A principled way to think about AI in education: guidance for educators and policy makers based on goals, models

Synthesis: Principled AI Education Framework — Rejecting both the promise-versus-peril binary and the rush to immediate implementation, Finkelstein (2025) advances a third path for AI in education: a scholarship-grounded framework that connects broad educational goals to actionable practice. It rests on three anchors that must be settled before any tool is chosen — What are our educational goals? What do we know about human learning? How can technologies serve those goals and models? The framework clarifies the respective roles of educators, learners, and technologies in shaping curricula, designing instruction, assessing learning, and cultivating community, and it defends a single through-line: AI must augment, not displace, human capacities. A July 2026 update argues that the decisive question is not whether machines will match humans at every task, but which skills, habits of mind, and practices must remain within human purview to preserve our humanity and societies — naming discernment, empathy, and sense of self as the outcomes that matter most. The principles of action, he notes, remain relevant a year later.

Noah Finkelstein, University of Colorado, Boulder — Drafted July 2025, updated July 2026.

Key Findings

1. Discernment, empathy, and sense of self. The capacities education must protect are distinctly human: framing appropriate questions and validating answers, understanding others' perspectives, and developing agency, purpose, and belonging within a field or culture.
2. Three goals, not one. Educational systems serve individual development, societal infrastructure — democracies depend on an educated populace — and the disciplines themselves; current AI discourse conflates these goals or collapses them into performance metrics.
3. Goals before tools. Principles for action derive from stated goals rather than from available products: aligning technology with learning goals, calibrating practice and feedback, supporting or blocking cross-disciplinary synthesis, co-designing with students, and collecting formative evidence.
4. Augment, don't displace. When AI generates essays or solves problems students should grapple with themselves, it undermines the goals education exists to serve; technology must extend human capacity rather than bypass it.
5. Scenarios expose the trade-offs. Ubiquitous AI access without training reduces education to performance without learning, blanket bans ("back to blue books") send contradictory messages about professional practice, and technology-enhanced skills work risks widening equity gaps if it displaces human interaction.
6. Let students train the tools. Having learners validate or train AI systems in their own subject domain builds Metacognition, discernment, and a practical understanding of where AI falls short.

Anchors: Goals, Learning, Technology

Finkelstein insists the three anchors be settled before any tool is selected. Goals: what should education accomplish — meaningful learning, democratic participation, preparation for dynamic futures. Models of human learning: how people actually learn, drawing on decades of learning-sciences research — active construction, Motivation, social mediation, Transfer of Learning, and self-regulation. Use of technologies: how AI can serve those goals and models rather than dictate them. In higher education, it treats Educational Development as the point of the enterprise, not a byproduct of tooling. It also separates responsibilities across four domains — curricula, instruction, assessment, and community — giving educators, learners, and technologies different jobs. Active learning and student agency sit at the center; the technology occupies a supporting role that must be deliberately chosen.

Core Mandate: Augment, Don't Displace

"AI must augment, not displace, human capacities" is the through-line of the framework. Technology use is judged against enduring educational values: advancing meaningful learning rather than efficient content delivery, supporting democratic societies rather than individual skill acquisition alone, and preparing students for dynamic futures rather than today's job market. In practice this means asking whether a tool does something the learner should do themselves — if so, redesign. The mandate converges with documented harms from displacing pedagogy, with the misattribution of competence to AI systems, and with Desirable Difficulties research showing that easier practice can harm long-term retention. It also reframes Human AI Collaboration: systems that reduce Collaborative Learning to isolated machine interaction, or that erode belonging and agency, fail the principled test regardless of technical sophistication. The human-in-the-loop stance keeps the learner as the agent and the model as an instrument.

Relationship to the Evidence Base

The framework reads as a normative complement to a thin empirical record. Where The Evidence Base on AI in K-12: A 2026 Review finds few causal studies because goals were left ill-defined, Finkelstein supplies the definition; where Self-Regulated Learning treats regulation as meaningful engagement, he warns that convenience tempts course design toward shortcuts. Distributed AI Literacy for citizenship answers his democratic goal, while authentic assessment answers his call for tasks that resist cheap measurement — a direct challenge to automated grading that scales cheaply but flattens what it scores. The framework also sits alongside Learning Theories and Sociocultural Learning, and it intersects with Mapping the Scaffolding of Metacognition and Learning by AI Tools in STEM Classrooms: A Bibliometric-Systematic Review on how AI can support or undermine metacognitive development depending on implementation.

Using the Framework as a Diagnostic

The framework works as a checklist for any AI-in-education initiative. First, state the goal clearly — not "use AI in my course" but "improve students' ability to evaluate evidence." Second, identify the learning model, drawing on research such as productive failure, deliberate practice, or social learning. Third, match the technology to the model — a Socratic dialogue partner for productive failure, peer matching for social learning, progressive Scaffolding for novice learning design. Fourth, evaluate against displacement: does the tool do something the learner should do themselves? If yes, redesign the curriculum. The questions also clarify the teacher's role in a Generative AI classroom, and connect to critical thinking scaffolds that keep judgment with the student.

What this means for practice

• Instructors. Settle the three anchors before choosing any tool: what your educational goals are, what is known about human learning, and how technology serves both. The framework derives its principles from stated goals rather than from what products offer.
• Instructors. Design so AI augments rather than displaces human capacity. The outcomes the paper names as the ones to protect are discernment, empathy and sense of self — framing appropriate questions and validating answers, understanding others' perspectives, and building agency, purpose and belonging in a field.
• Administrators. Keep the three goals of education distinct — individual development, societal infrastructure including democratic participation, and the disciplines themselves — rather than collapsing them into performance metrics that AI is easy to optimize against.
• Administrators. Plan for the conditions the paper anticipates: increased teaching demands and fewer support resources such as teaching assistants, which is the argument for the augment-not-displace position rather than a preference about tools.

Limitations

• Guidance, not tested practice. This is a single-author framing paper, drafted July 2025 and updated July 2026, drawing on learning-sciences scholarship and the author's own experience. It reports no trial, no adoption data and no measured outcomes.
• The update restates rather than adds evidence. The July 2026 revision argues the principles remain relevant a year on; that is the author's assessment, not new findings.
• Scope is higher education. The framework is written for colleges and universities, and the paper notes that for transformation practices specifically the evidence base is much sparser than for the tools themselves.
• Several levers sit above the instructor. The roles it identifies for making this work include department chairs, deans and centers for teaching and learning, so parts of the framework depend on institutional decisions a course-level reader cannot make alone.

Citation

Finkelstein, N. (2025). A principled way to think about AI in education: guidance for educators and policy makers on action based on goals, models of human learning, and use of technologies.