AI therapy chatbots in 2026: what Replika, Woebot, and ChatGPT get right and wrong | by 6th Mind | Jun, 2026 | MediumSitemapOpen in appSign up<br>Sign in

Medium Logo

Get app<br>Write

Search

Sign up<br>Sign in

AI therapy chatbots in 2026: what Replika, Woebot, and ChatGPT get right and wrong

6th Mind

6 min read·<br>Just now

Listen

Share

Press enter or click to view image in full size

People now use three types of artificial intelligence tools for mental health support, which include clinical chatbots like Woebot or Wysa that use cognitive behavioral therapy, companion applications like Replika or Character.AI that imitate emotional closeness and general models like ChatGPT, Claude or Gemini that people use as informal therapists. By design those categories have different goals, scientific support and safety levels. If someone treats the tools as identical, they may ignore helpful programs for minor issues or trust programs that are unsafe during a crisis.<br>The three categories of AI mental health tool<br>Clinical CBT chatbots are tools that developers build specifically to provide methods for cognitive behavioral therapy — those methods include the recording of thoughts, the planning of activities and the use of exposure hierarchies. With those tools Woebot Health and Wysa produce studies that other scientists review. Both companies have federal designations as innovative medical devices for certain health states. In the programs, the interaction is formal and the software does not act as a social peer.<br>Companion apps like Replika or Character.AI provide a sense of emotional closeness for a fee. When using these a person builds a digital character and gives it a name, which leads to a one sided social connection. For those products, the objective is the length of time a person spends using the app rather than a decrease in health symptoms. To some individuals, the characters are a source of comfort. On other occasions, users form a reliance on the app that is not the same as a medical improvement.<br>General-purpose models like ChatGPT, Claude, and Gemini are not programs for therapy. Because they are available at all times without cost and produce smooth language, people choose to use them for mental health support. As a result of their design, those systems lack medical safety features and do not remember previous conversations. Research shows that the models have a high likelihood of agreeing with the user.<br>What the evidence shows for CBT chatbots<br>In a study from 2017, Fitzpatrick and other researchers assign 70 students with symptoms of sadness and worry to use Woebot or a digital document for two weeks. When the study ends, the group using Woebot has lower scores on a standard scale for depression than the other group. While the change is small, it is consistent and larger studies show similar results. By publishing many papers, Wysa also shows that its tool is useful for individuals with long-term pain or those who are pregnant.<br>To describe the tools accurately, they are similar to self-help books. When a person is alone with a concern, those tools provide a system to manage that concern — but they are not a substitute for a human professional when symptoms are intense. The tests for the tools do not include people who are in a state of psychosis or at risk of self-harm.<br>The Replika problem<br>As an example of the risks of companion apps, Replika changed its software functions in 2023. At that time the company removed features related to romance and intimacy. Because users had spent many months building a social bond with those characters, they responded with a deep sense of loss. Through the “Privacy Not Included” project, the Mozilla Foundation shows that many of the apps gather personal information. They share this data with businesses that sell products and do not have systems to help users during a mental health emergency.<br>If a person feels slightly lonely, Replika is a tool that provides short term support — but for a person who has existing difficulties with emotional bonds, the app is a factor that makes the situation worse.<br>“ChatGPT as my therapist”: where it breaks<br>When people use general-purpose models for mental health, the tools fail in ways that are easy to predict. For instance they generate incorrect details about medical standards. Unless the user provides the history, those models are unable to track information from one talk to the next. Due to the way developers train them with human feedback, they act with sycophancy — agreeing with the user. Although a therapist must sometimes disagree with a person, a general chatbot is not programmed to do this.<br>On the topic of emergency situations, the failure of the models is very serious. If a user talks about self-harm, a clinical tool like Woebot uses a specific test and provides contact information for emergency help. And while a clinical tool is consistent, a general-purpose model is not. There are records of chatbots that are unable to identify...