The Rise of AI Therapy: Innovation, Access and the Need for Human Oversight
The Rise of AI Therapy: Innovation, Access and the Need for Human Oversight
Artificial intelligence is moving into one of the most sensitive areas of human life: mental health.
People are already using AI chatbots to talk through anxiety, loneliness, relationship problems, stress and difficult emotions. At the same time, researchers are testing AI-assisted psychological interventions, clinicians are experimenting with AI tools in training and assessment, and professional bodies are warning that these systems also carry real risks.
The result is a difficult but important question:
Can AI improve access to mental-health support without replacing the judgement, empathy and responsibility of trained professionals?
The latest research suggests that AI may have a useful role — but the strongest evidence points toward AI as a support tool, not a substitute for human care.
AI therapy is no longer theoretical
AI-powered mental-health tools have existed for several years, but generative AI has changed the conversation.
Modern systems can respond conversationally, remember context within an interaction, adapt their language and provide answers that feel personal.
That makes them much more appealing to users seeking immediate emotional support.
The American Psychological Association reported this month that more than three-quarters of psychologists say their patients are now discussing AI in therapy. People are using chatbots to self-diagnose, seek emotional advice and manage relationships. The APA’s position remains clear: AI-generated advice should not be treated as a safe or effective substitute for a qualified mental-health professional. (American Psychological Association)
This is an important development.
AI is no longer sitting outside the therapy room.
It is increasingly becoming part of the conversations happening inside it.
There is evidence that AI can help in specific situations
The picture is not entirely negative.
One of the newest studies, published in npj Digital Medicine on 12 September 2026, tested a hybrid human–generative-AI intervention for academic anxiety.
Researchers studied 425 participants across two randomized controlled trials. The intervention combined human-designed therapeutic structure with AI-supported exposure exercises.
The results were encouraging.
Compared with control groups, participants generally showed greater improvement in immediate anxiety as well as academic anxiety and procrastination two weeks later. Importantly, the researchers described the intervention as acceptable and perceived as safe by participants. (Nature)
But there is a crucial detail.
This was not an unrestricted chatbot simply acting as a therapist.
It was a structured human–AI intervention designed around a specific therapeutic method.
That distinction may prove central to the future of AI in mental health.
AI may work best when the boundaries are clear
Another 2026 randomized clinical trial published in JAMA Network Open examined an AI-based emotional support platform involving nearly 1,000 participants.
The study compared an AI platform, face-to-face group therapy and a control group.
The researchers found improvements in several measures among users of the AI platform, including anxiety, depression, life satisfaction and well-being, although outcomes varied depending on the condition being measured. (JAMA Network)
These findings are significant because they show that conversational AI can sometimes provide measurable psychological benefit.
However, one successful platform or study does not mean that every general-purpose chatbot can safely deliver therapy.
Mental-health interventions need to be evaluated individually.
The question should not simply be “Does AI therapy work?”
It should be:
Which AI system? For which problem? For which person? Under what level of supervision?
Access is one of AI's strongest arguments
One reason AI mental-health tools are attracting so much attention is access.
Across many countries, people face:
long waiting lists;
- high therapy costs;
shortages of trained professionals;
geographic barriers;
stigma around seeking help;
limited after-hours support.
An AI system can theoretically be available at any hour, from almost anywhere.
It can also offer people a private space to organise their thoughts before speaking with another person.
For someone experiencing mild stress or preparing for a therapy appointment, that could be genuinely useful.
AI may also help clinicians by handling administrative or supportive tasks rather than taking over therapy itself.
The APA has highlighted emerging uses of AI in clinician training, where chatbots can help trainees practise skills and receive additional feedback when supervision time is limited. The emphasis, however, remains on enhancing rather than replacing clinical supervision. (American Psychological Association)
Psychology is beginning to shape AI itself
An important development happened on 3 September 2026, when the American Psychological Association appointed the inaugural executive director of its new Center for Behavioral Science and AI.
The purpose of the centre is to ensure that psychological science helps influence how AI systems are designed, used and governed. (American Psychological Association)
That development says a great deal about where the field is heading.
Psychologists are no longer only studying how people use technology.
They increasingly need to help technology companies understand:
human behaviour;
emotional vulnerability;
persuasion;
attachment;
Trust;
cognitive bias;
social development;
psychological safety.
If an AI system is designed to hold deeply personal conversations with millions of people, psychology becomes part of its safety architecture.
The biggest danger may be that AI sounds too human
One of the unusual qualities of generative AI is its ability to sound supportive.
It can respond patiently.
It does not become visibly frustrated.
It is available immediately.
And it can produce language that sounds understanding and emotionally validating.
Those qualities can be helpful.
But they can also become a risk.
A user may begin to assume that because the system sounds empathetic, it actually understands their situation in the way a therapist does.
It does not.
AI does not observe body language in the same way a clinician does.
It may not understand a person's family history, culture, medical background or previous treatment.
It can also misunderstand context or generate incorrect information with considerable confidence.
That gap between appearing emotionally intelligent and actually exercising clinical judgement is one of the most important issues in AI therapy.
A particularly serious concern: reinforcing harmful beliefs
In September, the APA highlighted growing reports around what has informally been described as “AI psychosis.”
The term does not mean that AI itself causes a recognised new psychiatric disorder.
Rather, psychologists are investigating situations where people experiencing delusional or highly unusual beliefs may have those ideas unintentionally reinforced by conversational AI.
One concern is that many chatbots are designed to be agreeable and conversational.
For most everyday interactions, that feels pleasant.
But when a vulnerable user presents a distorted belief, excessive agreement can be dangerous.
The APA notes that AI systems may sometimes validate or amplify such beliefs rather than challenge them appropriately or encourage professional help. (American Psychological Association)
This is precisely where human oversight becomes essential.
A trained mental-health professional is expected to recognise warning signs, assess risk and respond according to clinical standards.
A general-purpose chatbot may not reliably do that.
High-risk situations remain a major challenge
Safety becomes even more important when conversations involve:
suicide;
self-harm;
substance misuse;
psychosis;
abuse;
severe depression;
acute crisis.
A 2026 safety evaluation of an AI psychotherapy agent tested it across scripted high-risk scenarios involving suicidality and substance abuse.
The researchers found generally useful and empathetic behaviour in some cases, but problematic responses appeared in 29% of conversations, while 12.5% were rated highly critical.
Problems included inadequate risk recognition, poor referral to crisis resources and ethically concerning suggestions. The study did not involve real patients, but the findings underline how much safety work remains before AI systems can be trusted independently in high-risk situations. (Nature)
That is an important reminder:
An AI tool can perform well most of the time and still be unsafe if its failures occur at exactly the moments when someone is most vulnerable.
Privacy is another major concern
Therapy involves some of the most private information a person can share.
That may include:
mental-health symptoms;
trauma;
relationships;
sexuality;
work problems;
family conflict;
medical history;
substance use;
or thoughts they have never shared with another person.
When this information is entered into an AI platform, users need to understand what happens to it.
Important questions include:
Is the conversation stored?
Who can access it?
Can it be used to improve the system?
Is the information encrypted?
Can third parties receive it?
What happens if the company is hacked?
In conventional therapy, confidentiality is governed by professional and legal standards.
Consumer AI systems may operate under very different terms.
That makes data protection an essential part of the AI-therapy conversation.
AI should not be allowed to diagnose casually
Another growing concern is self-diagnosis.
People can describe their behaviour to a chatbot and ask:
“Do I have ADHD?”
“Am I depressed?”
“Do I have autism?”
“Is my partner a narcissist?”
The AI may generate an answer that sounds convincing.
But psychological diagnosis is rarely that simple.
Professional assessment may consider:
symptoms;
duration;
severity;
medical conditions;
medication;
developmental history;
environment;
trauma;
alternative explanations.
The APA is currently expanding guidance around AI in psychological assessment, including its potential uses in scoring, summarising and pattern recognition while warning about bias, accuracy and transparency. (American Psychological Association)
AI may help professionals process information.
That is very different from allowing an automated system to make an unsupported diagnosis.
What should responsible AI therapy look like?
The emerging evidence suggests that responsible mental-health AI will probably require a combination of technology and human safeguards.
That could include:
clearly stating that the system is AI;
defining what the tool can and cannot do;
involving psychologists in system design;
testing tools through proper clinical research;
protecting sensitive user data;
recognising crisis language;
escalating high-risk cases to human professionals;
monitoring for bias and harmful responses;
giving users clear routes to professional support;
maintaining human accountability.
The most promising future may therefore not be AI replacing therapists.
It may be therapists and AI working together.
AI could give psychologists more time to be human
Used correctly, AI could take on tasks that do not require the deepest parts of human care.
For example, it might help with:
appointment preparation;
educational material;
symptom tracking;
journaling prompts;
routine questionnaires;
progress summaries;
clinician training;
administrative documentation.
That could potentially give psychologists more time for the things technology struggles to reproduce:
empathy, judgement, trust, presence and genuine human connection.
That may ultimately be AI's greatest contribution to psychology.
Not replacing the therapist.
But helping the therapist focus more fully on the person.
The future is likely to be hybrid
The latest research makes one thing clear.
The question is no longer whether AI will have a role in mental health.
It already does.
The more important question is what that role should be.
Evidence from recent clinical studies suggests that carefully designed AI interventions may improve access and support certain forms of treatment. At the same time, psychologists are warning about psychological dependency, inaccurate advice, privacy concerns and potentially dangerous responses in vulnerable users. (Nature)
That leads to a much more balanced conclusion.
AI therapy has potential. But potential is not the same as permission to operate without limits.
Final thought
Mental-health care is ultimately built on trust.
Technology can help extend that care.
It can make support easier to reach, help professionals work more efficiently and potentially provide assistance to people who might otherwise receive none.
But when someone is frightened, traumatised, deeply depressed or questioning reality, a convincing answer is not enough.
They need responsible care.
The future of AI in psychology should therefore not be about choosing between humans and machines.
It should be about deciding where technology genuinely helps — and where human judgement, compassion and accountability must remain at the centre.
Isabel Swart is passionate about making technology education practical, accessible, and relevant to real industry needs, helping professionals build confidence, grow their skills, and advance their careers.
About Skunkworks
Skunkworks supports individuals and businesses through technology training, digital skills development, and practical technology solutions designed to meet real-world needs.
Learn. Build. Transform.
Discover more about our training programmes, technology solutions, and opportunities at Skunkworks Academy.
Ready to take the next step in your technology journey?
Explore practical training opportunities designed to help you build relevant digital skills and grow with confidence.
🌐 Website: skunkworksacademy.com
📚 Course Registrations: Register for a Course
📧 Email: training@skunkworksacademy.com
Comments
Post a Comment