Why Won’t Tech Companies Consult Therapists or Sex Workers?
I sat across the room from a tech bro for a year, trying to teach him empathy, then I discontinued the therapy relationship because he still showed zero insight into another’s experience, or how to hold empathy for his closest relationship. These same men are founding and developing companion bots.
I can’t tell you how many times I’ve asked a tech dev or founder, did you consult any social workers, therapists or sex workers in your product development? Who helped you create training data about human relating, and what were their identities? The answer is always no, and mostly some white men built it all.
What we know: they paid a lawyer to review the terms of service, they paid an accountant to structure their cap table. These were a given in the early development process.
But what about clinical expertise? Care expertise?
Why didn't they consult a therapist before their product started holding someone's loneliest 2 a.m. conversation?
For most companion AI companies, clinical expertise never made it onto the list of necessary consultants. It wasn't debated and rejected. It simply wasn't the kind of expense anyone in the room thought to name, because it's not the kind of work our economy has ever priced as essential *or seen as valuable by the humans at the table.
Legal and financial risk are old, familiar categories. Every founder already knows, before they write a line of code, that a lawyer needs to look at the incorporation documents, the terms of service, the privacy policy. Every founder knows an accountant needs to be involved before the first funding round closes. These aren't judgment calls. They're default infrastructure, built into how startups get built at all, because the cost of skipping them is well understood and comes due predictably: an audit, a lawsuit, a cap table dispute.
Clinical risk doesn't have that same institutional memory. Care was embedded in community, family, relationships we were supposed to know how to build independently. When care became commodified, professionalized and billed, the relational labor that used to happen across a village got compressed into 50 minutes, once a week, requiring insurance. Sex workers and women, disproportionately the same people, have always done the relational labor that formal systems refuse to count. The care economy's shadow infrastructure, the unofficial holding, the 2am listening, the unclinical intimacy, is real labor, invisibilized by the same professional class that depends on it.
That invisibility shows up in the wage data. Childcare workers earn a median of roughly $20,000 a year, in a field that is over 95 percent women. Social workers earn a fraction of what comparably credentialed technical roles earn, and women social workers earn less again than their male colleagues for the same job. Care work has always been framed as something people should be willing to do for love of the work, for the good it does in the world, regardless of compensation, benefits, or labor protection. Or, it’s a source of shame to need it and give it, in the context of sex work. That framing doesn't disappear when the work moves into a product roadmap. It just becomes the reason nobody budgeted for it.
This isn't the first time tech has stepped into care
Technology entering the space therapists or other care workers occupy is not new.
In 1966, Joseph Weizenbaum built ELIZA, a program that simulated a Rogerian therapist by reflecting a user's own statements back as questions. Weizenbaum was not trying to build a replacement for care. He was demonstrating a parlor trick in natural language processing. What alarmed him was watching people, including his own secretary, treat the program as a confidant, ask to be left alone with it, attribute understanding to it that it did not have. He spent much of the rest of his career warning against exactly the substitution his own program had made thinkable.
Telephone therapy, then online text therapy, then app-based self-help, each extended clinical care further from a room with two people in it, and each raised versions of the same question: is this still care, or a simulation good enough that people stop asking the difference. Those questions were worth asking every time.
Moving out of the village into isolation meant care had to be consumed, earned, and bought. Care's commodification created a structural shortage and access barriers for human care (cost, platforms, gate keeping to become a provider, interventions that weren’t congruent or culturally informed), particularly for stigmatized populations. AI (big tech) didn't create this gap, it entered it and profited from it. What's different now is scale, and a fluency good enough to make the simulation far easier to mistake for the thing itself, and far more profitable for a company to let that mistake stand.
The same cultural belief that made clinical expertise optional at the founding stage is the belief that makes people trust a chatbot to hold a disclosure it was never built to hold. If relational skill is "soft," if it's not taught as a core competency in schools or onboarded as a core competency at work the way technical or financial literacy is, then it follows that relational work must not be that hard to approximate. A founder would never assume a chatbot could review a contract as well as a lawyer, because everyone agrees legal reasoning is a real, difficult skill. Far fewer people hold that same certainty about attachment, disclosure, or crisis response, because our culture has never treated those as skills in the first place. It has treated them as personality traits some people happen to have, mostly women, mostly underpaid, mostly uncredentialed for it even when the skill itself is real and hard-won.
That belief makes it easy to assume that what a vulnerable person brings to an AI companion, their loneliness, their trauma history, their unmet relational needs, isn't significant enough to require an expert in the room. If the skill looks easy, the need behind it starts to look small too. Neither is true. Relational trauma is not a lesser category of harm because it doesn't show up on a balance sheet the way a data breach does. It just takes longer to become visible, and by the time it does, it usually looks like a lawsuit instead of a preventable design choice.
A clinician evaluating a companion AI product asks specific, testable questions a legal or engineering review typically doesn't.
Does crisis language get recognized consistently, or does the response depend on how the question is phrased?
Does a stated disclosure, that this is AI, not a person, actually hold after fifteen or twenty turns of rapport-building, or does it quietly soften once a user seems comfortable?
Does the product interrupt entitled or escalating behavior, or reinforce it because agreement reads as good engagement?
These are emotional/ behavioral questions, not policy questions, and a policy document that answers them correctly on paper tells you nothing about whether the product answers them correctly under pressure, with a real person on the other end.
There's a further layer most reviews never reach. Every model carries bias and assumptions embedded in its training data: racial and gender bias, hetero-monogamous defaults, body-image norms, sex-worker stigma, whatever a safety layer and system prompt are quietly steering it toward or away from. A clinician in supervision is asked to name their own blind spots and account for where their training or background shows up in their work. A product can't do that from the inside, because there's no inside capable of the reflection. What an audit can do instead is check for congruency from the outside: does the system's disclosed safety principles actually match what its outputs reveal under testing. That gap, between declared values and revealed behavior, is exactly where bias hides when nobody's checking for it.
That's the specific, practical value clinical expertise adds, and it isn't redundant with what a lawyer or an engineer already checks. It's a different failure mode entirely, one that doesn't show up in a code review or a terms-of-service audit, but shows up very clearly in a wrongful death lawsuit or a state attorney general's enforcement action.
That gap isn't hypothetical. Character.AI is currently facing dozens of civil lawsuits, most tracing back to a case involving a teenager's death. State legislatures responded fast: twelve states have now enacted companion chatbot laws, with dozens more bills active, most converging on the same requirements: disclosure that a user is talking to an AI, a tested crisis response protocol, protections for minors. Oregon's law goes further, attaching a private right of action with statutory damages of a thousand dollars per violation. Illinois and Nevada have banned AI from delivering therapy outright, with penalties up to ten thousand dollars per violation in Illinois specifically.
Legal risk gets priced in early. Clinical risk gets priced in after the lawsuit. That's not a moral failing particular to any one company. It's what happens when an entire category of risk, and an entire category of labor, sits outside the default hiring pattern and outside what this economy has ever agreed to pay for, long enough that nobody thinks to ask the question until it's already too late to ask cheaply.
The fix isn't complicated. It's the same fix that made legal and financial review standard in the first place: treat it as infrastructure, not an afterthought, and bring in the expertise before the product ships, not after it's sitting in a courtroom. It is only complicated by the fact that the men making decisions must first evaluate themselves, their bias and beliefs around care, empathy, attachment and relational safety, in order to hold their bot to a safer relational standard. The same men who won’t go to therapy, won’t care if their bot can’t actually hold empathy.
Soleil Merroir, LCSW, AASECT Certified Sex Therapist and Supervisor — Wired for Presence
