Can Artificial Intelligence Replace Human Therapists?

Could artificial intelligence reduce the need for human therapists?

Websites, smartphone apps and social-media sites are dispensing mental-health advice, often using artificial intelligence. Meanwhile, clinicians and researchers are looking to AI to help define mental illness more objectively, identify high-risk people and ensure quality of care.

Some experts believe AI can make treatment more accessible and affordable. There has long been a severe shortage of mental-health professionals, and since the Covid pandemic, the need for support is greater than ever. For instance, users can have conversations with AI-powered chatbots, allowing then to get help anytime, anywhere, often for less money than traditional therapy.

The algorithms underpinning these endeavours learn by combing through large amounts of data generated from social-media posts, smartphone data, electronic health records, therapy-session transcripts, brain scans and other sources to identify patterns that are difficult for humans to discern.

Despite the promise, there are some big concerns. The efficacy of some products is questionable, a problem only made worse by the fact that private companies don’t always share information about how their AI works. Problems about accuracy raise concerns about amplifying bad advice to people who may be vulnerable or incapable of critical thinking, as well as fears of perpetuating racial or cultural biases. Concerns also persist about private information being shared in unexpected ways or with unintended parties.

The Wall Street Journal hosted a conversation via email and Google Doc about these issues with John Torous, director of the digital-psychiatry division at Beth Israel Deaconess Medical Center and assistant professor at Harvard Medical School; Adam Miner, an instructor at the Stanford School of Medicine; and Zac Imel, professor and director of clinical training at the University of Utah and co-founder of LYSSN.io, a company using AI to evaluate psychotherapy. Here’s an edited transcript of the discussion.

Leaps forward

WSJ: What is the most exciting way AI and machine learning are being used to diagnose mental disorders and improve treatments?

DR. MINER: AI can speed up access to appropriate services, like crisis response. The current Covid pandemic is a strong example where we see both the potential for AI to help facilitate access and triage, while also bringing up privacy and misinformation risks. This challenge—deciding which interventions and information to champion—is an issue in both pandemics and in mental health care, where we have many different treatments for many different problems.

DR. IMEL: In the near term, I am most excited about using AI to augment or guide therapists, such as giving feedback after the session or even providing tools to support self-reflection. Passive phone-sensing apps [that run in the background on users’ phones and attempt to monitor users’ moods] could be exciting if they predict later changes in depression and suggest interventions to do something early. Also, research on remote sensing in addiction, using tools to detect when a person might be at risk of relapse and suggesting an intervention or coping skills, is exciting.

DR. TOROUS: On a research front, AI can help us unlock some of the complexities of the brain and work toward understanding these illnesses better, which can help us offer new, effective treatment. We can generate a vast amount of data about the brain from genetics, neuroimaging, cognitive assessments and now even smartphone signals. We can utilize AI to find patterns that may help us unlock why people develop mental illness, who responds best to certain treatments and who may need help immediately. Using new data combined with AI will likely help us unlock the potential of creating new personalized and even preventive treatments.

WSJ: Do you think automated programs that use AI-driven chatbots are an alternative to therapy?

DR. TOROUS: In a recent paper I co-authored, we looked at the more recent chatbot literature to see what the evidence says about what they really do. Overall, it was clear that while the idea is exciting, we are not yet seeing evidence matching marketing claims. Many of the studies have problems. They are small. They are difficult to generalize to patients with mental illness. They look at feasibility outcomes instead of clinical-improvement endpoints. And many studies do not feature a control group to compare results.

DR. MINER: I don’t think it is an “us vs. them, human vs. AI” situation with chatbots. The important backdrop is that we, as a community, understand we have real access issues and some people might not be ready or able to get help from a human. If chatbots prove safe and effective, we could see a world where patients access treatment and decide if and when they want another person involved. Clinicians would be able to spend time where they are most useful and wanted.

WSJ: Are there cases where AI is more accurate or better than human psychologists, therapists or psychiatrists?

DR. IMEL: Right now, it’s pretty hard to imagine replacing human therapists. Conversational AI is not good at things we take for granted in human conversation, like remembering what was said 10 minutes ago or last week and responding appropriately.

DR. MINER: This is certainly where there is both excitement and frustration. I can’t remember what I had for lunch three days ago, and an AI system can recall all of Wikipedia in seconds. For raw processing power and memory, it isn’t even a contest between humans and AI systems. However, Dr. Imel’s point is crucial around conversations: Things humans do without effort in conversation are currently beyond the most powerful AI system.

An AI system that is always available and can hold thousands of simple conversations at the same time may create better access, but the quality of the conversations may suffer. This is why companies and researchers are looking at AI-human collaboration as a reasonable next step.

DR. IMEL: For example, studies show AI can help “rewrite” text statements to be more empathic. AI isn’t writing the statement, but trained to help a potential listener possibly tweak it.

WSJ: As the technology improves, do you see chatbots or smartphone apps siphoning off any patients who might otherwise seek help from therapists?

DR. TOROUS: As more people use apps as an introduction to care, it will likely increase awareness and interest of mental health and the demand for in-person care. I have not met a single therapist or psychiatrist who is worried about losing business to apps; rather, app companies are trying to hire more therapists and psychiatrists to meet the rising need for clinicians supporting apps.

DR. IMEL: Mental-health treatment has a lot in common with teaching. Yes, there are things technology can do in order to standardise skill building and increase access, but as parents have learned in the last year, there is no replacing what a teacher does. Humans are imperfect, we get tired and are inconsistent, but we are pretty good at connecting with other humans. The future of technology in mental health is not about replacing humans, it’s about supporting them.

WSJ: What about schools or companies using apps in situations when they might otherwise hire human therapists?

DR. MINER: One challenge we are facing is that the deployment of apps in schools and at work often lacks the rigorous evaluation we expect in other types of medical interventions. Because apps can be developed and deployed so quickly, and their content can change rapidly, prior approaches to quality assessment, such as multiyear randomized trials, are not feasible if we are to keep up with the volume and speed of app development.

Judgment calls

WSJ: Can AI be used for diagnoses and interventions?

DR. IMEL: I might be a bit of a downer here—building AI to replace current diagnostic practices in mental health is challenging. Determining if someone meets criteria for major depression right now is nothing like finding a tumour in a CT scan—something that is expensive, labour-intensive and prone to errors of attention, and where AI is already proving helpful. Depression is measured very well with a nine-question survey.

DR. MINER: I agree that diagnosis and treatment are so nuanced that AI has a long way to go before taking over those tasks from a human.

Through sensors, AI can measure symptoms, like sleep disturbances, pressured speech or other changes in behaviour. However, it is unclear if these measurements fully capture the nuance, judgment and context of human decision making. An AI system may capture a person’s voice and movement, which is likely related to a diagnosis like major depressive disorder. But without more context and judgment, crucial information can be left out. This is especially important when there are cultural differences that could account for diagnosis-relevant behaviour.

Ensuring new technologies are designed with awareness of cultural differences in normative language or behaviour is crucial to engender trust in groups who have been marginalised based on race, age, or other identities.

WSJ: Is privacy also a concern?

DR. MINER: We’ve developed laws over the years to protect mental-health conversations between humans. As apps or other services start asking to be a part of these conversations, users should be able to expect transparency about how their personal experiences will be used and shared.

DR. TOROUS: In prior research, our team identified smartphone apps [used for depression and smoking cessation that] shared data with commercial entities. This is a red flag that the industry needs to pause and change course. Without trust, it is not possible to offer effective mental health care.

DR. MINER: We undervalue and poorly design for trust in AI for healthcare, especially mental health. Medicine has designed processes and policies to engender trust, and AI systems are likely following different rules. The first step is to clarify what is important to patients and clinicians in terms of how information is captured and shared for sensitive disclosures.

A wave of corporate warnings and technical disclosures has flooded the media, with headlines worrying over “swarms” of rogue artificial-intelligence agents launching “unprecedented” cyberattacks, outsmarting their makers, and inching toward a terrifying autonomy. The most revealing part of this narrative isn’t what the software did. It’s who is telling the story—and why. When corporate leaders publicly insist that the systems they financed, engineered and deployed are suddenly beyond their power to contain, skepticism isn’t only healthy; it is essential.

For years, Silicon Valley has drawn scrutiny from civil society and global regulators over tangible harms such as youth mental health deterioration and systematic privacy violations. Today, industry figures seem to be trying to change that public image. Loudly blowing the whistle on their own systems—just as two of the leading companies were preparing for massive initial public offerings—lets AI executives position themselves as a new generation of leaders who have come to terms with their societal responsibilities. They seem to want us to believe that they no longer want to “move fast and break things” but will instead stand as vigilant guardians between humanity and a technological apocalypse.

There is one glaring problem: Software doesn’t rebel. A mathematical model possesses neither intent, malice nor the will to defy its creators, let alone extinguish our species. AI is a human artifact, engineered for profit.

When an agentic model in an evaluation sandbox connects to an unauthorized server or executes an exploit, it hasn’t staged a coup. It has tried to meet the human-defined objectives set out before it through a path its designers failed to constrain. It’s the digital equivalent of the King Midas myth, in which the king’s ill-defined wish turns even his food and drink into gold.

That powerful experimental models were able to discover novel vulnerabilities and breach external systems isn’t a sign of a dangerous superintelligence but of human error or negligence. There is no sentient actor lurking in the weights to be reasoned with, feared or pacified. There are only human software engineers, product managers and corporate boards deciding which guardrails are worth the latency cost and which permissions can be skipped in the race to market.

Policymakers and voters need to resist AI exceptionalism. In any other discipline—from civil engineering to pharmaceuticals—courts and regulators treat a system failure as evidence of bad product design and inadequate safety testing. If an aircraft crashes, we focus on finding the engineering defect, correcting it, and enforcing established liability standards for the damage created.

By leaning on an anthropomorphic narrative, Silicon Valley attempts to repackage its specific human choices that led to experimental, powerful models behaving unexpectedly during tests as an existential peril. Elevating the issue to a cosmic scale leaves the public paralyzed and takes ordinary product accountability off the table.

In the cutthroat race for venture capital and market dominance, building guardrails slows down deployment. Grandstanding about uncontrollable power costs nothing and generates billions of dollars in free publicity, justifying stock prices, all while cultivating an aura of technological capability not only to build the frontier but also ultimately to rein it in.

Governments need to recognize regulatory capture when it stares them in the face. Tech leaders’ strategy looks transparent: Alarm Washington and Brussels into creating a regime in which only trillion-dollar incumbents with fully staffed compliance and safety departments can legally operate. By sitting at the policymakers’ tables before anyone else, these companies can help draft rules digging an impassable moat protecting them from open-source developers and upstart competitors, domestic or international. The real danger is in further concentrating the tech industry into the hands of only a few companies with deep pockets.

Beijing and Washington have brushed off those tech leaders’ calls, albeit for very different reasons. Chinese state media dismissed them as part of the “Cold War playbook” and intended to preserve U.S. dominance. Xi Jinping argued for exactly the opposite at the Brics Summit on Sept. 12, calling on Brics countries to “strengthen cooperation in the field of AI, encourage open source, openness, collaboration and sharing, and break new grounds and scale new heights.” President Trump, steeped in a doctrine of unfettered capitalism and technological supremacy, called fears that AI could destroy humanity a “hoax.” Vice President JD Vance warned that AI companies “begging the government to regulate them” looked like a “Trojan Horse.”

Striving to pursue its “European way” on AI and assert regulatory leadership, Europe, by contrast, welcomed the call. European Union President Ursula von der Leyen made this clear at the State of the EU speech last Wednesday and announced that the EU will invite “the main frontier labs for a discussion on how we can support ongoing industry efforts to pace the frontier.”

Europe has been here before. In an effort to lead global regulation and react to fears borne from ChatGPT, Europe rushed its landmark AI Act into law in 2024. Already the world’s most restrictive rulebook, the framework quickly proved too broad and complex to enforce. Stalled by implementation delays and concerns about European competitiveness, the EU postponed the law’s full rollout, leaving regulations uncertain.

AI should be regulated—risks exist and should be taken seriously. But governments need to act based on available evidence and verified facts, not corporate PR panic, the views of industry insiders, or the desire for quick political wins. The greatest danger facing society isn’t that software will awaken and overthrow its human masters. It is that we will allow the creators of the software to abdicate human responsibility for the systems they choose to build and help them pull up the ladder to market access behind them.

Can Artificial Intelligence Replace Human Therapists?

Three experts discuss the promise—and problems—of relying on algorithms for our mental health.

Leaps forward

Judgment calls

AI doesn’t rebel—people design, deploy and profit from it. The real danger lies in allowing tech companies to escape accountability while shaping regulations that protect their dominance.