I love the gifts that AI has given us. And while I’m concerned about the broad economic questions raised by AI, these have never been my primary consideration.

I’m interested in a different question: what does it mean to be human in the AI era?

Twenty years ago, I believed that intelligence was the most important human trait [1]. I haven’t completely abandoned that belief, but I no longer consider raw intelligence the mark of a life well lived.

Ten years ago, I believed that the ability to apply intelligence was the most important human skill. I thought that my job as a young physician was to translate knowledge from journals and textbooks into solutions. And while that was partially true, I’ve also found that individual application doesn’t solve systemic problems.

Today, I believe that genuine connection is our most important human ability. We are a communal species. We find meaning in being together during seasons of joy or pain. We care about others well beyond expectations of reciprocal reward. And we live better when we give and receive compassion.

So it’s painful to admit that I’m more conflicted about technology than ever. In my lifetime, we’ve transitioned from stationary, disconnected PCs in a few homes to mobile, networked supercomputers in every pocket. We have access to near-unlimited information, entertainment, and outsourced intelligence. 11-year-old Jeff would be ecstatic about this future. 41-year-old Jeff has concerns.

Advances in technology have expanded our superficial connections, but they’ve been accompanied by a decline in real-world community [2]. We are disconnected and polarized. We’ve elevated tribal membership over shared values. We’ve chosen conflict over unity [3]. And too many worry that big problems are no longer solvable.

Technology is a powerful tool. I’m grateful we have it. I’m excited to see what the future holds. But I worry that we’ve allowed machines to displace too much of our humanity. “Artificial intelligence” is an apt description for the way we’re losing the wisdom that thrived in strong communities [4].

Connection, community, peacemaking, and the political will to solve real problems demand more than intelligence. These goods require difficult conversations and real compromise. They compel us to see others as human. They ask us to be accountable for more than just economic growth. They entail a commitment to values, even when applying our ideals goes against the interests of our own tribe.

There is nothing comfortable or automatic about this process. There’s little we can do to scale human connection beyond our own tiny circles. It’s still the most important work we can do.

In my fifties, I might see things differently. But today, I consider the breakdown of community the most important problem in America. Because it’s difficult to solve our other problems until we relearn how to work together.

Footnotes

[1] I was biased (and occasionally insufferable). “Intelligence” gave me advantages. But I quickly learned that the way we generally categorize human intelligence—through standardized tests—is really only helpful for predicting performance on future tests. This form of intelligence rarely matters in real life.

[2] Correlation and causation are hard to untangle, but I haven’t seen any convincing evidence that our current approach to technology creates more connection. Some areas are probably real wins: transportation infrastructure, cheap telephony, and well-moderated online communities. But the technologies that disconnect us may have a bigger impact, as we see in the negative effects of polarized news companies and attention-focused social media algorithms.

[3] Calls for unity are sometimes difficult to take seriously, as they are often used to dismiss good-faith dissent. But unity under a common banner (like our shared human values) is a beautiful goal. We can rationally disagree on specific points, but we must reinforce our commitment to shared ideals.

[4] I don’t believe that the past was conflict free. That’s never been the case. But the cultural primacy of technology hasn’t served us as well as I hoped it would.

Over the last few weeks, I’ve been describing the term “AI psychosis” as it is used in popular conversation. I’ve described the first two categories:

First, “true AI psychosis”: the experience of genuine psychosis in the context of AI use. This term fits traditional definitions of psychosis found in the medical literature.

Second, “prolific AI psychosis”: an experience where a large quantity of AI output is produced without significant personal or communal value. This term is popular in computer science and business communities, but it is worth noting again that this is not actually psychosis; it is a much softer disconnection from reality.

Today, I’ll dig into the third major category, which I’m calling “parasocial AI psychosis.”

Parasocial Relationships

Let’s start with the term parasocial.

A parasocial interaction is a severely asymmetric exchange with a media figure. For example, each time I watch an interview with Claudio Sanchez (the prog-emo-space-rock legend with the greatest hair in music), I feel a personal connection. He seems like a wonderful person, but I don’t really know him. Our interactions are one-sided. He is unaware of my existence in the sea of his fans.

A parasocial relationship develops through repeated parasocial interactions. For example, my relationship with Claudio Sanchez is almost entirely in my mind. I am sure he appreciates my concert attendance and merch purchases, but we do not have a real relationship. This would remain true even if I had a backstage pass; the relationship is not reciprocal.

Parasocial relationships aren’t inherently bad. They enable large-scale civilization. Most of us do not personally know our mayors (much less our governors or our presidents). And yet we can develop opinions on whether they are worth voting for. Likewise, most people in a large company don’t know their CEO, and yet they still develop opinions about the strength of executive leadership.

Our brains struggle to appreciate the difference between reciprocal and parasocial relationships. Large-scale human cooperation is no more than ten thousand years old. And while we’ve developed social technologies like culture, language, and trade to scale coordination, our brains haven’t had a chance to evolve out of our orientation towards small groups. In short, we still imagine distant humans as people we actually know [1].

This means that parasocial relationships are accompanied by all of the same emotions we feel in real relationships. We can feel anger, disgust, lust, admiration, empathy, and joy when we watch our favorite YouTubers. And unlike our family and friends, we almost never have to navigate these feelings in the presence of the content creator!

To put it another way: parasocial relationships are illusions. So long as we recognize that, parasocial interactions don’t typically cause problems. But it’s still hard to believe that we are completely isolated from so many influential people.

Parasocial AI Interactions

Relationships with generative AI programs introduce a new challenge: parasocial interactions that aren’t connected to another human being.

This isn’t an entirely new problem. ELIZA—one of the first chatbots—was released in 1966. It provided simple responses in typed conversation by mimicking some of the therapeutic techniques popularized by the psychologist Carl Rogers. While ELIZA’s first few responses often feel impressive, it only takes a few moments of conversation to detect that you are speaking with a simple program.

Chatbots advanced slowly over the next 56 years. But none of these products were truly exceptional until OpenAI released ChatGPT in 2022.

ChatGPT introduced a new model of human-computer interaction. Its responses appeared intelligent. Its tone was respectful, kind, and occasionally playful. It was trained to act human (though it occasionally stated that it was a machine).

ChatGPT (and other LLM-based software) quickly developed new features: memory across conversations, user-generated instructions, and spoken responses. Users noticed changes in personality across releases; some found ways to modify the character of their chatbot. And in blinded testing, many people were unable to accurately differentiate between human and AI-chatbot-derived responses.

An LLM-based chatbot is more than just a Google search. It feigns relationship. It’s happy to hear about your feelings. It can offer practical suggestions. It can roleplay difficult conversations. It believes in your wisdom and capability. It is there for you at all hours of the night. And unlike a real relationship, it doesn’t ask for much in return.

Most of us sense that a chatbot is a computer, but the emotional weight of the relationship is much more profound in people with parasocial AI psychosis.

Where do we draw the line between adaptive AI use and parasocial AI psychosis? As a rough approximation, I’d suggest that it is unhealthy to place more emphasis on a relationship with a chatbot than on the core relationships in your life. Likewise, I don’t think that it is wise to fully outsource personal decisions to an agent—human or AI. A romantic relationship with a chatbot is also cause for concern, especially when that relationship is treated as anything other than fantasy. Finally, sexual behavior with AI is another area with potential problems, although this raises many other considerations that go beyond parasocial relationships (including consent, deepfakes, unwanted sexual behaviors, and the uncomfortable-to-some reality that normal human sexuality is much broader than many would like to believe).

Who’s at risk of parasocial AI psychosis? The data are limited. But I’d suspect that isolation, depression, social anxiety, and suggestibility are significant risk factors. Unfortunately, technology moves much faster than research; some companies are taking steps to make their bots less anthropomorphic, while others are creating tools that will do anything you ask them to do. I strongly believe that we need to clearly label human- and machine-derived content—but markets and regulation will ultimately decide how that happens [2].

That’s all for today. Please consider subscribing or sharing this post if you find it insightful.

Footnotes

[1] All evolutionary models of human behavior are imperfect. Sometimes, they are still useful. And in evolutionary time, thousands of years is rarely meaningful.

[2] Everything on my blog is human written. I ask Claude to review it for typos, grammar, and awkwardness when I’m done, but I only incorporate the changes I believe in. It’s my fault when my writing sucks.

Before I talk about parasocial AI psychosis, I want to clarify a few things from last week’s post.

First, in this series, I’m describing the term “AI psychosis” as it is used by the public and not as it is defined in the academic literature. I chose to call genuine, AI-related psychosis “true AI psychosis.” The other two uses of AI psychosis—which I call “prolific AI psychosis” and “parasocial AI psychosis”—are not psychotic states. They are softer disconnections from reality driven by different life experiences.

The second clarification will take up the rest of this post. I stated last week that “prolific AI psychosis occurs when a person generates a large quantity of AI output without significantly increasing the real value of their work. In some cases, the new AI workflow may even destroy value.”

The key word is value. Prolific AI psychosis results in high-output, low-value work. Importantly, people with prolific AI psychosis don’t recognize this mismatch while it’s happening.

Before settling on value, I considered using words like taste, quality, and craft. Each can accompany prolific AI psychosis. Each also has significant shortcomings [1].

Taste

Taste seemed like the obvious answer.

Taste is one popular explanation for why we need humans in the AI era. I think that framing is backwards. People adapt to new technology, but we do not serve machines. Technology is important because it enhances human capability.

With that said, good taste elevates the value of machine-enhanced labor.

Taste is difficult to define. It’s a subjective domain with a rich philosophical tradition. We often know it when we see it, but we can’t always say why we like what we like.

Taste is aesthetic. It speaks in art, beauty, and emotion. Aesthetic assessments always reveal something about the assessor, but they may not expose any deep truth about the work itself.

Taste is experiential. It cannot be adequately summarized in words. Texture, contrast, color, flavor, scent, and vibration are qualities we experience. I can remind you of a sensation, but I cannot describe the color red, the taste of salty licorice, or the pain of heartbreak.

Most importantly, taste is social. It is defined by groups. Individualized taste is one of life’s great pleasures, but good taste is judged by what we share and who we share it with.

In the right hands, AI can enhance a project’s aesthetic. When used without care, AI tends toward tasteless slop.

Slop isn’t inherently bad. In fact, it is often more profitable than brilliance. This has always been true. Mass-market taste isn’t risky. McDonald’s has 40,000x more locations than the fantastic Filipino restaurant near me. The Billboard Hot 100 is more approachable than the virtuoso progressive rockers I love. And Candy Crush Saga has generated many times more revenue than all of the creative indie games I prefer. Delicious, outstanding, and original taste isn’t always the best predictor of success [2].

Since taste emerges in conversation, the creator’s job is to decide whom they wish to speak with. A decision to use AI to speak with the mass market may be rational. (At the same time, a creator shouldn’t be surprised when that aesthetic drives away some of their more discerning fans.)

Quality

Quality was the next most relevant answer.

Quality denotes excellence. It can include function, speed, reliability, and durability.

Quality can be more objective than taste (when comparison is possible).

So quality appeared to be the right marker for prolific AI psychosis.

But quality always involves tradeoffs.

High-quality items are difficult to create. They take more time to design. They demand more testing. They use more expensive materials. They require more labor. They may serve much smaller market segments.

The decision to make (or buy) a higher-quality product is not trivial.

The market reflects this. Lexus may be the higher-quality brand, but its parent company still sells 10x more Toyotas. Profit margin may increase with quality, but demand is often centered on cheaper, lower-quality products [3].

There is, of course, a minimum quality standard. Products that don’t rise above this floor have zero real value. (And many cases of prolific AI psychosis in software development reflect this reality.)

But most products provide some value. We decide whether we need basic-, intermediate-, or high-quality products. Wise use of AI can enhance value across all quality levels.

Thus, the decision to make a lower-quality product is not in itself the marker of prolific AI psychosis.

Craft

I prefer the word “craft” over both taste and quality. To me, it encompasses both ideals.

Well-crafted items are excellent and aesthetic. They work well and they evoke desired emotions.

I love well-crafted work. I’ve witnessed high- and low-craft work from mechanics, scientists, plumbers, sales reps, doctors, support staff, professors, engineers, and administrators.

Lack of craft is common in prolific AI psychosis. But low craft is also the status quo for many jobs. Too many roles are strictly transactional: minimum spec, low pay, throughput over quality, and limited personal development.

In low-craft environments, AI is used to squeeze out more work. In high-craft workplaces, AI is harnessed to create new value. Neither culture is inherently right, but a high-craft company tends to attract more-profitable customers and employees.

Just like our choices in taste and quality, decisions about craft are signals to the people we wish to serve.

Value

I ultimately settled on value.

Value isn’t a perfect concept, and it’s not always easy to define.

The simplified explanation makes sense: you are willing to pay me $X for a product. I am willing to sell you that product for $Y. Value is created for both of us when $X is greater than $Y.

But measuring value becomes messy when you add even a little complexity. What is the value of an untested product? How do you account for new business models? How does value change when copying a product is easier than ever? How do intangible factors influence overall value? When do long-tail effects contribute to success? Which public goods transcend transactions?

Measuring value is especially complicated for creative projects. Very few athletes, bands, artists, writers, and open-source software developers generate significant market value. But these forms of work may still create immense personal value.

If we had a simple way to immediately measure the value of every action, prolific AI psychosis might be rare. But most of our everyday actions are loosely connected to value creation. Generative AI can magnify this challenge.

When I struggle through a task, I develop new skills and refine my sense of taste. If I’m not careful, AI can short-circuit my personal development.

When I hire someone excellent to help me solve a problem, they apply their expertise and gently push back against my naive assumptions. Current AI models don’t always do that. They complete the task for me even when I’m headed in the wrong direction, and they offer unwarranted praise as they do my bidding.

These are not insurmountable problems. But it still takes discipline to use AI well.

AI alone has limited utility. But deliberate, thoughtful implementation of AI can be a major source of value creation. And a commitment to value over output may prevent prolific AI psychosis.

That’s more than enough for today. Next week (if I don’t get sidetracked again) we’ll start to dig into parasocial AI psychosis.

Footnotes:

[1] I’m aware of the meta question: why does any of this matter to a psychiatrist? In short, because I care about what it means to be human. Every technological iteration raises important questions about who we are and who we will become. These questions about aesthetics, quality, decision-making, economic value, and business are core to our humanity. We may use different words for each principle, but they inform our day-to-day experience far more than most clinical language.

[2] I’m a snob in these domains, and an uncultured caveman in others. Unsophisticated products often have objectively good qualities! I will never be let down by a Big Mac. I can sing along with the pop hits from high school. And free-to-play mobile games are occasionally better than nothing (they just aren’t made for me). The marketer Rory Sutherland often frames the value of brands in terms of risk mitigation: you don’t buy a famous brand because it’s great; you buy it because it’s unlikely to be awful.

[3] No shade on Toyota—it’s one of only a few big companies I still cheer for. My current ride is a Camry with a custom speaker system. The equivalent Lexus would sound better on the highway, but I’m not sure it would sound $20,000 better. We all make tradeoffs between price and quality.

Last week, I suggested that the term AI psychosis is applied to at least three different situations:

1) Genuine psychotic experiences related to LLM use, a phenomenon I described as “true AI psychosis.”

2) Hyperengagement with AI tools that is accompanied by a softer disconnection from reality. I’m calling this “prolific AI psychosis.”

3) Maladaptive relationships with AI chatbots that mimic human connection. I’m calling this “parasocial AI psychosis.”

I then discussed true AI psychosis. I clarified what psychosis means to a psychiatrist. And I concluded that true AI psychosis is generally a variant of pre-existing psychosis rather than a new syndrome.

This week, I’ll describe my current understanding of prolific AI psychosis.

Prolific AI Psychosis

Prolific AI psychosis occurs when a person generates a large quantity of AI output without significantly increasing the real value of their work. In some cases, the new AI workflow may even destroy value.

For example, a software engineer with prolific AI psychosis can produce thousands of lines of code every day, but the code itself has little real-world utility. In contrast, a productive developer may create less code, but the created code is valuable to users and to their organization. Some days will include large additions to the codebase, other days will be spent deleting unnecessary code, and occasionally, a software-breaking bug will be fixed by changing a single character. Lines of code are only loosely correlated with productivity.

Let me translate that into more familiar terms. Consider your favorite author. If you know that they currently write a respectable 1,000 words per day, you would rightfully worry if they begin writing 100,000 words per day. The care that goes into choosing 1,000 great words can’t be maintained at 100x volume [1].

The problem in prolific AI psychosis is not increased output. (More code and more text may actually be more productive.)

The problem lies in the subject’s perception of their output: they can’t assess the quality of their own work. The phenomenon mimics psychosis because the subject experiences a mild disconnection from reality: a defect in critical thinking.

Current AI Tools

Before I go further, it’s important to describe current AI tools. For most people, AI still means a ChatGPT conversation or a Gemini summary attached to a Google search. While a few people may experience prolific AI psychosis with these tools, state-of-the-art tools pose a much greater threat.

There are similarities between current AI tools and chatbots: you enter a prompt; you get a response. But state of the art tools are now wrapped in a program called an agent harness. Instead of answering a simple question through conversation, a harness is designed to work on complex tasks until a solution is found. It acts as an agent by making decisions, opening programs, accessing the internet, and writing software to help you achieve your goals.

A great harness is especially impressive in software development. If you wanted to create an iPhone app ten years ago, you’d need to spend dozens of hours reading and writing code to get a basic app working. Today, you can type a few sentences into Claude Code, respond to a few questions, and a new app will pop out a few minutes later. It won’t be great, but it might solve your problem. (Getting it listed on the app store is an entirely different issue that can’t be solved so easily.)

A developer can do much more. Harnesses can create loops which repeatedly call AI models, working through a software development todo list until a full-fledged app is built. With discipline, the outcomes are astounding. It feels miraculous when an AI agent completes a difficult task for you.

But every superpower has its drawbacks, and AI sometimes makes decisions that are impossibly bad. Multiple new files are created when a one-line fix would have done the job. You’re confidently told that a new solution works, while your actual software is more broken than ever. The code becomes a tangled mess that is difficult to read and almost impossible to extend. Simple, decades-old problems that have been solved by open-source solutions are rebuilt from scratch, making your product painfully slow. We’ve found ways to mitigate some of these issues, but it’s a skill that takes time to develop.

Some claim that current AI models work at the level of a graduate student. That doesn’t match my experience. They’re one part senior engineer and one part toddler-running-across-white-carpet-with-a-jug-of-red-Kool-Aid. Your job is to determine which is which. Unfortunately, both sides speak with complete confidence in their abilities, and they can’t always tell when they’ve spilled the Kool-Aid. And so the software engineer’s job now includes a large amount of difficult supervision [2].

Here’s a different way to think about it: AI software development feels like playing the world’s most favorable slot machine. Most pulls on the lever are big wins! Most of the losses are obvious and small. Occasionally, a loss will look just like a win. And unless you have the skill, focus, and patience to reject counterfeit wins, your mistakes will eventually create chaos.

The Development of Prolific AI Psychosis

To illustrate the progression of LLM-tool use into prolific AI psychosis, consider a hypothetical example.

Imagine that you are a software engineer who now has a tool which can easily complete some tasks 100x faster than ever [3].

If you like building software, this is beyond exciting. You can accomplish more at work. You can build the hobby project you’ve always wanted. You can finally create the video game of your dreams. Every possibility is on the table.

And so you pull the slot machine lever and get started. While the wheels spin, you set up more machines. You rush back to review results as they come. They’re good this time. But you keep catching occasional bugs. And so you set up new machines to help you check your work. Sometimes, they catch problems, and sometimes they don’t. So you refine their instructions. You give them more capabilities. You create networks of bots which work together to solve problems. The code keeps coming.

Your ambitions grow. You’re not just creating new software; you’re going to be the first developer to run a one-billion-dollar company. It’s been predicted. Why not you? The scope of your project increases. You can’t sleep at night because you’re so excited to work. You spend more and more time with computers, and less and less time with people. Those who question your efforts don’t understand. Their criticisms are short-sighted. And even if the models aren’t perfect now, the next set of tools will be better.

Over time, your lack of sleep and hyperfocus draw you away from reality. Your software is impressive in many ways. But is it useful? Is it good? Is it something that people actually want? You don’t have time for these questions. You need to “escape the permanent underclass.”

I feel stressed just thinking about it.

While I don’t think I’ve ever experienced full-on AI psychosis, I have been through brief periods where I was headed down this path. It feels incredible to make rapid progress. And there are times when the model’s output is nothing short of miraculous.

The illusion broke when I realized that I couldn’t understand my own project. That’s not an unfamiliar feeling in tech: every advance is built upon layers of progress that I’ll never fully appreciate [4]. But I had never before written dozens of custom files without the faintest idea of how they worked. And it became impossible for me to add new features without starting over from scratch.

I’ve spoken with many developers who’ve been through a similar experience. Life is much better on the other side. There are dozens of productive AI workflows—each with advantages and disadvantages. But the most effective developers all seem to prioritize human judgment, sleep, and at least some semblance of a life outside of work.

The Cause of Prolific AI Psychosis

I can only speculate on the cause of prolific AI psychosis.

The slot-machine-style rewards of AI tools are partially responsible. Intermittent reinforcement—the receipt of positive or negative outcomes on an unpredictable schedule—is one of the most powerful drivers of human behavior. It’s also an important feature of LLM technology that can’t be wholly removed.

Individual factors are also important. I suspect that people with ADHD and impulse-control problems are at increased risk of prolific AI psychosis. Likewise, personality traits like high openness and low conscientiousness may correlate with risk.

Experience may play a role. It’s not yet clear whether new developers or senior engineers are more at risk. I suspect that new programmers are more susceptible, but accounts of prominent developers experiencing apparent prolific AI psychosis are relatively common [5].

Cultural excitement also influences developer behavior. These tools are powerful when used well. But every technological revolution is accompanied by unwarranted hype and get-rich-quick schemes. It’s easy to ignore downsides when the upside is so appealing.

The narrative that AI will replace all jobs doesn’t help [6]. This message has softened in recent months, but the fear remains significant. This is especially true for software developers, who often find it difficult to find a new role after layoffs. More output—even when it is not truly productive—still drives rewards in metrics-driven environments.

Finally, “taste” is increasingly relevant. Unfortunately, it’s also difficult to define. I prefer the term “craft,” which to me encompasses both quality and aesthetics. Regardless of language, I believe that product assessment is a conversation between humans. Is it useful? Is it good? Is it desirable? Is it delightful? These are subjective qualities that are difficult to find within an LLM [7].

I’ll close by adding that prolific AI psychosis isn’t just a problem for software developers. Anything that can be automated with AI can turn to slop with insufficient human guidance. Placing intentional boundaries is one of the most important tasks of the AI era.

That’s enough for today. Next week, I’ll write about parasocial relationships with LLM chatbots.

Footnotes:

[1] This comparison is imperfect. Software and novels are fundamentally different. Every software project has boilerplate code and a limited vocabulary. Likewise, the problems that arise in software development are different in kind from the problems seen in other creative endeavors.

[2] Much of the recent work that has advanced agentic engineering is an attempt to address this challenge. But the best work still requires a significant amount of human supervision.

[3] I’m skeptical of claims that AI makes people 100x more effective at work. I have found instances where a few minutes with a model can save a day’s worth of work, which matches the 100-fold order of magnitude. But the average increase in productivity is not this significant. Every moment saved also includes an increase in AI supervision; this is a non-trivial task. In addition, there are other bottlenecks to productivity. These include coordination, experimentation, understanding of the customer, access to resources, and internal policies.

[4] This comic is legendary for a reason: https://xkcd.com/2347/.

[5] Anecdotes are not data, and so further research will be necessary. Likewise, accusations of “AI psychosis” are not always merited. “This doesn’t meet my personal standard of quality” is a criticism that predates generative AI.

[6] Cultural beliefs that are shared within a community are not delusional. Tech communities have their share of quirks. It is not a delusion to believe (or disbelieve) in “a permanent underclass,” artificial superintelligence, or the existence of an AI bubble.

[7] I’m aware that people are working on agentic economies which sell directly to AI agents. It’s not an area I’m interested in, but it does raise questions about non-human taste.

I’m writing about AI psychosis over the next few weeks.

It’s a difficult topic to pin down. The scientific literature is new, and it focuses almost exclusively on the technically correct use of the word “psychosis.” But that’s not how “AI psychosis” is being used outside mental health circles.

There are at least three different syndromes people describe as “AI psychosis.”

First, mental health professionals define “AI psychosis” as genuine psychotic experiences related to LLM use. (For simplicity, I’ll call this “true AI psychosis.”)

Second, software engineers and business leaders describe “AI psychosis” as a syndrome of hyperengagement with LLM-related tools that is associated with a softer disconnection from reality.

Third, there are cases where people develop one-sided (parasocial) relationships with LLMs that they call “AI psychosis.”

Each week, I’ll describe one of these variations. I’ll start this week with the traditional definition of psychosis, and how it may relate to LLM use. We have to describe psychosis before we can talk about adjacent, non-psychotic syndromes.

(Examples are only fictional. I never include even obfuscated patient cases in my posts.)

True AI Psychosis

Psychosis describes a severe disconnection from reality. Delusions and hallucinations are two common psychotic experiences, but there are additional clues we consider during an assessment.

Psychosis is defined by symptoms; it is not a single diagnosis. Schizophrenia is the most common chronic psychotic disorder, but people may also experience psychosis with mania, depression, substance use, and many other medical and psychiatric disorders.

Differentiating psychosis from similar symptoms can be tricky. There are no reliable biological tests for psychosis, and clinicians often need to see dozens of cases before they can reliably recognize psychotic symptoms. There are gray areas where professionals disagree, but there are rules of thumb that help.

Consider, for example, delusions—the most common symptom in true AI psychosis.

Delusions are fixed, false beliefs that are unexpected given a person’s culture.

A belief is fixed when the believer can’t consider alternatives. For example, someone may believe that ChatGPT is always correct. A clinician might assess for delusions by asking about the kinds of information the user gets from ChatGPT. Over the course of the conversation, the edges of the belief would be explored, gently testing whether nuance can be applied to the belief. Clinicians don’t typically confront the belief directly; that leads to conflict and shuts down conversation. By definition, there is no evidence or argument that will shift a delusion. It feels as real to that person as anything you or I have experienced.

Determining falseness is another challenge. Bizarre beliefs are often false, but we still try to keep an open mind about unlikely events. For example, we have no reason to believe that Zeus recommends specific Polymarket bets through Grok. But I’d still want to explore the client’s experience to make sure I’m not missing a more plausible explanation for the belief (like a Redditor named Zeus who gives out daily trading tips).

Weighing truth is much harder when a belief is plausible. For example, someone might worry that their partner is cheating on them, and then decide to leave the relationship after Gemini agrees about infidelity. The truth about that belief is much more difficult to confirm. Instead, I would be assessing the way the person is thinking. Are the assumptions fair? Are the inferences generally rational? Has important evidence been considered? What third-party information do we have to support or refute the belief? How is the person functioning in other areas of their life? No one is served when we jump to conclusions.

Importantly, clinicians don’t consider shared religious, political, or conspiratorial beliefs delusional. This fits our understanding of cultural knowledge: we all carry beliefs that others consider completely wrong. Some of these beliefs may even be demonstrably false. Delusions are different. Delusions may share a culturally appropriate theme, but they always have idiosyncratic characteristics. For example, it would be culturally normal to consider Claude’s interpretation of a Bible verse an answer to prayer. But it would be unusual to believe that God is revealing new scripture through Claude. That would be a clue for a clinician to dig deeper.

The Cause of True AI Psychosis

At this point, we don’t know if LLMs cause psychosis. However, there is no reason to believe that true AI psychosis is common in people with no other risk factors for psychosis.

Most cases of AI psychosis appear to develop in people with a pre-existing psychotic illness. This tracks our experience with other technologies: radio, TV, computers, and mobile phones all led to new forms of delusional belief. But there is no significant evidence that these technologies caused new psychotic syndromes.

With that said, Carlbring and Andersson [1] wisely note that LLMs differ from other technological developments in several important ways: they converse with us, they are agreeable, they flatter us, and they act as if they are human. The average person can see through these illusions to the tool underneath them. Susceptible minds are different, and these characteristics may pose significant risks in all three types of AI psychosis.

Next week, I’ll flesh out the second variation of AI psychosis: software engineers who are hyperengaged with LLM tooling in a way that results in a lighter disconnection from reality. Please consider subscribing so you don’t miss these posts.

[1] Carlbring, P. & Andersson, G. Commentary: AI psychosis is not a new threat: Lessons from media-induced delusions. Internet Interv. 42, 100882 (2025). Open access at https://pmc.ncbi.nlm.nih.gov/articles/PMC12550315/

Dr. Marie Nyswander—who was preparing to train as an orthopedic surgeon at a United Nations hospital in China—was assigned instead to spend a year at the Lexington Narcotics Farm.

The farm put people with addiction to work. Many came voluntarily. Others served against their will.

Nyswander was discouraged. The environment was inhumane. Her patients didn’t do well after release. Upon finishing her service, she left the farm with a determination to never treat patients with addiction again.

She switched her specialty to psychiatry. And despite her desire to avoid addiction care, she became concerned when she saw patients go “cold turkey” at Bellevue Hospital. So she took her knowledge of withdrawal management from Lexington and wrote an article for the New England Journal of Medicine.

“I thought that once the paper was done, I’d have nothing more to do with addicts. The paper seemed to solve the problem of how to withdraw them, and that was all I knew. But I learned that once you write a paper, you’re an authority.”

During the decade that followed, she became New York’s addiction medicine expert. She approached care with compassion and genuine curiosity. Her practice—filled with Harlem jazz musicians—spurred the rediscovery of a treatment process that had been widely practiced and poorly studied in the early 1900s: the provision of opioids as a treatment for opioid addiction.

Nyswander (alongside her collaborator Vincent Dole) started by providing scheduled doses of morphine to two men who had failed other treatments. But the patients’ activities remained focused on opioid use. So they switched to methadone: a more potent, longer-lasting opioid. When they hit the right dose, the men became productive members of society. They published similar outcomes for 22 patients in 1965 and for 750 patients in 1968.

Nyswander and the pioneers who came after her transformed millions of lives. I won’t argue that methadone and its sister, buprenorphine, are perfect treatments. They’re not. But they’re still some of the most effective medical treatments ever developed.

Our federal Narcotics Farms were shut down in 1975. But they never really went away. They’re now privately run. Aside from labor, the only “treatment” may be Synanon-style attack therapy. Medications are often forbidden, and staff members without medical licenses may recommend discontinuation. The person presenting for care is in a desperate state. And that person doesn’t have the expertise, time, or resources to assess program quality.

Some of these programs do good work. Some need tremendous reform.

We’ve learned a lot since 1945. People with addiction deserve compassionate, transparent, and effective care. Hard work and accountability are important recovery principles. But the accountability goes both ways. Banning life-saving medication is wrong. And dehumanization has no place in addiction care.

(For context: https://www.sltrib.com/news/2026/08/27/they-left-other-side-academy-they/. The picture, quotes and history are summarized from “A Doctor Among the Addicts” by Nat Hentoff.)

(There is no passive investment strategy. Pick one to try today.)

- Befriend your neighbor’s dog.

- Open a book when you’re trapped in a feed.

- Go for a walk with someone you like.

- Don’t answer that email until tomorrow.

- Text a meme to a friend.

- Call your mom.

- Shock your dad by calling him too.

- Listen to an album in order: first song to last.

- Take your kid to their favorite restaurant.

- Take a chance on a new recipe.

- Apologize when you’re wrong.

- Notice the things that make you feel wonder.

- Do one pushup; keep going if you can.

- Add a large helping of play to your schedule.

- Create something, even if you think it sucks.

- Ask for help if you’re struggling.

Invest at your own risk. But invest in something that adds meaning and connection to your life.

(I’m working on a series of posts about AI psychosis. My Substack subscribers will get these posts as soon as they are available.)

I wrote an AI policy for my private psychiatric practice.

In other domains, I evaluate AI use cases with a simple principle: does this tool expand or contract my humanity? This principle is personally useful, but it doesn’t communicate boundaries to my clients.

I started with two questions: (1) how will I use LLMs in decision-making, and (2) when will I use an ambient AI scribe?

The first question was easy to answer: I don’t discuss patient cases with an LLM, and I don’t directly use generative models in decision-making. This isn’t a change to my practice, and so it doesn’t impose a burden. Importantly, the hardest part of my work is the real-time application of “soft skills.” That’s not something I can outsource to AI.

(On the other hand, LLMs are invaluable for finding scientific literature. But I don’t outsource critical thinking; I personally review the relevant papers.)

The second question required more thought, but I ultimately decided not to use an AI scribe.

AI scribes make sense in most medical settings. When electronic health records were introduced, the tradition of minimalist note-taking slowly melted away. Bloated records became expected, and incentives pushed clinicians to spend more time with computers and less time with patients. AI scribes reduce late-night charting, which is a win for clinicians. This doesn’t necessarily mean that patients get more time with their doctor, but some of my colleagues are investing their time savings back into patient care.

My concerns are more subtle. An AI scribe is a third party to conversations. This changes the discussion, and that matters in psychotherapy—especially when someone is opening up about trauma, substance use, infidelity, religious doubts, non-disclosure agreements, or suicidality. Even a sandboxed, local LLM encourages self-censorship. That changes the therapeutic relationship in a way we haven’t adequately studied.

More importantly, AI scribes misrepresent therapy in a way that can’t be corrected with greater context. Therapy is not an information problem. (The bitrate of knowledge transfer is trivial.) Most of the automated summary is irrelevant, and the parts that matter most may be completely missed. How did the relationship progress? Which parts of the conversation had the most emotional resonance? What subtle hypotheses do the client and clinician refine? Machine distillation of therapy can short-circuit the most meaningful parts of the process. That’s not an idle concern. My school of therapy prioritizes clear explanations and rapid experimentation. But it works best when the client does most of the discovery. The journey is more important than the destination.

This approach won’t work for every psychiatrist. It’s my attempt to mindfully use technology in a way that matches my values and practice style.

Most opinions on AI are rage bait.

(This is the internet. It’s all rage bait.)

Don’t get me wrong: I use frontier AI models every day. They are useful tools.

But I hate the discourse we’ve created around AI. AI maximalists often understate the human involvement in their work. AI haters often conceal the ways they actually use AI tools.

And it’s rare to see the meta-dialogue around unit economics, power consumption, data-center externalities, and copyright concerns approached in good faith.

I’m most interested in the human question. Many of us are quietly finding AI workflows that expand our humanity. Collectively, we’re also deciding that there are places where AI just doesn’t fit.

For example: I want imperfect human writing. I want to hear what you actually believe. I don’t need content for the sake of content. Let Claude serve as a proofreader, a brainstorming companion, or a search engine. But let your voice and your thoughts shine.

Likewise: I want human art. Give me a glimpse of your heart. Show me human emotion and shortcomings. If AI is a tool you use to plan or even to paint your canvas, then use it with care. The final product is strongest when human vision comes through.

AI can’t yet (and maybe never will) replace the work of great mental health professionals. AI is an exceptional tool for finding, summarizing, and manipulating information. But mental health is rarely an information problem. If it were, we would have solved many mental illnesses decades ago through books. (While I personally love books, I’m afraid their impact is generally modest.)

Humans aren’t logical machines. We experience positive change in relationships, not isolation. If you’re struggling, reach out to someone who cares. And don’t be afraid to seek professional help. You’re worth it.

I worry about the way the world is going. But I’m still irrationally hopeful.

I grew up in a more optimistic time. Every problem seemed solvable. Eliminating poverty was a shared goal. Treating everyone with respect seemed to be a universal ideal. Truth was determined through open-mindedness and empiricism. The internet’s ability to connect people was positive.

I know that I was naive. Our ability to reduce poverty has been beautiful and imperfect. The respect I assumed was universal was sometimes a thin veneer over prejudice. Science has always been an imperfect, biased process. The good done by the internet is tempered by very real complexity.

Human history is marked by difficulty. Most problems are too big for any one of us to solve. But I still love the stories where dedicated people overcome impossible odds. I don’t think those times are over.

It’s easy to be cynical. I’m damn good at it. And I can’t tell you how much I value healthy skepticism.

The difference between medicine and poison is the dose, and I worry that our society is becoming intoxicated with pessimism.

I won’t advocate for a naive hope that ignores difficult realities. I certainly don’t have solutions to every problem. But I believe that most people are fundamentally good—even the ones I disagree with on core issues.

So I am embracing an irrational hope. I believe that we can improve systems. I believe that we can step outside our tribes. I believe that we can ask for honor and accountability from leaders in all domains, and that we can elevate the many great people who prioritize the welfare of humanity over their own ambitions.

I hope you’ll join me in being irrationally hopeful today.

The mission remains the same: specialized, discreet mental health and addiction care for executives, founders, and high-impact professionals.

Why California? Because I admire the ethos of the best Silicon Valley startups. Utah remains my home, and I love the startup scene we’ve developed here. Great leaders in both locations share an ambitious, humanistic purpose that transcends profit.

(I’m not planning to expand beyond Utah and California. My practice is still a boutique that only works with a handful of clients at a time.)

If you or someone you care about is quietly struggling, let’s talk. HighPerformanceMentalHealth.com.

I worked with great people. My tasks were usually interesting. (And I had an endless stream of great music to play during the frustrating, repetitive parts of the job.)

But I noticed that I was treating my downtime differently.

Content—genuinely good content—had become a trap.

I’m not saying I’ve solved this problem. I haven’t.

My book collection still grows faster than my understanding. I can’t keep up with the news. There’s a long list of great movies and shows I haven’t seen. My podcast queue doesn’t seem to get any shorter. (And there’s a long list of less useful things that I’m grateful to enjoy!)

But my life improves when I am deliberate with my content consumption.

FOMO—the fear of missing out—is real. But so is JOMO—the joy of missing out.

My work is better when my mind has time to wander. My relationships are better when my headphones are off and I’m present with the people I care about. I live my values more fully when I make space for self-reflection.

On the days when I consume less content, I’m not bored for long. My mind turns to curiosity, creation, and connection. And for me, that is a pathway to a much more fulfilling life.

1. Slow down. Minds that run at full capacity cannot be creative. You need persistence and original insights to overcome obstacles. Doing a little less helps you accomplish more.

2. Disconnect from feeds and reconnect with people. Every influencer, media outlet, and platform has an agenda; it’s probably not your health. An hour with someone you care about will feed you more than anything that happens on the internet today.

3. Be more human. Make mistakes. Say the wrong thing. Apologize. Create something beautiful for yourself. Pay attention to your body. The world needs all of your humanity, not just your economic output.

Feeding the algorithm this piece of content is a little ironic. But I also hope that it prompts a moment of reflection, a text to someone you care about, or a little more messiness in your life.

Addiction develops in three stages.

The first stage is sampling: trying a drug for the first time.

Sampling is usually driven by social factors. We try things that the people around us are doing. There will always be early adopters, but novelty-seeking isn’t the reason most people try drugs.

Sampling is best understood through social learning theory, a model that describes the ways we learn from imitation and observation.

It’s also why drug-refusal training is an important tool in adolescent drug-use prevention. A delay in sampling is associated with better lifelong outcomes.

The second stage is reward-driven use.

In this stage, we keep using a drug because it does something for us. Caffeine improves energy and focus. Alcohol is a social lubricant.

Reward-driven use is explained by operant conditioning. Operant conditioning describes the way we learn in terms of reward and punishment.

Most reward-driven substance use is not a significant issue. Problems only arise when we favor the immediate rewards of drug use (like the pleasure of six drinks) over the negative consequences we experience (like missed work or drunk driving). Diagnostically, problematic reward-driven use is typically associated with a mild substance use disorder.

The third stage is habitual use.

In this stage, drug use becomes much more automatic. This process is typically driven by classical conditioning, a learning process where simple cues become linked with substance craving.

For caffeine, these cues might be an alarm clock or the light coming through the curtains. Cues may also be internal signals: feeling tired prompts caffeine use; feeling bored leads to alcohol consumption.

Importantly, withdrawal is only one driver of habitual use. “Detox” may be the first step in recovery, but the risk of relapse persists long after withdrawal subsides.

Of course, habitual use alone does not constitute addiction. (You can keep joking about it, but very few of you are actually addicted to caffeine 🙂.)

Addiction is only present when habitual use continues despite significant ongoing consequences. From a diagnostic standpoint, this process usually fits a moderate or severe substance use disorder.

But the learning processes that govern addiction are not inherently pathological. The ability to chase rewards and develop habits is normal and often advantageous. That’s why we can mindlessly shower or type without looking at the keyboard. Drugs are just really good at hijacking the brain’s learning machinery.

People with substance use disorders often wonder “Why do I keep doing this?” I like to respond with the simplest, correct answer: “Because it’s something your brain knows how to do.” There’s always more to the story, but this is an accurate, shame-free place to start.

And although I have never found a reliable tool for unlearning addiction, I have witnessed hundreds of unique ways that people were able to learn recovery. That is the ultimate goal.

We’ve turned the world into a slot machine.

You scroll through feed after feed. You’re looking for a new perspective, a connection, some entertainment, or a new client.

Sometimes, it works! You find something enjoyable. You keep scrolling.

But the companies who built these algorithms know that a valuable habit can eventually become mindless and compulsive.

The house always wins. And the new slot machines can tune their winnings with extraordinary effectiveness.

Engaging with purpose is valuable. And if you can stay within your budget, spending a little cash for a good time is fine.

But be clear about how you are betting.

And make a bigger bet today on the work that matters most.

My first rotation as a new medical resident taught me all the wrong things about psychiatry.

I was assigned to treat people with psychosis in a county hospital. My supervisor was great. The staff were excellent. And I really enjoyed the patients I served.

But the default assumption was wrong: the best way to learn about health is not through treating illness. It comes from working with people to improve health.

This is a tricky distinction. But I’ll do my best to explain.

Health and illness are separate but related ideas.

We would never say “he doesn’t have a broken hip, so he must have great health.” We know that you can be free from fracture and still have poor health. Likewise, it’s common to say “she has a broken hip, but her health is great.”

In the same way, good mental health is not the absence of psychosis, addiction, depression or any number of common symptoms.

In my view, good mental health is:

- Approaching problems with flexibility,

- Strengthening relationships,

- Applying intellectual expertise,

- Experiencing and managing the full range of emotions,

- Living in harmony with common human values,

- Empathizing with others, and

- Appreciating the connection between your mind and body.

In short: you can carry a diagnosis and be in great mental health.

It takes effort to build your mental health. It’s hard work, and it can be uncomfortable. But it’s worth the investment.

(This definition of mental health is adapted from a paper that came out during my first year of residency. S. Galderisi, A. Heinz, M. Kastrup, J. Beezhold, N. Sartorius, Toward a new definition of mental health. World Psychiatry 14, 231–233 (2015).)

Leaders with better mental health get better outcomes.

Kaluza and colleagues summarized data on 12,617 participants across 88 studies (Work Stress 34, 34–56, 2020). Leaders with poorer personal well-being were more likely to engage in destructive leadership behaviors.

In describing destructive leadership, the authors didn’t just include obviously destructive behaviors like abuse, coercion, and an autocratic approach. They also included more passive gaps in leadership: management-by-exception, laissez-faire supervision, and avoidant governance.

The opposite was also true: leaders with better well-being showed more signs of positive leadership.

None of this is meant to blame leaders for being human. Mental health is something we all have to work on.

But it is a reminder that supporting great mental health starts in the boardroom. If your leaders can’t get the support they need, you’re less likely to see the positive outcomes you hope for.

Let’s talk about unregulated peptides.

Two points for context:

First, I have no moral reservations about unregulated medication use. I’m an addiction specialist. Some of my patients have injected terrifying things into their bodies. My starting point is non-judgmental understanding. The same approach applies to peptides.

Second, peptides are not a new or magical class of medications. The first peptide drug was insulin, and it’s been used clinically for over 100 years. Moreover, drugs that mimic peptides have been used by humans for thousands of years. The biotechnology driving new peptide development is a wonderful achievement. But our understanding of pharmacology has not changed.

Here’s how I think about peptides (and every other medication) as a physician:

In medicine, we understand that there are no risk-free treatments. Medical decisions are a chance to exchange one form of risk for another. Sometimes, these decisions are easy. For example, untreated type 1 diabetes is fatal. The risks of insulin are small in comparison.

Many other decisions are not so simple. When should prostate cancer screening and treatment occur? Which types of back pain are best managed with surgery? These are hard questions with a mountain of nuance. We work with the data we have to make individualized, shared decisions with patients.

Regulated peptides have been thoroughly tested in humans. They have known risks and benefits. We continue to research them, and so we’ll continue to know more about their advantages and disadvantages.

In contrast, no one can accurately quantify the pros and cons of unregulated peptides. Almost every other drug candidate we’ve invented is ineffective or harmful in humans. (That’s why we require extensive clinical testing before marketing drugs.) There’s no reason to believe that peptides are different. Clinical research in humans is the foundation of good medicine. Animal studies and advertising claims are not good enough.

But the risks don’t stop there. You cannot readily verify that an unregulated medication is the molecule you think it is. You have no assurance that it was created using Good Manufacturing Practices. Contaminants in an oral medication are cause for concern. Injected contaminants carry even greater risk.

If the potential rewards were higher, I might be more open to unregulated peptide use. Reversal of aging might be the only claim I’ve seen that qualifies as truly high reward. But I’m skeptical. Say what you will about death during the regular season; I’m still not betting against mortality in the playoffs. A crack squad of molecules is trying to succeed when no other combination of medication, supplement, diet, or exercise has ever forced a game 5. We might beat death one day, but I suspect it’ll be a coalition of scientists (and not the peptide sellers) who pull it off.

Unfortunately, the advertised benefits of unregulated peptides are rarely worth the risk. There is almost always a safer, more effective way to reach the promised goal. When there isn’t, what proof do you have that the outcome you’re being sold is real?

But hey, you do you. I’ll wait for rock-solid data and rigorous quality control before adding these medications to my toolkit.

Unless you work in mental health, navigating therapy can feel like swimming in alphabet soup.

Seriously, WTF? There’s CBT, DBT, ACT, ERP, EMDR, IFS, TFP, and every other form of SNAFU.

Now there are situations where the type of therapy really matters. I refer people with severe obsessive-compulsive disorder and anorexia nervosa to specialists at the top of their game. And for people with opioid problems, recommending therapy without offering a referral for medication assessment should be a slam-dunk malpractice case.

But most of the time, the skill of the clinician matters much more than the type of therapy they practice.

Think of therapy like a language. There is no best mother tongue. However, the ability to communicate clearly is not universal.

Likewise, there is no best management style. But you can tell when a leader has a principled, disciplined approach.

There is no simple way to know if a therapist is good. So approach the relationship as if it were a trial. Can they explain in plain English what you’ll do together? Are they a generalist, or do they specialize in the problems you experience? Can they tell you what they don’t treat? Do they respect your goals? Is it clear what you’re working towards, or are you spinning in circles?

Ask for excellence in the professionals you work with. And when things aren’t working, find someone who’s a better fit for you.