Risks
The documented ways AI-assisted learning can go wrong, ordered by how well each is established — from replicated patterns to claims that currently rest on almost nothing.
Established patterns
Assisted performance mistaken for learning
Large performance gains while AI is available do not certify learning: the assisted-vs-unassisted dissociation is replicated across subjects and designs, and is the reason this site badges every finding with its assessment conditions.
- Who it applies to
- Applies to both secondary and university learners; the size of the dissociation varies by intervention design.
- Generative AI without guardrails can harm learning: Evidence from high school mathematicsCluster-randomized trial · 2025 · Study tier 1 · Secondary direct · University different
+48%/+127% assisted practice against -17%/null unassisted exams; students did not perceive the harm.
- Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performanceRandomized trial · 2025 · Study tier 2 · Secondary different · University direct
Best essay improvement with the tool; no knowledge, transfer, or motivation differences without it.
- ChatGPT as a Learning Tool for Medical Students: Results From a Randomized Controlled TrialRandomized trial · 2025 · Study tier 3 · Secondary different · University partial
Open-resource quiz advantage vanished on the closed-book retest.
Signals under investigation
Dependence on AI scaffolds
When AI assistance is withdrawn, performance can fall back toward baseline — learners may rely on rather than learn from the assistance. Evidence comes from scaffold-withdrawal phases, not yet from long-horizon trials.
- Who it applies to
- Observed in university course deployments; untested in secondary settings.
- Impact of AI assistance on student agencyRandomized trial · 2024 · Study tier 2 · Secondary different · University direct
Withdrawal effects d 0.28-0.61 on platform metrics; checklist partially compensated; measurement circularity (High RoB).
Metacognitive offloading
Trace and self-report evidence suggests learners offload monitoring and evaluation to the AI, with lower mental effort during the task and weaker reasoning in some tool-assisted products. Causal, delayed consequences are not yet measured.
- Who it applies to
- Evidence is university-lab based.
- Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performanceRandomized trial · 2025 · Study tier 2 · Secondary different · University direct
Trace data: revision loop centered on ChatGPT with relatively fewer metacognitive processes (descriptive process mining).
- Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiryLaboratory experiment · 2024 · Study tier 2 · Secondary different · University direct
Lower cognitive load across all three subscales with the LLM; quality deficit fully mediated by reduced germane load (exploratory).
Widening gap for low-prior-knowledge learners
Higher-prior-knowledge learners appear to benefit more from AI assistance, and struggling learners' difficulties can compound — including an illusion of competence. Moderation evidence is thin and partly observational.
- Who it applies to
- Signals in secondary-age programming novices and university CS1 students.
- The Widening Gap: The Benefits and Harms of Generative AI for Novice ProgrammersQualitative study · 2024 · Study tier 3 · Secondary different · University direct · Partly vendor funded
Observational: struggling students' difficulties compounded; illusion of competence (n=21, no control).
- Studying the effect of AI Code Generators on Supporting Novice Learners in Introductory ProgrammingRandomized trial · 2023 · Study tier 2 · Secondary direct · University different
High-prior learners retained significantly more with Codex training; low-prior did not.
Poorly supported claims
Erosion of previously acquired skills
Claims that AI use erodes existing skills currently rest on cross-sectional correlations and small preliminary lab work; no longitudinal within-person evidence exists in either direction.
- Who it applies to
- Untested in both launch populations.