Holmes Bialik Fadel Artificial Intelligence in Education 2019

Most education technology writing ages badly because it describes products. The 2019 work by Wayne Holmes, Maya Bialik and Charles Fadel, published through the Center for Curriculum Redesign as Artificial Intelligence in Education: Promises and Implications for Teaching and Learning, has held up because it argued about purpose rather than tools.

Quick Answer: Holmes, Bialik and Fadel argued in 2019 that artificial intelligence in education should be evaluated by whether it improves learning and supports teachers, not by technical novelty. They distinguished AI used to deliver existing curricula more efficiently from AI that prompts schools to reconsider what should be taught at all.

How WebPeak Builds Learning Platforms Around Pedagogy

Educational platforms frequently digitise a textbook and add a quiz engine, which automates delivery without improving learning. WebPeak, a worldwide full-service digital agency, approaches learning products from the instructional model outward: mapping the intended learning progression first, then designing the data model, feedback loops and teacher-facing views that support it. Their teams build the analytics educators actually need — where a cohort stalled, which misconceptions recur — rather than engagement dashboards that measure time spent. That distinction between measuring activity and measuring learning is precisely the one the 2019 authors emphasised. Institutions building such platforms typically need MERN stack development for the application and AI services for adaptive components, both available through WebPeak.

The Central Distinction the Authors Drew

The framework's most durable contribution is separating two very different ambitions that get conflated in education technology discussions.

The first is using artificial intelligence to teach existing curricula more effectively — intelligent tutoring systems, adaptive practice, automated assessment and dialogue-based tutors. These accept the current curriculum as given and try to deliver it better, with faster feedback and pacing adjusted to individual learners.

The second is using the existence of artificial intelligence as a reason to reconsider curriculum content itself. If certain cognitive tasks can be automated, the argument runs, education should shift emphasis toward capabilities that remain distinctly human and toward the literacy needed to work alongside these systems. This is a curriculum design question, not a technology deployment question, and the authors treated it as the more consequential of the two.

They were also consistently sceptical of claims that technology would replace teachers, arguing instead for augmentation: automating administrative and assessment burden so educators spend more time on the relational and diagnostic work that machines handle poorly. That framing directly informs how institutions should evaluate the tools available now, including the study-path questions raised in this guide to choosing AI programs.

What the Framework Asks Schools to Evaluate

Translating the argument into procurement and classroom practice produces a short list of questions worth asking of any educational AI tool.

• What learning outcome does this improve? Named specifically, and measurable in a way that is not simply engagement time.
• What does it free teachers to do instead? If it adds work rather than removing it, the augmentation case fails.
• How does it handle learners it models poorly? Adaptive systems perform worst on students whose patterns are least represented in their training data.
• What data does it collect and who can access it? Student data governance is a first-order concern, not a procurement footnote.
• Does it make pedagogical assumptions explicit? Every adaptive system encodes a theory of learning, and that theory should be inspectable.
• Can teachers override it? Systems that cannot be overruled by a professional judgement deskill the profession over time.

Categories of Educational AI and What Each Actually Delivers

The 2019 categorisation still maps cleanly onto the tools schools encounter today.

Category	Function	Strongest Evidence Base	Main Risk
Intelligent tutoring systems	Step-level guidance in structured domains	Mathematics and procedural subjects	Poor transfer to open-ended subjects
Adaptive practice platforms	Adjusts difficulty and sequencing	Retrieval practice and fluency	Narrows curriculum toward measurable items
Automated assessment	Scores responses at scale	Structured and short-answer formats	Rewards features correlated with quality, not quality
Dialogue-based tutors	Conversational explanation and questioning	Emerging and uneven	Confident incorrect explanations
Teacher support tools	Planning, grading and administrative load	Time savings widely reported	Quality drift without review discipline
Learning analytics	Identifies at-risk learners	Institutional retention contexts	Proxy measures encoding existing inequities

What Has Changed Since 2019 and What Has Not

The technical landscape changed substantially after 2019 with the arrival of widely accessible large language models, which made dialogue-based tutoring and automated feedback dramatically more capable than the systems available when the work was written. The authors could not have anticipated the specific capability jump. What did not change is the evaluation problem. A more fluent tutor is not automatically a more effective one, and the field still lacks the volume of rigorous classroom trials needed to distinguish tools that improve learning from tools that improve satisfaction. The 2019 insistence on pedagogical grounding reads as more relevant now, not less, because fluency makes weak pedagogy harder to detect.

In practice, schools that apply the framework well run small controlled pilots with a comparison group, measure a named learning outcome over a term, and require teachers to document where the tool failed. Those that skip this tend to renew licences based on usage statistics, which measure adoption rather than benefit. The same evidentiary caution applies to claims about AI capability generally, a theme running through this discussion of realistic AI governance.

Key Takeaways

• The 2019 framework separates using AI to teach existing curricula better from rethinking what curricula should contain.
• Augmentation of teachers, not replacement, was the authors' central position and remains the practical standard.
• Every adaptive system encodes a theory of learning, and schools should require that theory to be made explicit.
• Fluent AI tutors are not automatically effective ones, and fluency makes weak pedagogy harder for reviewers to detect.
• Usage statistics measure adoption, not learning, and should never be the basis for renewing an educational licence.

Frequently Asked Questions

What is the main argument of the Holmes, Bialik and Fadel work?

That artificial intelligence in education must be evaluated against learning outcomes and pedagogical grounding rather than technical capability. The authors also argued that AI's existence should prompt reconsideration of curriculum content, not just more efficient delivery of what is already taught.

Is the 2019 analysis still relevant?

Its technical examples are dated, but its evaluative framework is arguably more useful now. The core questions about pedagogical grounding, teacher augmentation and evidence of learning gains apply directly to current generative AI tools, which are far more capable but no better evidenced.

Does AI in education replace teachers?

The authors argued clearly against this framing. Teaching involves relational, diagnostic and motivational work that current systems handle poorly. The realistic value is in removing administrative and repetitive assessment burden so teachers can spend more time on work requiring human judgement.

What is an intelligent tutoring system?

A system that models a learner's knowledge state and provides step-level guidance within a structured domain, adjusting hints and problem selection based on inferred understanding. They have the strongest evidence base in mathematics and other procedurally structured subjects with clearly defined correct steps.

How should a school evaluate an educational AI product?

Define the specific learning outcome it should improve, run a term-length pilot with a comparison group, require teachers to log failures, and review student data handling before deployment. Base renewal decisions on the outcome measure rather than on usage or engagement statistics.

Conclusion

The enduring insight is that educational technology decisions are curriculum decisions wearing technical clothing, and treating them otherwise hands pedagogical authority to product design. Before your institution's next procurement, write down the single learning outcome the tool must improve and how you will measure it over one term. If you are advising students on where to study the underlying field, continue with the guide to evaluating artificial intelligence programs.