By Joni Gutierrez, Ph.D., and Ronald Lethcoe, M.Ed.
Every model release resets the tactical conversation in education. A new tool arrives, a policy gets drafted, a detector gets purchased, a syllabus gets rewritten, and by the following term the ground has moved again. That cycle is exhausting, and it has a structural flaw: the technology sets the agenda and the institution responds.
The alternative is to decide first what learning is for, and only then ask what a tool is permitted to do inside that purpose. The Eight Essential Principles for education in the age of AI were written to support that second posture (Gutierrez & Lethcoe, 2026). They are not a checklist to be satisfied once and filed, and they are not a list of forbidden products. They are eight standing questions that outlive the tool you have not seen yet.
This article puts the framework under load. Each of the eight principles is set against a real case from the last two years of AI in education, and the principles are grouped into four pairs so that each one is read alongside the commitment that keeps it honest.
Why four pairs
The framework has two lenses. The first four principles name the irreducible human core of education: flourishing, relationships, creativity, and accessibility. The last four translate that core into operational commitments: equity, inclusion, openness, and universality (Gutierrez & Lethcoe, 2026).
Read the two lenses as a mirror, and the pairs fall out on their own. The first human commitment meets the last operational one, and so on inward:
- Human Flourishing and Universality: what learning is for, and whether it travels.
- Human Relationships and Openness: the bond, and the evidence that keeps it honest.
- Human Creativity and Inclusion: who makes the work, and whose knowledge counts.
- Accessibility and Equity: who gets in, and who gets somewhere.
Nothing is being merged. All eight principles keep their own claims and their own ways of failing. What the pairing offers is a seam: placed side by side, each principle shows what its partner cannot protect alone.
Pair one: Human Flourishing and Universality
Growth needs resistance
The purpose of education is the holistic growth of a person. AI can deliver faster answers, but intelligence is not the same as wisdom, and wisdom includes knowing which problems are worth solving. Intellectual growth requires resistance. Remove every difficulty from a course and you remove the struggle through which students learn to think.
A study from the MIT Media Lab put a measurement on that intuition. Researchers assigned 54 participants to write essays with a large language model, a search engine, or no external tool while recording EEG activity. Across the first three sessions, the LLM group showed the weakest and least distributed functional connectivity, reported the lowest sense of ownership over their essays, and had more difficulty accurately quoting their own work (Kosmyna et al., 2025). The paper is a preprint with a small, task-specific sample, so it should be treated as suggestive rather than definitive. Even with that caution, it reinforces a question worth asking: what happens when speed removes the effort through which learning takes hold?
It has to work everywhere, and without the machine
Universality asks a different question of the same design. Education serves learners in conditions that are nothing like a well-funded pilot: thin bandwidth, older devices, shared spaces, no paid subscription, and competing demands on the day. If a learning design only works at the top of the resource curve, it is not finished.
Los Angeles Unified School District offers a cautionary case. In March 2024 the district launched Ed, an AI assistant intended for roughly half a million students under a five-year contract worth about six million dollars. Within months, vendor AllHere furloughed most of its staff amid financial difficulties. Because the chatbot depended in part on human moderation from the company, LAUSD took Ed offline (Young, 2024). The lesson is not that districts should avoid ambition. It is that a capability whose continuity depends on a single vendor is not yet durable infrastructure.
Together, the pair is a durability test. A vision of intellectual growth that runs only on a paid frontier model and a strong signal is a promise made to some students and quietly withheld from the rest. The call to preserve the ability to read, write, calculate, and create by hand is the same claim as the call to preserve intellectual friction, viewed from the infrastructure side.
Three moves
- Name the friction. For each assignment, write down the one cognitive move the student must make unaided.
- Run the low-resource test. Could a student on a phone, on shared data, with no paid account, earn full credit?
- Keep an unplugged path. Every AI-assisted assignment gets a by-hand version that earns equal respect.
Pair two: Human Relationships and Openness
Teaching is a relationship, not a delivery
Content delivery was never the whole of teaching. A tool can generate practice, explain a concept five different ways, and answer at midnight. It can respond to cues in language, but it does not participate in the human relationship through which an educator notices hesitation, builds trust, or understands the context of a learner’s life. The proper role of the tool is a bridge, not a wall: it should absorb administrative background noise so that educators can spend their attention on relational work.
Estonia’s national AI Leap program is instructive here because of its sequencing. The pilot year began in August 2025 with a support program for upper secondary teachers across 154 participating upper secondary schools, alongside professional learning communities in which teachers developed subject-specific approaches. Students did not receive the learning application until late January 2026, and decisions about how and how often to use it remain with teachers and students. The program’s chief executive has been blunt about the goal: it is not to increase students’ AI use, but to offer a pedagogically guided alternative to commercial models (AI Leap, 2026).
Fluent is not the same as true
Openness begins with what the machine actually is. Large language models generate outputs by estimating patterns in data and can produce fluent responses that are incomplete, misleading, or wrong. They are not oracles. Every output should therefore be treated as a draft awaiting human judgment. The practical form of the principle is process: what was asked, what came back, what was kept, what was rejected, what was verified.
The most instructive recent case in higher education runs in the direction faculty rarely anticipate. A Northeastern University senior noticed signs of AI use in her business professor’s lecture notes, including a stray ChatGPT citation, after students had been told not to use AI in the course. She filed a formal complaint and requested a tuition refund of just over eight thousand dollars for the course. Northeastern declined the refund, and the professor later acknowledged using several AI tools and said instructors should be transparent with students about their use (Nolan, 2025).
Openness is how the relationship survives AI. The alternative to disclosure is surveillance, and surveillance dissolves trust faster than any tool. It also has to run in both directions. A rule an instructor will not apply to their own practice is not a rule; it is a posture, and students read postures accurately.
Three moves
- Ask for process, not proof. A short note on each submission: what you used, where, and why.
- Grade the judgment. Have students defend one output they rejected.
- Spend the saved time on people. Automate the noise, then reinvest the minutes in conferences and feedback.
Pair three: Human Creativity and Inclusion
Pattern generation is not creativity
Generative AI produces new outputs by modeling statistical patterns learned from existing data. That is not the same as human creativity, which is shaped by intention, memory, emotion, identity, boredom, necessity, culture, and lived experience. When lived experience leads, AI can become a creative partner. When AI leads, creative work risks becoming fluent without being meaningfully authored. This is also why the humanities become more essential rather than less: they ask why something matters, not only how it was produced.
Admissions essays make the stakes concrete, because the personal essay is one of the few parts of an application designed to carry the idiosyncrasies of a life. Lee and colleagues analyzed 81,663 applications to a selective U.S. university from 2020 through 2024. After 2023, surface-level linguistic features converged, with the largest shifts among fee-waived applicants and students who were rejected; estimated LLM use also rose sharply in 2024 and increased more among lower-SES applicants (Lee et al., 2026). The authors caution that these patterns may standardize student voice and complicate equitable evaluation. The point is not that AI assistance erases individuality by definition. It is that, at scale, the same kinds of assistance can pull distinctive writing toward common forms.
The most common answer is not the whole of knowledge
AI systems learn from records of human activity that have been collected, digitized, and made available for training. Those records do not represent every community, language, or experience evenly, and model outputs can reproduce those imbalances. Inclusion therefore requires more than representation added after the fact. It requires asking whose histories and ways of knowing are missing or underrepresented, and refusing the single default learner.
A Stanford study published in 2026 shows how quietly this can go wrong. Researchers generated feedback on the same 600 eighth-grade persuasive essays using four language models, changing student descriptors in the prompt while holding the essay text constant. The feedback shifted with those descriptors. Essays described as written by Black students drew more validation; Hispanic and English learner labels prompted more attention to English usage; white-student labels more often elicited critique of argument, evidence, and clarity. The authors describe broader patterns of positive feedback bias and feedback withholding bias (Tan et al., 2026; Barshay, 2026).
Both principles defend the particular against the average. Creativity protects the maker. Inclusion protects the material the maker draws on. A student can be the undisputed author of work that still erases her own voice, because the course rewarded the register the machine produces by default.
Three moves
- Human spark first. An idea on paper before a prompt is typed.
- Interrogate the default. Ask students who the model assumed it was talking to.
- Assign what only this student can make. Ground the work in local, personal, or community context that the model cannot know from the assignment prompt alone.
Pair four: Accessibility and Equity
Access is designed in, or it is missing
Accessibility is not a compliance task added at the end of course design. It is the foundation for meaningful participation, and it begins with a refusal to design for an imagined ideal learner. AI can help with parts of this work: draft captions, image descriptions, translation, multimodal explanation, and text simplification can reduce barriers when they are reviewed and used intentionally. Those gains only materialize when the design is intentional.
The regulatory floor is now explicit. The U.S. Department of Justice adopted WCAG 2.1 Level AA as the technical standard for the web content and mobile applications of public entities, including public colleges and universities. In April 2026, the Department extended the compliance date for public entities with populations of 50,000 or more to April 26, 2027 (U.S. Department of Justice, 2026). The deadline changed; the underlying duty to provide accessible services did not.
What institutions are doing in that window is the more interesting story. Cal State LA offers faculty an AI-assisted PDF remediation service that addresses missing tags, incorrect reading order, alternative text, and table structure. The campus documentation is admirably plain about the limits: the tool improves accessibility but will not produce a fully accessible file, results vary with complexity, and every document should be verified before it is used in instruction (California State University, Los Angeles, n.d.). That is the correct shape for this kind of adoption. AI can clear the backlog. It cannot certify the result.
Equal access does not mean equal outcomes
Equity asks what happened after the door opened. AI will not level the field on its own, and where institutions ignore existing inequality, it can widen it. The discipline is critical adoption rather than blind adoption: teaching students to question these systems rather than only to operate them, and assessing whether students can reason, verify, explain, and take responsibility.
The clearest warning remains the detector literature. Researchers at Stanford tested seven widely used GPT detectors against TOEFL essays written by non-native English speakers. On average, the detectors misclassified 61.22 percent of those essays as AI-generated, while the same tools classified essays by U.S.-born eighth-grade students almost perfectly. A major reason was reliance on text perplexity: writing with more predictable word choices can look more machine-like to detectors, a pattern that can disadvantage non-native writers (Liang et al., 2023). A tool presented as an objective check produced a sharply unequal result, and it fell hardest on the students least able to absorb a false accusation.
Read together, the pair forms the full sequence. Accessibility opens the door and is answered in the build, before the term begins. Equity asks who actually got somewhere and is answered in the results, after the term begins. Captioning the video is accessibility. Noticing who used the captions well is equity.
Three moves
- Design first, remediate second. Headings, alternative text, captions, and structure before the term starts.
- Disaggregate before you celebrate. Look at who is using the AI supports and who is not.
- Teach the skill you assumed. Prompting, verification, and critique are taught skills, not background knowledge.
Eight questions to carry back
If the principles are hard to hold in memory, the questions underneath them are not. They survive every model release.
- What helps students flourish?
- What relationships make learning possible?
- How do students become authors of their own ideas?
- Who has access?
- Who is included?
- How do we make thinking visible?
- What must students still be able to do for themselves?
- What kind of world are we preparing them to build?
The framework does not ask anyone to choose between humanity and technology. It asks us to put the technology in its proper place. AI can make education more responsive, more accessible, more creative, and more equitable. It will do so only if human judgment stays at the center, and that is a choice institutions make on purpose or lose by default.
This article is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
References
AI Leap. (2026, March 9). Estonia launches nationwide AI Leap education program and introduces learning-supportive AI in upper secondary schools. https://tihupe.ee/en/estonia-launches-nationwide-ai-leap-education-program-and-introduces-learning-supportive-ai-in-upper-secondary-schools/
Barshay, J. (2026, April 27). AI gives more praise, less criticism to Black students. The Hechinger Report. https://hechingerreport.org/proof-points-ai-bias-feedback/
California State University, Los Angeles. (n.d.). PDF accessibility remediation. Retrieved August 12, 2026, from https://www.calstatela.edu/accessibility/pdf-accessibility-remediation
Gutierrez, J., & Lethcoe, R. (2026, May 11). Eight essential principles: A model for human-centered learning in the age of AI. CHAIRES. https://chaires.center/2026/05/11/eight-essential-principles-a-model-for-human-centered-learning-in-the-age-of-ai/
Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task (arXiv:2506.08872). arXiv. https://arxiv.org/abs/2506.08872
Lee, J., Borchers, C., Alvero, A. J., Joachims, T., & Kizilcec, R. F. (2026). The digital divide in generative AI: Evidence from large language model use in college admissions essays. Proceedings of the Thirteenth ACM Conference on Learning @ Scale (L@S ’26), 179-190. https://doi.org/10.1145/3774398.3811620
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7
Nolan, B. (2025, May 15). Northeastern college student demanded her tuition fees back after catching her professor using OpenAI’s ChatGPT. Fortune. https://fortune.com/2025/05/15/chatgpt-openai-northeastern-college-student-tuition-fees-back-catching-professor
Tan, M., Phalen, L., & Demszky, D. (2026). Marked pedagogies: Examining linguistic biases in personalized automated writing feedback. Proceedings of the LAK26: 16th International Learning Analytics and Knowledge Conference, 568-577. https://doi.org/10.1145/3785022.3785113
U.S. Department of Justice. (2026, April 20). Extension of compliance dates for nondiscrimination on the basis of disability: Accessibility of web information and services of state and local government entities. 91 Fed. Reg. 20902. https://www.federalregister.gov/documents/2026/04/20/2026-07663/extension-of-compliance-dates-for-nondiscrimination-on-the-basis-of-disability-accessibility-of-web
Young, J. R. (2024, July 15). An education chatbot company collapsed. Where did the student data go? EdSurge. https://www.edsurge.com/news/2024-07-15-an-education-chatbot-company-collapsed-where-did-the-student-data-go
License: This article is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0). You are free to share and adapt it, including for commercial purposes, with attribution.
Suggested attribution: Gutierrez, J., & Lethcoe, R. (2026). Eight Essential Principles in Practice: Human-Centered Learning in the Age of AI. CHAIRES: Center for Human–AI Research, Ethics, and Studies. Licensed CC BY 4.0.
