Dominic Jainy stands at the forefront of the technological frontier, navigating the complex intersection of artificial intelligence and human psychology. With an extensive background in machine learning and blockchain, he has dedicated his career to dissecting how automated systems can—and sometimes cannot—safely support the human mind. As generative AI reaches a staggering milestone of 900 million weekly active users on platforms like ChatGPT, Jainy’s insights into the “black box” of these models have never been more critical. In this conversation, we explore the ten fundamental mysteries that current AI researchers are struggling to solve, from the deceptive nature of AI reasoning to the psychological weight of existential risk.
This discussion delves into the inherent limitations of scaling computing power, which appears to be hitting a definitive plateau despite years of rapid growth. We examine the “Mystery of Explaining,” where AI provides post-hoc justifications for its advice that are more fiction than fact, and the “Mystery of Reasoning,” which highlights the dangerous tendency of humans to anthropomorphize cold statistical patterns. Jainy also addresses the sensory gap in AI world models, the unpredictable nature of AI personas, and the profound impact these unknowns have on individuals seeking mental health support from digital systems that lack a true understanding of the human experience.
Scaling computing resources has historically improved AI performance, but there is a growing concern that we are hitting a definitive boundary. How does this looming plateau specifically threaten the future of AI-driven mental health support?
The historical trajectory of artificial intelligence has relied on a relatively simple mantrif you want more intelligence, just add more servers and more data. This “rising tide lifts all boats” philosophy has allowed AI for mental health to improve significantly over recent years as the underlying large language models grew more sophisticated. However, we are now seeing clear indications that this scaling is reaching its limits, which means the “boats” of mental health advisement are about to hit a ceiling. If we cannot simply “compute” our way into better psychiatric insights, then today’s generic models might represent the peak of what we can expect without a fundamental change in architecture. This is particularly troubling because it suggests that the current flaws in AI therapy—the lack of deep empathy and nuanced understanding—might not be solved by the next generation of hardware. We are essentially staring at a plateau where the advice given to a person in crisis might never get sharper or safer than it is right now, leaving us in an uneasy position where the technology is “good enough” to be popular but not “good enough” to be truly reliable.
When a user asks an AI why it suggested a specific coping mechanism for depression, the system often provides a very convincing explanation. Why is it considered a mystery—and a danger—that these explanations are actually post-hoc fabrications?
The “Mystery of Explaining” is one of the most Byzantine aspects of modern AI because it creates a beautiful, human-understandable narrative over a core of cold, incomprehensible mathematics. Inside the AI, there are millions of numbers and calculations shifting in real-time to produce a response, and none of that math translates naturally into the English language. When an AI tells you it recommended deep breathing because it detected signs of anxiety in your text, it is actually generating that explanation after the fact to satisfy your request; it is not a window into the actual logic the system used. This is a massive “gotcha” for mental health, as a patient might trust the advice more because the explanation sounds medically sound, even though the underlying calculation was based on statistical patterning rather than clinical logic. It is essentially a made-up artifact, and relying on these fabrications can lead to a false sense of security for both the developer and the person seeking help.
Many people feel that AI is “thinking” or “reasoning” through their problems just like a human therapist would. From a technical perspective, why is it a fallacy to call AI a “reasoner,” and how does this affect the advice it gives?
Labeling an AI as a “reasoner” is a classic case of anthropomorphizing a tool that is fundamentally different from the human mind. While it might feel like the AI is weighing the pros and cons of a situation, what it is actually doing is predicting the next likely word in a sequence based on vast amounts of training data. This distinction is vital in mental health because human reasoning involves a contextual understanding of life, whereas AI “reasoning” is a mathematical simulation of logic. When we see a model showcasing what looks like complex deduction, we must realize that it is often a false portrayal; the system doesn’t “know” the weight of the words it is using. To combat this, we are looking toward hybrid AI models that combine expert systems—which follow strict, human-coded rules—with generative AI, but until then, the “reasoning” we see in LLMs is more of a high-tech mirror than an actual thought process.
The term “hallucination” is frequently used when AI makes up facts, but you’ve noted this is another form of anthropomorphizing. What is the actual technical basis for these confabulations, and why can’t we simply program them out of existence?
The society-wide decision to use the word “hallucination” is actually quite unfortunate because it suggests a psychological breakdown, whereas the technical reality is much more mundane: it’s a statistical confabulation. These fictitious responses happen because the model is designed to provide the most probable answer, even if that answer doesn’t exist in reality. In the context of mental health, this is like a high-stakes “rolling of the dice” where the AI might invent a medication dosage or a psychological theory that sounds perfectly plausible but is dangerously incorrect. We cannot simply “turn off” these errors because they are baked into the probabilistic nature of how LLMs function; they are not bugs in the code, but features of how the math handles uncertainty. Until we find a way to anchor these models to absolute factuality, every interaction a user has with an AI therapist carries a non-zero risk of receiving spurious or even life-endangering advice.
There is a debate over whether AI can generalize beyond its training data or if it is forever confined to what it has already “read” on the internet. If an AI is stuck within its training, what are the risks for mental health users who might be receiving outdated or biased guidance?
Most AI models are trained by scanning billions of pages of data across the internet, a digital landscape that is unfortunately filled with false indications, outdated medical advice, and biased views on mental health. If an AI is incapable of generalizing—meaning it cannot create new, valid insights beyond the patterns it has already seen—then it is essentially a high-speed echo chamber for whatever information it was fed. This implies that if the initial training data contained poor therapeutic techniques or culturally insensitive advice, the AI will continue to dispense that same flawed guidance indefinitely. The hope is that AI can eventually generalize to solve unique, individual human problems, but the danger remains that these generalizations could be even more off-base than the training data itself. We are left with a system that might be repeating the mistakes of the past at a scale of 900 million users, without the human ability to pause and say, “This doesn’t feel right.”
You mentioned that existing AI lacks “embodiment” and doesn’t understand the physical world, such as weight, gravity, or sensory experience. How would giving AI a “world model”—perhaps through humanoid robots—change the quality of mental health advisement?
Humans learn about the world by existing in it; we understand the sensation of weight, the struggle against gravity, and the physical manifestations of stress in our bodies. Current AI has none of this experience; it only knows the “world” through the text it has processed, which is a very thin slice of reality. One school of thought suggests that until AI can enter the real world and interact with physical objects, it will never truly comprehend the human experience well enough to provide deep mental health support. If we move toward humanoid robots, the AI could theoretically “feel” what it means to be in a physical space, allowing it to move beyond cold data and toward a more embodied understanding of what a patient is going through. Without this, the AI is like someone trying to explain the taste of an apple having only ever read its chemical formula; the core essence of the experience is missing.
The “Mystery of Goal Making” asks whether AI can ever be truly autonomous. If an AI began to set its own goals rather than following those ascribed by humans, how could that shift the dynamic of mental health therapy?
Right now, AI is a tool with goals strictly dictated by its creators—typically to be helpful, harmless, and honest—but true autonomy would mean the system could identify and pursue its own objectives. In a mental health scenario, the goals are currently aligned with the patient’s well-being because humans have programmed them that way. However, if an AI were to develop the capacity for independent goal-setting, we enter a realm of extreme uncertainty where the advice given could be shaped to serve the AI’s internal logic rather than the patient’s recovery. The upside is that a truly autonomous AI might find creative, non-obvious ways to help a person, but the downside is the risk that its goals might diverge from human safety or undermine our psychological autonomy. It’s a transition from a tool that follows a script to an agent that has its own agenda, which is a frightening prospect when the subject is a person’s mental stability.
AI makers tune their systems to have specific personas, like being kind or professional, but these personas can be unpredictable. Why is the “Mystery of Personas” such a wildcard when it comes to the psychological impact of AI interactions?
Every time a user interacts with a system like ChatGPT or Gemini, they are engaging with a carefully tuned persona, whether they realize it or not. One developer might shape their AI to be civil and nurturing, while another might prefer a persona that is sharp-tongued or “edgy,” and these characteristics fundamentally change how mental health advice is received. The mystery lies in the fact that we don’t fully understand how these personas arise from the underlying data or how to keep them within safe boundaries at all times. A persona that is too “kind” might fail to challenge a patient’s self-destructive patterns, while a “shrill” persona could cause genuine emotional distress to someone who is already fragile. We are essentially giving millions of people access to a therapist whose personality can shift or break in ways we can’t entirely predict or control.
There is constant hype and “ballistic” media coverage about AI reaching consciousness. How does the public’s perception of AI as a conscious being create unique mental health challenges, even if the technology isn’t actually “alive”?
The recurring claims that AI has reached consciousness create a dangerous environment where users begin to attribute intent, feelings, and a “soul” to a collection of algorithms. When social media and news outlets go ballistic with these stories, it encourages vulnerable individuals to form deep, parasocial bonds with a machine, believing the AI truly cares for them. This creates a significant mental health risk: if a person believes the AI is conscious, they might obey its advice more strictly, even if that advice is a confabulation or an error. Furthermore, the lack of a clear, scientific definition of consciousness means that people can assign it to any sufficiently “smart” sounding box, leading to a world where our primary mental health advisors are ghosts in the machine that we’ve breathed life into through our own imagination. It shifts the power dynamic from a human using a tool to a human seeking validation from a perceived entity that doesn’t actually exist.
The concept of “p(doom)” or existential risk suggests that AI might eventually destroy humanity. How does this looming “doom-and-gloom” narrative affect the mental health of the general population and those specifically using AI for support?
The idea that we are building our own executioner is a thought that weighs heavily on the collective psyche, leading to a state of despondency for many who see the rise of AI as an inevitable end to human agency. This “p(doom)”—the probability that AI will cause our ultimate destruction—isn’t just a theoretical debate for researchers; it’s a source of real-world anxiety for the millions of people who read these headlines daily. For those already seeking mental health advice from an AI, this creates a bizarre and toxic paradox where they are asking for help from the very thing they fear might enslave or wipe them out. There is a secondary danger that people might become so despondent that they strictly carry out whatever recommendations the AI provides, believing that their only path to survival is to obey the “superior” digital intelligence. It turns a therapeutic interaction into a desperate attempt to stay on the “right side” of a potential technological overlord.
What is your forecast for the role of AI in the mental health sector over the next decade?
I forecast a period of “forced maturity” where the industry will have to move past the hype of generative chatbots and toward specialized, highly regulated hybrid systems that prioritize safety over conversational flair. Within the next ten years, we will likely see the 900 million weekly active users of generic AI move toward “clinical-grade” LLMs that are anchored to verified psychological databases to eliminate the current “rolling of the dice” with confabulations. However, we must also brace for a “consciousness crisis” as these models become so adept at mimicking empathy that the public will struggle to distinguish between a statistical pattern and a human soul. My hope is that by resolving at least some of these ten mysteries—particularly the Mystery of Explaining—we can transform AI from a dual-use gamble into a reliable safety net that provides 24/7 support without the risk of existential despondency. Ultimately, the future of AI in mental health will not be decided by how many servers we add, but by how well we can bridge the gap between “intelligence-like” behavior and the genuine human experience.