Classrooms are entering a new phase of the AI era—one where the debate is no longer limited to whether students can use tools like chatbots, but whether schools should redesign assessment around what those tools make possible. In recent months, educators have found themselves pulled in two directions at once: on one side, pressure from AI labs and vendors eager to integrate their systems into learning platforms; on the other, mounting anxiety that the very presence of AI could weaken students’ thinking habits, reduce productive struggle, and turn evaluation into a game of surveillance.
At the same time, a third concern has become harder to ignore. Several “cheating detection” products—often marketed as safeguards against AI-generated work—have come under scrutiny for reliability. The implication is uncomfortable for schools: if detection systems are prone to false positives, then the cost of trying to police academic integrity may be paid not only in student trust, but also in fairness. That combination—AI’s expanding role plus doubts about enforcement—has led some professors and education researchers to argue for a different strategy: stop treating AI as something to catch, and start treating it as something to learn with. The goal is to give students “eval powers”: the ability to evaluate claims, sources, reasoning quality, and output limitations, rather than simply producing answers.
This shift is not a call to ignore misconduct or to lower standards. Instead, it reframes what standards should measure. If AI can generate plausible text quickly, then the question becomes: what evidence will demonstrate genuine learning? Many educators now believe the answer lies less in detecting authorship and more in assessing judgment—how students decide what to trust, how they verify, and how they revise when they discover errors.
The pressure on schools is real, and it’s arriving from multiple angles. AI labs are increasingly positioning their tools as “learning companions,” “tutors,” or “productivity layers” that can be embedded into homework, classroom management, and even grading workflows. For administrators, these offerings can look like a solution to staffing constraints and a way to personalize instruction at scale. For teachers, they can appear as both opportunity and threat: opportunity because AI can provide explanations, practice prompts, and feedback; threat because it can also replace the very cognitive work that makes learning stick.
That tension has fueled a growing conversation about “cognitive atrophy”—a phrase used by critics to describe the risk that students may rely on AI outputs instead of building internal skills. The concern isn’t that AI is inherently harmful; it’s that students may gradually outsource the mental steps that lead to understanding. When the fastest path to a submission is to ask a model for an essay, the slow path—reading, outlining, grappling with ambiguity, checking facts—can become optional. Over time, optional becomes default.
But there’s another layer to the atrophy argument that educators often mention privately: it’s not only about effort. It’s about identity. Students learn who they are as thinkers through repeated experiences of making sense of material. If AI reduces those experiences, students may lose confidence in their own reasoning. Even when they do well on assignments, they may struggle later when the scaffolding disappears. In other words, the fear is not just that students will cheat; it’s that they will stop practicing the skills that cheating bypasses.
Yet the response—tightening surveillance—has its own problems. Cheating detection tools have been widely discussed, and the article-level debate has intensified as educators report inconsistent results. Some systems flag student writing that appears “too AI-like,” while others fail to detect AI assistance at all. The most troubling issue is that these tools can treat writing style as a proxy for integrity, even though style varies widely across age, language background, disability accommodations, and individual writing development. A student who writes clearly and concisely might be misread as suspicious. A student who struggles with grammar might be misread as “human” even if they used AI to structure their work. The result is a system that can punish the wrong behaviors and miss the ones that matter.
This is why the “eval powers” approach is gaining traction. It doesn’t rely on guessing whether a model wrote a paragraph. Instead, it asks students to demonstrate evaluation in ways that are difficult to outsource. Evaluation is not a single skill; it’s a cluster of abilities: judging credibility, spotting logical gaps, comparing sources, recognizing uncertainty, and revising conclusions when new evidence emerges. These are the kinds of competencies that AI can assist with—but not fully replace—because they require context, domain understanding, and accountability for decisions.
In practice, teaching evaluation powers means changing the shape of assignments. Traditional tasks often reward the final product: the essay, the problem set, the lab report. Under AI conditions, the final product can be generated quickly, which makes it a weaker signal of learning. The alternative is to assess the process and the reasoning behind it. That can include requiring students to submit evidence trails: drafts with tracked revisions, annotated sources, short “reasoning memos” explaining why they chose certain claims, and reflection prompts that ask them to identify what they were uncertain about and how they resolved it.
One unique angle in the current debate is that evaluation training can be designed to be explicitly AI-aware without being AI-dependent. Students can be taught to use AI as a sparring partner rather than a ghostwriter. For example, they might ask an AI to produce multiple competing arguments, then evaluate which argument is stronger and why. Or they might request a summary of a research paper, then check the summary against the original text and identify inaccuracies. The key is that the student’s grade is tied to verification and critique, not to the ability to generate fluent prose.
This approach also addresses a common misconception: that evaluation skills are abstract and therefore hard to teach. In reality, evaluation can be made concrete through routines. Students can learn checklists for source credibility, methods for cross-verifying claims, and structured ways to test whether an explanation actually matches the evidence. They can practice identifying common failure modes in AI outputs—confident errors, missing context, invented citations, and oversimplified causal claims. When students learn these patterns, they become less vulnerable to persuasive but incorrect content.
Importantly, evaluation training can reduce the need for adversarial policing. If students know they will be asked to justify decisions and verify information, then the incentive shifts. Using AI to brainstorm or to draft is still possible, but the student must engage with the substance. The assignment becomes less about producing text and more about demonstrating judgment. That’s a fundamentally different model of integrity: not “prove you didn’t cheat,” but “prove you can think.”
There is also a cultural shift implied by this strategy. Many school systems have treated assessment as a gatekeeping mechanism: pass or fail, correct or incorrect, authorized or unauthorized. The eval powers approach treats assessment as a learning mechanism. It assumes that students will use tools, and it builds the curriculum around teaching them how to use those tools responsibly. That doesn’t mean surrendering standards. It means redefining what standards look like.
For teachers, this can be both empowering and demanding. Empowering because it offers a clearer pedagogical mission: teach evaluation, not detection. Demanding because it requires time to redesign assignments and rubrics. Teachers must learn how to evaluate reasoning quality and evidence use at scale. They also need support to ensure that evaluation tasks are fair across student backgrounds. For instance, students with limited access to tutoring or advanced reading materials might struggle with verification tasks unless scaffolding is provided. The solution is not to abandon evaluation, but to implement it with instructional supports: guided practice, exemplars, and incremental complexity.
Administrators and policymakers face a parallel challenge: how to align institutional policies with classroom realities. Many schools have issued blanket bans on AI tools, often framed as “no use.” But blanket bans can be difficult to enforce and can push usage into hidden channels. A more workable policy might be “use with accountability,” where students are allowed to use AI for specific purposes—brainstorming, outlining, translation, practice—while being required to cite, verify, and document their evaluation steps. This kind of policy is harder to communicate than a simple ban, but it is more consistent with how students actually behave in a world where AI is already embedded in everyday software.
The “eval powers” argument also resonates with a broader educational trend: competency-based learning and authentic assessment. In many fields, professionals are not judged by whether they wrote every sentence themselves. They are judged by whether their work is accurate, defensible, and aligned with evidence. Teaching students to evaluate is essentially preparing them for that professional reality. In workplaces, AI tools are likely to be ubiquitous. The differentiator will be the ability to decide what to trust, how to validate, and when to escalate uncertainty.
This is where the debate becomes more than a classroom issue. It’s about the future of knowledge work. If AI can draft, summarize, and simulate, then human value shifts toward oversight: verifying outputs, checking assumptions, and applying context. Education that trains evaluation powers is therefore not merely a workaround for cheating detection failures. It is a direct response to the changing nature of expertise.
Still, skeptics worry that evaluation training could become performative. Students might learn to write “reasoning memos” that sound convincing without truly reflecting understanding. That risk is real, and it’s why evaluation tasks must be designed carefully. Good evaluation assignments include opportunities for students to encounter contradictions and to correct them. They also include questions that require specificity: not just “is this claim true?” but “what evidence supports it, and what would change your mind?” When students are asked to defend decisions with concrete references, superficial compliance becomes harder.
Another safeguard is to incorporate iterative cycles. Instead of a one-shot submission, students can be required to revise after receiving feedback—either from a teacher, peers, or structured automated checks that focus on evidence alignment rather than writing style. Revision forces students to engage with evaluation as an ongoing practice. It also mirrors real-world workflows, where drafts are rarely final on the