Ways Generative AI Can Support and Threaten Learning in K–20 US Education

Synthesis: Kendeou, Greene and Nixon led 30 learning scientists through a preregistered Delphi consensus exercise that rated 65 learning processes for their likelihood of being enhanced by generative AI, their vulnerability to disruption by it, their importance to learning, and the strength of the evidence behind each judgment. Forty-two processes reached consensus. The pattern they call the GenAI Paradox is a sequence: initial acquisition of information is highly vulnerable, consolidation is significantly enhanced, and higher-order thinking is vulnerable again — so learners should acquire ideas first without GenAI, use it to strengthen what they have learned, then prioritize their own thinking when applying it. Risk in the higher-order, self-regulatory and social processes was one-sided, with enhancement ratings below threshold, and the report's most consequential meta-finding is that the experts rated these processes important while rating the evidence for the effects of GenAI on them as thin. The experts also judged elementary-level learners more vulnerable to disruption than secondary or higher-education students.

Key Findings

1. The paradox has a mechanism, and it is offloading. GenAI tempts learners to forgo the effort of acquiring information and the effort of thinking in a higher-order manner, while helping them consolidate what they have already acquired.
2. The foundations of memory are vulnerable, with one exception. Encoding reached consensus on vulnerability only (mean 3.30; 75% rated ≥3), as did Retrieval (3.30; 82%) and Information Acquisition (3.00; 75%). Mental Load was the category's single enhancement signal (3.40; 82%).
3. Higher-order thinking is at one-sided risk. Ten of eighteen items reached consensus on vulnerability only, including Creativity, Critical Thinking, Epistemic Agency, Epistemic Cognition, Epistemic Metacognition, Evaluation, Inquiry Learning, Intellectual Virtues, Learning From Diverse Perspectives and Sourcing. Critical Thinking was the strongest single vulnerability in the report (mean 4.00; 96% rated ≥3; enhancement 2.40) and Intellectual Virtues the weakest enhancement (2.10; 32%).
4. Four higher-order processes are enhancement-only exceptions. Brainstorming (3.60; 79%), Curiosity (3.40; 82%), Debugging (3.50; 79%) and Information Search (3.10; 79%) reached consensus on enhancement but not vulnerability.
5. Self-regulation is where using GenAI well is hardest. Of the seven processes in that category, only Help-Seeking reached consensus on enhancement (3.70; 86%). Metacognition (3.50; 86%), Self-Regulated Learning (3.50; 93%) and Motivation Regulation (3.10; 75%) reached consensus on vulnerability — so the skills needed to use GenAI productively are among the ones experts think it erodes.
6. The social fabric of learning shows disruption with no countervailing gain. Five collaboration processes reached consensus on vulnerability only: Social Connection (3.80, against enhancement 1.90), Socio-Emotional Development (3.60, against 2.00), Social Learning (3.30), Collaborative Learning (3.30) and Sense of Belonging (3.30). Mean enhancement ratings across the cluster were consistently low.
7. Affective-motivational processes are vulnerable, and motivation is double-edged. Agency (3.60; 79%), Engagement (3.50; 82%) and Persistence (3.40; 79%) reached consensus on vulnerability only, while Motivation reached consensus on both (enhancement 3.30, vulnerability 3.20) — personalized instruction can strengthen it, replacement of learner effort can starve it.
8. Writing is the only domain-learning process the experts agreed on, and it is high risk (3.80; 93% rated ≥3). Math Learning, STEM Reasoning, Reading to Learn and Language Proficiency did not reach consensus at all.
9. Elementary learners were judged more vulnerable than secondary or higher-education students, even though the same processes were endorsed across all three levels — the risk is the same in kind and more acute in younger learners.
10. Enhancement concentrates in study strategies, and it is the strongest signal in the report. Learning From Feedback (4.00; 100% rated ≥3), Learning From Examples (3.90), Tutoring (3.90), Self-Testing (3.80; 89%), Learning From Questioning (3.70), Scaffolding (3.60) and Deliberate Practice (3.40) all reached consensus on enhancement with low vulnerability ratings.
11. The evidence does not support the confidence. Across the 42 consensus processes, importance ratings were high (most above 3.0 of 4) while evidence ratings were low (most below 2.5) — the report's own framing of this meta-finding is "big feelings, little evidence". The authors attribute the gap to funding that prioritizes tool development over rigorous study, and to research incentives that reward speed.
12. Eleven recommendation themes, in three tiers. High impact and high feasibility: Brainstorming and Generative Ideation Support, Adaptive GenAI Scaffolding, and GenAI Literacy. Above the midpoint on both: Epistemic Agency First, Engagement Struggle Process Over Product, GenAI as Critical Partner, Critical GenAI Literacy, and Norm-Guided GenAI Practice. Highest value but hardest: Competency-Gated GenAI Use, Authentic Assessment Redesign, and Inclusion and Equity-Centered GenAI Interaction.

The paradox, stated precisely

The sequence matters more than any single rating. GenAI threatens the processes that commit ideas to memory, enhances the middle step where learners strengthen and consolidate those ideas, and threatens the higher-order thinking needed to use them in new contexts. GenAI's facility at provoking productive thought appears limited to comprehension and consolidation, not higher-order thinking. That is also why the report resists the question everyone asks — its authors argue that "is GenAI good or bad" is the wrong question, given that effects track use.

Where the experts' optimism stops

The consensus on higher-order thinking is the report's sharpest departure from other 2026 syntheses. The OECD's Digital Education Outlook and Fesler et al. both allow that pedagogical design can recover some of the deep-process risk; these experts, at the level of consensus, were "not yet optimistic", while conceding that the empirical literature is too thin to settle the question either way. The collaboration cluster is the one place where the report finds disruption without any countervailing enhancement, and the guidance section explains the asymmetry that makes it a design target: collaborative problem solving is anchored to a shared product, so partners can tell whether they are progressing, while collaborative learning has no such visible referent and can erode undetected.

How the consensus was produced

The procedure was preregistered on AsPredicted before data collection, with the anonymized rating dataset, instruments and analysis code promised to the Open Science Framework, and all contributing experts disclosing financial interests in AI and educational technology companies. Experts were recruited from the Oregon Learning Sciences meeting and by the lead authors' nomination, against criteria of active research in learning sciences, cognitive or educational psychology, a peer-reviewed publication record, and experience with educational applications of technology, with representation sought across cognitive science, educational technology, K–12 and higher-education US practice, and equity and accessibility.

In the first survey, 27 experts plus three lead authors nominated up to ten learning processes each, with a short description, roughly three supporting references, the levels where each applied, and whether it came from the nominator's own research. The lead authors consolidated the nominations into 65 processes in eight categories and 11 recommendation themes. In the second survey 29 participants rated every process on Enhancement Likelihood (1–5), Vulnerability Likelihood (1–5), Importance to Learning (1–4) and Strength of Evidence (1–4), and rated the recommendations for perceived impact and implementation realism. Consensus required a mean of at least 3.0 and at least 75% of respondents rating 3 or higher for enhancement and vulnerability, and 75% rating 3 or 4 for importance and evidence — a double criterion that keeps a high mean from standing in for a divided panel of experts. The report defines GenAI narrowly as general-purpose large language models such as ChatGPT, Claude and Gemini, excluding GenAI features embedded in other applications, though its authors expect the recommendations to transfer.

What this means for practice

• Instructors. Sequence GenAI to the phase of learning: keep initial encoding and retrieval with the learner, use GenAI for quizzing, feedback and probing explanations, and require the learner's own reasoning when ideas are applied in new contexts.
• Instructors. Protect productive struggle explicitly — Learning From Feedback was rated the most likely enhancement in the report (mean 4.00, 100% rated ≥3) and struggle-dependent processes the most vulnerable, so feedback without effort leaves the risk untouched.
• Course and program designers. Treat collaboration as a design problem rather than a delivery default: make the process of thinking, regulating and collaborating visible, since collaborative learning can erode without a visible product to reveal it.
• Assessment designers. Put process evidence alongside products, which the report's authentic-assessment recommendation asks for, because the same submitted artifact can represent either the work or its outsourcing.
• Administrators, policymakers and funders. Calibrate policy to the process-level findings rather than to blanket permission or prohibition, and read the evidence ratings as the report's audit of the field: importance above 3.0 of 4 with evidence below 2.5 on the same processes leaves open what GenAI supports, for whom and under what conditions.

Limitations

• Consensus is expert opinion, not evidence. The panel's central meta-finding is that the evidence behind their own ratings is thin, with most consensus processes rated below 2.5 of 4 — so it maps where experts expect risk, not where effects are measured.
• Twenty-three of the 65 processes reached no consensus at all. Their status is unresolved rather than null, and the unresolved set includes Math Learning, STEM Reasoning, Reading to Learn and Language Proficiency, which are exactly the domain questions educators ask about.
• The panel is small, invited and concentrated in one national context. 30 experts entered the study of whom 28 provided quantitative ratings, drawn from one meeting and the lead authors' nominations, and their judgments are about US K–12 and higher education rather than about education generally.
• The ratings are predictions about the future, made in a moving target. The report's scope is general-purpose large language models as they stood during the project, so ratings could date with the models even if the underlying processes do not.
• Interest disclosure is a statement, not a control, and level claims rest on one comparison. The report records that all contributing experts disclosed financial interests in AI and educational technology companies without testing whether ratings varied with them, and its finding that elementary learners are more vulnerable comes from the experts' initial ratings by level rather than from process-level results by level.
• The recommendations are rated for realism by the panel that proposed them, and the report's convergence with other 2026 syntheses is argued in prose rather than established by pooling their estimates with these ratings, so both the feasibility tiers and the agreement story describe expert judgment rather than tested implementation.

Citation

Kendeou, P., Greene, J. A., Nixon, N., Choi, J. Y., Fulsher, A., Choi, J., Aguilar, S., Allen, L., Azevedo, R., Bernacki, M., Butler, A., Carpenter, S., Crompton, H., Doroudi, S., Fiorella, L., Fyfe, E., Graesser, A., Hutt, S., Karumbaiah, S., Kizilcec, R., Lee, V., McNamara, D., Miele, D. B., Mills, C., Muis, K. R., Roschelle, J., Roscoe, R. D., Stewart, A., Walker, E., Wang, X. C., Witherby, A., Xu, Y., & Zhai, X. (2026). Ways generative AI can support and threaten learning in K-20 US education (Expert Consensus Report). University of Minnesota, College of Education and Human Development.