Agentic AI in Schools: When the AI Does the Research AND Writes the Essay

A middle schooler in Austin has a history essay due Monday. On Sunday night, she opens her laptop and types one instruction into her AI agent: “Write a five-paragraph essay about the causes of World War I, using at least four historical sources, in MLA format, for a 7th-grade class.” Eleven minutes later, she has a 750-word essay, four cited sources from reputable history databases, correct MLA formatting, and a title she’d be comfortable putting her name on. She didn’t read a single source. She didn’t write a sentence. She barely looked at the output before submitting. Meanwhile, at a school across town, a teacher has spent the weekend redesigning her assignments because she noticed that every essay her class submitted last week was suspiciously coherent, unusually well-structured, and oddly similar in tone. Welcome to the central educational crisis of 2026—not AI-assisted writing, but AI-autonomous completion of academic work. The question is no longer “did the student use AI?” It’s “what was the student supposed to learn, and did they?”

Why This Is Genuinely Different From Previous “Cheating” Technologies

Every generation has had its cheating technologies. CliffsNotes. Wikipedia. Essay mills. Calculators, which schools eventually accepted and integrated. Spell-check, which schools eventually stopped penalizing. Each time, the debate followed the same arc: moral panic, partial acceptance, eventual integration with modified assessments.

Agentic AI is categorically different for three reasons:

Scale. Previous cheating tools required effort—finding a CliffsNotes, buying an essay, remembering to use Wikipedia. Agentic AI requires one sentence. The marginal effort to avoid doing an assignment is essentially zero.

Quality. Previous cheating tools produced recognizably low-quality work—essays that were obviously copied, analyses that were shallow. Modern AI agents produce work that is often superior to what the student would have produced themselves. This makes detection harder and consequences more complicated.

Breadth. Previous cheating tools were domain-specific. AI agents work across all subjects simultaneously. The same agent that writes an English essay will also solve math problems, generate science lab write-ups, produce social studies presentations, and complete coding assignments. There’s no subject left that requires effort that an agent can’t reduce to a prompt.

This isn’t a cheating problem in the conventional sense. It’s a structural problem with how schools have defined academic work—and it requires a structural solution, not just stricter enforcement.

What the Research Says About What Learning Actually Requires

The Retrieval Practice Effect

Decades of cognitive science research have established that the act of retrieving and applying information—writing, solving, explaining, arguing—is what consolidates learning into long-term memory. This is called the “testing effect” or “retrieval practice effect,” and it is one of the most replicated findings in educational psychology. A 2013 meta-analysis by Roediger and Karpicke reviewed over 100 studies and confirmed that active retrieval—not re-reading or passive review—produces the strongest retention and transfer.

When a student uses an AI agent to produce an essay they don’t read, they perform no retrieval. The assignment, from a learning science perspective, produces nothing. The grade is assigned to work that contains zero cognitive engagement from the student.

Generative Processing and Understanding

Educational psychologist Merlin Wittrock introduced the theory of “generative learning”—the idea that understanding is constructed when learners actively connect new information to existing knowledge. This generation process—even when imperfect, even when producing bad first drafts—is the mechanism of comprehension. When an AI agent generates the connections for the student, the student bypasses the mechanism. The output may demonstrate understanding; the student does not possess it.

This has a concrete consequence: students who offload essay writing to AI agents routinely fail oral examinations on the same material they “wrote” about. Research from several European universities (Porsdam Mann et al., 2023) found that approximately 35% of students whose AI-generated essays scored highly could not answer basic clarifying questions about their own essay’s arguments.

The Expertise Reversal Effect

Here’s where it gets complicated: the research on AI assistance is not uniformly negative. A phenomenon called the “expertise reversal effect” (Kalyuga et al., 2003) shows that instructional supports that help novice learners become counterproductive for experts—once a learner has internalized a skill, external scaffolding reduces learning. This applies to AI agents: a student who already knows how to write a research essay and uses an AI agent to handle the tedious citation formatting is having a different experience than a student who has never written a research essay and uses an agent to produce the whole thing.

The implication: the harm of agentic AI in education is most acute for novice learners who haven’t yet internalized the skills the assignment is designed to develop. For students who have mastered a skill, AI assistance in applying it is far less harmful—and potentially beneficial.

What Schools Are Actually Testing

The fundamental problem is that most current school assignments were designed to serve as learning activities (the process of doing them was the point) and assessments (submitting them revealed what the student knew). AI agents have decoupled these functions: an agent can produce the assessment output without any of the learning activity having occurred.

The assignments that are most vulnerable: research essays, structured writing, reading responses, problem sets with worked solutions, lab reports, and any assignment where the student can provide a high-level goal and receive a polished output. The assignments least vulnerable: oral presentations, hands-on laboratory work, real-time performance tasks, portfolio reviews with defense, collaborative problem-solving with observation.

The Policy Landscape: Battles Schools Are Fighting

School AI policies fall into three broad categories in 2026:

Restrictive policies: “No AI tools of any kind, for any school work.” These are the easiest to write and the hardest to enforce. AI agents are now accessible through browsers, phones, embedded in Microsoft and Google products that schools provide. Blanket prohibition is largely unenforceable and may be educationally counterproductive.

Permissive policies: “AI tools are allowed as long as work is ‘original.’” These are well-intentioned but practically incoherent—“original” is undefined in an agentic context. Is an essay “original” if a student wrote the prompt that generated it? If they edited one paragraph?

Skill-defined policies: The most educationally sophisticated approach. These define which skills an assignment is assessing, specify what level of AI assistance is consistent with demonstrating that skill, and modify assessments accordingly. A writing assignment designed to assess argumentation skills might permit AI research assistance while requiring a handwritten outline and in-class drafting. A math assignment designed to assess problem-solving might permit AI verification of final answers while requiring documented work steps.

The third approach requires significant teacher training and assignment redesign—resources most schools don’t have.

Comparison: Assignment Types and Agentic AI Vulnerability

Assignment Type	AI Vulnerability	What Learning Is at Risk	Low-AI Alternative
Research essay	Very high	Research skills, synthesis, argumentation	Oral defense, in-class writing, process portfolio
Reading response	Very high	Comprehension, critical thinking	Socratic seminar, annotated copy, book talk
Math problem set	High (worked solutions)	Procedural fluency, problem setup	Whiteboard explanation, varied problem types
Lab report	High (write-up)	Scientific communication	Real-time lab notebook, lab video
Coding assignment	High	Computational thinking, debugging	Live coding session, code explanation interview
Oral presentation	Moderate (prep)	Communication, synthesis	In-person delivery with Q&A
Hands-on project	Low	Physical skills, design thinking	More of this
Portfolio with defense	Low	Reflection, self-assessment	Requires time investment

What to Do as a Parent

Have the “For Whom?” Conversation

The most powerful conversation you can have with your child about agentic AI and school isn’t about rules or cheating. It’s about purpose:

“Who is school for? When you have an AI do your homework, who benefits? What happens in five years if you can’t explain the things your transcript says you know?”

Kids who understand that grades are signals and that the skills underlying those grades have real future value are more likely to make good decisions about AI use—not because they fear consequences, but because they understand the stakes.

Understand What Assignments Your Child’s Teachers Have Updated

Some teachers are ahead of this. Ask your child: “Has any teacher changed how they assign work because of AI?” The answers will tell you which classrooms are thoughtfully engaging with the issue. For those that aren’t, it’s worth asking whether the school has professional development support for assignment redesign.

Establish Home Rules That Distinguish Process from Output

A useful household framework:

• AI is allowed to help you understand something you don’t understand (teaching function)
• AI is allowed to check work you’ve already completed (review function)
• AI is not allowed to produce work you present as your own without you having done the underlying cognitive work (replacement function)

The third category is the hardest to monitor. The most effective safeguard isn’t surveillance—it’s building genuine interest in the subjects your child studies, so the work feels worth doing. For more on when AI helps vs. hurts, see our piece on AI homework tools—when it helps vs. hurts.

Support Teachers Who Are Redesigning Assessments

Teachers who are redesigning assignments to be AI-resistant—oral exams, in-class writing, process documentation, project defenses—are often doing extra work with no additional support. A parent email saying “we appreciate the effort to make assignments meaningful in the AI era” goes further than most parents realize. Teachers need to know that parents support the harder assessments, not just the easier ones.

Understand the Learning Science, Not Just the Rules

This is the deepest intervention. When your child understands why retrieval practice and generative processing matter for their own brain development—when they understand that the discomfort of writing a bad first draft is where learning happens—they have an internal motivation to engage that external rules can’t fully provide. A useful resource to read together: the research summary from the Learning Scientists (learningscientists.org) on retrieval practice. It’s designed for general audiences and explains the cognitive science in accessible terms.

For broader context on AI’s effects on kids’ writing development, see our analysis of AI writing tools and kids’ brain learning.

Know What AI Cheating Detection Is and Isn’t

Several AI detection tools (Turnitin AI, GPTZero, Originality.ai) are used by schools, but all carry significant false positive rates—they flag student writing as AI-generated when it isn’t, and increasingly fail to detect sophisticated AI outputs as detection-avoidance strategies improve. Relying on detection creates adversarial dynamics and false accusations. The better solution is designing assignments that require genuine learning processes—which makes AI replacement demonstrably irrelevant. For more on the school-side perspective, see our piece on AI cheating: what parents should know before assuming their kids did something wrong.

What to Watch for Over the Next Three Months

Major assessment redesigns at the district level. Several large U.S. school districts and international systems (UK, Australia, Singapore) are conducting major reviews of their assessment frameworks in light of agentic AI. Watch for guidance documents from your state’s education department.

Oral examination revival. Multiple higher education institutions have announced returns to oral examinations for courses where AI completion of written assignments is undetectable. Expect K-12 systems to follow, especially in high-stakes assessment contexts.

“AI disclosure” requirements. Some schools are piloting policies that require students to document what AI assistance they used, similar to bibliographic citations. This shifts from prohibition to transparency—a policy approach more likely to be sustainable long-term.

Your child’s own experimentation. Every middle and high schooler is currently running informal experiments: which assignments can I get AI to do? Which teachers notice? What happens if I submit AI work? The most effective parental intervention is ongoing conversation about what they’re discovering—curiosity, not judgment.

Frequently Asked Questions

Is using an AI agent for homework cheating?

It depends on what the assignment is designed to teach. If using an AI agent to complete the work means the student didn’t develop the skill the assignment was designed to build, that’s academically dishonest regardless of whether the school’s policy explicitly covers it. The better question is: “What was I supposed to learn from this, and did I learn it?”

My child’s teacher hasn’t updated assignments for AI—should I say something?

Yes, thoughtfully. Rather than criticizing, ask: “I’ve been reading about how schools are redesigning assignments for the AI era. Has the school given teachers any guidance or support on this?” This opens a productive conversation without accusing anyone of failing.

Will all homework eventually be AI-generated?

Probably some percentage of submitted work already is. The question is whether schools evolve their assessment systems to focus on what can’t be easily automated: demonstrated understanding, oral fluency, physical creation, collaborative problem-solving, and real-time performance. Schools that don’t evolve will face a credentialing crisis—grades that don’t predict skill.

What’s the difference between AI assistance and AI completion?

AI assistance preserves the student’s cognitive engagement—explaining a concept the student didn’t understand, checking an answer they calculated, suggesting structural improvements to a draft they wrote. AI completion replaces it—producing the initial work the student would otherwise have done. The cognitive science literature is clear that the distinction matters enormously for learning.

How do I talk to my child if I find out they submitted AI-completed work?

Start with curiosity, not accusation: “Walk me through how you approached this assignment. What did you learn about the topic?” The answer will tell you immediately whether they engaged with the material. If they didn’t, shift to the future-focused conversation: “What happens when you need to demonstrate this knowledge in a job interview or on a standardized test?” Shame is less effective than forward-looking consequence mapping.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

Sources

1. Roediger, H.L. & Karpicke, J.D. (2013). “The power of testing memory: Basic research and implications for educational practice.” Perspectives on Psychological Science, 1(3), 181–210. https://doi.org/10.1111/j.1745-6924.2006.00012.x
2. Porsdam Mann, S., et al. (2023). “Combining human and large language model feedback: Learning to resolve conflicts ethically.” Journal of Medical Ethics. https://doi.org/10.1136/jme-2023-109197
3. Kalyuga, S., et al. (2003). “The expertise reversal effect.” Educational Psychologist, 38(1), 23–31. https://doi.org/10.1207/S15326985EP3801_4
4. Wittrock, M.C. (1990). “Generative processes of comprehension.” Educational Psychologist, 24(4), 345–376. https://doi.org/10.1207/s15326985ep2404_2
5. Mollick, E. & Mollick, L. (2023). “Assigning AI: Seven Approaches for Students, with Prompts.” The Wharton School, University of Pennsylvania. https://ssrn.com/abstract=4475995
6. U.S. Department of Education. (2023). Artificial Intelligence and the Future of Teaching and Learning. Office of Educational Technology. https://tech.ed.gov/ai/
7. Cognitive Science Society. (2023). Learning Scientists: Overview of Effective Learning Strategies. https://www.learningscientists.org/