Rethinking AI Evaluation in Education: The TEACH-AI Framework and Benchmark for Generative AI Assistants

AI summary

Overview

Research area: Human-centered AI evaluation, AI ethics and safety, educational technology, and learning sciences (cs.CY, with ties to HCI and AIED).

Technical level: Intermediate. The paper is conceptual and framework-driven rather than algorithmic or mathematical, but it assumes familiarity with HCI evaluation methods, learning theory (ZPD, constructionism), and generative AI capabilities. Educators and designers can engage with the toolkit portion without a technical background.

Scope: The paper proposes TEACH-AI, a ten-component benchmark framework and companion reflective checklist for evaluating generative AI assistants in educational settings along human-centered, pedagogical, and ethical dimensions rather than accuracy alone.

What This Paper Is About

Most evaluations of AI in education measure technical performance—accuracy, task completion, efficiency—while ignoring learner identity, agency, context, and ethics. The authors argue that these metrics fail to capture what actually makes an AI assistant effective and responsible in a classroom. Their goal is to define what "effective, value-aligned human–AI collaboration" means in education and to provide a usable instrument that designers, educators, and policymakers can apply to generative AI tools.

Key Contributions

1. The TEACH-AI framework: A domain-independent, pedagogically grounded benchmark built from ten core evaluation components—explainability, helpfulness, adaptivity, consistency, learning exploration, system usability, responsibility and ethics, accessibility, workflow and stakeholder coordination, and refinement—each decomposed into subcomponents, indicators/metrics, and supporting literature.

2. A literature-grounded synthesis: A scoping review of 126 sources spanning three historical phases (pre-LLM/pre-2017, transformer era 2017–2022, and the generative AI phase 2023–present) used to derive the recurring components of human-centered evaluation.

3. A practical toolkit: A preliminary set of reflective checklist questions mapped to each of the ten components, designed to be answered with simple Yes/No judgments or progressive scales, so the framework can be applied in classrooms, design reviews, and prototyping.

4. A bridging structure for human and automated evaluation: The framework is explicitly designed to accommodate both human raters and LLM-as-judge evaluation methods, positioning it as a foundation for scalable automated benchmarking in education.

Main Findings

• Existing evaluation is misaligned with education: Current benchmarks for large language models emphasize general reasoning or factual recall, with few targeting pedagogical efficacy in real learning contexts, and little stakeholder validation or alignment with teaching needs.

• Ten recurring evaluation components emerged from the literature: Iterative coding of 126 sources converged on the ten components above, which the authors group into three interconnected arguments: (1) agent capability (explainability, adaptability, helpfulness, consistency); (2) fostering creative exploration, engagement, and deep thinking; and (3) responsible, accessible, refinable operation.

• Traditional tutoring systems are rule-bound: Pre-generative Intelligent Tutoring Systems deliver individualized instruction through predefined decision trees and have shown positive outcomes via immediate feedback and adaptivity, but their reliance on rules limits responsiveness to dynamic, diverse student behavior—a gap generative tutors are positioned to fill.

• Evaluation approaches are typically domain-specific: UX and human-AI evaluation research tends to focus on narrow systems (e.g., coding platforms, writing tutors), leaving no standardized framework for cross-disciplinary environments like Scratch, Teachable Machine, or EarSketch.

• The framework is selective, not prescriptive: Components can be applied according to research goals, stakeholder roles, and context—single-agent studies might emphasize helpfulness or explainability, while multi-agent settings prioritize workflow and coordination; the checklist is explicitly framed as a tool for reflective practice rather than a to-do list.

• The current version is conceptual: The authors acknowledge TEACH-AI has not yet been empirically validated or tested at scale; the index and tech-eval depth are described as forthcoming.

Methodology in Plain English

The authors conducted a scoping review—a structured way of mapping a field rather than answering a narrow empirical question—following the framework established by Arksey and O'Malley. They asked one guiding question: "How are AI agents evaluated in educational environments?" They searched major venues (CHI, NeurIPS, IDC, AIED) and Google Scholar, ultimately reviewing 126 sources: 27 conference papers, 78 journal articles, and 21 books or gray literature items. They sorted these sources into three eras—pre-LLM (pre-2017, 37 papers), transformer era (2017–2022, 43 papers), and generative AI phase (2023–present, 36 papers)—then used iterative coding to identify recurring evaluation themes. Those themes became the ten components. The scoring structure for the toolkit was inspired by Donella Meadows' concept of leverage points, and weekly meetings with a senior faculty advisor were used to validate themes and refine interpretations.

Why This Matters

Research impact: The paper shifts the evaluation conversation from model performance to sociotechnical and pedagogical outcomes, offering a shared vocabulary that spans HCI, learning sciences, AI ethics, and AIED. It also provides a structured way to connect human evaluation with automated LLM-as-judge methods, which is an open problem in AI evaluation research.

Real-world applications:

• Classroom deployment review: Teachers and administrators could use the checklist to assess whether a chatbot or AI tutor is actually supporting learning, is accessible, and behaves ethically before or during adoption.
• Edtech product design and procurement: Companies and school districts could use the ten components as shared acceptance criteria when building or purchasing generative AI learning tools.
• Accessibility and equity audits: The accessibility and responsibility components give auditors concrete prompts for testing assistive technology integration, multimodal compatibility, fairness stress tests, and privacy compliance.
• Policy and standards development: Policymakers writing guidelines for AI in schools could adopt the framework as a value-aligned rubric for what "responsible AI in education" concretely requires.

Industry relevance: Edtech vendors, LLM providers targeting education markets, and school systems all face growing pressure to demonstrate that their AI tools are safe, equitable, and pedagogically sound. A domain-independent, stakeholder-aligned benchmark gives them a defensible structure for internal review, external reporting, and regulatory alignment—and could eventually feed into automated evaluation pipelines for model releases.

Future Directions

• Co-design and empirical validation: The framework needs testing with diverse stakeholders (teachers, students, administrators, developers) across varied educational contexts to verify that its components and indicators hold up in practice.

• Scalable digital prototype: The authors plan to integrate TEACH-AI into a digital tool capable of running large-scale benchmarks, moving it from a paper checklist to an operational evaluation platform.

• Automated assessment via LLM-as-judge and RLAIF: A key open question is whether the framework's components can be reliably scored by AI evaluators, which would enable continuous, low-cost monitoring of educational AI systems.

• Calibration and reliability questions: How to weight the ten components, how to handle disagreement between human raters, and how to adapt components per domain remain unresolved and are flagged as necessary future work.

Target Audience

Educators and instructional designers who need a practical way to judge AI tools before classroom use; HCI and AIED researchers studying human-centered evaluation; AI ethics and policy researchers focused on education and safety; edtech developers and product teams building generative tutoring assistants; and school administrators or district technology leaders responsible for procurement and governance decisions involving AI in learning environments.

Authors’ abstract

As generative artificial intelligence (AI) continues to transform education, most existing AI evaluations rely primarily on technical performance metrics such as accuracy or task efficiency while overlooking human identity, learner agency, contextual learning processes, and ethical considerations. In this paper, we present TEACH-AI (Trustworthy and Effective AI Classroom Heuristics), a domain-independent, pedagogically grounded, and stakeholder-aligned framework with measurable indicators and a practical toolkit for guiding the design, development, and evaluation of generative AI systems in educational contexts. Built on an extensive literature review and synthesis, the ten-component assessment framework and toolkit checklist provide a foundation for scalable, value-aligned AI evaluation in education. TEACH-AI rethinks "evaluation" through sociotechnical, educational, theoretical, and applied lenses, engaging designers, developers, researchers, and policymakers across AI and education. Our work invites the community to reconsider what constructs "effective" AI in education and to design model evaluation approaches that promote co-creation, inclusivity, and long-term human, social, and educational impact.