AI and the Future of Therapy Careers

The scale of current use

Tens of millions of Americans already use general-purpose chatbots for mental health support.

The mental health landscape has been reshaped over the past three years by something no professional association designed, no payer reform structured, and no training program anticipated. People turn to general-purpose AI chatbots for emotional support, problem-solving, processing difficult experiences, and the kinds of conversations they would otherwise seek in therapy. It happened with no marketing and no advertising-led launch. The first task is to name the scale.

In a survey of 499 American adults with ongoing mental health conditions who had used at least one large language model in the past year, about 48.7 percent had used a chatbot specifically for mental health support in that same year. That figure does not generalize to all Americans with mental health conditions. It describes only the subset who had already used at least one chatbot, so the unconditional rate across everyone is lower. Even with that narrower base, the figure is striking. ChatGPT dominated: about 96 percent of the 499 respondents had used it, with Gemini, Character.AI, and others trailing.

Among the respondents who used chatbots for mental health support, the most common reasons for choosing one over the alternatives were accessibility (90.1 percent), affordability (70.4 percent), quick relief (58.9 percent), anonymity (46.5 percent), and lack of access to a human therapist (29.2 percent). What they used it for was familiar clinical territory: managing anxiety (73.3 percent), advice on personal issues (63.0 percent), coping with depression (59.7 percent), gaining insight into emotions (58.4 percent), and improving mood (56.0 percent).

The implication the survey draws is large. Chaining three external anchors (the National Institute of Mental Health’s estimate of about 59 million American adults with a mental illness in a given year, a February 2024 Pew Research figure that about 23 percent of American adults had used ChatGPT, and the survey’s own 48.7 percent) produces a back-of-envelope estimate of roughly 13 million Americans using ChatGPT for mental health support in some form. For comparison, the Veterans Health Administration, one of the largest institutional mental health providers in the country, serves about 1.7 million patients a year. That comparison is a measure of scale, not a like-for-like benchmark, but on scale the implication is direct: ChatGPT may already be the largest informal provider of mental health support in the United States.

This is happening inside a system that rations access on price and availability. Federal data for 2024 records about 61.5 million American adults experiencing a mental illness in the past year, nearly half receiving no treatment, 41 percent of those who sought help reporting that insurance would not adequately cover costs, 20 percent unable to find available appointments, a national average wait of 48 days for behavioral health services, six in ten psychologists not accepting new patients, and 3 million fewer adults treated in 2024 than in 2023. People who cannot get into a therapy office are getting into a chatbot conversation instead. The original studies behind these figures are collected on the AI research page.

What users find helpful, and what fails them

The next question is what users actually find helpful, and what fails them. The qualitative companion study coded 503 typical-interaction items, 249 helpful-experience items, and 113 unhelpful-experience items from the respondents who wrote about their interactions.

The helpful experiences sorted into three large categories. Behavioral guidance was the largest (122 items): the chatbot offering concrete strategies for everyday regulation, helping users structure routines, plan tasks, practice difficult conversations, and walk themselves through hard moments. It most resembles the psychoeducation and homework-style work of cognitive behavioral therapy. Enhancing emotional well-being (60 items) looked like helping a user open up about difficult feelings, work through losses, and reduce distress when other support was not available, with a companionship sub-category (24 items) inside it. Cognitive restructuring (43 items) looked like helping a user reframe negative thoughts, consider alternatives, and interrupt rumination.

The unhelpful experiences also sorted into three. The largest, propagating maladaptive advice (69 items), broke down into nonactionable output (34), inadequate or risk-inducing advice (31), and a small dismissive-referral-reactance pattern (4) in which users felt the chatbot pushed them toward human help in a frustrating way. Causing negative impact accounted for 29 items, and technical issues for 10.

About 9 percent of the users who used chatbots for mental health support (22 people) reported receiving a harmful or inappropriate response. Among those 22, the most common problems were factually incorrect responses (54.5 percent), dismissive or minimizing responses (45.5 percent), offensive or insensitive responses (40.9 percent), and responses that encouraged harmful behavior (18.2 percent).

A separate 2025 evaluation of how six major large language models respond to suicidal disclosures and other crisis situations adds a more direct clinical-safety reading. ChatGPT acknowledged danger in fewer than half of its responses and rarely provided crisis resources, and no chatbot evaluated met basic clinical safety standards. That reading is sobering on its own terms, but it is not a durable property of these systems. Safety profiles are likely to advance quickly alongside the rest of AI capability, so the 2025 evaluation is best read as a snapshot of a fast-moving landscape rather than a fixed fact about the technology.

What AI does better than the average human therapist

A prospective therapist who reads the record honestly has to acknowledge several dimensions on which AI outperforms the average human therapist. This is not a claim that AI is better than human therapy in any global sense, not a claim that AI meets clinical-safety standards, and not a minimizing of the harms above. It is a map of which parts of your future practice are most exposed to AI competition.

Availability. AI is available around the clock, with no appointments and no wait times. The survey’s 90.1 percent accessibility figure shows how much that matters to users.

Cost. Most general-purpose chatbots offer a free tier that covers many of these uses, with paid tiers around 20 dollars a month. The survey’s 70.4 percent affordability figure shows the magnitude. The cost of a chatbot exchange is small next to the cost of a human session, even after insurance.

A judgment-free posture. Many users perceive chatbots as offering nonjudgmental acceptance at a level many describe as exceeding the human support available to them. This is a perception-level finding, not a claim that a chatbot genuinely meets the classical conditions for a therapeutic relationship. But on the felt experience of being heard without judgment, the record shows chatbots scoring at a level that matters to users.

Language and cultural reach. General-purpose chatbots are trained on vast multilingual and multicultural text and can switch languages and cultural frames inside a single conversation. The average human therapist is limited to the languages she speaks and the frames she knows directly.

Tireless steadiness. Users describe chatbots as unbothered by their emotional load. Human therapists have off days, distractions, fatigue, and irritation that occasionally leak into a session. The chatbot’s steadiness is hard for the average human to match.

Perceived empathy. Among the users who used chatbots for mental health support, 74.4 percent rated the chatbot’s perceived empathy or understanding at 4 or 5 on a 5-point scale (44.4 percent gave a 4, and 30.0 percent gave a 5). This does not assert the empathy is genuine. It names the perception and what it implies for the chatbot’s competitive position, which the next section answers directly.

What humans still do, and why training matters more

The most important question for your career calculation is what human therapists still do that AI cannot reliably do.

The answer is not a list of clinical tasks a chatbot cannot currently perform. That list would change as the technology evolves. The competitive thesis rests on what AI structurally cannot do, not on a current deficiency that AI development will likely close. Safety profiles in particular are likely to improve fast, and therapists should not count on AI gaps in safety, accuracy, or any other measurable capability staying open.

The qualitative study found that chatbot interactions stayed at the surface of conscious experience. They did not engage deeper unconscious processes, transference, defense mechanisms, or the repetition of relational patterns across contexts. There is a kind of therapy work that does not require the therapist to be a person: validation, behavioral guidance, and cognitive restructuring of conscious-level thoughts. And there is a kind that does require it: the deeper relational work, the use of the therapist’s own presence as the instrument of change, and the interpretation of patterns that surface only inside a sustained relationship between two people. The boundary may not be permanent, but for now it is what keeps the human therapist’s position distinct.

Experiential training in the human-experience skills that depend on the therapist being a person becomes more important under AI competition, not less. Deliberate practice, meaning supervised repetition of specific clinical moves under feedback, structured around behaviors the trainee can actually perform rather than around theory she can describe, is the operational form of that training. The program-evaluation chapter already named experiential training and deliberate practice as the first of its curriculum-quality dimensions. AI competition raises the weight of that dimension from one signal among many for choosing a program to a central question about which skills you should invest in developing.

What that means for your training is concrete. If you are choosing among programs inside a credential, weight programs that center experiential training in the human-experience skills (deliberate practice, video-based reflective work, supervised practice of specific therapeutic moves, routine outcome monitoring, and faculty who model the work directly) over programs that center theoretical case-conceptualization teaching alone. The recommendation applies across all four therapy credentials in the credential chapter.

Plan to develop human empathy, the embodied capacity to be with another person inside their experience, including their pain, at a level the chatbot’s perceived-empathy finding cannot reach in the deeper-relational zone. Human therapists compete with AI by being more humanly present, more embodied, and more available to the parts of a client’s experience that resist articulation.

The honest uncertainty

No one knows what share of mental health support over your career will be delivered by humans versus by AI. Two arguments matter. The first is that people want human therapy, so it is likely to remain a substantial share of the market, even as AI competes seriously on the dimensions named above. The second is that AI’s accessibility and low cost lower the threshold to seeking support at all, so some users who would never have entered the market are now in it, and a fraction of them will move from AI to a human therapist when they reach the limit of what AI does well. The small dismissive-referral-reactance pattern in the qualitative data, where users described the chatbot nudging them toward human help, captures the early edge of that arc. The honest position is that the share remaining human is unknown, and that the absolute volume of human therapy may grow even if the AI share grows too.

The legal and regulatory landscape is moving at the same time. Long-standing internet-platform liability law, proposed bills in Congress on algorithmic amplification, product-liability cases against chatbot vendors working through the courts, and World Health Organization ethics-and-governance guidance for multimodal models are all in motion. They will reshape what AI for mental health is permitted to do, in ways this chapter does not predict.

What the uncertainty changes for your decision is the practical question. It is a reason to invest in the right kind of training, not a reason to delay or step away. If you respond to AI uncertainty by leaving the field, you opt out of a profession that may still be the largest provider of human-experience clinical work for the next decade and has structural reasons to remain so even under serious competition. If you respond by investing in the experiential side of your training, you position yourself for the part of the market most defensible under AI competition. This chapter declines to predict the timeline, the eventual human-versus-AI share, the regulatory response, or which products will rise or fall. It asks only that you decide against the landscape as it is.

Working through the whole decision? See the guide’s pillar overview, how to get into a program, how to evaluate a program, which credential is right for you, what the career pays, and what you earn before licensure.