• AI’s transforming how lessons are planned and taught, offering huge potential for innovation when guided by frameworks that keep teachers’ judgment and inclusivity at the center.
• Co-design and feedback loops let teachers stay in charge—embedding reflection and autonomy into AI workflows so technology adapts to educators, and not the other way around.
• Transparency tools like model cards help teachers see how AI makes choices. Visibility turns compliance into confidence, empowering educators to shape fairer, more inclusive classrooms.

Artificial intelligence isn’t a side tool in education anymore — it’s becoming part of the system’s wiring. From chatbots to adaptive learning platforms, AI now shapes how lessons are planned, delivered, and evaluated. These tools can personalize instruction, automate grading, and save teachers’ valuable time.³ Yet they also raise difficult questions about what’s lost when algorithms start mediating the teacher–student relationship. As one team of researchers observed in a qualitative study, AI risks reducing teachers to “mere technology operators,” where commercial platforms dictate classroom choices and turn education into a kind of service economy.⁴

Every teacher plans their lessons differently. Some experienced educators can map a lesson in their heads, while those newer to the profession might spend evenings piecing together resources, activities, and assessments.⁵ Many blend district materials with ideas found online, constantly adjusting for their students’ needs. Into this already complex process, generative AI has arrived almost overnight. With a single prompt — “Create a lesson plan on the Declaration of Independence” — a teacher can get a ready-to-use plan, complete with learning goals, discussion prompts, and homework.

However, these systems don’t just reproduce information; they also reproduce assumptions. Pedagogical bias refers to the implicit beliefs and preferences embedded in AI design — what the model “thinks” good teaching should look like.⁶ When those biases privilege Western examples, standardized pacing, or lecture-style instruction, they risk narrowing what counts as effective learning. A recent study of pre-service teachers in South Korea found that while AI boosted creativity and efficiency, it often failed to capture inclusive pedagogy or adaptive strategies for diverse classrooms.⁷

Even so, use is climbing fast. A U.S. national survey found that 37% of teachers use AI at least monthly to prepare lessons and 33% for creating worksheets.⁸ Those using it weekly save an average of 5.9 hours per week — nearly six full weeks a year. For many, it’s not about replacing expertise; it’s about reclaiming time to teach. The opportunity ahead lies in embracing AI not as a shortcut, but as a responsibly used catalyst that works best when teachers are empowered to guide it with ethics, context, and care.

AI-driven platforms are increasingly built into High-Quality Instructional Materials (HQIM)—the lesson banks, adaptive quizzes, and analytics systems districts purchase to standardize teaching quality. These tools recommend pacing, personalized learning content, and can even generate daily plans. The problem isn’t the technology itself—it’s how it reallocates decision-making authority. When an algorithm decides which examples a student sees, the teacher’s professional judgment may shift from designer to implementer.

Field data illustrate how significant that shift has become. In a cross-country qualitative study of more than a hundred secondary-school teachers, most described AI in the classroom as a direct challenge to their instructional authority.⁴ Students now compare teacher explanations with AI answers and expect the system to arbitrate truth. One respondent put it bluntly: “There’s a third participant in my class, and I’m competing with it for trust.”

The Alberta Teachers’ Association, after eight years of AI-in-education research, also brings up the significant concern of moral passivity with overreliance on AI by educators.³ As AI systems deliver increasingly confident outputs, decision-makers—both teachers and administrators—risk deferring to machine recommendations without questioning their rationale. Relying too heavily on machine-generated recommendations can lead to serious adverse consequences for students while also contributing towards a gradual erosion of professional agency, decision-making, and accountability.

To understand why this erosion matters, it helps to look through the lens of Self-Determination Theory, one of the most established frameworks for human motivation.⁹ The theory suggests that three psychological needs—autonomy, competence, and relatedness—form the foundation of engagement and professional satisfaction. Each is being tested by the way AI’s entering classrooms.

• Autonomy: Teachers’ sense of control diminishes when algorithms dictate lesson flow or when predictive dashboards frame their next instructional move. What used to be professional discretion becomes compliance with automated recommendations.
• Competence: Confidence suffers when AI-generated explanations outperform or contradict teachers’ own. Some educators describe feeling undermined by the fluency and precision of machine responses, even when those responses lack nuance or empathy.⁴
• Relatedness: Teaching’s always been relational work—built on trust with students, a shared sense of purpose in the classroom, and peer networks that sustain professional growth. When AI begins to mediate those relationships—by interpreting performance data, providing automated feedback, or filtering communication—teachers can start to feel one step removed from the people they teach. They might spend more time managing dashboards than understanding the student behind them. Over time, that distance can chip away at the sense of connection and belonging that keeps many educators in the profession.

When any of these needs go unmet, intrinsic motivation begins to slip.⁹ Teachers might disengage from reflective practice, creativity wanes, and instructional quality could ultimately suffer. Systems can’t maintain high performance if the professionals running them feel reduced to operators.

This dynamic carries measurable organizational consequences. Teacher autonomy is closely tied to instructional quality, retention, and well-being.^10,11 When educators lose control over how material is selected or sequenced, engagement declines. 

AI won’t ever replace teachers—but it’s starting to reshape how decisions are made. The opportunity now is to build systems that don’t just include teachers but rely on them. That means designing AI tools with educators, not for them, and ensuring human judgment remains the final checkpoint in every instructional loop.

Employing Co-design and Override Architecture

Before an algorithm enters the classroom, teachers should shape it. This isn’t consultation after the fact—it’s co-creation. In Sociotechnical Systems Theory, technology and human processes evolve together; when one dominates, performance suffers.¹² Co-design means teachers participate in system design, data labeling, and feedback cycles long before deployment. Their context defines the parameters, rather than being adapted to them later.

Then there is the importance of utilizing override architecture in AI educational platforms, which refers to the built-in friction that keeps humans thinking. Every recommendation, from lesson sequence to student placement, should include a visible checkpoint: “Confirm or adjust this suggestion?” paired with a short notes field. In the context of lesson planning, instead of a “next” button, the system might present two AI-generated slides for a single slideshow, requiring the teacher to choose one of the two options and edit it before moving to the next slide. Each small act of review makes human reasoning part of the workflow, not an optional add-on.

Behaviorally, this design matters because reflection changes the cognitive mode. Passive workflows invite procedural compliance; decision points trigger intentional thought. From a Self-Determination Theory perspective, autonomy isn’t freedom from structure—it’s choice within structure. Each override reinforces competence (“my judgment matters”) and autonomy (“my context counts”). Over time, these micro-choices rebuild agency, converting automation into collaboration.

Embedding Agency Loops

Keeping teachers visible in AI-mediated learning also requires what we might call “agency loops.”¹³ When students interact with adaptive content, the system can explicitly link the experience to teacher guidance: “This question builds on Mrs. Cruz’s lesson on energy transfer.” If a student’s reading level is automatically adjusted downward, the interface should pause: “Confirm or revise this recommendation?” That simple design cue institutionalizes agency. The machine doesn’t guess what the teacher thinks—it asks.

Agency loops work because they shift the default posture of AI from directive to deferential. The system expects human input. Teachers no longer need to fight for control; the structure itself preserves it. When decision rights are visible, accountability becomes shared instead of displaced.

The model of embedding teachers’ judgements into AI systems already shows promise. In Commonwealth of Learning initiatives, teachers in the Pacific region co-created over 300 lesson plans using generative AI.¹⁴ The tools produced drafts; teachers refined, localized, and published them. This approach could extend to national curriculum alignment, professional learning networks, and AI-driven assessment tools—anywhere algorithms intersect with expertise. The principle stays constant: technology proposes, teachers dispose.

Keeping teachers in the decision loop isn’t about slowing progress; it’s about keeping progress accountable. 

AI-driven education tools don’t generate knowledge in isolation. They learn from historical data — and with it, the same social patterns, exclusions, and stereotypes that shaped earlier curricula. In effect, they may automate bias under the guise of innovation.

A recent evaluation of more than 300 AI-generated lesson plans from ChatGPT, Gemini, and Copilot revealed that 94 percent lacked multicultural or diversity content when analyzed with Banks’ Multicultural Integration Framework.⁵ The framework measures four levels of inclusion — from token references to full integration of multiple cultural perspectives and structural equity. Most AI-generated materials never progressed beyond the first level. The lessons also rarely asked whose knowledge was being represented or which perspectives were missing.

This absence isn’t cosmetic. In education, representation defines what students see as attainable and what teachers view as “normal.” In a complementary study of language-model narratives, more than three-quarters of stories that mentioned Native American individuals portrayed them as historical figures rather than contemporary learners.¹⁵ When they did appear in modern contexts, success was described as exceptional — “against all odds.” That phrasing doesn’t celebrate achievement; it marginalizes it. Decades of social-psychology research show that such portrayals can trigger stereotype threat — a subtle pressure that reduces performance and belonging among students who see their identities framed as anomalies.¹⁶

Bias isn’t also confined to lessons involving vignettes. In a 2025 audit by Common Sense Media, AI “teacher assistants” such as Gemini proposed more punitive behavior plans for students with Black-coded names and more supportive interventions for those perceived as White.² These discrepancies weren’t visible case by case; they emerged statistically across hundreds of prompts — the kind of pattern teachers under workload pressure might never detect.

Behavioral science explains this through the representativeness heuristic: we judge category fit by similarity to a prototype. For most large-language models, that prototype of a scientist, engineer, or leader mirrors the historical majority groups in their data — typically white, male, and able-bodied. When algorithms learn from that skewed archive, they unconsciously treat other identities as outliers. Women, Indigenous learners, and students with disabilities don’t vanish entirely; they reappear as “exceptions that prove the rule.” The pattern looks statistical, but its impact is psychological — reinforcing an implicit hierarchy of who belongs in advanced or leadership roles.

This bias carries operational consequences. Teachers relying on AI-generated materials may unknowingly replicate inequities embedded in the training data. Students exposed to those materials may absorb narrow cues about who is likely to succeed, which gradually shapes self-concept, participation, and even subject choice.^17,18 When oversight is limited, these feedback loops normalize disparity. Over time, AI systems don’t just reflect inequality, but could institutionalize it.

Banks’ model underscores why this matters. Multicultural integration isn’t about adding diverse examples; it’s about reconstructing curriculum so multiple knowledge systems inform how subjects are taught. Without that layer, AI remains trapped in reproduction mode — sophisticated at generating text but perhaps tone-deaf to context. A math problem about engineering can teach computation or it can also teach belonging, depending on who’s visible in the example.

There’s another governance risk worth considering — what psychologists refer to as moral disengagement.¹⁹ It describes moments when people begin to feel less responsible for the ethical outcomes of their choices, particularly when decisions pass through systems that appear impartial. In education technology, this dynamic may emerge as automated planners or dashboards take on more influence. Teachers might come to rely on these tools without fully questioning how recommendations are generated. Over time, the line between human and machine judgment can start to blur, and with it, our sense of where ethical responsibility truly lies.

In the context of lesson planning, that can mean accepting an AI-generated unit because it looks polished or data-aligned, without fully questioning whose voices or perspectives it leaves out. Over time, that mindset can normalize compliance that feels efficient but may reproduce bias in what instructional materials are regarded as being “high quality.”

If algorithms are going to help design lessons, then teachers should be able to see how those decisions are made. That’s the simplest way to keep professional judgment where it belongs — in human hands.

High-quality instructional materials (HQIM) are already defined by rigor and alignment. What AI versions of these materials rarely include is transparency. As AI takes on a larger role in education planning, that gap starts to matter. Models draw from huge datasets, but teachers can’t always tell what’s in them or whose perspectives they represent. When the process isn’t visible, it’s hard to challenge what’s missing. So maybe the next generation of “high quality” isn’t just better content — it’s better visibility.

Transparency as a Design Standard

In AI research, model cards are short documents that explain what a model does, where its data come from, and how performance differs across groups.²⁰ They’ve been used in other industries to reduce bias and increase trust. In computer vision, for instance, public model cards helped close gender and skin-tone accuracy gaps by nearly 30% within a year.²⁰ Education can take a page from that playbook.

Imagine that every AI-generated lesson came with a small, expandable summary card. A teacher could click to see:

• which datasets informed the content,
• whether examples include diverse, contemporary voices, and
• a clear statement of what settings the model’s built for — and what it isn’t.

That kind of visibility doesn’t slow teachers down; it lets them make more informed judgement calls. During planning, they could adjust materials based on who’s in the room and not just who’s in the dataset. District teams might also compare vendors using inclusion metrics, while teacher training could include sessions on how to read and flag issues in those cards. Transparency like this turns equity into something practical. It helps counter salience bias, where people focus on what’s visible and miss what’s hidden.

Over time, “high quality” could mean something broader — materials that are rigorous, responsive, and representationally fair. That shift doesn’t just protect students; it strengthens teacher autonomy by giving them data to support their own judgment.

Accountability as a Cultural Norm

Transparency’s only powerful when people use it. If bias-audit results for curriculum models were shared through a central database, vendors might fix representational imbalances faster. However, accountability shouldn’t just flow downward. Teachers themselves could rate the transparency quality of AI-generated lessons — confirming whether the content feels balanced and contextually fair. Those peer verifications, once aggregated, could serve as social proof that a tool earns professional trust. In behavioral science, social proof describes how people look to others’ actions to determine what’s credible or appropriate. When teachers see colleagues endorsing transparent systems, it signals reliability and helps establish ethical, well-documented tools as the professional norm. Incentives could help that culture take root. Education ministries, unions, or professional networks might recognize educators who consistently participate in transparency audits or resource reviews. Even symbolic acknowledgment — such as accreditation points or public commendation — can turn responsible practice into shared professional pride.

Behavioral models like COM-B help explain how this kind of cultural change actually happens. The framework starts with behavior — what people do — and argues that any sustained change depends on three conditions: capability, opportunity, and motivation.²¹

Capability means teachers have the knowledge and skill to interpret algorithmic decisions, question bias, and adjust materials accordingly. That might involve AI-literacy modules in teacher education or ongoing workshops that let educators practice reviewing model outputs and flagging issues.

Opportunity is about creating the space and support for that behavior to occur — time in the schedule, feedback channels within AI platforms, and policies that require human-in-the-loop oversight so teachers can shape system updates rather than just react to them.

Finally, motivation reflects both incentives and internal values. Recognition helps, but so does moral purpose — the belief that transparency protects students and promotes equity. When capability, opportunity, and motivation align, accountability doesn’t feel imposed; it becomes something teachers choose to sustain.

Overall, transparent systems can turn accountability from surveillance into understanding. When teachers can see how algorithms generate lessons—what data they draw from and what assumptions they make—they don’t lose authority; they strengthen it. Clarity empowers them to ask better questions, flag ethical concerns, and make sure AI supports inclusive, evidence-based instruction. In that sense, transparency isn’t just a safeguard—it’s what allows teachers to use AI ethically and still teach in ways that reflect their professional judgment and their students’ realities.

AI’s promise in education still collides with a familiar barrier: unequal access. Installing, maintaining, and repairing AI tools isn’t cheap, and those costs may quickly create fault lines between schools. Well-funded districts might afford adaptive platforms and predictive analytics that refine instruction, while under-resourced schools struggle to keep basic devices running. The result isn’t innovation everywhere — it’s concentration, where technological progress clusters around affluence.

The inequity isn’t just local; it’s global. Some systems, like South Korea’s, are beginning to curate AI-based classroom solutions in a systematic way.²² Others haven’t yet established metrics to evaluate digital investments. Without a coordinated effort to close this gap, an AI divide could easily deepen the digital divide it was meant to solve — widening disparities in learning quality and digital literacy. In practice, that could mean students in wealthier regions benefit from AI-assisted feedback and adaptive instruction, while those in resource-limited schools remain excluded from these innovations altogether. Additionally, since much of the research on AI in education still comes from the United States, global perspectives on equity and cultural context are likely to remain being under-represented.²³

Even in higher-income countries, access to quality content may lag behind the technology meant to deliver it. Only one in five U.S. superintendents and principals report being “very familiar” with HQIM and more than a third say they’re “not familiar” at all.²⁴ Yet HQIM implementation has been linked to measurable learning gains. In Duval County, Florida, math scores rose six percentage points in grade 3 and three points in grade 4 — twice the statewide improvement within a single year of adopting HQIM.²⁵ Unless AI and HQIM evolve together, schools might gain faster tools but lose stronger foundations.  

Ultimately, without thoughtful structural guardrails, AI might inadvertently reinforce the same inequities it aims to solve, but at a greater scale.

The rise of AI in education has revealed two defining challenges. First, automation is gradually redrawing the contours of teacher autonomy—transferring instructional decisions once grounded in professional discretion toward algorithmic reasoning. Second, bias is taking on new forms within the very systems built to foster equity, influencing how educators instruct and how students engage through hidden design choices. Together, these shifts risk weakening the trust, expertise, and integrity that uphold the teaching profession itself. Both trends threaten to erode trust, agency, and fairness at the heart of education.

However, the same forces that threaten progress can also renew it. When teachers help craft the tools they use, AI stops dictating instruction and begins elevating it. Co-design and feedback loops convert data into dialogue, keeping educators in command of how lessons develop. And when transparency becomes standard practice, accountability evolves from oversight into understanding. Seeing how algorithms assemble materials doesn’t weaken teachers’ authority—it reinforces it with evidence.

Globally, the question isn’t whether AI fits into education, but how wisely it’s integrated. Technology can accelerate progress, but only teachers create meaning. If we empower them to lead, question, and adapt, AI can move from experiment to essential—without sacrificing what makes learning human.

At The Decision Lab, we work with education systems, policymakers, and learning organizations, applying behavioral insights to build AI and digital tools that make classrooms and institutions work better. Sometimes that means rethinking how online learning’s designed; other times, it’s weaving transparency into the systems teachers use every day. Either way, we turn innovation into something practical, evidence-based, and built to last. Partner with us to reimagine what smarter, fairer, and more empowered learning can look like.

The COM-B Model for Behavior Change 

We briefly highlighted how the COM-B model can empower teachers to use AI in supporting the next generation of learners. Its roots run deeper. Built on Capability, Opportunity, and Motivation as the three drivers of Behavior, COM-B has been applied in healthcare, public policy, and organizations to encourage ethical action, build capacity, and sustain meaningful change.

How Might Behavioral Science Transform Education? 

It’s a fair question—how can behavioral science actually change education? Two of our senior consultants explore how its principles uncover hidden barriers: cognitive biases, environmental limits, and structural frictions that shape decisions. By making the invisible visible, behavioral science offers cost-effective, evidence-based ways to improve outcomes in areas like financial aid, academic success, and digital learning.

1. Armitage, K. L., Bulley, A., & Redshaw, J. (2020). Developmental origins of cognitive offloading. Proceedings of the Royal Society B, 287(1928), 20192927.
2. AI Risk Assessment Team. (2025, August 6). AI teacher assistants. Common Sense Media. https://www.commonsensemedia.org/ai-ratings/ai-teacher-assistants?gate=commsdistributionlink
3. McRae, P. (2025, October). Artificial intelligence is revolutionizing education. ATA Magazine. The Alberta Teachers’ Association. https://teachers.ab.ca/news/artificial-intelligence-revolutionizing-education-0
4. Alasgarova, R., & Rzayev, J. (2025). The implications of artificial intelligence for teacher agency and teacher–student relationships through the Technology Acceptance Model. International Journal of Technology in Education and Science, 9(3), 450–473.
5. Trust, T., Maloy, R., Xu, C., & Pelletier, K. (2025). Civic education in the age of AI: Should we trust AI-generated lesson plans? CITE Journal of Educational and Social Studies, 25(3-25). https://citejournal.org/volume-25/issue-3-25/social-studies/civic-education-in-the-age-of-ai-should-we-trust-ai-generated-lesson-plans
6. Chen, B., Cheng, J., Wang, C., & Leung, V. (2025). Pedagogical biases in AI-powered educational tools: The case of lesson plan generators. Social Innovations Journal, 30.
7. Kerr, R. C., & Kim, H. (2025). From prompts to plans: A case study of pre-service EFL teachers’ use of generative AI for lesson planning. English Teaching, 80(1), 95–118.
8. Walton Family Foundation. (2025, June 25). The AI dividend: New survey shows AI is helping teachers reclaim valuable time. https://www.waltonfamilyfoundation.org/the-ai-dividend-new-survey-shows-ai-is-helping-teachers-reclaim-valuable-time
9. Deci, E. L., & Ryan, R. M. (2012). Self-determination theory. In Handbook of theories of social psychology (Vol. 1, pp. 416–436). Sage.
10. Wang, X., Yang, L., Chen, K., & Zheng, Y. (2024). Understanding teacher emotional exhaustion: Exploring the role of teaching motivation, perceived autonomy, and teacher–student relationships. Frontiers in Psychology, 14, 1342598.
11. Worth, J., & Van den Brande, J. (2020). Teacher autonomy: How does it relate to job satisfaction and retention? National Foundation for Educational Research. https://www.nfer.ac.uk/media/r1ubcb4s/teacher_autonomy_how_does_it_relate_to_job_satisfaction_and_retention.pdf
12. Trist, E. L., & Bamforth, K. W. (1951). Some social and psychological consequences of the longwall method of coal-getting: An examination of the psychological situation and defences of a work-group in relation to the social structure and technological content of the work-system. Human Relations, 4(1), 3–38.
13. Leonardi, P. M. (2025). Homo agenticus in the age of agentic AI: Agency loops, power displacement, and the circulation of responsibility. Information and Organization, 35(3), 100582.
14. Balaji, V., Ogange, B. O., & Mays, T. (2025). Teacher in the loop AI (TiL-AI): A strategy for empowering educators in developing countries through OER adaptation. Journal of Learning for Development, 12(2), 439–445.
15. Kannan, P. (2024, October 3). How harmful are AI’s biases on diverse student populations? Stanford HAI. https://hai.stanford.edu/news/how-harmful-are-ais-biases-on-diverse-student-populations
16. Osborne, J. W., & Walker, C. (2006). Stereotype threat, identification with academics, and withdrawal from school: Why the most successful students of colour might be most likely to withdraw. Educational Psychology, 26(4), 563–577.
17. Shapiro, J. R., & Williams, A. M. (2012). The role of stereotype threats in undermining girls’ and women’s performance and interest in STEM fields. Sex Roles, 66(3), 175–183.
18. McGee, E. (2018). “Black genius, Asian fail”: The detriment of stereotype lift and stereotype threat in high-achieving Asian and Black STEM students. AERA Open, 4(4), 2332858418816658.
19. Moore, C. (2015). Moral disengagement. Current Opinion in Psychology, 6, 199–204.
20. Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., … & Gebru, T. (2019, January). Model cards for model reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 220–229).
21. Willmott, T. J., Pang, B., & Rundle-Thiele, S. (2021). Capability, opportunity, and motivation: An across-contexts empirical examination of the COM-B model. BMC Public Health, 21(1), 1014.
22. Ministry of Education, Republic of Korea. (2024, November 29). AI digital textbooks for 2025 to realize personalized education for all [Press release]. https://english.moe.go.kr/boardCnts/viewRenewal.do?boardID=265&boardSeq=102075
23. Varsik, S., & Vosberg, L. (2024). The potential impact of artificial intelligence on equity and inclusion in education (No. 23). OECD Publishing.
24. Marken, S., & Malek Ash, A. (2025, April 1). Defining “quality” is hard, making math progress harder: Most schools and districts lack an official definition for high-quality instructional materials. Gallup. https://news.gallup.com/poll/658484/defining-quality-hard-making-math-progress-harder.aspx
25. Steiner, D. (2024, January). The unrealized promise of high-quality instructional materials: Overcoming barriers to faithful implementation requires changing teacher and leader mind-sets.State Education Standard, 24(1). https://www.nasbe.org/the-unrealized-promise-of-high-quality-instructional-materials