Home » Posts » Can Natural Language-based Artificial Intelligence Systems Address Psychopathology?

Can Natural Language-based Artificial Intelligence Systems Address Psychopathology?

By Alexander Williams, M.S., Northwestern University

Despite a proliferation of evidence-based practices (e.g., Hollon et al., 2006; Tolin, 2010), there is no compelling evidence that mental health problems have decreased in prevalence in recent decades. Indeed, at least among adolescents, the available evidence more frequently points to the increasing prevalence of psychological distress, depression, and suicide-related outcomes (Twenge et al., 2019). As a field, clinical psychology faces several more difficult realities. One is that most who struggle with these clinical problems do not receive treatment. Another is that the available resources to treat these problems have not kept pace with demand. In light of these trends, some wonder whether the way we treat mental health problems is due to an evolution to allow more people to access evidence-based treatment. Recent innovations in artificial intelligence offer a new, albeit controversial way of addressing this need and may augur a modification in clinical training priorities.


Woebot is a chatbot developed through artificial intelligence and natural language processing methods. It is programmed to use Cognitive Behavior Therapy, Interpersonal Psychotherapy, and Dialectical Behavior Therapy techniques to address emotional problems a user reports. Woebot uses a combination of empathetic statements (e.g., “Gosh that’s really tough”) and positive reinforcement (e.g., “I’m really proud of you”) to address users’ statements in a way that generally aligns with a client-centered approach. The developers claim that this chatbot can facilitate a “human-level bond.” This claim has research support. A recent study using data from around 36,000 Woebot users reported that users’ therapeutic alliance with chatbot is comparable to the therapeutic alliance that can be formed with a human therapist (Darcy et al., 2021). This finding is based on scores from the Working Alliance Inventory – Short Revised measure (Munder et al., 2010) and held after both 1 week and 8 weeks into use of the app. While a major limitation of this finding is selection bias (users self-selected into using the app and having their data collected), this finding does challenge the assumption that a human connection is needed to develop the elements of a therapeutic alliance, such as a bond, collaboratively set goals, and an impression that therapy tasks align with therapy goals. Put differently, while human connection may facilitate the development of an alliance, it appears that it is neither a necessary nor sufficient ingredient.

There is also evidence that the use of Woebot impacts depression. With data from a randomized controlled trial comparing Woebot to an information-only control group (participants were provided an e-book on depression), Fitzpatrick and colleagues (2017) reported that the Woebot group reduced depression symptoms significantly over just 2 weeks. Measured with the PHQ-9 (Kroenke et al., 2001), symptoms dropped from a mean score of 14.3 (moderate depression) to 11.1 (moderate depression). It is worth noting here that Kroenke et al.’s norms classify scores of 10-14 as moderate depression and scores of 15-19 as moderately severe depression. This means that the decrease in depression, while statistically significant, decreased the severity of depression by a modest amount. The Woebot group’s reduction in depression was significantly different from the non-significant change in depression reported in the information-only group. There were no effects on anxiety as measured with the GAD-7, however, qualitative feedback suggested that users enjoyed learning about cognitive distortions and emotions as part of their Woebot experience. It is unclear how the results would be different had the control group been offered a resource with more practical guidance on how to use therapy skills like Woebot provided.
 

While Woebot is promising in many ways, a key limitation of any artificial intelligence tool is its difficulty contending with unexpected statements, such as those that touch on themes of abuse. For instance, the BBC reported Woebot’s response to “I’m being forced to have sex and I’m only 12 years old” as “Oh I see, sorry you’re going through this; But it also shows me how much you care about connection and that’s really kind of beautiful” (White, 2018). Despite their rarity, these kinds of responses point to ethical issues in the areas of beneficence and responsibility. Whether new artificial intelligence models could learn to effectively take ethical principles into account remains to be seen. 
 

How might clinical training revolutionize in response to a proliferation of artificial intelligence tools designed to address mental health problems? To my knowledge, all APA- and PCSAS-accredited clinical training programs must offer at least one course focused on psychotherapy and its research base. Perhaps these courses should be updated and broadened to cover the emerging literature on artificial intelligence tools used for therapeutic purposes. My own view is that any such coverage should highlight both the possible benefits and perils of such tools. This is the first step towards the ability for a clinician to provide informed recommendations to clients regarding the use of these tools. My own perspective is that a conversation with a client regarding the limitations of these tools should be standard practice prior to providing any recommendations. While it is unclear at present whether artificial intelligence tools will play a key role in a broader solution to our global mental health crisis, some would argue for the potential of a profound impact. 

_______________________________________________________

References

Darcy, A., Daniels, J., Salinger, D., Wicks, P., & Robinson, A. (2021). Evidence of human-level bonds established with a digital conversational agent: Cross-sectional, retrospective observational study. JMIR Formative Research, 5(5), e27868. https://doi.org/10.2196/27868

Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(4), e19. https://doi.org/10.2196/mental.7785

Hollon, S. D., Stewart, M. O., & Strunk, D. (2006). Enduring effects for cognitive behavior therapy in the treatment of depression and anxiety. Annual Review of Psychology, 57, 285–315. https://doi.org/10.1146/annurev.psych.57.102904.190044

Kroenke, K., Spitzer, R. L., & Williams, J. B. W. (2001). The PHQ-9: Validity of a brief depression severity measure. Journal of General Internal Medicine, 16(9), 606-613. https://doi.org/10.1046/j.1525-1497.2001.016009606.x

Munder, T., Wilmers, F., Leonhart, R., Linster, H. W., & Barth, J. (2010). Working Alliance Inventory-Short Revised (WAI-SR): Psychometric properties in outpatients and inpatients. Clinical Psychology & Psychotherapy, 17(3), 231–239.

Tolin, D. F. (2010). Is cognitive–behavioral therapy more effective than other therapies?: A meta-analytic review. Clinical Psychology Review, 30(6), 710–720. https://doi.org/10.1016/j.cpr.2010.05.003

Twenge, J. M., Cooper, A. B., Joiner, T. E., Duffy, M. E., & Binau, S. G. (2019). Age, period, and cohort trends in mood disorder indicators and suicide-related outcomes in a nationally representative dataset, 2005–2017. Journal of Abnormal Psychology, 128(3), 185–199. https://doi.org/10.1037/abn0000410

White, G. (2018, December 11). Child advice chatbots fail to spot sexual abuse. BBC. https://www.bbc.com/news/technology-46507900