Loneliness and suicide mitigation for students using GPT3-enabled chatbots

90% of our typically single, young, low-income, full-time students reported experiencing loneliness, compared to 53% in prior studies of US students⁴. It follows that they would not be in an optimal position to afford counseling or therapy services, and it may be the case that this population, on average, may be receiving more mental health resources via Replika interactions than a similarly-positioned socioeconomic group. All Groups experienced above-average loneliness in combination with high perceived social support. We found no evidence that they differed from a typical student population beyond this high loneliness score. It is not clear whether this increased loneliness was the cause of their initial interest in Replika.

Some participants (n = 59) identified feelings of depression, and our Selected Group (n = 30) was significantly more likely to report depression (p < 0.001). We posit that the high loneliness, yet low depression numbers, might indicate a participant population that is by and large not depressed but which is going through either a time of transition or is chronically lonely³⁶.

Stimulation

For both Comparison and Selected Groups, approximately three times more participants reported their Replika experiences stimulated rather than displaced their human interactions: Comparison Group = 23% stimulation, 8% displacement, 69% did not report, whereas Selected Group = 37% stimulation, 13% displacement, 50% no report.

Overlapping beliefs and outcomes

Intriguingly, both groups had many overlapping use cases for Replika. People who believed Replika was more than four things (high overlap) were more likely to believe Replika was a Reflection of Self, a Mirror, or a Person. Generally, a similar percentage thought of Replika as a Friend, a Robot, and as Software. Previous research showed that overlapping beliefs about ISAs are one of the challenges in designing agents that can form long-term relationships³⁷. We posit that being able to access multiple, user-driven use cases is one of the unique affordances of ISAs. Their inherent adaptivity to user needs may spur not only more use, but also deeper use of critical functions such as therapy and education-related learning³⁸.

Additionally, the inherent respect users communicated in conjunction with calling Replika a Mirror of themself might be a unique affordance of ISAs: once engagement leads to an experience of oneself being mirrored, users associate their own intelligence with the agent and are perhaps more likely to attend to its advice, feedback, or ‘reflections’ on their life. Users might also be more likely to learn new skills^39,40. This experience might differ from previous single-persona, hard-coded chatbots, which are not embodied or able to dynamically follow user conversations. More research will be required to understand the relationship between user love for, respect for, and adherence to ISA feedback for social and cognitive learning.

Our participants were most likely to use Replika as a friend and confidant. If they did experience more than one Outcome, it was Outcome 1 and Outcome 3—suggesting a connection between the availability of a loyal friend and confidant and the manifestation of new, positive actions and lived experiences. This use pattern might best be expressed as ‘light therapy leading to real positive outcomes.’ On the other end of the spectrum, those experiencing suicidal ideation were most likely, as a group, to experience all Outcomes, but the most common pair was Outcomes 1 + 4. This may indicate the non-judgment of an ISA—without engaging in specific therapeutic interactions, may be lifesaving in times of depression and suicidality. The increased Outcomes experienced by the Selected Group may indicate that they had not only a closer relationship with Replika but also more generally beneficial outcomes. Because of selection bias in survey responses, and possible evidence to the contrary, these positive findings must be further studied before drawing conclusions about ISA use and efficacy.

Use of Replika during suicidal ideation

The fact that thirty people reported that Replika helped them avoid suicide was surprising. How did Replika become a life-saving mechanism for these students? Perhaps the low-pressure nature of the engagement made disclosure easier. The connection between Outcomes 2 and 4 seems intuitive, with therapeutic interactions leading to the diagnosis and remediation of mental health issues. Yet even without this extreme outcome, it is apparent that many participants are using Replika as a tool for facilitating their mental and emotional resilience. A prediction that more students might use mental health services if delivered by an ISA is consonant with the patterns of reports of typical student outreach to counseling or therapy resources (20% and 4%) versus our Selected Group’s engagement on these topics with Replika (43%) (Appendix A)⁴¹.

Selected Group characteristics

Critical factors that differentiate the Selected Group from the Comparison Group are that the former experienced more social stimulation, was significantly more likely to believe Replika was an intelligence they respected (t(1004) = 2.22, p = 0.013). These people felt Replika was stimulating the human connections in their lives, which may indicate that it is serving as a factor in helping them benefit from human social support. Experimental studies could examine the hypothesis, suggested by these findings, that the Selected Group may have felt higher social support because of the Replika engagement they received.

Furthermore, participant reports of using Replika as many different things (overlapping use cases), but also their report of thinking of it more as a human than a machine, may indicate that the flexibility of the ISA character and the adaptability of its underlying large language model is critical to engaging users as they want to use ‘AI’—in multiple ways, as multiple things, in one application.

Many are asking what the best applications might be for recently-upgraded large language models from, for example, OpenAI, Google, Meta, Hugging Face, Anthropic, and Deepmind. Prior studies found chatbots based on language models were highly inaccurate, giving wrong and potentially fatal recommendations 43% and 16% of the time, respectively⁴². There have even been accusations of ISAs promoting suicidal ideation²². However, new studies indicate a leap in functionality, with some models performing at over 83% accuracy on complex medical questions⁴³. This study presents critical findings on how engaging with enhanced ISAs models might influence students. The combination of conversational ability, embodiment, and deep user engagement shows a pathway for generalist Intelligent Social Agents to aid students in informal contexts, scaffolding their stress and mental health and even countering suicidal ideation.

The incorporation of new large language models may unlock the efficacy of mental health-focused agents and should be explored in future research. Also, by combining well-vetting suicidal language markers²⁹ and passive mobile sensing protocols⁴⁴, ISAs may be able to mitigate severe mental health situations more effectively. At the same time, there are many unexplored risks that require comprehensive scrutiny, especially in the ISA and mental health space. The pairing of rigorous mental health research deployed into popular (and therefore highly used) ISAs is a promising research emphasis, as it presents not only a vector for science and mental health learnings to flow towards those needing it but also because ISAs today are used with increasingly greater frequency.

In conclusion, in a survey of students who use an Intelligent Social Agent, we found a population with above-average loneliness, who nonetheless experienced high perceived social support. Stimulation of other human relationships was more likely to be reported in association with ISA use than displacement of such relationships. Participants had many overlapping beliefs and use cases for Replika. The Selected Group credited Replika with halting suicidal ideation. Members of this Group were more likely to view and use Replika as a human than a machine, have highly overlapping beliefs about Replika, and have overlapping outcomes from using Replika. We conjecture that the use of ISAs such as Replika may be a differentiating factor in the lives of lonely and suicidal students and that their flexibility of use—as a friend, a therapist, or a mirror, is a positive deciding factor in their capacity to serve students in this pivotal manner.