Why MentorPRO Chose Human-at-the-Helm AI — and Why It Matters for Students

by Mia Lamont May 7, 2026

Why MentorPRO Chose Human-at-the-Helm AI — and Why It Matters for Students

Why MentorPRO Chose Human-at-the-Helm AI — and Why It Matters for Students

Human-at-the-Helm: Why MentorPRO Builds AI That Serves Mentors, Not Students

There is a version of AI-enhanced student support that is spreading rapidly across higher education right now. It looks like this: a student types a question into a chat window, a large language model generates a response, and a human reviewer spot-checks a small fraction of those exchanges after the fact. The vendor calls this human-centered.

The research calls it something else.

At MentorPRO, we made a different choice — a choice grounded in evidence, shaped by what students actually want, and validated in peer-reviewed research. We call it human-at-the-helm. And it changes everything about what AI can and should do in a mentoring relationship.

The Problem With Campus Chatbots

Colleges are facing genuine resource constraints. Median academic advisor caseloads exceed 300 students per advisor (Carlstrom and Miller, 2013), and over 60% of college students now meet criteria for at least one mental health condition (Lipson et al., 2022). The pressure to scale is real.

The dominant response has been to deploy student-facing AI chatbots for advising, financial aid, wellness, and increasingly for emotional support. The University of Central Florida, for example, reports that its Knightbot independently resolves 85% of student queries without any human contact. Vendors present this as a sign of technological sophistication. The research suggests it is a warning sign.

Growing evidence links overreliance on AI chatbots with:

  • Lower grades and cognitive disengagement (Abbas et al., 2024; Kosmyna et al., 2025)
  • Heightened loneliness and reduced socialization (Fang et al., 2025; Gerlich, 2025)
  • Feedback loops in which social isolation drives chatbot use, which deepens isolation (Luo et al., 2024)
  • Gains that dissipate over time, particularly when human accountability is absent (Xu and Ma, 2025)
  • Sycophantic validation loops that can reinforce harmful thinking patterns in vulnerable students (APA, 2025)

And since around half of all jobs are secured through social networks (Mouw, 2003; Rajkumar et al., 2022), a chatbot that substitutes for a human mentor does not just provide a worse experience. It actively withholds the social capital that determines long-term career outcomes. A chatbot might provide information about an internship, but it cannot vouch for a student’s character or work ethic to a potential advisor or employer.

“A chatbot might provide information about an internship, but it cannot vouch for a student’s character or work ethic to a potential advisor or employer.” — Rhodes, J. E. (under review). Applied Developmental Science.

The Human-in-the-Loop Fallacy

Most AI vendors offering student-facing tools claim they use a “human-in-the-loop” design. In theory, this means humans provide oversight of AI outputs. In practice, it means something much weaker. When a system handles millions of student interactions daily, human reviewers can monitor only a tiny fraction — and they do so retrospectively, after exchanges have already taken place.

The research identifies at least three structural failures in this model:

  1. Performative accountability — The appearance of oversight without substantive control (Elish, 2019)
  2. Automation bias — Human reviewers defer uncritically to AI recommendations, especially under high workload and time pressure (Goddard et al., 2012)
  3. Skill erosion — Supervisors who review exceptions rather than routine interactions gradually lose the professional judgment they are supposed to provide (Autor, 2015; Bainbridge, 1983)

These are not hypothetical concerns. They are documented patterns in the academic literature on AI governance, and they are now appearing in research on higher education specifically (Rhodes, under review, Applied Developmental Science).

Human-in-the-loop places AI at the center and humans on the periphery. MentorPRO does the opposite.

What Students Actually Think About AI in Mentoring

Before MentorPRO built its AI approach, the research team asked the people it would affect most directly. What do students actually want from AI in a mentoring relationship? The findings — from a peer-reviewed study currently under review — were unambiguous, and they validate every design decision MentorPRO has made.

Student Attitudes Toward AI-Assisted Peer Mentoring (under review)

Students rated traditional peer mentoring significantly higher than AI-assisted versions across all measures. Perceived logic decreased 23.4%, expected success decreased 16.3%, and confidence in recommending decreased 20.8% when AI assistance was introduced — all at p < .001.

Key findings from the student attitudes study (N = 60 undergraduates):

  • Logic ratings for mentoring dropped by 23.4% (Cohen’s d = 0.92, large effect) when students learned AI was involved
  • Expected success dropped by 16.3% (Cohen’s d = 0.54)
  • Confidence in recommending the program to a friend dropped by 20.8% (Cohen’s d = 0.52)
  • Students showed highest comfort with AI for career guidance (M = 3.84 out of 5)
  • Students showed lowest comfort with AI for mental health (M = 2.54) and social connection (M = 2.68) — exactly the domains where chatbots are being deployed most aggressively
  • When asked whether AI would help mentors provide better responses than mentors could provide alone, students disagreed, averaging M = 2.92 on a 5-point scale

Student comments were equally direct:

“I would feel tricked because I could easily just find AI and use it myself. I wouldn’t need a peer mentor at that point.”

“When it comes to peer mentoring, I feel like it is a way to be able to connect with someone about your experiences. It wouldn’t feel the same if they were using AI.”

These findings are not a reason to abandon AI in mentoring. They are a precise and actionable map for where AI belongs — behind the scenes, serving the mentor, invisible to the student.

The Human-at-the-Helm Framework

The human-at-the-helm model — described by Dr. Rhodes in a manuscript currently under review at Applied Developmental Science, supported by NSF Award No. 2450833 and the Axim Collaborative — is built on a simple premise: machines excel at processing data; humans excel at relationships.

The framework assigns each category of task to the entity best suited for it.

AI handles:

  • Synthesizing student background, goals, challenges, and check-in history into instant summaries
  • Surfacing evidence-based guidance from a curated library of peer-reviewed mentoring research
  • Identifying relevant campus resources in real time
  • Flagging patterns in engagement data that a mentor might miss across a large caseload

Humans handle:

  • Every direct interaction with a student
  • Interpretation, context, and judgment
  • Relationship-building, trust, and genuine care
  • Referrals, introductions, and the irreplaceable act of vouching for someone

By positioning AI as a tool for scaling — rather than replacing — human connection, the human-at-the-helm approach addresses resource constraints without sacrificing authentic relationships. The Check-In and Goal Setting features feed directly into this framework, giving mentors a continuously updated picture of each mentee before every conversation.

“This is not a compromise. It is the correct division of labor — one that the science of human development, the research on AI risk, and the expressed preferences of students all point to simultaneously.” — Rhodes, J. E. (under review). Applied Developmental Science.