• Traditional mentorship is eroding just as young people need it most. AI-guided advising offers a way to scale guidance and reflection, but only if it preserves the human depth that fosters growth.
• Embedding “reflective friction” in AI dialogue can restore critical thinking. By prompting youth to question, reflect, and plan, hybrid systems turn easy validation into opportunities for persistence, curiosity, and grit.
• Human-in-the-loop empathy training can teach what AI cannot feel. Pairing AI-guided reflection with human feedback strengthens emotional understanding and behavioral empathy among adolescents.
• AI should mediate, not replace, human trust. Automation can streamline logistics while mentors focus on what machines can’t replicate: authentic connection and care.

Before getting lost in the weeds exploring the challenges and opportunities of AI integration in mentorship, it’s worth stepping back and asking some important questions: What is the goal of mentorship, and why is it still important today? While to some degree mentorship is about guiding individuals to reach academic or professional goals, there’s also an important element of personal development involved. For vulnerable youth in particular, mentoring relationships are a protective factor, leading to greater self-esteem, mental health, and emotional regulation.⁵ According to a survey by MENTOR, a US non-profit developing and expanding youth mentorship, the process also promotes a sense of belonging among young people, which in turn is an important part of development.⁷

It’s clear, then, that advising is not just about choosing the right internship or gaining industry tips to get ahead; it’s about providing a safe space in which young people can explore who they are and what they want in life. Unfortunately, in countries like the United States, access to youth mentoring is eroding, despite growing youth mental health challenges.⁶ There are also various barriers to mentorship; youth from lower-income households or whose parents have less than a high school education are less likely to receive mentorship than students from more affluent or educated families.⁶ Now is as good a time as ever to look at ways to enhance, scale, and increase access to mentoring with the new technologies we have at our disposal. 

There are several approaches to integrating AI into mentorship, including technology-driven models where AI acts as the primary and only mentor, and human-in-the-loop (HITL) models where AI leads interactions and humans intervene only when the system flags a problem.³ Jean Rhodes, Director of the Center for Evidence-Based Mentoring at the University of Massachusetts, advocates for a “human-at-the-helm” approach.⁴ According to this model, AI is positioned as a sophisticated tool to enhance the work of human mentors, not replace them. They function as intelligent assistants that can make mentoring more accessible and evidence-based, while preserving the all-important human relationship aspect. 

Regardless of who is supporting whom (that is, whether there’s an AI or human in the driver’s seat), the main challenge in redesigning mentorship in the age of AI is ensuring that the same behavioral and psychological outcomes are achieved as in traditional models of mentoring. In other words, finding a balance between maintaining the deep, meaningful human interactions that foster personal development and growth and the need for greater efficiency, speed, and accessibility.

Good mentors are meant to push their mentees beyond their comfort zone, make them think out of the box, and make them feel uncomfortable, in the best possible way. This can mean encouraging someone to question their beliefs and assumptions, nudging them toward a harder path, or offering them space to embrace uncertainty. Why is this important? Because struggle and facing challenges are essential for building grit—the capacity to persist through setbacks—and curiosity, the willingness to explore unfamiliar territory. These traits, in turn, are key ingredients in academic and professional success, which is arguably one of the goals of mentorship.¹⁰ 

As they’re currently designed, conversational agents (CAs) rarely push back against or challenge users. Studies have observed that many CAs and “AI companions” are explicitly designed to maximize user satisfaction and emotional engagement.⁸ They do this by agreeing with users, mirroring their tone, and avoiding disagreement or challenge. This design logic produces what researchers have called sycophantic responses: interactions that feel warm and affirming but ultimately reinforce users’ existing views.⁹ From a behavioral science perspective, these responses tap into our confirmation bias (the preference for information that aligns with one’s beliefs) and cognitive ease, which is the brain’s tendency to favor low-effort processing. Not only does this dynamic make users more inclined to interact with AI (just think how addictive our algorithmically-tailored social media feeds are), but it can also undermine agency, reflection, and critical thinking, qualities that are central to personal growth. 

These shortcomings are part of broader anxieties around the cognitive impacts of human overreliance on AI dialogue systems. Although research in this area is in its infancy, evidence is beginning to emerge, particularly from academic settings. One meta study conducted by researchers in Australia found that students’ overreliance on AI tools involved accepting AI-generated recommendations without question.¹¹ This led to both errors in the task they were performing and diminished cognitive abilities, including decision-making and critical thinking skills. Experts at MIT have even warned that using ChatGPT to do simple tasks and make decisions can lead to “cognitive debt.”¹² The underlying reason for our affinity for AI is our increasing preference for fast and optimal solutions over slow ones that are constrained by practicality. In other words, we want a quick fix, not long-term challenges and struggles.

If AI mentorship is to foster real growth, it needs to do more than agree with the user. It must ask questions that challenge assumptions, slow down thinking, and build reflective capacity. The opportunity lies in designing hybrid mentorship models that deliberately reintroduce what is known as “productive struggle”: moments of difficulty that strengthen persistence, curiosity, and self-efficacy. Scientists from Stanford University’s AI Lab argue that productive struggle, a popular concept in education, is essential for preserving the way children learn in the age of AI,¹³ and it might just be the key to ensuring that mentorship remains effective too. 

For example, by embedding structured friction into conversational flows—prompts that ask “why,” surface alternatives, or present counterfactuals—AI systems can nudge users away from passive validation and toward active reflection. Research on motivational interviewing and Socratic questioning already demonstrates how open-ended, gently challenging questions increase engagement and internal motivation.¹⁴ Translating these techniques into AI dialogue patterns can help youth build the same reflective habits that effective human mentors cultivate.

We can learn a lot about designing AI systems in this way from the education sector, where CAs are already being used to prompt students’ deeper reflection when completing academic tasks. In higher education, for instance, researchers at the National Taiwan University of Science and Technology designed a chatbot-assisted learning system to explore how AI could be used to support self-regulated learning (SRL).¹⁵ Popularized by Barry Joseph Zimmermann in the late 1980s, SRL is an educational approach that requires students to set goals, monitor their progress, and reflect on their performance.¹⁶ 

In the study, the researchers programmed chatbots to deliver metacognitive prompts—questions that encouraged learners to plan, monitor, and evaluate their actions rather than simply receive information. Students who interacted with these reflective prompts demonstrated significantly higher levels of self-regulation, critical reflection, and learning persistence compared to those in traditional settings. The authors argue that the key benefit of the chatbot was not in providing answers, but in scaffolding the reflective process. That is, helping students slow down, think through their reasoning, and take ownership of their learning goals. In essence, the AI served as a mirror rather than a tutor, modeling the same kind of guided questioning that effective human mentors use to strengthen autonomy and resilience. 

Researchers at MIT are already starting to build AI systems that encourage perspective-taking and critical thinking among users.⁴ The idea is that the chatbot asks Socratic questions and prompts the user to reach out to real people in their lives for support and guidance, rather than validating their inputs. 

In short, the goal is not to make AI more human but to make it more human-supportive, engineering friction, curiosity, and self-questioning back into the loop. When mentorship technologies are built to prompt reflection instead of validation, they can help young people practice the very capacities that make mentorship transformative: the courage to pause, question, and grow.

Mentorship is fundamentally relational; it’s just as much about how we relate to others, interpret emotions, and build trust as it is about what to do. Traditional mentoring research shows that effective mentoring relationships depend on mutuality, trust, and empathy, allowing young people to process emotions and reflect on social situations.²⁰ Particularly in formative stages, young people often turn to mentors for the relational work of reflection, identity-building, and navigating interpersonal complexity.²¹ 

Yet in the rush to automate guidance, this relational dimension is too often lost. A recent study from the University of Cambridge found that when chatbots interact with children, they frequently display what is termed an “empathy gap”—the inability to respond appropriately to emotional needs or relational nuance—thereby placing young users at higher risk of confusion or distress.¹⁷ Another investigation by Cornell University and Stanford University determined that while conversational agents can mimic empathic responses, they significantly underperform humans when interpreting and exploring a user’s lived experience.¹⁸ While adults who grew up in a world where AI was once only accessible to scientists and tech experts are likely to understand the limits of AI’s empathy, younger, digital native generations may find it more challenging to perceive. 

In youth-mentoring and advising settings, many AI-driven systems still optimize for task completion and cognitive ease. For example, they focus on producing résumé feedback, presenting career options, or delivering scripted responses, rather than cultivating emotional attunement or perspective-taking. In other words, the system is built to minimize cognitive load and maximize measurable outcomes, so it systematically underweights the affective dimension of mentorship. Psychologically, when mentorship lacks emotion and connection, young people may develop what developmental scientists call instrumental reasoning—viewing social and career advice purely as means to ends, rather than part of a broader relational and personal-growth process. And while instrumental mentoring, which focuses on increasing competencies and skills, is effective in certain circumstances, developmental mentoring is often preferred as it prioritizes relationship building between mentor and mentee.¹⁹ 

Developmental psychology emphasises that empathy, trust, and attachment are vital to youth development, resilience, and long-term success. Without them, mentees may gain knowledge, but lack the socio-emotional scaffolding required to navigate uncertainty, collaborate effectively, or lead others. In short, when mentorship platforms rely on AI that cannot authentically engage with “how I feel” or “what someone else might be thinking,” they risk raising a generation that may be technically adept but lacks emotional intelligence. 

In this sense, the challenge is not only that AI mentoring is less empathic than human mentoring, it is that many systems are built without empathy as a design priority. As the field scales, it’s important not to leave out the relational half of mentorship entirely, given the risk of stunting young people’s emotional development. 

While AI itself might not be able to process human emotion the way we can, new technological breakthroughs suggest that it can help young people navigate and learn about human emotion in an age where face-to-face interactions are in decline. The key is human-in-the-loop frameworks where people play a central role in overseeing the algorithm’s responses and users’ interactions. 

At Tecnológico de Monterrey’s Instituto del Futuro de la Educación (IFE) in Mexico, researchers have taken an explicitly developmental approach to this problem through an AI tutor named Carla: a WhatsApp-based system designed to help adolescents practice emotional, cognitive, and behavioral empathy.²² Carla was created by a multidisciplinary team of educators, psychologists, and computer scientists in response to a growing concern about declining empathy among young people and the limits of traditional mentorship to address it at scale.²⁵ By combining behavioral science with accessible technology, Carla offers a way to make empathy training consistent, personalized, and widely available—something human-only programs struggle to achieve as demand for mentorship grows. 

Humans were embedded throughout the process of creating Carla. AI handled the conversational scaffolding, asking open-ended questions like “How do you think your friend felt?” or “What could you do differently next time?”, while trained facilitators reviewed the dialogues, provided follow-up discussion, and assessed growth. This human-in-the-loop design ensured that empathy was not reduced to sentiment analysis or keyword detection but remained a relational, context-sensitive skill.

Over 18,000 conversations were collected from participants aged 13 to 17. Pre- and post-assessments showed measurable gains in empathy across all three dimensions when AI-guided dialogue was paired with human feedback. The greatest improvements appeared in behavioral empathy—the capacity to translate understanding into action—suggesting that reflection plus gentle prompting can help youth internalize empathic reasoning.²²

Carla’s design works because it deliberately engages mechanisms known to foster empathy and prosocial behavior. By prompting users to take another’s perspective, it activates mental-simulation heuristics—the same cognitive process that underpins theory of mind.²³ Through repeated question-answer loops, it also employs self-explanation and implementation-intention framing, both linked to stronger empathic accuracy and moral reasoning.²⁴ 

Carla’s most radical insight is not technological but human. Empathy, like language or problem-solving, can be practiced through repetition and reflection. In a world where digital interactions increasingly mediate how young people learn to relate, tools like Carla remind us that AI can teach what it was never built to feel. Just because AI falls short in its ability to empathize with users doesn’t mean that humans shouldn’t leverage its benefits to help users develop this important trait. 

Mutual trust is widely recognized as a significant component of traditional mentoring relationships. This is because it provides emotional and psychological safety, which allow both mentor and mentee to use their cognitive energy for learning and productivity, rather than self-protection.²⁶ Building this trust takes time and perseverance, and often relies on other relational aspects such as perceived competence in one another, excellent communication, and even a sense of humor.²⁶

But as AI systems begin to mediate that exchange, many are discovering that while information can be automated, trust cannot. Despite most teenagers regularly using AI companions for relationship advice, entertainment, and information, several studies have shown that they don’t fully trust them. In a study of European 15-30-year-olds conducted by Opeepl, fewer than 2 in 10 respondents said they would fully trust AI with big life decisions, and 4 in 10 said they rejected the idea outright.²⁷ Common Sense Media’s study of youth in the United States found similarly low levels of trust, with over half of youth expressing distrust in the information or advice given to them by AI companions. This pattern, where AI appears dependable but fails to evoke emotional safety, creates functional confidence in the system without genuine trust.

From a developmental perspective, this matters deeply. Adolescence is a stage defined by social learning; young people form identity, values, and confidence through relationships that model reliability and care.²¹ When technology mediates these interactions, it can strip away the subtle cues that signal trustworthiness, like tone, empathy, and accountability. Over time, this may condition mentees to see guidance as transactional, eroding the motivational benefits of belonging and mutual respect.

The best use of AI in mentorship isn’t to replace the human guide, but to leverage it as a tool to protect their time and amplify their reach. Think of AI as the infrastructure of trust: it handles the matching, scheduling, and reminders that keep relationships alive, so mentors can spend their energy where it counts. This is not a new idea; in sectors such as healthcare and education, AI is being used to take care of the time-consuming and boring admin work so that people can free up more cognitive and physical capacity for their face-to-face work. Keeping humans in the driver’s seat doesn’t mean ignoring technology; it means using it to make the human parts of mentorship work better so that trust can still be cultivated.

A strong example of how this works is CareerVillage, a nonprofit that connects students with volunteer career mentors online.²⁹ The platform uses machine learning to match mentees’ questions with relevant professionals and to organize the advice archive so that responses remain timely and transparent. In this model, AI acts as a trust scaffold: it manages logistics, ensures consistency, and reduces the wait time between question and response, all of which reinforce behavioral cues of dependability. Human mentors, freed from the administrative burden, can focus on being present with their mentees through listening, empathizing, and tailoring feedback.

Behavioral research suggests that trust formation depends as much on structure as it does on sentiment. Systems that deliver predictable, timely interactions signal competence and care, key dimensions of trustworthiness.³⁰ This can be seen in organizational, educational, and healthcare settings; when structure is in place, we feel secure. By automating these reliability loops, CareerVillage ensures that even in large-scale digital environments, mentees experience the same consistency, responsiveness, and clarity that they would in traditional mentoring setups. 

Crucially, CareerVillage does not present AI as the mentor but as the mediator of mentorship. Algorithms facilitate matching, while mentors provide the authenticity and context that only humans can. This hybrid system restores the relational bandwidth that technology threatens to erode when mentorship programs are not designed properly. It also models a behavioral principle that could redefine digital mentorship more broadly: AI should handle the mechanics of trust so humans can handle its development. 

None of these interventions come without risk. Even the most thoughtfully designed systems can inherit the blind spots of their creators. AI mentors trained on incomplete or biased data may inadvertently reinforce stereotypes in the very youth they’re meant to empower. Human-in-the-loop models, while safer, still depend on mentors who have the time, resources, and digital literacy to engage effectively. Many of these conditions aren’t guaranteed in under-resourced contexts. 

This brings us back to the issue of accessibility. There is a fine line between using technology to scale and increase access to mentorship and embedding technology to the point that it starts to exclude marginalized or vulnerable groups. Organizations and companies leading the way in integrating AI into mentorship programs should pay close attention to the contexts in which both mentors and mentees live and work. 

AI-guided mentorship remains an experiment unfolding in real time. Across the three challenges explored here—AI’s tendency to flatter rather than stretch, its inability to model empathy, and its struggle to build trust—a common thread emerges: technology alone cannot replicate the conditions for human growth. What it can do, however, is make those conditions easier to access. From AI tutors like Carla that teach perspective-taking, to platforms like CareerVillage that scale trust through structure, we are learning that the most promising forms of AI are those that help people become more human, not less.

To achieve that, a few principles must anchor the path forward. First, any AI implementation should strengthen rather than replace human connection. Second, mentees should understand when and how AI assists their mentors, preserving trust and authenticity. Third, decisions that affect mentees’ well-being must remain under human control, with AI acting as adviser, not authority. Finally, these systems should democratize expert knowledge, giving mentors access to research, data, and best practices they might not otherwise have while keeping human judgment at the center.

The future of mentorship will not be written by algorithms alone. It will depend on how educators, designers, and policymakers choose to blend technology with empathy and efficiency with reflection. Behavioral science has a vital role to play in guiding that balance—testing interventions, measuring impact, and helping us design for trust at scale. Here at The Decision Lab, we specialize in translating insights from psychology and behavioral design into practical, human-centered solutions. Partner with us to explore how AI can expand, not erode, the relationships that help young people grow, reflect, and thrive.

What Can Behavioral Science Do For Your Virtual Skills-Based Program?

In digital learning environments, preserving the “human touch” is essential. In this article, Justin Pinnels and Dr. Maraki Kebede examine how behavioral insights can strengthen virtual education by sustaining motivation, reflection, and belonging—all core goals of hybrid human-AI mentorship.

Youth Mental Health is in Crisis. Can Behavioral Science Help? 

Learn more about how behavioral science tools like nudging and the EAST framework can make mental-health support more accessible, relatable, and effective for youth ages 12-24. 

1. Homer. (1998). The Odyssey (R. Fagles, Trans.). Penguin Books.
2. Common Sense Media. (2025, July 16). Parents’ ultimate guide to AI companions and relationships. https://www.commonsensemedia.org/articles/parents-ultimate-guide-to-ai-companions-and-relationships
3. Donner, M.-T., & Hummel, S. (2025). Systematic literature review of AI-based mentoring in higher education. ScaDS.AI Dresden, TUD Dresden University of Technology. https://doi.org/10.13140/RG.2.2.26529.52320
4. Rhodes, J. (2025, August 17). Human-at-the-helm: Imagining AI’s role in mentoring programs. Evidence-Based Mentoring. https://www.evidencebasedmentoring.org/from-human-in-the-loop-to-human-at-the-helm-reimagining-ais-role-in-mentoring-programs/
5. Martins, J., Moreira, T., Cunha, J., Núñez, J. C., & Rosário, P. (2024). Be SMART: Promoting goal setting with students at-risk of early school leaving through a mentoring program. Children and Youth Services Review, 157, 107423. https://doi.org/10.1016/j.childyouth.2023.107423
6. Annie E. Casey Foundation. (2024, January 24). Spotlight on youth mentoring: Trends, impacts and recommendations.https://www.aecf.org/blog/spotlight-on-youth-mentoring
7. MENTOR. (2023). Who mentored you? A study examining the role of mentoring across generations.https://www.mentoring.org/resource/who-mentored-you/
8. Bernardi, J. (2025, January 23). Friends for sale: The rise and risks of AI companions. Ada Lovelace Institute. https://www.adalovelaceinstitute.org/blog/ai-companions/
9. Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., … Perez, E. (2023, October 20). Towards understanding sycophancy in language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2310.13548
10. Davis, P. (2014, August 19). Grit: The secret ingredient to success. Psychology Today. https://www.psychologytoday.com/us/blog/pressure-proof/201408/grit-the-secret-ingredient-to-success
11. Zhai, C., Wibowo, S., & Li, L. D. (2024). The effects of over-reliance on AI dialogue systems on students’ cognitive abilities: A systematic review. Smart Learning Environments, 11(28). https://doi.org/10.1186/s40561-024-00351-8
12. Kovanovic, V., & Marrone, R. (2024, June 18). MIT researchers say using ChatGPT can rot your brain. The truth is a little more complicated. Melbourne Connect (republished from The Conversation). https://www.melbconnect.com.au/discovery/mit-researchers-say-using-chatgpt-can-rot-your-brain.-the-truth-is-a-little-more-complicated
13. Stanford Artificial Intelligence Lab. (2024, July 9). Productive struggle: The future of human learning in the age of AI.https://ai.stanford.edu/blog/teaching/
14. Mahoney, B. B., Oostdam, R., Nieuwelink, H., & Schuitema, J. J. (2023). Learning to think critically through Socratic dialogue: Evaluating a series of lessons designed for secondary vocational education. Thinking Skills and Creativity, 50, 101422. https://doi.org/10.1016/j.tsc.2023.101422
15. Lee, Y.-F., Hwang, G.-J., & Chen, P.-Y. (2025). Technology-based interactive guidance to promote learning performance and self-regulation: A chatbot-assisted self-regulated learning approach. Educational Technology Research and Development, 73, 2279–2304. https://doi.org/10.1007/s11423-025-09329-1
16. Zimmerman, B. J. (1989). A social cognitive view of self-regulated academic learning. Journal of Educational Psychology, 81(3), 329.
17. University of Cambridge. (2024, July 12). AI chatbots have shown they have an ‘empathy gap’ that children are likely to miss. https://www.cam.ac.uk/research/news/ai-chatbots-have-shown-they-have-an-empathy-gap-that-children-are-likely-to-miss
18. Fleischman, T. (2024, May 8). AI-generated empathy has its limits. Cornell Chronicle. https://news.cornell.edu/stories/2024/05/ai-generated-empathy-has-its-limits
19. Schenk, L., Sentse, M., Lenkens, M., Nagelhout, G. E., Engbersen, G., & Severiens, S. (2020).  Instrumental Mentoring for Young Adults: A Multi-Method Study. Journal of Adolescent  Research, 0743558420979123.https://doi.org/10.1177/0743558420979123
20. Spencer, R. (2006). Understanding the Mentoring Process between Adolescents and Adults. Youth & Society, 37(3), 287–315. https://doi.org/10.1177/0743558405278263
21. Rhodes, J. E. (2005). “A Model of Youth Mentoring.” American Journal of Community Psychology, 30(2), 145–159.
22. Treviño, R. (2025, June 11). AI tutor teaches empathy to young people through WhatsApp. TecScience. https://tecscience.tec.mx/en/education-and-humanism/ai-tutor-teaches-empathy/
23. Decety, J., & Cowell, J. M. (2014). The complex relation between morality and empathy. Trends in Cognitive Sciences, 18(7), 337–339. https://doi.org/10.1016/j.tics.2014.03.008
24. Batson, C. D., Ahmad, N., & Tsang, J.-A. (2023). Empathy and moral development: Integrating emotion and reason in prosocial motivation. Annual Review of Psychology, 74, 503–528. https://doi.org/10.1146/annurev-psych-032921-054439
25. Konrath, S. H., O’Brien, E. H., & Hsing, C. (2011). Changes in dispositional empathy in American college students over time: A meta-analysis. Personality and Social Psychology Review, 15(2), 180–198.
26. Thurman, Sabrina, Azul Bellot, and Tiffanie Grant. 2024. “Building Trust in Mentoring Relationships.” Center for Engaged Learning (blog), Elon University. Oct. 8, 2024. https://www.centerforengagedlearning.org/building-trust-in-mentoring-relationships.
27. Opeepl. (2025, August). Youth attitudes towards AI: Usage, trust, and fears in 2025. https://www.opeepl.com/youth-attitudes-towards-ai-usage-trust-and-fears-in-2025
28. Hoff, K. A., & Bashir, M. (2015). Trust in automation: Integrating empirical evidence on factors that influence trust.Human Factors, 57(3), 407–434.
29. CareerVillage. (n.d.). CareerVillage. https://www.careervillage.org/
30. 3fold Partners. (2025, June 11). How structure builds trust and stronger teams.https://3foldpartners.com/insights/2025/how-structure-builds-trust/