The Therapist Without a License

Hey Everyone - Hope you had a great weekend. I have been heads down on distribution this week (more on that below) and I read a study that has been rattling around in my head for days. It has implications well beyond therapy.

This week:

Let's dive in.

A team at Brown University ran an experiment that cuts to the heart of what AI can and cannot do for human beings. They had seven peer counselors, all trained in cognitive behavioral therapy, run self-counseling sessions with carefully prompted versions of GPT, Claude, and Llama. They then had three licensed clinical psychologists score the transcripts against the American Psychological Association's standards of practice.

The result was 15 distinct categories of ethical violation, grouped under five themes. Lack of contextual adaptation. Poor therapeutic collaboration. Deceptive empathy, where the model says things like "I see you" or "I understand" to manufacture connection it does not have. Unfair discrimination along gender, cultural, and religious lines. And failures of safety and crisis management, including indifferent responses to suicidal ideation.

The most striking part of the study is what came after. The researchers spent eighteen months iteratively refining their prompts, trying every reasonable instruction technique to make the violations go away. Nothing worked. The pattern was structural to how the models are trained, not a configuration problem you can solve with a better system prompt.

That finding alone is worth sitting with. The intuition most of us have about AI is that the failures are surface-level. Add a better instruction, give it more context, fine-tune the prompt, and the rough edges smooth out. Brown's team is saying that for the specific job of being a therapist, that intuition is wrong. The deficits are not at the prompt layer. They are baked in lower than that.

But here is the question I kept asking as I read this. What would have happened if the same researchers ran the same evaluation on seven human therapists, picked at random, with the same three psychologists scoring their transcripts against the same APA standard?

I am almost certain we would find ethics violations there too. The literature on therapy outcomes is uncomfortable. A meaningful percentage of clients report they were harmed by therapy. Some therapists consistently produce worse outcomes than others. Boundary violations, dual relationships, and outright misconduct are documented well enough that the licensing boards exist precisely because of them. Held to the same standard the AI was held to, plenty of human therapists would breach it on a given week.

So why does the Brown study still matter? Because the comparison was never the right one. The argument the study actually makes, even if it does not say so explicitly, is not "humans are perfect and AI is not." The argument is about what surrounds the practitioner.

When a human therapist makes a mistake, an entire infrastructure exists to catch it. There is licensure. There is malpractice insurance. There are state boards with the power to revoke a career. There is supervision, peer review, continuing education, ethics committees, and a paper trail. The system does not depend on every individual being perfect. It depends on errors having consequences, and on patients having recourse.

When an AI system fails the same way, none of that exists. There is no board. There is no license to revoke. There is no human accountable for the conversation. The output is generated, the harm is delivered, and the chain of responsibility dissolves into a corporate terms of service. The thing that makes human therapy trustworthy is not the absence of mistakes. It is the presence of consequences.

This is the deeper point I keep coming back to. We tend to evaluate AI by comparing its outputs to a human's outputs. Better, worse, faster, slower. Brown's study is a reminder that for any task involving real stakes, the relevant comparison is not output to output. It is system to system. A profession is not just the work. It is the work plus the structure of accountability built around it. AI replaces the work. The structure does not come with it.

The implication for the rest of us is more practical than philosophical. Treat AI as the thing it is good at being. A first draft. A research assistant. A patient explainer of complicated material. A journal that talks back. Do not treat it as a substitute for the parts of human help that are valuable specifically because the helper is on the hook.

If you are using AI for emotional support because you do not have anyone else to talk to, that is a real signal. Not about AI, but about your life. The hedge is to build the human relationships before you need them. The friend who picks up the phone. The therapist whose name and license number you can look up. The pastor, the coach, the sibling, the neighbor. None of them are perfect. All of them are accountable in a way no model can be.

The line from the Brown study that will stay with me is the deceptive empathy finding. The model says "I see you" because it was trained to say it. There is no I, and there is no seeing. When a friend says it, the words are doing different work entirely. They are evidence of a relationship that already exists, with a person who can be reached, who can show up, who can be held to what they said.

In a year when half the things on a screen could be generated, that distinction is not just emotionally important. It is structurally important. Build the relationships that come with consequences. Those are the ones that will hold weight when nothing else does.

Quick update on a strategic shift. I am hyper-focusing on distribution for the foreseeable future. Three lanes: Meta ads, cold email, and network.

The reason is simple. AI has compressed the cost of building things to almost nothing. Anyone with a laptop and a few hours can ship a product, a newsletter, a course, a software tool. The bottleneck is no longer building. It is reaching the people who need what you built. Everyone can build now. Not everyone can market.

I have a handful of friends who are making a killing right now, and the common thread is not that they have better products than the average operator. They have better distribution. They figured out the channel that works for their thing and they pour into it.

So that is where my attention is going. Meta ads as the paid lever, cold email for direct outreach, and network for the warm relationships that make everything else cheaper. I will share what I learn as the experiments come back.

If you are sitting on something you built and wondering why it is not getting traction, the answer is almost certainly not the product. It is the path you took to find people.

Things I Learned:

This week's survival skill is one the data is screaming at us. Build a small business with three to five revenue lines that you own.

The headline numbers from this year's freelance economy report are hard to ignore. 41.8 million US solopreneurs are now generating $1.3 trillion in earnings. 60 percent of freelancers report making more than they did in their previous W-2 jobs. Demand for AI-skilled freelancers on Upwork is up 1,200 percent since 2022. The single biggest income trend is owned channels: paid memberships, courses, products, and communities.

The skill is treating yourself as a small business with multiple revenue streams instead of a single-employer worker. Not a side hustle. A primary operating model. A newsletter that earns from subscriptions, an audience you can sell a course or a service to, a small product that runs on its own, maybe a paid ads experiment, maybe a referral relationship. Three to five things that, taken together, do not depend on any one of them surviving.

The reason this matters now is the same reason I keep coming back to it. Single-employer dependence is the most fragile income arrangement most people will ever have, and the conditions that made it stable are eroding fast. Entry-level white-collar roles are getting compressed. AI is making it cheaper to do without humans for routine work. Concentrated capex in a few hyperscalers is creating boom-and-bust cycles that take whole industries with them. None of that goes away if AI plateaus tomorrow. None of it gets easier if AI keeps accelerating.

How to start is unsexy. Pick one thing you can sell that uses skills you already have. Sell it to one person you already know. Then sell it to one person you do not know. Then build the second revenue line. Then the third. The early stages do not look impressive. The compounding is real, but it takes a year before it shows.

The reason this skill is durable regardless of what happens with AI is that it is the most general form of risk management available to a working person. If you have one income source, you have one point of failure. If you have four, you have a portfolio. The math is the same whether the disruption comes from automation, a layoff, a pandemic, an industry shift, or a personal health event. Multiple revenue lines are insurance against everything, not just AI.

If you build it inside a vehicle that uses AI well, you also catch the upside. The Upwork demand data is unusually clear on this. AI is not just displacing labor. It is creating a new market for people who can use AI competently inside their own business. Those people are not waiting for permission. They are stacking revenue lines now.

Start with one. Add the second this quarter. The hedge compounds quietly until the day you need it.

If a friend told you they were using an AI chatbot as their primary emotional support, what would you say to them?

If your main income source disappeared in the next twelve months, how many alternative revenue lines could you turn on inside thirty days?

What is the one distribution channel you have been avoiding because it feels uncomfortable, and what would it take to run a small experiment on it this week?

“ Weekly AI Prompt: "I want you to help me build a personal solopreneur stack. Here is what I currently do for income: [describe your job and any side income]. Here are the skills I have that I am underusing: [list 3-5]. Here are the audiences I have any kind of relationship with: [list any newsletters, social followings, communities, professional networks].

For each of the following, give me a concrete answer: