Learning by Collaborating in the Age of AI: Research Mentorship for Polymathic, Workflow-Centered Science

If AI can run the analysis, what is a research mentor for? Zheng's answer is the end-to-end workflow: which question is worth asking, what evidence suffices, how the data were generated, when an AI-produced result is wrong, and how a finding means something in context. The chapter proposes Everyone's Backyard, a model in which disciplinary experts, trainees and community partners mentor one another in both directions, organized around four capacities — statistical thinking, data ethics, collaboration, and AI for social good and society. Its pilot, the Spring 2026 aiX Faculty Fellows Program at Columbia, brought 20 faculty fellows and 15 student interns from 14 schools and departments together for 17 projects across one semester. Evidence is descriptive, but the pattern is the interesting part: interns mentored faculty on tools while learning disciplinary context, and faculty mentored one another across fields.

Key Findings

• Technical execution is no longer the scarce part. Trainees can obtain definitions, examples and basic technical guidance independently, so mentoring shifts to judgment, evidence sufficiency, validation and interpretation.
• The unit of training is the whole workflow. Because a choice in data cleaning or prompting constrains what can be claimed at the end, stages cannot be mastered in isolation and early errors can pass unnoticed through later steps.
• Mentoring runs in both directions. Disciplinary experts need help adopting AI without losing expertise and Learner Agency, while AI-native trainees need help becoming workflow developers who can frame questions and recognize wrong results.
• Four capacities anchor the curriculum. Statistical thinking, data ethics, collaboration, and AI for social good and society organize the model rather than a tool-specific skill list.
• The pilot ran at scale across disciplines. 20 faculty fellows and 15 student interns from 14 schools and departments developed 17 projects spanning law, health sciences, language education, social science and policy, business, psychology and STEM education.
• Recurring contact, not a single training event. Across 105 weekly intern reflections in 15 program weeks, self-reported progress and discussion with the faculty lead dipped together in Week 8 and recovered, a descriptive association the author reads as support for steady touchpoints.

The Everyone's Backyard model

The design premise is that AI collapses the cost of the middle of a research project while leaving the ends expensive. Literature search, coding, cleaning, Visualization and drafting can now be assisted; framing a worthwhile question, judging when evidence suffices, diagnosing how data were produced, catching a wrong result and interpreting a finding in context cannot. A mentor who keeps teaching the middle is training for work the tools now do, while a trainee who never learns the ends inherits a workflow they cannot defend.

The model responds by treating the workflow as a design problem, decomposed into stages from problem formulation and stakeholder engagement through evidence synthesis, data acquisition, study design, AI-assisted analysis, validation, communication and reflection. Mentors and trainees specify where AI may assist, where a human check is mandatory, and where the failure modes concentrate. This is a Pedagogies and Teaching Strategies of design rather than of content coverage, and it explains why the "backyard" is shared: community partners bring problems and consequences, faculty bring disciplinary standards, and interns bring tool fluency.

What the pilot looked like

The Spring 2026 program ran from January to early May with a full-day bootcamp, bi-weekly cohort meetings, weekly intern studio hours, individualized faculty support and a public showcase where 13 teams presented. Within teams, the dominant collaboration pattern was a Translator model, with several pairs reaching a more equal Co-Pilot partnership: interns taught tools while absorbing disciplinary context, and faculty taught one another through peer-led tool demonstrations and shared troubleshooting.

Evidence came from end-of-semester faculty and intern reviews, weekly research reflections, project status surveys and a department-leader survey. Faculty rated the fellowship's effect on their professional network and community positively (n = 12) and rated bi-weekly cohort meetings as a meaningful source of connection (n = 13). The reflections give the program its useful diagnostic: progress and contact moved together, which suggests that reciprocal mentoring needs a rhythm rather than an orientation session.

Reading a descriptive pilot

The author is explicit that the analysis is descriptive. Participants self-selected, faculty reviews were anonymous and unmatched to baseline responses, and the weekly reflections are self-report. No comparison group exists, so the reported growth in faculty confidence and intern competence cannot be attributed to the model rather than to the cohort, the novelty of the program or the semester's support.

What survives that caution is the structural argument, which does not depend on the pilot's effect size: if AI absorbs execution, mentorship has to be re-specified around judgment, and the people best placed to teach tool fluency may be the trainees. For graduate programs and centers for teaching and learning, the transferable pieces are the four capacities, the workflow-decomposition exercise, and the pairing of a disciplinary expert with a tool-fluent partner. The AI Literacy implication is that faculty development is not a one-time upskilling, because the tools and the workflow both keep moving.

What this means for practice

• Decompose the research workflow before designing mentorship. Map stages, mark where AI may assist, and name the human checks that cannot be delegated.
• Pair in both directions deliberately. Assign a tool-fluent intern or trainee to a disciplinary expert and give both explicit teaching responsibilities, rather than treating the trainee as support staff.
• Protect recurring contact. Weekly studio hours carried the pattern in this pilot, and a bootcamp alone will not sustain workflow-level mentoring.
• Assess judgment, not tool use. Ask trainees to defend evidence sufficiency, catch a planted error and explain a limitation, which are the capacities AI does not supply.
• For faculty developers, run this as a cohort program with a showcase. The public artifact creates accountability, and the cross-disciplinary mix is what produced peer mentoring.

Limitations

• The evidence is one semester of one pilot (Spring 2026, Columbia University) with 20 fellows and 15 interns, and the analysis is explicitly descriptive.
• Participants were self-selected, faculty reviews were anonymous and not matched to baseline responses, and reflections and surveys are self-reported.
• No comparison group or pre-registered measures are reported, so growth in confidence and competence cannot be attributed to the model.
• The program depends on faculty time and intern labor at a research-intensive university, and transfer to teaching-focused or under-resourced institutions is untested.
• The four capacities are proposed rather than validated, and no instrument is offered for measuring them.

Citation

Zheng, T. (2026). Learning by Collaborating in the Age of AI: Research Mentorship for Polymathic, Workflow-Centered Science. arXiv:2610.34606.