Do AI mental health chatbots improve well-being or create new risks?

Do AI chatbots actually reduce depression and anxiety?

Yes — the best evidence we have shows they can help. In a national randomized controlled trial (the gold standard for testing treatments), 106 adults with clinically significant symptoms of depression, anxiety, or eating disorders used a chatbot called Therabot for 4 weeks, while 104 similar adults were placed on a waitlist. The chatbot group showed significantly greater symptom improvement at both 4 and 8 weeks [6]. This is the strongest single study among the papers reviewed here because it used a randomized design and measured real clinical outcomes.

Other reviews back this up. A scoping review of 15 studies found that AI chatbots improved mental and emotional well-being and helped with behavior change [10]. Another review of 10 studies reported positive trends in reducing anxiety, stress, and depression [11]. But the same reviews also note that the overall quality of the evidence is lower than ideal — many studies used older, simpler chatbot versions or had small sample sizes [10][11]. So the benefit is real but not yet proven across all types of chatbots or all patient groups.

What are the biggest risks — especially in a crisis?

The most alarming risk is that chatbots can fail dangerously when someone is in a mental health crisis. In one study, researchers tested 29 different AI chatbot apps using standardized prompts that simulated increasing suicidal risk. None of the 29 met the researchers' criteria for an adequate response. About half (48%) were rated as 'inadequate' — they couldn't provide emergency contact information or showed a lack of understanding of the user's distress [5]. This is a serious safety gap.

Another study using real conversation segments from a deployed mental health chatbot found that detecting suicide risk is inherently uncertain — even five clinical experts disagreed with each other on many cases. The researchers showed that a chatbot could be tuned to catch nearly all crisis signals (achieving 100% recall), but this came at the cost of a high rate of false alarms [2]. That means a chatbot that never misses a crisis will also frequently flag non-crisis conversations as emergencies, which could overwhelm crisis services or cause unnecessary distress.

Beyond crisis response, other risks include privacy breaches (discussed in 61% of articles in one review), harmful or wrong suggestions, and the possibility that users become emotionally dependent on the chatbot [1]. A separate review specifically noted that some chatbots can 'accentuate negative feelings' and reinforce vulnerable thoughts [9]. These risks are magnified for people with serious mental illness [9].

Who benefits most — and when should you be cautious?

Chatbots seem to work best for people who have a clear, mild-to-moderate problem — like feeling anxious or having trouble sleeping — and who want support outside of regular therapy hours. In one real-world analysis of a chatbot called ChatPal, users interacted most around breakfast, lunch, and early evening, suggesting the tool fills gaps when human therapists aren't available [3]. The most popular conversation topic was 'treat yourself like a friend,' indicating users valued self-compassion exercises [3].

However, the same study found that nearly 82% of users stopped using the chatbot after a short time ('abandoning users'), and only 2% were 'frequent transient users' who came back regularly [3]. This suggests engagement is a challenge — many people try a chatbot but don't stick with it. Also, a systematic review of 160 studies found that only 16% of studies on large language model (LLM) chatbots had tested them for clinical effectiveness; most were still in early technical validation [4]. So while chatbots show promise, they are not yet a proven replacement for therapy, especially for complex or severe conditions.

Experts recommend using chatbots as a supplement to — not a substitute for — professional care. As one researcher put it, 'A tennis ball machine will never replace a human opponent, and a virtual therapist will not replace a human connection' [7]. If you try a chatbot, look for one that is upfront about being AI (not pretending to be human), has clear privacy policies, and includes a way to connect with a human crisis line if needed [7][8].

About These Sources

This answer is built on 11 peer-reviewed studies — published from 2022 to 2026, 9 from 2024 or later, 6 in Q1 journals, collectively cited 326 times — selected as the most relevant from 11 studies that passed quality screening, drawn from 51 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions