The ability of AI tools to accurately assess the cognitive complexity of educational tasks is a critical benchmark for their practical utility in classrooms. A new study reveals significant limitations in current models, finding they correctly classify math tasks only 63% of the time and exhibit systematic biases, raising urgent questions about their readiness to support instructional planning without sophisticated oversight.
Key Takeaways
- Eleven AI tools—six general-purpose and five education-specific—were evaluated on classifying math tasks by cognitive demand, achieving an average accuracy of only 63%.
- No tool exceeded 83% accuracy, and education-specific models performed no better than general-purpose ones like ChatGPT or Gemini.
- All tools showed a systematic bias, over-classifying tasks into middle categories (Procedures) and struggling with extremes (Memorization and Doing Mathematics).
- Error analysis found tools overweighted surface textual features and provided plausible but incorrect reasoning, a significant risk for novice teachers.
- The core failure was not ignoring task dimensions but incorrectly reasoning about multiple aspects simultaneously, highlighting a fundamental reasoning gap.
Evaluating AI's Ability to Classify Cognitive Demand in Math
The study aimed to test whether current AI tools could reliably perform a high-leverage teaching task: classifying the cognitive demand of mathematical problems using a established, research-based framework. This framework categorizes tasks into four levels: Memorization, Procedures Without Connections, Procedures With Connections, and Doing Mathematics. Researchers tested eleven AI tools, providing them with straightforward prompts to approximate how a time-pressed teacher might use them.
The six general-purpose models included ChatGPT, Claude, DeepSeek, Gemini, Grok, and Perplexity. The five education-specific tools were Brisk, Coteach AI, Khanmigo (Khan Academy's AI), Magic School, and School.AI. The aggregate performance was poor, with an average accuracy of 63%. Strikingly, the education-focused tools, ostensibly fine-tuned for pedagogical contexts, showed no advantage over their general-purpose counterparts.
The failure pattern was consistent and revealing. Tools exhibited a strong center-clustering bias, frequently misclassifying tasks at the highest (Doing Mathematics) and lowest (Memorization) levels into the intermediate procedural categories. Furthermore, when tools made errors in judging the broad level of demand (high vs. low), they consistently prioritized surface features—like the presence of key words or visual elements—over a deep analysis of the underlying cognitive processes required from the student.
Industry Context & Analysis
This study provides a crucial reality check amidst booming investment and hype around AI in education, a market projected to reach over $30 billion by 2032. Companies like Khan Academy (Khanmigo) and numerous edtech startups are aggressively marketing AI as a teacher's co-pilot for lesson planning and differentiation. However, this research suggests these tools may be generating a "plausibility trap"—providing confident, well-articulated rationales for incorrect classifications that could mislead educators, especially novices lacking the expertise to spot flawed reasoning.
The performance gap revealed here is stark when compared to AI benchmarks in other domains. While large language models (LLMs) can achieve over 90% on the MMLU (Massive Multitask Language Understanding) benchmark for general knowledge, they fall to 63% on this specialized pedagogical task. This underscores that raw linguistic capability and broad knowledge do not translate directly into expert-level disciplinary judgment, a challenge also seen in legal or medical AI applications where reasoning precision is paramount.
Technically, the error analysis points to a fundamental limitation in current transformer-based models. The failure stemmed not from a failure to identify relevant task dimensions, but from an inability to correctly reason about the interplay of multiple aspects to reach a holistic judgment. This is a known weakness in LLMs, which can struggle with multi-step, constrained reasoning tasks that require holding several variables in balance—precisely what expert teachers do intuitively. Unlike a retrieval-augmented generation (RAG) system that might simply fetch a similar problem, cognitive demand classification requires synthesis and applied theory.
The lack of outperformance by education-specific tools is particularly telling. It suggests that current fine-tuning on educational corpora or adding pedagogical guardrails may not be sufficient to build true "educational reasoning" models. This contrasts with domains like coding, where models fine-tuned on GitHub data (e.g., specialized variants of Codex) show marked improvements on benchmarks like HumanEval for code generation.
What This Means Going Forward
For educators and school districts, this study serves as a vital caution. AI cannot yet be trusted as an autonomous tool for curriculum adaptation or task classification. Its utility is currently limited to that of a brainstorming assistant, with all outputs requiring rigorous expert review. Professional development for teachers must now include "AI literacy" focused on identifying these plausibility traps and understanding the models' systemic biases, such as the tendency to default to procedural thinking.
For developers and edtech companies, the path forward is clear but challenging. Simply collecting more educational data for training is unlikely to solve the core reasoning problem. The next generation of educational AI will likely require novel architectures that explicitly model cognitive processes and pedagogical frameworks, moving beyond pure language modeling. Hybrid systems that combine LLMs with structured knowledge graphs of learning sciences research, or tools that force chain-of-thought reasoning through specific rubrics, may be necessary to break the 80-90% accuracy barrier required for trustworthy classroom deployment.
The market will shift as these limitations become widely understood. Tools that offer transparency—explaining not just a classification but the weight given to different task features—and those that are designed for human-in-the-loop collaboration will gain trust. The watchpoint will be whether any provider can release a model or feature that demonstrably outperforms the 83% ceiling found in this study, validated through independent, peer-reviewed research. Until then, the promise of AI as a true instructional partner remains on the distant horizon, awaiting a fundamental advance in its capacity for nuanced, context-sensitive reasoning.