Psychological Effects of Human-AI Interaction under Exam Stress

Introduction 

Humanity’s continuous goal to replicate the cognitive functions of the human brain through machines has evolved alongside the development of science and technology. After many years of scientific progress, the 20th century marked the beginning of major breakthroughs that laid the foundation for Artificial Intelligence.

After several decades of development, Artificial Intelligence technology has been used in many fields of today’s society. Specifically, there are several typical applications of Artificial Intelligence with future implementations. In this study, we will look closely into Machine Learning, Neural Networks and Natural Language processing. 

Machine Learning emerged through key milestones beginning in the 1940s, including the first neural network model (1943), Arthur Samuel’s checkers program introducing the term “Machine Learning” (1959), and its resurgence in the 2000s with deep learning and large datasets (Yu et al., 2023). As a core branch of artificial intelligence, it enables computers to emulate human learning and transform information into structured knowledge that improves performance. Among its techniques, Artificial Neural Networks (ANNs) are the foundation of modern deep learning, consisting of interconnected neurons that replicate cognitive mechanisms for problem-solving and pattern recognition. Building on ML and ANNs, Natural Language Processing (NLP) allows computers to understand and generate human language (Khurana et al., 2022). Advanced NLP models like ChatGPT, based on the Generative Pre-trained Transformer (GPT) architecture, use billions of parameters trained on diverse text to produce coherent, context-aware responses (Ray, 2023). 

Stress is a natural adaptive response, but when prolonged or excessive, it becomes harmful. Chronic stress disrupts neurochemical balance, impairing cognition, emotion, and motivation. Neuroscientifically, it activates the hypothalamic-pituitary-adrenal (HPA) axis, triggering cortisol release (Herman et al., 2017). While short-term cortisol improves alertness, long-term exposure weakens neural communication and synaptic plasticity - the brain’s capacity to form and reorganize connections crucial for learning and memory. Sustained stress particularly affects the hippocampus and prefrontal cortex, which regulate memory, attention, and emotional control (McEwen, 2017). Consequently, students under academic pressure often experience anxiety, fatigue, and loss of motivation, reinforcing cognitive decline. Thus, stress functions not only as a biological response but also as a major barrier to learning and emotional well-being. 

In recent years, researchers have begun exploring the potential of AI-driven chatbots as accessible tools for managing academic and psychological stress. By combining principles of machine learning, natural language processing, and affective computing, these systems are designed to recognize emotional cues, provide empathetic responses, and encourage healthy coping strategies. Chatbots such as ChatGPT can engage users in reflective dialogue, offer motivational support, and deliver evidence-based techniques rooted in cognitive-behavioral approaches. Through continuous interaction, such AI systems may help alleviate stress by offering emotional support and can assist students in developing self-awareness and more efficient learning strategies. This integration of technology, neuroscience, and psychology marks a promising direction for understanding how artificial intelligence can enhance human resilience and academic performance (Aramaki et al., 2022). 

Although previous research has demonstrated the potential of AI-based interventions, such as conversational agents, to provide cognitive and emotional support, the precise mechanisms through which these systems alleviate academic stress, enhance learning, and promote psychological well-being remain unclear. Building on this, the present study aims to investigate the psychological impact of AI-human interaction during periods of academic hardship, with a focus on how AI-based support influences stress and anxiety. A secondary objective is to evaluate the effectiveness of different mental health AI chatbots, considering their unique approaches and response strategies, and to compare these effects with traditional human-human support. To achieve this, previous studies will be evaluated based on AI chatbots’ structured interactions in terms of stress levels and emotional states. It is hypothesized that receiving supportive responses from AI will lead to reduced anxiety and improved mental well-being, demonstrating the potential of AI to serve as an effective tool for academic and psychological support.

Methods 

1.1 - Main Focus 

This study employs a structured literature review to examine current knowledge regarding the use of Artificial Intelligence, particularly conversational agents and chatbots, for psychological and emotional support. The review focuses on empirical evidence concerning the effectiveness of AI interventions in reducing anxiety, distress, and related symptoms, with special attention to student populations and academic stress contexts. 

1.2 - Literature Search Settings 

Relevant literature was identified through PubMed, Web of Science, EMBASE, and PsycINFO using combinations of keywords such as “chatbot”, “conversational agent”, “mental health”, “anxiety”, “stress”, “student”, “academic”, and “intervention”. 

Search date: October 2025. 

Example search string: (“chatbot” OR “conversational agent”) AND (“mental health” OR “stress” OR “anxiety”) AND (“student” OR “academic”). 

The search was limited to English-language publications from 2020 to 2025. 

1.3 - Eligibility Criteria 

Studies were included if they reported empirical outcomes of AI-based interventions targeting mental health or stress reduction, employed validated psychological measures (e.g., STAI-S, GAD-7), and involved populations comparable to students or young adults. Both randomized controlled trials and pilot studies were considered. Exclusion criteria were non-empirical articles, studies lacking outcome measures, or interventions not involving conversational AI. 

1.4 - Data Extraction and Categorization 

Data were extracted systematically, focusing on study design, sample characteristics, chatbot type and architecture, intervention content, duration and frequency, outcome measures, and ethical or methodological considerations. Studies were categorized according to intervention format (e.g., short structured dialogues, micro-CBT techniques), target population, and assessment methodology. Ethical considerations such as privacy, transparency, informed consent, and limitations on crisis response were recorded to guide interpretation. 

1.5 - Data Synthesis 

The extracted information was synthesized narratively to map the landscape of AI-based mental health interventions, identify common methodological approaches, and highlight gaps in research, particularly regarding short-term support for exam-related stress and student populations.

Results 

Study selection and characteristics.

The search yielded N₁ records; N₂ full texts were screened; five studies met inclusion criteria (PRISMA flow). Samples were predominantly university students or young adults. Interventions lasted 1–8 weeks and were delivered via mobile or web chatbots (rule-based or generative). Primary outcomes included STAI-S or GAD-7; secondary outcomes included mood, perceived empathy/quality, and intention to reuse. The combined sample size was approximately 1,760 participants. 

2.1- Eficacy of AI Chatbots in Reducing Anxiety and Depression 

A systematic review of recent studies examining AI-based interventions for mental health support among students and young adults reveals consistent evidence regarding their potential efficacy. Research indicates that AI chatbots can reduce symptoms of anxiety and depression while providing empathetic and high-quality responses. For instance, a randomized controlled trial evaluating the Fido chatbot reported significant reductions in both depressive and anxiety symptoms in the intervention group compared to baseline, with these effects maintained at a one-month follow-up. Participants who engaged more frequently with the chatbot experienced a greater decline in loneliness, although overall differences in loneliness between groups were not statistically significant. (Karkosz et al., 2024). 

Similarly, a study involving 181 Argentinian college students assessed the effects of Tess, an AI chatbot designed to provide mental health support. Although no significant differences were observed between the experimental and control groups from baseline to eight weeks, intragroup analyses revealed that the experimental group experienced a significant decrease in anxiety symptoms. (Klos et al., 2021). 

Generative AI chatbots such as Therabot have also been evaluated for their therapeutic potential. Participants using Therabot demonstrated substantial reductions in symptoms of major depressive disorder, generalized anxiety disorder, and chronic health-related functional disability compared to controls. Mean changes in symptom scores ranged from -6.13 to -10.23 for depression, anxiety, and related conditions, with effect sizes between 0.627 and 0.903. Engagement with Therabot was high, with average use exceeding six hours, and participants rated the therapeutic alliance with the chatbot as comparable to that of human therapists (Heinz et al., 2025).

2.2 - Perceived Empathy and Quality of AI Responses 

Evidence further suggests that AI chatbots are perceived as highly empathetic and capable of delivering high-quality responses. In a study comparing responses from AI chatbots to physician-generated replies on public forum queries, evaluators preferred chatbot responses in 78.6% of cases. Chatbot responses were longer, more detailed, and rated as significantly higher in quality and empathy than those provided by physicians. The proportion of chatbot responses deemed empathetic or very empathetic was 45.1%, compared to only 4.6% for physician responses. (Ayers et al., 2023). 

2.3 - Intervention Characteristics and Methodology 

Across studies, interventions typically employed short, structured dialogues incorporating evidence-based techniques, including cognitive-behavioral strategies and psychoeducation. Immediate situational anxiety was measured using the State-Trait Anxiety Inventory - State version (STAI-S), while general anxiety over longer periods was assessed with the Generalized Anxiety Disorder scale (GAD-7). Additional self-report measures evaluated mood, perceived clarity of next steps, and subjective stress reduction, providing insights into the short-term effectiveness of AI interventions. 

2.4 - Ethical and Methodological Considerations 

AI-based tools are supplements, not substitutes, for psychotherapy. Responsible use requires privacy protection, informed consent, transparency about AI involvement, and clear crisis-response limitations, especially for students and minors. Consistent with WHO and professional-society guidance, deployment should be supervised to avoid misrepresentation as clinical therapy.

Discussion 

The findings from this research suggest that AI-based interventions, particularly conversational chatbots, can offer meaningful support for students experiencing academic stress. The observed reductions in situational anxiety and depressive symptoms across several studies indicate that short, structured AI interactions may serve as an effective mechanism for mitigating acute stress responses. Interestingly, in the study evaluating the Fido chatbot, the control group spent more time on their intervention (mean 117.57 minutes) than participants using Fido (mean 79.44 minutes), yet both groups demonstrated reductions in anxiety and depression. This observation highlights the potential efficiency of brief, targeted interactions with AI, suggesting that engagement quality and structure may be more critical than sheer duration of use (Karkosz et al., 2024). 

Further insights can be drawn from the study involving Tess, where no significant differences were observed between experimental and control groups in depressive symptoms after eight weeks. However, a significant intragroup decrease in anxiety was observed for participants using Tess, indicating that AI interventions may have a more immediate impact on anxiety than on depressive symptoms in short-term applications (Klos et al., 2021). These findings align with meta-analytic evidence showing that AI chatbots can significantly reduce distress (Hedges’ g ≈ 0.70) and depressive symptoms (Hedges’ g ≈ 0.64), particularly when delivered via mobile or generative platforms (Li et al., 2023). 

The mechanisms underlying these effects may be partially explained by the perceived empathy and quality of AI responses. In a comparative evaluation of AI and human responses, participants perceived chatbot interactions as significantly more empathetic and of higher quality than physician responses, with 78.5% of chatbot responses judged as good or very good quality versus 22.1% for physicians, and 45.1% rated as empathetic or very empathetic compared to 4.6% for physicians (Ayers et al., 2023). This heightened perception of empathy may contribute to the immediate psychological benefits observed in AI interventions, providing users with emotional validation, guidance, and support, even in the absence of direct human interaction. 

The broader literature on human-AI interaction contextualizes these findings within a complex psychological framework. Research indicates that interactions with AI can influence self-confidence through social comparison and may impact social affiliation and feelings of loneliness, particularly among individuals with high attachment anxiety (Tang et al., 2023; Reich & Teeny, 2025). Meta-analytic reviews further suggest that AI is appreciated when perceived as competent and capable, while personalization requirements or unmet expectations can induce aversion (Qin et al., 2025). Additionally, exposure to AI can carry risks of dehumanization, potentially affecting self-regard and interpersonal dynamics (Dang & Liu, 2025). 

In the context of academic stress, AI chatbots may function as immediate, accessible sources of social support, normalizing anxiety, guiding task prioritization, and delivering micro-cognitive-behavioral techniques. Person-centered messaging can increase perceived social presence and quality of support, which likely mediates reductions in situational anxiety (Heinz et al., 2025). Neurocognitive evidence also suggests that the human brain engages differently with AI-generated versus human-generated advice, with implications for decision-making, authenticity, and behavioral adaptation (Tang et al., 2023; Reich & Teeny, 2025). These dynamics may explain why brief, structured chatbot interventions can produce meaningful short-term improvements in stress-related outcomes. 

However, it is important to note the limitations of the current evidence base. Most studies are short-term (1-8 weeks), rely on online or convenience samples, and focus primarily on self-reported outcomes rather than clinical diagnoses. Long-term effects, durability of benefits, and the impact on overall well-being remain underexplored. Ethical considerations, including privacy, transparency, informed consent, and appropriate disclaimers regarding AI as a complement rather than a replacement for professional care, are critical, particularly for student populations (WHO, 2024-2025). 

Overall, these findings suggest that AI chatbots can provide efficient and empathetic support, complementing human interaction in managing academic stress. The differential effects on anxiety versus depressive symptoms, coupled with high perceived empathy, indicate a promising complementary role for AI interventions in acute, short-term contexts, though further research is needed to establish long-term outcomes and optimal deployment strategies.

---

Written by Arina Potapenko 

References 

1. Cao, Zheng. (2017). Development and Application of Artificial Intelligence. 10.2991/icmeit-17.2017.79. 

2. Yu, Liqiang & Zhao, Xinyu & Huang, Jiaxin & Hu, Hao & Liu, Bo. (2023). Research on Machine Learning with Algorithms and Development. Journal of Theory and Practice of Engineering Science. 3. 7-14. 10.53469/jtpes.2023.03(12).02. 

3. Khurana, Diksha & Koli, Aditya & Khatter, Kiran & Singh, Sukhdev. (2022). Natural language processing: state of the art, current trends and challenges. Multimedia Tools and Applications. 82. 3713-3744. 10.1007/s11042-022-13428-4. 

4. Aramaki E, Wakamiya S, Yada S, Nakamura Y. Natural Language Processing: from Bedside to Everywhere. Yearb Med Inform. 2022 Aug;31(1):243-253. doi: 10.1055/s-0042-1742510. Epub 2022 Jun 2. PMID: 35654422; PMCID: PMC9719781. 

5. Ray, P. P. (2023). ChatGPT: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope. Internet of Things and Cyber-Physical Systems, 3, 121–154. https://doi.org/10.1016/j.iotcps.2023.04.003 

6. Herman, J. P., McKlveen, J. M., Ghosal, S., Kopp, B., Wulsin, A., Makinson, R., Scheimann, J., & Myers, B. (2016). Regulation of the 

Hypothalamic-Pituitary-Adrenocortical Stress Response. Comprehensive Physiology, 6(2), 603–621. https://doi.org/10.1002/cphy.c150015 

7. McEwen B. S. (2017). Neurobiological and Systemic Effects of Chronic Stress. Chronic stress (Thousand Oaks, Calif.), 1, 2470547017692328. 

https://doi.org/10.1177/2470547017692328 

8. Karkosz, S., Szymański, R., Sanna, K., & Michałowski, J. (2024). Effectiveness of a Web-based and Mobile Therapy Chatbot on Anxiety and Depressive Symptoms in Subclinical Young Adults: Randomized Controlled Trial. JMIR formative research, 8, e47960. https://doi.org/10.2196/47960 

9. Klos, M. C., Escoredo, M., Joerin, A., Lemos, V. N., Rauws, M., & Bunge, E. L. (2021). Artificial Intelligence-Based Chatbot for Anxiety and Depression in University Students: Pilot Randomized Controlled Trial. JMIR formative research, 5(8), e20678. https://doi.org/10.2196/20678

10. Heinz, M. V., Mackin, D. M., Trudeau, B., Bhattacharya, S., Wang, Y., Banta, H. A., Jewett, A. D., Salzhauer, A. J., Griffin, T. Z., & Jacobson, N. C. (2025). Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI. https://doi.org/10.1056/AIoa2400802 

11. Ayers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., Faix, D. J., Goodman, A. M., Longhurst, C. A., Hogarth, M., & Smith, D. M. (2023). Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum. JAMA internal medicine, 183(6), 589–596. https://doi.org/10.1001/jamainternmed.2023.1838 

12. Li, H., Zhang, R., Lee, Y. C., Kraut, R. E., & Mohr, D. C. (2023). Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. NPJ digital medicine, 6(1), 236. https://doi.org/10.1038/s41746-023-00979-5 

13. World Health Organization. (2024). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. World Health Organization. https://www.who.int/publications/i/item/9789240084759 

14. No Person Is an Island: Unpacking the Work and After-Work Consequences of Interacting With Artificial Intelligence. Tang PM, Koopman J, Mai KM, et al. The Journal of Applied Psychology. 2023;108(11):1766-1789. doi:10.1037/apl0001103. 

15. Does Artificial Intelligence Cause Artificial Confidence? Generative Artificial Intelligence as an Emerging Social Referent. Reich T, Teeny JD. Journal of Personality and Social Psychology. 2025;:2026-09140-001. doi:10.1037/pspa0000450. 

16. AI Aversion or Appreciation? A Capability-Personalization Framework and a Meta-Analytic Review. Qin X, Zhou X, Chen C, et al. Psychological Bulletin. 2025;151(5):580-599. doi:10.1037/bul0000477. 

17. Dehumanization Risks Associated With Artificial Intelligence Use. Dang J, Liu L. The American Psychologist. 2025;:2026-09700-001. doi:10.1037/amp0001542.