Would You Have a Relationship with a Robot? AI isn’t Therapy

Shirley Katz, Ph.D., RP, CCC, Clinic Director

Use of AI is rapidly growing and working its way into conversations about mental health.  Clients often tell us they use chatbots to support their work in therapy or compare it to their therapy.  At our practice, we do not use AI to generate ideas or treatment suggestions.  Some therapists may use it with client permission to write their notes, but only a confidential, legally compliant health related system.  Most of us prefer to use the writing process to reflect on our therapeutic relationship, goals and progress and to make plans for follow up sessions. This also consolidates memory.  Its how we spend more time processing a case even when we are not client facing.  Overall, we discourage the use of AI tools to substitute or compare psychotherapy experience.  This is not based on fear of technology but more so about ethics, regulatory responsibility and what the science actually shows about the topic.  There are good reasons.
Psychotherapy is a regulated health profession, not a wellness or self-help tool like a workbook.  Therapy is relational and interactive.  It is also controlled by the College of Registered Psychotherapists of Ontario (CRPO) and also has standards related to client safety, informed consent, confidentiality and privacy, competence, scope of practice and limits, crisis and risk assessment and intervention, accountability and professional liability.  AI systems are not regulated professionals.  These systems cannot hold licensure, be held accountable for competency, or be disciplined for ethical or privacy breaches.  They are not human and do not offer attuned health attachment-based experiences.  Therapeutic relationships are the cornerstone of mental health treatment.   
Research on the topic has shown that when AI systems take on therapist‑type roles, they operate outside existing health‑care regulation demonstrating serious gaps in protection for clients.  Studies further found that even when AI was asked to follow evidence-based models it violated core ethical standards showing problems with boundaries, inappropriate reassurance and failures in handling potential risk or crisis.  Licensed clinicians are required to practice in ways that protect the public.  Offering or endorsing tools that have known ethical problems does not meet that obligation.  Therapists at our clinic would likely ask clients to refrain in using these tools. 
Another issue that comes up with the use of AI actually pre-dates it.  Historically, there has always been concern about clients seeking therapy from multiple, non coordinated sources. This emerged from an ethical principle called nonmaleficence (do no harm) and supported by ethics of fidelity, integrity and responsibility. When client care is uncoordinated, they can get conflicting input leading to potential confusion or overwhelm which can undermine the effectiveness of treatment.  At the very least, if clients are seeking input from AI they should be discussing how it is impacting them and ensure that it is coordinated information, which supports reflection or skills practice in line with the treatment goals of therapy and not harmful, confusing or contradictory.
It is essential to remember that AI systems do not do well at identifying and managing risk.  Crisis care requires human accountability and judgement.  AI is inconsistent and unreliable in this regard and does not escalate proper supports. Conversely, a trained therapist knows how to assess risk over time, due to the emphasis on alliance and an ongoing therapy relationship, noting any subtle shifts or incongruence for example. They are obligated to respond and accountable if harm occurs and was mismanaged.  For example, therapists have a duty‑to‑warn when there is impending harm, a duty‑to‑protect, a duty to report.  Emergency response protocols are not optional.
Therapy happens by a real human relationship that requires tuning in, building rapport, assessment, repair of rapport if there has been a rift, reflection on that relationship strength, embodied presence over time, genuineness, etc.  It’s a central mechanism of change.  Most clients have experienced their hurts in the context of relationships, and healing and relearning happens in the context of a human relationship.  The relationship is central to the treatment in all modes of therapy.  This is evidence based.  Conversely, research has demonstrated that AI’s empathy is not relational, it’s performance. Clients can feel comforted in the moment or reassured, while being misunderstood or misguided in meaningful ways that impact how they cope or respond in human circumstances.  This can further create a problem if people do not access real care or if they have trauma, as they can mistake an AI response as a replacement for real human relational witnessing and safety.  
Further problems include how these systems are based on data generated by humans and reflect bias, such as demonstrating greater stigma towards certain diagnoses.  For clients who are marginalized this can be harmful.  In regulated practice ethics emphasize justice, equity, inclusion and cultural safety. 
Overall, the evidence base for the utility of AI in mental health is very limited and large, peer reviewed studies conclude that it is not supported as a replacement or adjust for professional, human relationship-based therapy.  It is also not confidential.  AI systems do not meet Canadian consent and privacy standards for mental health data use and storage.
Our position on AI is that it may have some use in administrative support when used cautiously and with human oversight.  Psychotherapy is human, relational, ethical and regulated.  AI systems currently do not offer that.   Choose therapy not AI.  Human healing happens in a relationship with all its potential imperfections.  Change that happens in such relationships reflects what it is to be human.  If you want to discuss how your use of AI may fit into your mental health plan or progress please ask your therapist. You are not a robot, and neither are we.