Artificial intelligence tools show significant limitations in evaluating the cognitive complexity of mathematics tasks, raising critical questions about their immediate utility in classroom planning and curriculum development. A new study reveals that both general and education-specific AI models frequently misclassify tasks, demonstrating a systematic bias toward middle-difficulty levels and often providing misleadingly confident explanations that could misguide educators.
Key Takeaways
- Eleven AI tools, including six general-purpose (ChatGPT, Claude, DeepSeek, Gemini, Grok, Perplexity) and five education-specific (Brisk, Coteach AI, Khanmigo, Magic School, School.AI), were tested on classifying math tasks by cognitive demand.
- Average accuracy across all tools was only 63%, with no single tool exceeding 83% accuracy. Education-specific tools did not perform better than general-purpose models.
- All tools struggled most with tasks at the extremes of cognitive demand (Memorization and Doing Mathematics), showing a bias toward middle categories (Procedures with/without Connections).
- Error analysis revealed tools consistently overweighted surface textual features over underlying cognitive processes and showed weaknesses in reasoning about what makes a task high or low demand.
Evaluating AI's Ability to Classify Math Task Complexity
The study aimed to determine whether current AI tools can reliably classify the cognitive demand of mathematical tasks, a core competency for teachers adapting curricula. Researchers tested eleven prominent AI tools using a research-based framework with four levels of cognitive demand: Memorization, Procedures Without Connections, Procedures With Connections, and Doing Mathematics. The goal was to approximate the performance a teacher might achieve using straightforward, practical prompts rather than highly engineered ones.
The results were sobering. On average, AI tools accurately classified tasks only 63% of the time. The best-performing tool reached just 83% accuracy, leaving a significant error rate. Notably, the five tools marketed specifically for education—Brisk, Coteach AI, Khanmigo, Magic School, and School.AI—showed no advantage over general-purpose models like ChatGPT or Claude. This suggests that current "education-specific" fine-tuning may not adequately address the nuanced reasoning required for pedagogical classification.
A clear pattern of systematic bias emerged. All tools performed poorly on tasks at the framework's extremes—the routine Memorization tasks and the complex, non-routine Doing Mathematics tasks. Instead, they exhibited a strong tendency to misclassify these into the middle categories of Procedures With or Without Connections. Furthermore, the AI tools often generated plausible-sounding rationales for their incorrect classifications, which the study notes would be particularly persuasive to novice teachers lacking deep content pedagogy knowledge.
Industry Context & Analysis
This study arrives amid a surge of investment and adoption of AI in education technology. Companies like Khan Academy (with Khanmigo) and a host of startups have raised significant funding—Khan Academy's AI initiative is backed by over $10 million from donors like Microsoft—promising to reduce teacher workload. However, this research reveals a stark performance gap between marketing promises and pedagogical capability. Unlike benchmarks for code generation (HumanEval) or general knowledge (MMLU), there is no standardized public benchmark for educational task classification, making direct model comparisons difficult and allowing claims to go largely unverified.
The failure of education-specific models to outperform general ones is particularly telling. It indicates that current fine-tuning methods, which often rely on curated educational datasets, may not be capturing the deep, situated reasoning a master teacher employs. This contrasts with successes in other specialized domains where fine-tuning on high-quality data has led to superior performance, such as in medical or legal applications of large language models. The AI's tendency to rely on surface features (e.g., keyword spotting for "explain" or "calculate") mirrors known limitations in general-purpose LLMs that lack true reasoning about intent and context.
The systematic bias toward middle categories is a critical failure mode for classroom application. Misclassifying a simple memorization task as a procedure with connections could lead a teacher to waste valuable instructional time. Conversely, misclassifying a high-demand "Doing Mathematics" task as a low-level procedure could result in missed opportunities for deep student engagement and critical thinking. This error pattern suggests the AI models are essentially performing a form of textual pattern matching rather than the cognitive task analysis required.
What This Means Going Forward
For educators and school administrators, this study serves as a crucial caution. While AI can be a valuable assistant for generating ideas or drafting content, it currently cannot be trusted as an autonomous auditor of curricular rigor or cognitive demand. Teacher professional development must now include "AI literacy" components that focus on critically evaluating AI output for pedagogical soundness, especially for novice teachers who are the target market for many of these tools.
For edtech developers and AI companies, the path forward is clear. Simply packaging a general-purpose model with an educational interface is insufficient. Significant research and development is needed to create models that genuinely understand pedagogical frameworks. This will likely require novel training approaches, such as reinforcement learning from human feedback (RLHF) with master teachers, or the development of hybrid systems that combine LLMs with structured knowledge bases of curriculum standards and learning sciences research.
The immediate watchpoint is whether this research catalyzes the creation of a public, standardized benchmark for educational task classification. Such a benchmark would drive healthier competition and transparency in the edtech AI market. In the short term, the most effective use of AI in lesson planning may be in a collaborative, human-in-the-loop role, where the teacher uses the AI's output as a starting point for their own expert judgment, rather than as a final authority. The tools that succeed will be those that augment, rather than attempt to replace, the irreplaceable pedagogical reasoning of a skilled educator.