Skip to main content
Looking to try us out? Request a demo

No Generic Person

Culture, Power, and the Ethical Use of AI in Clinical Training

Paul Hoard, PhD··5 min read
No Generic Person

In developing Praxplay, we have been wrestling directly with both the exciting and the more concerning sides of AI. We have worked hard to refine our persona engine so it simulates realistic clients rather than eloquent sycophants, and along the way we have learned a great deal about engaging this technology ethically.

One of those lessons arrived early. When we first tested persona portraits internally, one image depicted a Native American client wearing a feathered headdress without any cultural or personal context to support that choice. The problem was not the regalia itself, but the model’s use of it as a default marker of Native identity. The image reduced distinct Native nations and individual identities to a familiar stereotype.  It was a vivid reminder that AI can turn cultural difference into caricature. If we aren’t careful, we end up training on representations of people that fit stereotypes instead of deepening our appreciation of diversity.

AI Is a Distorted Mirror

AI models are trained on the data available to us, which means the biases and assumptions that permeate our literature, writing, and public discourse are inevitably reflected in them. In that sense, they are mirrors of us, though uneven ones. They reflect some of us more than others, because the groups that have produced the most text and images carry the most weight. They are also refined using human preferences, which shapes them to please the user. One unintended consequence is that users’ own biases and prejudices, at whatever level of awareness, can be reinforced in what the system produces. Far from being fair, neutral, or objective, these models reflect our biases. Engaging with them ethically requires deliberate attention from users and developers alike.

There Is No Generic Person

Praxplay uses AI to generate and then simulate clients so helping professionals can practice and learn skills. But there is no “generic” person. No persona, simulated or human, exists outside a particular context. So when I type something like “client with anxious symptoms and history of complex trauma,” the AI fills in the blanks from somewhere. It relies on unexamined defaults, and those defaults reinforce existing power structures and dominant narratives. They reflect the cultures that have been most represented in shaping them. This is true not only of race and ethnicity but also of class, religion, disability, gender, and sexuality.

Three Ways It Shows Up

As we have built and tested Praxplay, this problem has shown up in three clear ways.

Standardizing. The default client is always drawn from somewhere, and usually that somewhere is the dominant culture and its assumptions. The default never announces itself as a culture. It just presents as “normal.”

Erasing. Sometimes culture is noted, but only on the surface. A persona may be assigned a cultural background while the simulation ignores the experiences, relationships, and context described in that individual’s story. The answer is not to prescribe how members of a group should behave, but to avoid treating one worldview as universal. 

Stereotyping. Other times, culture appears as caricature or costume. Rather than reflecting the complexity and diversity within a community, the persona reflects the dominant culture’s assumptions about it. The headdress is the obvious example.

The Harder Problem Is in the Text

The portraits were easier to catch. The same process happens at the level of text, though. How a persona relates to their family, talks about their parents, and understands faith, distress, and authority are all deeply shaped by cultural background. Stereotyping is just as present there, and often more insidious, because it is subtler and harder to notice.

For those of us whose experiences align with the dominant assumptions in a model’s responses, those assumptions can be especially difficult to recognize. A persona built on those defaults will often align with my own worldview. I don’t feel the offense or misalignment as quickly, because the problem stems from my own cultural view of others. 

As a company, this has required far more care as we test and refine our persona engine. We can’t simply outsource our responsibility to the AI models that power our software, as if it weren’t ours too.

Clients Who Push Back

There is a related risk: sycophancy, or excessive agreement with the user. A simulation that consistently validates the counselor can create a misleading picture of the therapeutic relationship. If a simulated client always agrees, never expresses uncertainty or discomfort, and never challenges a counselor’s assumptions, including when the counselor misreads them culturally, then practice can cultivate bad habits. We aim to create plausible, varied responses while recognizing that no simulated persona can represent a community or establish how a real client would respond.

What Ethical Use Asks of Us

For anyone using Praxplay or other AI simulation tools, a few considerations matter:

  • Reflect on your own social location and how it shapes your practice and your interpretation of clients.

  • Use practice to support supervision, never to replace it. Bringing a Praxplay transcript to supervision can help you explore both the interventions you practiced and what you may have missed about the persona, including sycophancy or relational ruptures.

  • Question the feedback. We have worked hard to make Praxplay’s evaluation robust, constructive, and helpful, but no system’s feedback should be accepted thoughtlessly.

  • Treat culture as relational, not a checklist. Where I teach, drawing on Stanley Sue and Nolan Zane’s work, we talk about cultivating cultural credibility, not just competence. Credibility is always navigated relationally. It can be gained or lost regardless of credentials, because it’s tied to how we actually show up, not just what we know.

Met in Relation

Every client, in whatever helping profession we serve, is a human being to be met in relation, not a list of attributes and diagnoses. Software like Praxplay can’t make the encounter ethical or ensure that we always act ethically. It can help us practice, but how we practice matters as much as whether we practice. We must always be thoughtful about who we are becoming when we do.