AI therapy bots (part 2): Survey of non-AI online psychotherapy offerings

Part 2 in a series on psychotherapeutic chatbots that include artificial intelligence capabilities

The ubiquity of computers and, more important, the Internet has brought psychotherapy online.  In this post I’ll take a rambling look at the following two kinds of “online therapy”:

  • Online therapy guided by a trained psychotherapist
  • Self-guided therapy with no or very minimal trained therapist involvement

One-on-one face-to-face psychotherapy is expensive and often inconvenient. Internet-based psychotherapy can potentially overcome both of these challenges. Self-guided approaches will usually be less expensive and more convenient than therapist-guided approaches, but do they provide effectiveness similar to therapist-guided programs?

How effective is online self-guided psychotherapy?

Some results from meta-analyses give us an idea of the effectiveness of self-guided psychotherapy:

Across meta-analyses, higher quality studies are generally associated with smaller effect sizes. A higher quality study will generally use effective randomization procedures, a no-treatment control group, comparator groups that address other potential reasons you might see improvements, and enough participants to achieve reasonable statistical power (among other things). The scientific research base is generally considered to overestimate effects of treatments, both because of publication bias and researcher degrees of freedom. It’s probably safe to say that self-help psychotherapy with minimal human involvement shows small but positive effects on average, not moderate or large. Findings of treatment effect for face-to-face therapy are more consistently in the medium effect size range.

Effectiveness of therapist-guided online psychotherapy

What about putting therapist-guided CBT online? A 2014 meta-analysis found no significant difference in treatment effects for guided internet-delivered CBT compared to face-to-face treatments. My focus is less on this modality because it doesn’t overcome one of the most important drawbacks of face-to-face counseling: the heavy labor involvement of a trained therapist with its concomitant expense as well as the inconvenience of scheduling with a human therapist.

Internet self-guided approaches show lower effectiveness than therapist-guided interventions: Why?

Understanding the way self-guided internet based CBT and other psychotherapy techniques have failed or can be strengthened will be important in designing AI-based approaches that overcome such limitations. The research literature in this space is tremendous so I’ll just hit a few interesting results. My purpose is not to make well-researched conclusions but rather survey the space before moving on to considering AI-based therapy bots.

One huge reason self-guided approaches don’t work as well as therapist-guided approaches is dropout. In a meta-analysis of ten RCTs of self-guided web-based interventions for depression, researchers calculated that almost 70% of participants dropped out of the interventions before completing at least 75% of the treatment modules. Participants were more likely to drop out if they were male, had lower educational level, and/or had co-morbid anxiety symptoms. Older participants were less likely to drop out.

I imagine that another reason that self-guided approaches don’t work as well as therapist-guided approaches is that a good therapist can respond to exactly what a particular person needs at a particular point in time. Looking for research-based evidence of this would require cracking into individual programs to see how personalized programs were, whether therapist-guided or self-help guided. I’m not going to do that right now, but it’s an important area to consider.

It seems reasonable to assume that AI-based interventions may be able to achieve better results with psychotherapeutic techniques than non-AI based interventions for a couple reasons:

  • The AI component can function as the guide, reproducing whatever “something” a human therapist is providing in achieving better results than self-guided approaches — perhaps personalization, but also some sense of feeling understood and cared for.
  • AI techniques beyond those reproducing therapist guidance can be used to achieve better engagement and lower attrition.

Is the personal guide what matters (AI or human)? Can you develop automated approaches without AI that achieve effectiveness similar to human-guided approaches?

Ways to improve self-guided online therapy (without AI)

One cheap intervention that might improve results would be to provide automated reminders (for example by email) to keep participants engaged and prevent them from dropping out. This 2013 RCT of a self-guided treatment for anxiety and depression found that automated email reminders twice weekly and triggered when participants appeared at risk of dropping out promoted completion as well as lowered symptoms at follow-up. Participants with and without the email reminders showed response to the treatment in the medium effect size range compared to participants in the control group. This is generally higher than seen in meta-analyses of self-guided online therapy. The treatment evaluated was The Wellbeing Course, described as follows:

The Wellbeing Course is a new five lesson transdiagnostic online intervention based on models of cognitive behavioural and interpersonal therapies. The original Wellbeing Program developed by the researchers [8][9] was not available when the researchers moved institutions and, consequently, a new intervention was developed. This new course is based on a pragmatic model of psychotherapeutic change that assumes that symptoms of anxiety and depression are the result of unhelpful habits of thought and actions, that is, maladaptive cognitions and behaviours. Such unhelpful habits often develop over months or years and may become an entrenched part of a person’s life. This model also assumes that interventions that are structured, systematic, and which require adherence and commitment over several months are more likely to facilitate sustained improvements than sporadic or unstructured therapy sessions, which may only result in short-term symptom relief.

Thus, the Wellbeing Course is a structured intervention that participants complete over an 8 week period. Participants are strongly encouraged to learn about and practice the psychological skills taught in the course, and to adopt these into their everyday lives. The course systematically teaches core psychological skills that aim to increase the frequency of cognitions and behaviours that promote emotional health, and reduce those that maintain distressing symptoms. Examples of the former include realistic thinking skills, planning, and problem solving skills, assertive communication, behavioural activation, and graded exposure. Examples of the latter include patterns of catastrophic and self-defeating thinking, passive or aggressive communication styles, avoidance, and behavioural inhibition. By doing so, this Course targets symptoms common to anxiety disorders and depression. The Wellbeing Course was designed as a low intensity intervention, which could be used as a standalone intervention, an intervention for those on waiting lists for traditional therapy, as an adjunct to traditional therapy, or to facilitate treatment gains post-treatment.

Another way to improve treatment effectiveness might be to use interactive techniques in the lessons to maintain engagement and promote learning. Here’s an online program Deprexis that has shown medium-size effects in RCTs (comparable to effects seen in RCTs of face-to-face psychotherapy). From an RCT evaluating Deprexis:

The Web-based intervention consists of 10 content modules representing different psychotherapeutic approaches, plus one introductory and one summary module, each of which can be completed in 10 to 60 minutes, depending on the user’s reading speed, interest, motivation, and individual path through the program (see Figure 1 and Multimedia Appendix 1 for screenshots). Modules are organized as simulated dialogues in which the program explains and illustrates concepts and techniques, engages the user in exercises, and continuously asks users to respond by selecting from response options. Subsequent content is then tailored to the users’ responses, resulting in a simulated conversational flow. All modules are accompanied by illustrations (eg, drawings, photographs, flash animations). The program version that was evaluated in this study did not include audio or video features in order to increase accessibility by reducing the requirements for broad bandwidth and specialized hardware or software.

Figure 1

Example screenshot (see Multimedia Appendix 1for additional examples)

The modules cover a variety of therapeutic content that is broadly consistent with a cognitive-behavioral perspective, although the program is not restricted to one CBT manual. Instead, an effort was made to design the program as an integrative treatment tool that provides a variety of relevant therapeutic approaches and fits within the broad array of contemporary CBT. The modules’ theoretical rationale and content draws from theories like (1) Behavioral Activation, (2) Cognitive Modification, (3) Mindfulness and Acceptance, (4) Interpersonal Skills, (5) Relaxation, Physical Exercise and Lifestyle Modification, (6) Problem Solving, (7) Childhood Experiences and Early Schemas, (8) Positive Psychology Interventions, (9) Dreamwork and Emotion-Focused Interventions, and (10) Psychoeducation.

Self-guided online psychotherapy generally consists of lessons to be learned — it appears to be fundamentally an educational endeavor. CBT, one of the psychotherapeutic techniques generally considered evidence-based and effective, does involve a set of techniques that need to be learned and practiced. DBT, newer and with mindfulness grafted on, as well consists of skills that can be learned and developed over time, through effortful attention and practice.

Psychodynamic therapy is a little different from CBT and DBT and seems less likely to be successfully put online as an educational program. It focuses on “[revealing] the unconscious content of a person’s psyche in an effort to alleviate psychic tension.” It relies heavily on the relationship of the client to the therapist.

Here’s a fundamental question I’ve arrived at: are psychotherapeutic techniques effective because of what the client learns? Or how the psychotherapist interacts with them? Perhaps a combination of both, with different approaches leaning more heavily towards one modality than the other.


Tentative conclusions

Self-guided online psychotherapy on average shows lower treatment effects that therapist-guided online psychotherapy. Therapist-guided online psychotherapy has shown, on average, equivalent effectiveness with face-to-face psychotherapy. Of course underneath these averages there is substantial heterogeneity, depending on particular therapeutic programs, patient characteristics, therapists (in the case of therapist-guided therapy), etc.

In general, self-guided psychotherapy might be improved by engaging participants better in a way that human therapists do naturally. This includes reminding someone to participate, noting if they seem to be lacking in engagement and motivating them to engage, personalizing the feedback and lessons, helping them understand concepts that they find difficult, and (perhaps) making them feel cared for.

In the next post in this series, I’ll take a look at two AI-based psychotherapeutic chatbots, Woebot and Tess (from X2AI). I will need to come back to the landscape of electronic CBT and other online psychotherapy offerings. I have only scratched the surface!