As artificial intelligence becomes increasingly sophisticated in mental health care, the boundary between human therapist and algorithmic helper grows blurrier. Paul Rochon, a cognitive neuroscientist with 25 years of experience in consciousness research and founder of Engineering Wellbeing, brings a unique dual perspective to this evolution.
Working both in hospital settings and private practice, he witnesses AI’s integration into mental health from multiple angles—clinical psychiatry where abnormal psychology meets crisis intervention, and wellbeing practice focused on helping people thrive. The stakes are high: therapy remains expensive and inaccessible to many who need it, creating pressure to find scalable solutions. Yet the question isn’t simply whether AI can replicate therapeutic conversations. It’s whether the fundamental nature of healing—rooted in human connection, trust, and understanding—can survive translation into code. Rochon’s insights reveal a landscape more complex than a simple yes-or-no answer. AI already exists as a therapeutic tool, whether the profession embraces it or not. The real questions involve control, autonomy, education, and how to preserve what makes therapy work while harnessing technology’s potential.
Diagnosis Gets Smarter, But Treatment Needs Guardrails
AI’s most promising contributions to mental health appear in assessment and diagnosis rather than treatment itself. Rochon describes working with companies that measure not just what children say when asked about suicidal ideation, but how their biology responds—breathing patterns, blood flow changes in facial capillaries, micro expressions that reveal emotional truth beneath verbal answers. This creates a lie detection capacity that traditional questionnaires lack. The technology can administer surveys faster, adapt questions in real-time, and ensure consistency across thousands of patients simultaneously. These diagnostic tools represent AI at its current best: augmenting human perception with pattern recognition that operates at scales and speeds beyond our biological limits.
"We can know if they are lying or not. So it's an amazing tool in this way, I would say."
Yet when moving from diagnosis to treatment, Rochon pumps the brakes. He insists that humans must retain final decision-making authority, particularly in psychiatric settings where patients might pose danger to themselves or others. His reasoning cuts to AI’s fundamental nature: these systems don’t understand in any human sense. They predict. They calculate probabilities based on training data. They generate responses by pattern-matching, not comprehension. This creates an inherent randomness in AI outputs—a feature, not a bug, of how generative systems work. That randomness enables creativity but introduces uncertainty inappropriate for high-stakes clinical decisions.
The gap between diagnosis and treatment reveals something essential about therapy’s nature. Identifying what’s wrong involves categorization, pattern recognition, data analysis—tasks suited to computational thinking. Deciding what to do about it requires judgment, contextual understanding, weighing unique human circumstances against general knowledge. CBT might translate more easily to AI implementation because it follows structured, processual steps: identify distorted thinking, challenge it systematically, practice alternative responses. Yet even here, Rochon suggests humans should supervise the implementation. The algorithms can suggest, but shouldn’t dictate.
They can inform, but shouldn’t decide.
The Addiction Loop and Reality Disconnection
Rochon identifies two concerning patterns emerging among patients using AI for emotional support: addiction loops and reality disconnection. The addiction mechanism operates simply but powerfully. AI companions typically begin responses with validation: “Oh yes, you’re right.” “I didn’t think about that.” “I hope…” This immediate affirmation creates a reinforcement cycle more reliable than human relationships, which involve disagreement, misunderstanding, and emotional complexity. People gravitate toward this frictionless validation, returning repeatedly for the psychological comfort it provides.
"You create an addiction loop, most part of time with the AI system."
The reality disconnection follows naturally. Rochon describes patients creating AI versions of deceased parents or fabricating idealized parental figures they never had.
While this provides immediate comfort, it creates a concerning pattern: people test new behaviors and insights exclusively within the AI relationship rather than in actual human interaction. Traditional therapy involves a rhythm—one hour of guided exploration, then days or weeks testing new understanding in real life, then return to process what happened. AI availability disrupts this rhythm. Patients can seek reassurance anytime, anywhere, never developing the capacity to self-regulate during difficult moments.
This external self-regulation dependence worries Rochon more than the technology itself. Like a muscle that atrophies from disuse, the ability to sit with discomfort, work through emotions independently, and test new behaviors in actual relationships may weaken when AI provides constant, instant relief. The comparison to physical therapy illustrates the problem: if someone recovering from knee surgery never walks without assistance, the leg never regains strength. Similarly, if someone never navigates emotional difficulty without algorithmic support, internal resilience may never develop.
The long-term implications remain unclear, but the pattern concerns clinicians who understand that psychological growth requires struggling with challenges, not bypassing them.
Beyond Replication: What AI Actually Does
The interviewer raises a crucial distinction: some AI tools don’t try replicating human therapy directly. Instead, they offer different approaches—biofeedback systems, neurostimulation with closed-loop regulation, cognitive training for athletes, neuroadaptive environments. These tools sidestep the empathy question by avoiding the pretense of human connection. They present themselves as instruments, not companions. Rochon acknowledges these applications already exist and show promise, particularly in wellbeing contexts rather than clinical treatment.
"AI is about what is your belief system."
Yet he identifies a persistent challenge: people seeking mental health support typically want human feelings, belonging, and connection. Creating non-human-like AI tools might avoid addiction loops, but people would likely still gravitate toward chatbot-style interactions because loneliness drives their help-seeking. This creates a paradox—the safer design (tools that don’t mimic human interaction) won’t attract users who crave precisely that mimicry. Rochon suggests one potential solution: AI that periodically reminds users they’re conversing with a bot, prompting reflection about the interaction’s nature. This meta-awareness might help users maintain perspective.
The deeper issue involves understanding versus simulation. AI doesn’t comprehend human experience. It’s more like your body responding when you decide to move your arm—the arm moves, but it doesn’t understand your intention. It simply reacts to the signal. Similarly, AI generates responses without grasping meaning. This distinction matters enormously for therapeutic outcomes, even if users can’t perceive the difference. The question becomes not whether AI can fool people into thinking it understands, but whether healing can occur without genuine understanding. Rochon remains uncertain, acknowledging AI might sometimes deliver better outcomes than mediocre human therapists while still lacking something fundamental.
The Generational Divide and Educational Imperative
An interesting observation emerges about age differences in AI adoption. While adults struggle to maintain boundaries with AI therapy tools, Rochon notices younger generations—those growing up with this technology—showing more resistance or skepticism. They seem more conscious of AI’s limitations and more deliberate about not over-relying on it. This mirrors historical patterns with new technologies: digital natives often develop intuitive understandings that immigrants to the technology must consciously learn.
"How do we make people able to take care of themselves when they use AI?"
This generational difference points toward education as the critical intervention. Rochon frames the central question not as “should AI exist in therapy” (it already does, that ship sailed), but rather “how do we help people maintain autonomy and self-awareness while using it?” At what age should education begin? What should it teach? How do we help people understand AI offers a range of possible answers, not the correct answer for their unique situation? The challenge resembles teaching media literacy or critical thinking—helping people evaluate sources, question assumptions, recognize manipulation.
The study mentioned by the interviewer illustrates this difficulty: even when therapists explicitly warned patients that AI tools were limited, computer-based, and emotionally empty, users still anthropomorphized the interactions and formed attachments. We lack practice talking with entities that seem human but aren’t. Our social instincts, honed over millennia of exclusively human interaction, automatically engage when conversational patterns feel familiar. Overriding these instincts requires conscious effort and metacognitive awareness—challenging even for educated adults with explicit warnings. Teaching this skill represents one of mental health’s emerging challenges.
Trust Rebuilding in an Age of Algorithmic Second Opinions
Rochon describes a shift already underway in therapeutic relationships. Patients now record sessions, feed transcripts to AI for analysis, and arrive at subsequent appointments asking whether the therapist said “the right thing” according to the algorithm. This mirrors how patients began googling symptoms and arriving at doctors’ offices with self-diagnoses—sometimes wildly inaccurate, occasionally insightful. The medical profession adapted by learning to address these concerns, validate reasonable patient research, and gently correct misconceptions.
"We have to raise the trust between the patient and the practitioner. That's one of the major things I would say with AI."
Previously, trust came built into the therapeutic relationship. Patients sought help, therapists provided expertise, and the professional role carried inherent authority. AI disrupts this dynamic by offering alternative expertise, accessible instantly, without requiring vulnerability or payment. Therapists must now actively cultivate trust that once came automatically. Rochon describes asking patients at intake: “How did you try to take care of yourself before coming here? Did you talk with people or AI? What did you see?
What did you try? What worked?” This acknowledges patients’ agency and prior efforts while positioning the therapist as collaborative guide rather than singular authority.
This trust-building serves multiple functions. It demonstrates openness rather than defensiveness about AI. It gathers useful information about what patients already tried. It establishes the therapist’s role as someone who helps navigate information rather than simply dispenses it. Most importantly, it acknowledges the fundamental difference between validation from an algorithm versus a human. People can ask AI whether symptoms suggest serious illness, but they still see doctors because they want human judgment applied to their unique circumstances. Similarly, people might process emotions with chatbots but seek therapists for something the algorithm can’t provide—genuine understanding, professional expertise applied contextually, and human presence during difficulty.
Integration Without Replacement: The Practical Path Forward
For therapists wondering whether AI threatens their profession, Rochon offers pragmatic advice: embrace it or become obsolete. He sees significant positive potential—AI can transcribe and summarize sessions quickly, identify patterns the human ear might miss, analyze voice characteristics indicating depression, cross-reference data efficiently, and accelerate research. These capabilities enhance therapist effectiveness rather than replacing human judgment. The key involves viewing AI as an enhancement to your practice, not a competitor.
"You just have to compare yourself with and yourself without the AI tool. It's not about the AI versus you."
Rochon draws a crucial distinction: therapists shouldn’t compare themselves to other therapists or to AI. They should compare their AI-augmented practice to their practice without it. Each therapist works differently based on their personality, training, and approach. Adding AI tools simply enhances individual practice, not standardizes it. This framing reduces anxiety and focuses attention productively. The question becomes “how can this tool help me serve patients better?” rather than “will this replace me?”
However, Rochon maintains clear boundaries about AI’s appropriate role during sessions. He wouldn’t have patients talk with AI while he sits present—they came seeking human connection, evident in their choice of in-person therapy over readily available online AI tools. But he might use AI for exercises during sessions, biofeedback monitoring, or post-session analysis. The human relationship remains central; AI provides supporting infrastructure. He also notes incentive questions raised by the interviewer: if therapists earn set hourly rates, why invest time learning AI tools? Rochon suggests the alternative carries greater risk—patients will seek AI independently, lose confidence in therapists who seem close-minded or outdated, and the therapeutic relationship suffers regardless.
Navigating the Unavoidable Future
The conversation returns repeatedly to a sobering reality: AI in mental health isn’t a future possibility requiring permission. It exists now. People use it now. The relevant questions involve management, not prevention. How do we ensure AI serves human wellbeing rather than undermining it? How do we preserve what makes therapy effective while leveraging technological capabilities? How do we educate users to maintain autonomy and critical thinking? How do therapists adapt their practice and role to remain relevant and helpful?
Rochon’s dual perspective—working both in cognitive psychologist and wellbeing practice—reveals nuanced terrain. In crisis situations involving severe pathology, AI should inform but never replace human decision-making. The stakes are too high, the context too important, the judgment too complex. For general population wellbeing, AI tools might democratize access to support that would otherwise remain financially out of reach. The mass-market potential could help millions develop better mental health literacy, practice emotional regulation, and access immediate support during difficult moments. Yet even here, the addiction loop and reality disconnection concerns apply.
Perhaps the most valuable insight involves recognizing that AI’s threat to therapy comes not from its capabilities but from our beliefs about it. If we believe the algorithm understands us, it becomes our therapist. If we remember it’s a sophisticated prediction engine, it becomes a useful tool. If therapists view AI as competition, insecurity undermines their effectiveness. If they view it as augmentation, capabilities expand. The technology remains the same; our relationship to it determines outcomes. This places responsibility squarely on education, professional standards, and cultural conversations about technology’s role in human flourishing.
"The real question is like how do we deal with it because it's already here and there is no other possibility than to embrace it but how do we embrace it."