Sheet 12 of 17 · Research agenda
Open questions
The questions that matter most are about writing, thinking, relationships, persistence and careers. Almost none have longitudinal answers yet.
Fig. 12.1: Seven questions the evidence can't yet answer
Register 12.2: Evidence gaps
Next step
A longitudinal study of the Fall 2026 entering class
Most of these questions need the same students followed from matriculation through graduation and into work, measuring AI exposure, competency, independent learning, achievement, belonging, persistence and post-graduation outcomes. No such study exists in this evidence base.
Seven questions the evidence can’t yet answer
- Does sustained AI use affect independent writing ability? 12% of UK undergraduates put AI text directly into assessed work, and many students worry about their writing skills. No study tracks writing development over time.
- Does AI change how students develop critical thinking? Structured interventions show gains in higher-order thinking (g = 0.58); everyday use is unstudied.
- Does AI-mediated help-seeking affect relationships with faculty and peers? About 22% of teens’ first moves when stuck are AI-mediated. Effects on relationships are unmeasured.
- Does AI improve persistence among students with less prior preparation? Tutoring gains were largest for students with lower prior knowledge (IZA 2025). No persistence data.
- Do students from different socioeconomic backgrounds develop different levels of AI competency? Adoption rises with parental education (53% to 66%); competency itself is not measured.
- Does AI use change students’ sense of academic agency? Gen Z excitement about AI fell and anger rose between 2025 and 2026. Agency is not measured directly.
- Which AI competencies predict successful employment after graduation? Employers rank critical thinking first (92%); no study links student AI competencies to job outcomes.
Evidence gaps
- Longitudinal behaviour: How the same students’ practices change from high school to graduation.
- Independent learning: Retention, transfer and writing development when AI is unavailable.
- First-generation students: Little disaggregated national evidence on use, access or consequences.
- Intersectional outcomes: Who benefits across income, race, language, disability and discipline.
- Institution comparisons: Community colleges, regional publics, R1s, liberal-arts and online colleges.
- Graduate education: Research, authorship, coding and scholarly verification practices.
- Observed behaviour: How closely surveys match actual workflows, studied without surveillance.
- Policy effectiveness: Which rules improve understanding rather than stated compliance.
- Human-support substitution: When AI extends support and when it replaces relationships.
- Metacognition: Meta-analytic effect not significant (g = 0.078); evidence thin.
- Agentic & embedded AI: Research still centres on chatbots, not delegated workflows or LMS features.
- Long-term costs: Whether time saved exceeds time spent verifying, redesigning and resolving disputes.