Artificial intelligence can crank up the difficulty on your child's homework in seconds, but asking it to simplify a problem often backfires. A new study shows that large language models are structurally wired to make assignments harder, not easier.
AI models struggle to dial down the complexity of schoolwork for beginners, even though they easily turn simple problems into advanced challenges. If your child is stuck on a concept, leaning on an AI tutor to explain it at an elementary level is more likely to confuse them than help.
Edtech companies sell AI tutors as infinitely patient coaches that can meet your child wherever they are. This paper shows that raw intelligence in an AI does not translate into teaching skill. An AI capable of solving graduate-level math or complex code cannot reliably calibrate an assignment downward for a beginner.
When a child asks a bot to "make this simpler," the model frequently hands back something that is still too abstract, or worse, technically inaccurate. Using AI to help a struggling student could quietly dig them into a deeper hole.
Researchers wanted to measure "educational control"—the ability of an AI not just to solve a problem, but to systematically calibrate its cognitive difficulty along Bloom's Taxonomy. While tech companies measure AI progress by how hard of a test a model can pass, parents and teachers need models that understand how learning actually builds step by step.
Evaluating 2,520 programming tasks across three major technical benchmarks revealed a striking "directional asymmetry" in how AI handles difficulty:
- It scales up, not down: The tested models reliably ramped up the cognitive demand of basic tasks into multi-step problems, but routinely failed when instructed to strip advanced tasks down to foundational concepts.
- Execution does not equal education: Being able to generate correct, complex output did not correlate with knowing how to teach that material to someone starting from scratch.
- Generalists beat specialists: General-purpose AI models showed a better grasp of difficulty progression than models trained specifically on coding, though both struggled with simplification.
Language models are trained on internet-scale text where technical concepts naturally build toward complexity, jargon, and density. Asking a model to add constraints, edge cases, and nuance matches its underlying predictive training. Asking it to subtract complexity while keeping the core concept intact requires pedagogical empathy—something token-prediction algorithms lack.
This paper is a preprint and has not yet undergone formal peer review. The experiment was tested exclusively on computer programming tasks using a single family of models (Qwen3), rather than across subjects like reading comprehension or history. Crucially, the researchers evaluated the models against technical benchmarks, not with real students in a classroom.
- If your child is bored and cruising through homework... use AI prompts to stretch them by asking for harder variations, added constraints, or real-world problem scenarios.
- If your child is completely stuck on a foundational concept... step in with a human explanation, textbook example, or teacher video rather than asking an AI to "explain it to a beginner."
- If an educational app promises AI-driven "adaptive difficulty"... treat the claim skeptically unless a human educator curated the progression of lessons behind the scenes.
- If you do use AI to break down a topic... verify the simplified version yourself first to make sure the bot did not introduce errors while attempting to dumb down the prompt.
Treat AI like an enrichment coach for a child who needs a challenge, not a remedial tutor for one who is falling behind. Until models learn how to scaffold downward reliably, breaking things down simply still requires a human brain.
Yi Zhang, Julia Rayz (2026). From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs. arXiv (preprint). — arxiv.org