The care-full craft of feedback in an age of generative AI

Synthesis: A conceptual/position paper arguing that feedback in an age of GenAI must be understood as "matters of care" — ethical, relational practices rather than information transmission. It builds on a ten-principle Manifesto for Feedback in the Age of GenAI (Winstone et al. 2025, Copenhagen Feedback Symposium) and distils four core values for integrating GenAI into a Multimodal AI feedback landscape: (1) feedback processes should support meaning-making, (2) build educative relationships, (3) be trustworthy, and (4) be respected as a professional craft.

Core arguments

• Feedback as more than comments. Feedback information ("bits and bytes") is only input; without student sense-making and uptake it does not constitute feedback. GenAI comment-generators risk replicating Sadler's "dangling data" critique — and may entrench an information-centric view the field spent decades moving past.
• "AI slop" and "human slop." GenAI's mass-produced, speedy content (Madsen & Puyt 2025's "AI slop") has a human analogue: comment banks, quick marks, and standardized templates that already weaken feedback practice. Quantity and speed ≠ quality.
• Relational recognition. Only human feedback can be recognitive — "the mutual acknowledgment of agency, vulnerability, and shared humanity" (Corbin, Tai & Flenady 2025); GenAI feedback is "extra-recognitive." Students in Otaki et al. (2026) described GenAI interactions as "each interaction feels like a new beginning," lacking the continuous timeline of human relationships.
• The safe-space paradox. GenAI feedback reduces perceived personal risk (asking "stupid" questions, exposing weaknesses) — a genuine benefit, especially in power-hierarchical cultures — but may come at the cost of priming students for the friction of professional feedback encounters.
• Epistemic trust. Trust is dynamic and relational ("a verb, not a noun"): students ascribe less trust to AI-generated than human feedback (Ruwe & Kuklick 2026; Henderson et al. 2025 — see Comparing Generative AI and teacher feedback: student perceptions of usefulness and trustworthiness), algorithm aversion makes AI trust decline sharply after errors, and GPT-4 limitations (outdated data, over-generalization, hallucinations) strike at expertise, integrity, and benevolence — the dimensions of epistemic trust.
• Feedback as professional craft. Tuck's ethnography shows marking is non-linear, context-sensitive, and entangled with "readings of students" — including their "back stories." Multimodal feedback (written + dialogic + digital) is where this craft lives; GenAI lacks genuine dialogue. Risk: outsourcing feedback production erodes the craft (skill atrophy, "unthinking") and even human-in-the-loop approval can internalize AI bias.

Agenda for the future — four respects

• Respect for scholarship: design GenAI integration starting from known feedback challenges (e.g. ipsative feedback across modules) rather than from comment generation; research beyond self-report.
• Respect for equity: equitable access to meaningful feedback encounters; scaffold feedback literacies and evaluative judgment to avoid amplifying Matthew effects; respect conscientious objectors to GenAI.
• Respect for professional craft: distinguish what GenAI cannot replicate (relationship-rich, multimodal, dialogic feedback); protect against de-prioritizing care-full feedback because it is labor-intensive.
• Respect for human connection: preserve situated, meaningful, care-full elements of feedback encounters; study how trust, connection, and engagement evolve as GenAI enters the ecosystem.

What this means for practice

• Instructors. Treat generated text as feedback information rather than feedback: design the sense-making step — a dialogic follow-up, a comparison task, an explicit action plan — because "bits and bytes" without uptake do not constitute feedback.
• Instructors. Start GenAI integration from a known feedback challenge (for example, ipsative feedback that shows progress across modules) instead of from comment generation, which is where mass-produced speed invites the same weaknesses as comment banks and quick marks.
• Instructors. Keep the dialogic and Multimodal AI encounters human-led and in the loop: GenAI lacks genuine dialogue, and outsourcing comment production risks eroding the craft even when a human approves the output.
• Instructors. Coach epistemic vigilance rather than blanket trust or distrust: students ascribe less trust to AI feedback than to human feedback and algorithm aversion makes that trust fall sharply after errors, so pair AI comments with human feedback and evaluative judgment tasks.
• Instructors. Protect care-full feedback from being de-prioritized because it is labor-intensive, and respect students who conscientiously object to using GenAI at all.

Limitations

• A conceptual/position paper (theoretical analysis) built on a ten-principle symposium manifesto and a synthesis of others' research: it conducts no study of its own, so it cannot show that the four respects (scholarship, equity, craft, human connection) improve feedback uptake.
• Its trust evidence is perception-based and largely self-report — students rating identical comments labeled human versus AI (Ruwe & Kuklick 2026) — and the headline comparison of 90.5% versus 60.1% trustworthiness comes from one Australian survey (Henderson et al. 2025), not from behavior in a feedback encounter.
• Constraints it attributes to GPT-4 (outdated training data, over-generalization, repetition and verbosity, hallucinations) are drawn from other authors' reviews and are not measured against student learning in this paper.
• The claim that feedback is a professional craft rests on a single ethnography of academics' marking (Tuck's research), with no cross-institutional or cross-disciplinary comparison offered.

Citation

Winstone, N. E., Gravett, K., Bearman, M., Noble, C., Jensen, L. X., Jones, A., & Nicola-Richmond, K. (2026). The care-full craft of feedback in an age of generative AI. Assessment & Evaluation in Higher Education, 51(5), 911–927