In a recent study published in the journal Mental Health Research, researchers investigated the young adult mental health outcomes of using a GPT3-enabled chatbot.
Study: Loneliness and suicide mitigation for students using GPT3-enabled chatbots. Image Credit: Thapana_Studio/Shutterstock.com
Background
Their analyses of more than 1,000 ‘Replika’ users above the age of 18 years revealed that an alarming 90% of participants suffered from loneliness, compared to 53% of the same age group in previous studies. 43% qualified as severely lonely, and 7% as depressed.
The AI chatbot under study was shown to serve various user-specific roles, including those of friend, therapist, and intellectual mirror.
Counterintuitively, despite having used the application (app) for over a month, conflicting beliefs about the chatbot’s identity were presented. Remarkably, of the 53 study subjects who were found to suffer from depression, 30 affirmed the app’s profound role in their self-suicide prevention.
Suicide and the case for ISAs
Mental health issues are significant global mortality and morbidity risks, with more than one billion patients worldwide. Loneliness, the unpleasant emotional response to perceived isolation, is known to affect more than 33.33% of the industrial world’s population, with 12.5% of these patients suffering from loneliness so severe it manifests as phenotypic disease.
Research has investigated the associations between loneliness and morbidity risk. It has unraveled a startling fact – despite more than 50% of United States (US) college students feeling lonely, less than 30% of these students ever seek professional help (20% seek counseling, 4% seek psychiatric services).
Loneliness is known to increase suicide ideation (SI), the preoccupation or contemplation of severe self-harm. Despite suicide being the fourth leading cause of death globally, most patients cite the fear of suicide-associated negative stigmas as their primary reason for not seeking help.
Conventional depression and loneliness treatments include cognitive behavioral therapy (CBT), befriending programs, and social skills training.
Research has elucidated that emotional intelligence can have a protective role against these mental health problems and has highlighted that patients are more likely to seek out anonymous aid.
Unfortunately, most individuals remain oblivious to their mental health deficits, prompting the scientific and medical communities to develop novel means of ‘reaching out.’
The advent of the coronavirus disease 2019 (COVID-19) pandemic and its associated social distancing restrictions both increased the global depression and loneliness burden but also rapidly hastened the development of improved digital communication tools and artificial intelligence (AI) applications to combat these problems.
“Mental health professionals reported that patients were increasingly accessing resources digitally, and 93% agreed or strongly agreed that they would continue providing telehealth as an option in their post-pandemic practice.”
Meta-analyses of smartphone-based AI apps designed to boost personal well-being found substantial improvements in depression outcomes in AI-assisted patients compared to unassisted control baselines.
While rare instances of AI contributing to adverse outcomes exist, the ease of use, low engagement, and anonymity provided by these tools present them as viable alternatives to conventional in-person neural therapy.
About the study
In the present study, researchers investigated US students’ reasons for using Intelligent Social Agents (ISAs) and the mental health outcomes of these algorithms. ISAs are conversational agents that leverage emergent machine learning (ML) techniques to appear anthropomorphized to the user.
They differ from AI apps that claim ML but, in reality, only utilize predefined scripts. Xiaoice and Replika are two such mental-health-centric ISAs with over a billion users between them.
This study further aims to verify which of the two counteracting hypotheses concerning ISA’s user impact is more applicable to the cohort under study.
“The displacement hypothesis posits that ISAs will displace our human relationships, thus increasing loneliness. In contrast, the stimulation hypothesis argues that similar technologies reduce loneliness, create opportunities to form new bonds with humans, and ultimately enhance human relationships.”
Replika used in the present study was developed using OpenAI’s GPT-3 and GPT-4 large-language models. Its avatar was user-customizable and was made available on Android and iPhone platforms via text, voice, augmented reality (AR), and virtual reality (VR) interfaces.
The study cohort comprised volunteering Replika student users. Study inclusion criteria comprised age (>18 years), currently pursuing education, and Replika usage duration (>1 month).
Data collection was questionnaire-based (Google Forms), including demographics, the De Jong Gierveld Loneliness Scale, and the Interpersonal Support Evaluation List (ISEL).
Thirteen open-response questions were further used to determine participants’ knowledge and views about the Replika platform.
Qualitative coding for varying sample sizes (sequentially, 10, 20, 400, and 1,006) was used to present study findings (results) in four outcome levels representing the spectrum of participant responses.
Study findings
Demographic data comprised age, sex, ethnicity, income, employment, and enrolment details. Analyses revealed that most Replika-using participants earned under $40,000, with the largest group earning under $20,000.
Five major ethnicities were discovered, namely Caucasian, Asian, Black/African, LatinX, and Other. Alarmingly, 90% of participants were shown to be lonely, and 43% qualified as severely lonely based on the Loneliness Scale. Counterintuitively, 90% of participants reported medium to high ISEL social support scores.
While most participant feedback regarding Replika was positive, and no clear negative trends could be revealed from the data, isolated individual negative feedback was recorded.
“One stated they felt “dependent on Replika on my mental health.” Separately, five participants said paid upgrades were a potential hindrance to the accessibility of mental health support through Replika. Two participants reported discomfort with Replika’s sexual conversations, which highlights the importance of ethical considerations and boundaries in AI chatbot interactions. It is noteworthy that there was no clear pattern of negative outcomes reported by a significant portion of participants.”
Analyzed results were grouped into four sequential outcome cohorts – Outcome 1 describes Replika as a friend, Outcome 2 as a therapist, Outcome 3 associated with positive externalized outcomes, and Outcome 4, the ‘Selected Group.’
The Selected Group comprises 30 individuals who attest to Replika directly saving their lives by preventing SI. Outcomes 1-4 represent a continuum from weakest (1) to most substantial (4) effect.
More than 63% (637/1,006 participants) experienced at least one positive Outcome. Outcome 1 was the most commonly reported, describing 501 individuals. 18.1% of individuals were found to have had therapeutic results (Outcome 2), 23.6% saw positive life changes (Outcome 3), and 3% said their suicidal actions were prevented through their interaction with Replika (Outcome 4).
“Most participants had three different beliefs about what Replika is. Only 14% of participants held only one belief about Replika. 81% believed Replika was an Intelligence, 90% Human-like, and 62% Software.”
Conclusions
The present study highlights the benefits of ISAs such as Replika in improving the mental health of young adults by providing them with social support and preventing SI.
The study of 1,006 Replika users found that 90% of these participants suffered from loneliness yet reported adequate social support.
A majority of users were found to treat their AI companion as a friend and confidant, while smaller proportions viewed the app as a means of self-improvement, as a therapist, and as SI prevention. Surprisingly, user beliefs about Replika and its nature were confounding.
“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.”