation.

The research examined synthetic patients for medical education, not a particular psychotherapy training platform. It does not prove that every AI tool has the same biases. It does show why automatically generated characters should not be assumed to represent the broader population.

Three kinds of problem

Underrepresentation: Some age groups, social settings or cultural experiences rarely appear in training scenarios.

Stereotyped narratives: A character is assigned predictable difficulties simply because they belong to a particular group.

Biased assessment: Automated feedback favours certain ways of speaking without adequately accounting for linguistic or cultural context.

What does a practical audit involve?

Start by listing the available training characters. Look for repeated demographic features, missing life experiences and cases where identities are automatically linked to dysfunctional behaviour.

Next, test the same learning objective across different profiles. Challenge should arise from scenario design and the intended skill, not assumptions about who a person is.

Invite qualitative review from educators with appropriate expertise and, where feasible, people representing different perspectives. Balanced demographic counts alone do not guarantee accurate, non-stereotyped stories.

Diversity without treating people as categories

The objective is not to create a catalogue of “difficult populations” for trainees to practise on. It is to build humanly plausible characters with personal histories, needs and agency—without reducing identity to a symptom.

Virtual clients are educational constructions, not representatives of every person in a demographic group. That limitation should be explicit.

From theory to practice: When selecting simulated cases, ask not only, “How many scenarios are available?” but also, “Whose experiences are represented, and how?”