TL;DR
This paper synthesizes the collective wisdom of 50 interdisciplinary experts to map out how Generative AI (GenAI) can revolutionize educational feedback. It positions AI not just as a tool for automation, but as a "Thermomix" capable of synthesizing complex pedagogical strategies into personalized, real-time dialogues while warning against the risks of student "cognitive offloading."
What Problem Does This Paper Address?
While feedback is a cornerstone of effective learning, providing high-quality, personalized guidance at scale is a massive challenge for human educators. Existing research on feedback is often "siloed" across different fields like educational psychology and computer science, leading to fragmented evidence and a lack of clear design principles.
With the sudden rise of GenAI, there is a gold rush to automate feedback, but this brings new risks: Will students stop thinking for themselves if the AI provides all the answers? Will they trust a machine's critique? The authors aim to bridge these disciplinary gaps and provide a roadmap for making AI-enhanced feedback engaging, effective, and ethically sound.
Methods
This paper is a meeting report summarizing the "Future of Feedback Symposium" held in June 2025. Rather than presenting a single experiment, it synthesizes the interdisciplinary perspectives of 50 scholars through structured small-group activities. The authors categorize the current state of the field into three pillars:
- Determinants of effectiveness (what makes feedback work).
- Engagement and perception (how students interact with and feel about feedback).
- Scalability (how to move from the lab to the global classroom).
Datasets
Not discussed in the paper. As a synthesis of a symposium, the report references numerous existing meta-analyses and studies (over 100 citations) rather than introducing a new primary dataset.
Key Insights and Ideas
- The "Thermomix" Analogy: The authors describe GenAI as a "Thermomix for feedback." Just as the kitchen appliance integrates cooking, baking, and stirring, GenAI can synthesize previously separate functions—like immediate guidance, psychological motivation models, and real-time engagement analysis—into a single, unified architecture.
- The Clark-Kozma Debate Revisited: The report touches on a classic educational debate: does the media (the AI) influence learning, or is it just the instructional method? The consensus is that the method must remain the driver, with AI acting as a facilitator to enable strategies that were previously impossible at scale.
- Cognitive Offloading: A major concern is "metacognitive laziness." If AI scaffolds are too effective, students might offload their critical thinking to the machine, undermining long-term learning gains and self-regulation skills.
- Credibility vs. Objective Quality: Research suggests that students' use of feedback depends more on their perception of the provider's credibility than the actual quality of the advice. If students perceive AI praise as "less authentic," its motivational impact may vanish.
- AI as an Experimental Tool: Beyond just delivering feedback, AI allows researchers to systematically vary features like tone, timing, and detail to scientifically map out what triggers student trust and engagement.
Key Results
The report outlines several consensus points for the future of the field:
- Elaborated > Simple: Meta-analyses confirm that "elaborated feedback" (explanations and hints) consistently outperforms simple "correct/incorrect" notifications. AI is uniquely positioned to generate this complexity rapidly.
- Human-AI Co-design: The experts emphasize that AI should not replace teachers but enhance them. They advocate for "human-in-the-loop" systems where educators help co-design and review AI-generated feedback.
- Multimodal Potential: AI can move beyond text to provide audio or video feedback, which could significantly improve accessibility and engagement for diverse learners.
Limitations and Future Work
- The "Black Box" Problem: The lack of interpretability in GenAI remains a hurdle. If a teacher or student doesn't know why the AI gave a specific piece of feedback, they are less likely to trust it.
- Data Sovereignty: The report raises critical questions about who owns learner data when commercial EdTech providers are involved.
- Call for "Megastudies": The authors call for international, high-powered, multi-site randomized controlled trials (RCTs) to identify how contextual factors (age, gender, subject matter) influence AI feedback effectiveness.
- Open-Source Infrastructure: To avoid corporate silos, the authors suggest developing open-source platforms for large-scale experimentation and the systematic deployment of feedback tools.
Summary updated