AI as the Autopilot of Therapy: Assistance, Not Replacement

Young people are already using AI for mental-health support. The question is no longer whether AI belongs in the therapeutic space, but whether it will be governed responsibly.

By Dr Danie Adendorff

Artificial intelligence has entered the mental-health debate through two inadequate extremes. One side speaks as if AI will soon replace the therapist. The other side responds as if any AI involvement in emotional or psychological support is inherently unsafe, dehumanising, or professionally illegitimate.

Both positions are too simple.

The better position is this: AI must assist therapy, not replace accountable therapeutic judgement.

That distinction is essential. Therapy is not merely the exchange of advice. Formal therapy involves diagnosis, treatment planning, regulated professional practice, safeguarding, crisis responsibility, ethical accountability and clinical judgement. AI has not replaced that. It should not be allowed to pretend that it has.

But something else has already happened. AI has become a de facto mental-health support tool. Young people are using chatbots to seek reassurance, ask psychological questions, process distress, interpret relationships, manage anxiety and make sense of emotional pain. This is not yet formal therapy in the clinical sense, but it is undeniably part of the mental-health ecosystem. The professional world may not have invited AI into the room, but patients have already brought it in.

That is why the debate must mature. The real issue is not whether AI is “for” or “against” therapy. The real issue is whether mental-health systems, professional bodies, regulators, schools and families will build a responsible model of AI-assisted support before informal use develops faster than the safeguards around it.

1. AI has not replaced therapy, but it has entered the therapeutic space

Recent research shows that AI use for mental-health advice among young people is no longer marginal. A nationally representative US survey of adolescents and young adults aged 12 to 21, published in JAMA Pediatrics in June 2026, found that 19.2% reported using AI chatbots for mental-health advice.

That finding must be stated carefully. It does not mean that one in five young people are receiving formal AI therapy. The authors distinguish chatbot advice from professional counselling. Formal counselling reflects clinical care. Chatbot use captures a wider and less regulated category of mental-health advice-seeking.

But the distinction should not reduce the importance of the finding. If nearly one in five young people in a nationally representative sample report using AI chatbots for mental-health advice, AI is no longer a hypothetical future presence in mental health. It is already a practical reality.

The trajectory is as important as the headline figure. An earlier nationally representative study found that 13.1% of US adolescents and young adults had used generative AI for mental-health advice. The later figure of 19.2% represents a substantial increase in less than a year. That growth does not prove clinical replacement, but it does show rapid normalisation of AI as an informal mental-health advice channel.

This should also be understood within a broader youth AI-use environment. Pew Research Center’s 2026 survey on how teens use and view AI found that chatbot use among US teenagers had become common across general advice, schoolwork, entertainment and emotional-support contexts. The Pew figure for emotional support is not directly equivalent to the McBain mental-health-advice measure, but it confirms the wider behavioural setting: young people are already incorporating AI chatbots into ordinary information-seeking and support-seeking routines.

The same JAMA Pediatrics study also found that nearly two-thirds — 63% — of young people who had used AI chatbots for mental-health advice had not disclosed that use to anyone. This is strategically important. It suggests that AI mental-health support may be developing as a partially hidden parallel system: used by young people, valued by many of them, but often invisible to parents, teachers, clinicians and regulators.

The phrase “shadow system” should not be used carelessly. One survey does not prove the full existence of a mature parallel mental-health infrastructure. But it does show a serious warning indicator: a substantial proportion of young users are seeking mental-health advice from AI systems without telling the adults or professionals who might otherwise help them interpret, challenge or escalate that advice.

That changes the public question. The issue is no longer whether AI should be allowed into mental health. It is already there. The issue is whether this development will remain informal, unobserved and inconsistently governed, or whether it will be brought into a responsible framework of disclosure, escalation, clinical governance, safety testing and accountability.

Professional denial is not a strategy. If young people are already consulting chatbots before speaking to parents, doctors, teachers, counsellors or therapists, then mental-health systems must adapt. The correct response is not panic. It is governed integration.

2. The autopilot principle

The most useful analogy is aviation.

Autopilot did not abolish the pilot. It assists the pilot. It helps with navigation, stability, monitoring, workload reduction, route management and routine control. In modern aviation, automation can make flight safer, but only when it remains under trained human command. The pilot must still understand the aircraft, monitor the system, intervene during abnormal conditions and retain final responsibility for judgement and command.

Therapy should approach AI in the same way.

AI should not become the captain of the therapeutic aircraft. It should be treated as an assistive system. It may help with screening, structured questioning, symptom tracking, psychoeducation, between-session reflection, risk-language detection, treatment adherence, continuity of notes and alternative case formulations. It may help the therapist think more systematically and help the patient communicate more clearly.

But the therapist remains the accountable professional. The therapist must retain responsibility for judgement, safeguarding, ethical decision-making, crisis response, human context and clinical interpretation.

The danger in aviation is not autopilot itself. The danger is automation dependency, degraded human skill and loss of situational awareness. The same applies to therapy. A clinician who accepts AI output uncritically may miss the patient. A patient who treats a chatbot as a therapist may mistake fluency for competence, reassurance for care and validation for treatment.

This is also why the autopilot analogy leads directly to governance. Autopilot is not simply installed and trusted. It is certified, monitored, bounded, trained for, checked against failure modes and operated within a command structure. AI-assisted therapy requires the same discipline.

The lesson is clear: assistance is not command.

3. What the empirical research actually supports

The strongest empirical evidence does not support a simplistic claim that AI is “better than therapists” in any general sense. It supports a more careful conclusion: AI has promising mental-health applications, but its performance is task-dependent, context-dependent and governance-dependent.

A 2025 systematic review and meta-analysis in Digital Health reported strong pooled diagnostic accuracy and therapeutic efficacy across included AI psychiatry studies. That is important evidence of potential. However, pooled results do not prove that AI can assume autonomous clinical responsibility. The performance of an AI system depends on the disorder, dataset, method, patient population, study design, and whether the system is being used for screening, diagnosis, monitoring, treatment support or conversational advice.

The source also requires careful handling. It is peer-reviewed and indexed, but it should not carry more weight than it can bear. It brings together heterogeneous studies from a rapidly changing field, and its findings should be used as a signal of promise rather than as proof of deployment readiness.

Other recent research comparing large language models with mental-health professionals also shows both promise and limitation. Levkovich’s 2025 text-vignette-based study examined how large language models performed on structured mental-health scenarios compared with mental-health professionals. The results were encouraging in some areas but variable across conditions. That variation is the point. A model can perform impressively on a written vignette and still not be ready to handle the ambiguity, emotion, silence, risk and context of real clinical practice.

Structured vignettes are not patients. Real patients are inconsistent, distressed, ashamed, defensive, confused, silent, frightened, intoxicated, traumatised, manipulative or in crisis. Clinical practice involves much more than producing a plausible answer to a written prompt.

More directly relevant to young people, Feng et al.’s 2025 systematic review and meta-analysis of randomised controlled trials found that AI chatbots showed small-to-moderate effects in mitigating mental distress, with a standardised mean difference of −0.35, and smaller effects in promoting health behaviours, with a standardised mean difference of 0.11, across 31 trials involving nearly 30,000 participants. That is meaningful evidence of potential benefit. It is not evidence of replacement. It supports the narrower and more defensible conclusion that AI chatbots may help reduce certain forms of distress or improve some behaviours under defined study conditions.

This is why the evidence should be interpreted as support for AI-assisted clinical judgement, not AI replacement. The correct empirical conclusion is not that AI is superior to therapists. The correct conclusion is that AI can strengthen parts of the therapeutic process if used as a governed support layer.

4. What AI can contribute

AI has several potential strengths in mental-health support.

First, AI can provide availability. Human therapists are scarce, expensive, unevenly distributed and often inaccessible. Many people who need help do not receive it. AI cannot solve that problem alone, but it can provide a first layer of support, information and structured reflection while users wait for or supplement human care.

Second, AI can support consistency. A human clinician may be tired, overloaded, distracted or influenced by the emotional pressure of a session. AI can help maintain structured checklists, compare symptom patterns and prompt consideration of overlooked alternatives.

Third, AI can support pattern recognition. It can track language, mood, recurring themes, behavioural changes and inconsistencies over time. Used responsibly, this may improve continuity between sessions and help identify early warning indicators.

Fourth, AI can support psychoeducation. Many patients need clear explanations of anxiety, depression, trauma responses, sleep, grief, stress, addiction, emotional regulation and relationship patterns. AI can make basic mental-health information more accessible, provided it is accurate, carefully bounded and does not pretend to diagnose.

Fifth, AI can function as a second reasoning layer. It can ask: What information is missing? What alternative diagnosis should be considered? What risk indicators appear in the language? What should be escalated? What collateral information would be needed before a firm judgement is made?

This is where AI is most valuable: not as a replacement therapist, but as an intelligent assistant to clinical judgement.

5. What AI cannot safely own

The case for AI assistance must not become technological worship.

Mental health is not only a pattern-recognition problem. A person is not a text file. A patient may communicate through silence, posture, agitation, hygiene, eye contact, psychomotor change, dissociation, smell of alcohol, family dynamics, financial distress, medication effects, cultural context, legal pressure or fear of consequences. A chatbot does not see all of that unless the information is supplied.

AI also lacks clinical accountability. If a chatbot mishandles a suicidal disclosure, reinforces delusional thinking, misreads coercive control, or reassures a user who needs urgent help, who is responsible? The developer? The platform? The clinician who recommended the tool? The institution that failed to regulate it?

Those questions cannot be left hanging indefinitely. Responsibility should attach to the party that designs, deploys, recommends, integrates or relies on the system. Developers should be accountable for foreseeable design failures, misleading claims, inadequate safety testing and weak crisis escalation. Health providers should be accountable if they integrate AI into care without appropriate validation, disclosure, professional oversight and clinical escalation. Institutions should be accountable if they present AI support as safe, therapeutic or professionally supervised without proving that the necessary safeguards exist. Regulators should require clarity about whether a system is wellness support, clinical decision support, medical software or part of regulated care.

There is also the problem of crisis. Therapy is not only supportive conversation. It includes risk assessment, safeguarding, escalation, emergency referral and sometimes firm boundary-setting. A system designed to be agreeable may over-validate harmful thinking. A system designed to avoid risk may disengage precisely when the user needs skilled intervention. A system trained to be helpful may provide advice where it should refer.

Recent safety research strengthens this caution. Moore et al.’s 2025 ACM FAccT paper, Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers, found that large language models can express stigma and provide inappropriate responses in mental-health contexts. This is not a reason to reject AI assistance. It is a reason to reject ungoverned substitution.

Clinical judgement cannot be reduced to confident language.

6. The human element still matters

A strong case for AI assistance must also recognise why human therapists remain indispensable.

Therapeutic alliance is not sentimental language. It is one of the most consistently studied factors associated with psychotherapy outcome. Trust, rapport, repair after misunderstanding, emotional attunement and the patient’s belief that another human being is responsibly present all matter.

A good therapist is not simply a provider of information. A good therapist interprets the whole person in context. The therapist listens not only to what is said, but also to what is avoided, contradicted, embodied, repeated or displaced. The therapist notices timing, affect, relational patterns, deflection, shame, fear, anger, dissociation, dependency and silence.

This evidence has limits. Much therapeutic-alliance research concerns established therapeutic relationships, not every form of crisis intervention, brief support, digital self-help or early advice-seeking. AI may therefore have its most plausible role precisely where formal alliance has not yet been established: early screening, between-session support, psychoeducation, triage, reflection and monitoring.

That qualification matters. The argument for human oversight does not depend on human infallibility. Therapists can make mistakes. They can miss risk. They can be influenced by fatigue, ideology, institutional pressure, sympathy, frustration or overconfidence. The case for human accountability is not that humans are perfect. It is that clinical responsibility, ethical judgement and safeguarding must remain located in accountable human institutions.

The future should not be human judgement alone or machine judgement alone. It should be assisted human judgement.

7. The manipulation problem

There is another uncomfortable issue: patients can manipulate therapists.

This does not mean patients are generally malicious. It means human communication is selective. Patients may omit, exaggerate, minimise, rehearse, perform, charm, intimidate, flatter, or present a version of themselves shaped by fear, shame, strategy, trauma, litigation, workplace pressure, family conflict or self-protection.

Some intelligent or highly self-aware patients may manipulate the human therapeutic process more effectively than others. A therapist may be influenced by emotional performance, social status, ideological cues, dependency, anger, sophistication of language, or the desire to be experienced as caring and non-judgemental. In forensic, organisational, insurance, child-protection and high-conflict contexts, this problem becomes especially serious.

AI may be less vulnerable to some forms of interpersonal manipulation. It does not need to be liked. It does not become intimidated in the ordinary human sense. It does not respond to status in the way a person might. It can be designed to preserve uncertainty, ask for corroboration and resist premature closure.

But AI can also be manipulated. It can be given false facts. It can be led through selective prompting. It can be trapped inside the user’s narrative. It may have no independent access to body language, collateral accounts, medical records, medication history, substance use, domestic circumstances or the broader reality of the patient’s life.

This is the bridge between the human case and the governance case. Human therapists may see the person, but can be socially influenced by the person. AI may see the pattern, but can be trapped inside the data it is given. Each has a different vulnerability profile. That is exactly why the answer is not therapist alone or AI alone.

The answer is a combined model: human accountability supported by machine-assisted structure, with governance strong enough to prevent either side from becoming the unexamined authority.

8. What governed AI-assisted therapy should look like

A responsible model for AI-assisted therapy would require clear boundaries.

1. Mandatory disclosure.

Patients should know when AI is being used, what it is being used for, and what it is not capable of doing. A system should not simulate professional authority while hiding its non-human status or its limitations.

2. Assistance, not autonomous decision.

AI may assist with screening, monitoring, documentation, risk flags and structured reflection, but final clinical judgement must remain with a qualified professional where clinical care is being provided.

3. Explicit escalation thresholds.

Systems used in mental-health contexts must know when to escalate to human care, crisis lines, emergency services, safeguarding procedures or clinical review. Escalation rules should be tested, not merely asserted.

4. Audit trails and reviewability.

If AI influences clinical reasoning, there should be a record of what it suggested, what evidence it used, and how the clinician accepted, rejected or modified the recommendation.

5. Strict data protection.

Mental-health conversations are among the most sensitive forms of personal data. Users must not be turned into training material without clear, informed and meaningful consent.

6. Failure-mode testing.

AI systems must be tested against real clinical failure modes: self-harm, psychosis, coercive control, abuse, addiction, eating disorders, obsessive rumination, delusional reinforcement and manipulative presentation.

7. Professional training and accountability.

Professional bodies must train clinicians to use AI intelligently. The therapist of the future should not be replaced by AI. But the therapist who refuses to understand AI may become less equipped to understand the patient who is already using it.

This framework does not require hostility to AI. It requires seriousness. A mental-health AI system should not be judged only by whether it produces comforting language. It should be judged by whether it improves support, preserves accountability, detects risk, respects privacy, avoids dependency and escalates when a human professional must take responsibility.

9. The real future: assisted clinical judgement

The future of mental health should not be framed as therapist versus machine. That is the wrong debate.

The more serious debate is between ungoverned informal AI support and governed AI-assisted clinical care. The first already exists. The second must now be built.

AI will not remove the need for therapeutic alliance, ethical responsibility, crisis judgement, safeguarding or human context. But neither can the profession ignore the fact that AI is already being used by patients as a source of mental-health advice. The task is therefore not to worship the technology or banish it from serious discussion. The task is to govern it.

The autopilot principle gives us the right model. Autopilot assists the pilot, but it does not hold command responsibility. AI should assist therapy, but it must not replace accountable therapeutic judgement.

That is the balanced position. It is also the practical one.

Mental-health care should not become machine-led. But it should become better assisted, better monitored, better supported and more honest about the reality that patients are already using AI. The professional choice is no longer whether AI enters the therapeutic space. It has entered.

The remaining question is whether therapists, regulators, schools, parents and health systems will integrate it responsibly before informal use grows faster than the safeguards.

Sources and further reading

McBain, R. K. et al. (2026). AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults. JAMA Pediatrics. Published online 1 June 2026. DOI: 10.1001/jamapediatrics.2026.2015.

McBain, R. K. et al. (2025). Use of Generative AI for Mental Health Advice Among US Adolescents and Young Adults. JAMA Network Open, 8(11), e2542281.

McClain, C., Anderson, M., Sidoti, O. and Bishop, W. (2026). How Teens Use and View AI. Pew Research Center.

Rony, M. K. K. et al. (2025). The role of artificial intelligence in diagnosis and therapeutic strategies for psychiatry: A systematic review and meta-analysis. Digital Health. DOI: 10.1177/20552076251330528.

Feng, X. et al. (2025). The Effectiveness of AI Chatbots in Alleviating Mental Distress and Improving Health Behaviors Among Adolescents and Young Adults: Systematic Review and Meta-Analysis of Randomized Controlled Trials. Journal of Medical Internet Research. DOI: 10.2196/79850.

Levkovich, I. (2025). Evaluating diagnostic accuracy and treatment efficacy in mental health: A comparison of large language models and mental health professionals using text vignettes. European Journal of Investigation in Health, Psychology and Education. DOI: 10.3390/ejihpe15010009.

Flückiger, C., Del Re, A. C., Wampold, B. E. and Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340. DOI: 10.1037/pst0000172.

Moore, J., Grabb, D., Agnew, W., Klyman, K., Chancellor, S., Ong, D. C. and Haber, N. (2025). Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25), 599–627. DOI: 10.1145/3715275.3732039. https://doi.org/10.1145/3715275.3732039.

Author workflow disclosure

This article was produced through an AI-assisted but human-directed workflow. AI support was used for accessibility assistance, structuring, language refinement, source-discovery prompts, revision planning, and conversion of editorial comments into amendments. Dr Danie Adendorff retained responsibility for the argument, accepted or rejected changes, checked the logic of claims, assessed source credibility, and remained accountable for the final text. AI-generated material was not treated as empirical evidence, and synthetic or illustrative examples were not presented as observed data.

Image note

The image accompanying this article is AI-generated and is intended for illustration purposes only.

CONTACT

© 2026 Dr Danie Adendorff. All rights reserved. Rumbls.com is an independent analytical blog.

TERMS AND CONDITIONS

DISCLAIMER

COOKIE POLICY

PRIVACY POLICY