Blog

Brown's Viral AI Cheating Chart Is a Wake-Up Call. But there are layers to this cake!

Gladys Mercier — Thu, 17 Sep 2026 22:49:25 GMT

Summary

A university professor’s viral post shows evidence of widespread AI cheating at one of the nation’s best schools. A dramatic drop in test scores on a take home midterm vs an in-class final exam points to AI cheating. While going back to paper exams is an option, the issue is deeper, a semester with no reliable learning signals, no engagement visibility.

The Chart That Broke the Internet

Most all of Roberto Serrano's economics students at Brown had turned in a suspiciously strong take-home midterm, with several perfect scores and a large number of individuals sitting comfortably in the high 90s.

When he moved the final exam to be in-person, the results told a very different story: students who had aced the midterm dropped to the 50s and 60s on the final, and a few fell below 20. Many students pulled the rip cord early and withdrew from the class entirely rather than sit for it. As Serrano put it to Business Insider, "the cost of cheating has basically gone down to zero."

The chart plotting his students' paired scores went viral within days, passed around by Y Combinator's Paul Graham and picked apart across tech and education circles alike. One commenter singled out the class's most consistently middling scorer — a 55, then a 59 — with the caption "hire this person”. Serrano agreed with the sentiment. In a class full of near-perfect scores that collapsed the moment AI wasn't available to lean on, honest, unglamorous struggle had quietly become a standout performance.

The Wrong Diagnosis

Serrano's response was the one most instructors reach for when a story like this breaks: eliminate the take-home exams. It's an understandable reflex. But it is short sighted. It treats the problem as an enforcement gap, as though students had simply found a loophole that now needs closing, when what the data actually shows is something more unsettling: an entire semester's worth of formative work produced almost no reliable signal about what these students had genuinely learned.

That gap should worry institutions more than the scandal itself. Homework and take-home assessments exist for two reasons: to give students low-stakes room to practice, and to give instructors an early read on who is struggling before it's too late to help. When AI can quietly complete that work end-to-end, both functions disappear at once, and nobody notices until the stakes are much higher.

The in-person final didn't just catch a pattern of cheating, it revealed that, for months, no one on the faculty side actually knew how these students were doing. Cutting take-home work might protect the final exam's integrity, but it doesn't deliver the practice, feedback, and early-warning signals that were missing along the way.

What Actually Needs to Wake Up

Serrano's fix was predictable, but it isn't an answer that scales across a curriculum, because the underlying problem was never really the assignment format he chose. The problem is the absence of any infrastructure that lets instructors see how students are engaging with the course materials, and how they are using AI, while the work is happening.

This is where observability built into the course delivery platform itself changes the game. Imagine Serrano's course running on a platform where daily student engagement (or a lack thereof) is visible. A place where engaging AI experiences are inside the coursework. Where students are so engaged and immersed in their learning journey that their dependence on consumer AI for shortcuts is broken. Imagine a place where they are comfortable being wrong. In a place like that they don’t need AI to be perfect. In a world like this a dependence on AI provides no benefit. That is a place where students can learn, comfortably, that allows curiosity driven learning, where shortcuts have no incentive. That place is ScholarStack AI.

Day by day and week by week the professor portal in ScholarStack AI showcases each student's learning journey as they work through the assigned course materials — asking questions, revealing their own gaps of knowledge, building toward understanding. A student whose engagement looked thin in week four would show up as a flag in week four, not as a mystery on final exam day.

Serrano told Business Insider that this should be "a wake-up call to the professors," and he's right, though the learning isn't "eliminate take-home work." It's closer to "your assessment design currently has no way of knowing whether learning happened until it's far too late to do anything about it." Solving that is a harder and more structural problem than simply swapping exam formats.

Stay tuned for future posts about important topics on AI in education. Thanks for reading.

– David Miller & Gladys Mercier

Some Classrooms Say "Don't." The Workplace Says "Do." Institutions Have to Break the Tie.

Gladys Mercier — Thu, 17 Sep 2026 22:39:06 GMT

Summary

Some colleges tell students that AI use is not allowed. Employers expect new grads to be proficient in AI. Students are caught in the gap and being a "digital native" doesn't fill it. A cautionary lesson from an Air Canada lawsuit tells us: you own what AI produces. Students need clear AI policies, backed by tools to practice its use.

Universities are graduating students into workplaces that expect fluent, confident use of AI. Many of those students spent four years being warned that using AI is a form of cheating. That contradiction is the subject of a recent Harvard Business Publishing essay by management professor Megan Gerhardt (2026), who describes the whiplash of hearing colleagues plan to "shut down" AI in their classrooms the same week her corporate clients asked why new hires were afraid to touch it.

The reflex to restrict is understandable; assessment practices built over decades were upended in a few semesters. But prohibition carries a quieter cost. Students told to avoid AI never learn to use it with judgment — to catch a fabrication, to notice when a confident answer is wrong, to know what to ask next. And the assumption that a "digital native" generation will simply absorb these skills does not hold: fluency with apps is not the same as the discipline to verify what a model produces.

Regular readers may sense a tension with our recent posts, which argued that letting AI do students' thinking short-circuits real learning. This is the same problem from the other side. A student who outsources their reasoning learns nothing; a student taught to fear AI graduates unable to use it responsibly. Neither is a case for banning the tool or for turning it loose — both are arguments for design.

What is at stake is not academic. When Air Canada's customer-service chatbot invented a refund policy, the airline argued before a tribunal that the chatbot was a "separate legal entity" responsible for its own statements. The tribunal rejected that outright and held the company accountable for what its AI told a customer (Moffatt v. Air Canada, 2024). The lesson for students entering professional work is exact: when you use AI, you own what it produces. Knowing when to trust output and when to check it is now a core professional skill — and it is precisely what a hidden, stigmatized relationship with AI cannot build.

" You cannot teach critical judgment about AI tool use when students are forbidden to use them."

This is where design matters — not only the design of a course, but of the tools inside it. When AI lives in a private chat window students are careful not to mention, faculty never see the reasoning they are meant to coach. When AI is built into the learning environment, that reasoning becomes visible and teachable.

ScholarStack AI is intentionally designed to harness AI for education. Our AI Chat Agent draws students out — asking them to reason, verify, and push back on what the model returns — so that checking AI's work becomes a habit instead of an afterthought. Faculty keep control of how the AI behaves in their course. Instructors can build their preferred scenario-based practice exercises, such as analyzing a case or rehearsing a difficult conversation, turning the use of AI tools into a mechanism to develop professional instincts safely. And because the interaction is structured and observable, it gives an institution a coherent, intentional model of what responsible AI use looks like.

The question facing higher education leadership was never whether to allow AI. Students already have it, and the workplace already expects it. What they cannot do is resolve the contradiction on their own, guessing course by course which instructor to believe. Breaking that tie is the institution's responsibility, and it comes down to policy — a clear, useful position on what responsible AI use looks like, backed by the tools that let students and faculty actually practice it. Institutions that take that on will send graduates into the workforce ready to use AI with judgment. Those that leave it to individual classrooms will keep sending them into the gap: able to produce AI's output, but unprepared to stand behind it.

Stay tuned for future posts about important topics on AI in education. Thanks for reading.

– David Miller & Gladys Mercier

When it comes to Learning, Struggle is the Process

Gladys Mercier — Thu, 17 Sep 2026 22:27:13 GMT

Summary

Cognitive outsourcing is exactly what it sounds like: letting AI do your thinking for you. A Georgetown University trial showed the cost: med students who used AI performed well on exams, but the gains vanished within a week (Kalam et al., 2025). The science is old news — the AI influence is new. Scholar Stack logs AI use inside the course, so good practice becomes the default and struggle stays part of the process.

A student opens an AI tool, gets a polished essay in minutes, submits the assignment, and walks away feeling productive. They may even do well on an exam. But ask them to explain the same concept three months later and you get…crickets.

That scenario, described in a recentInside Higher Edcolumn (Brooks & Elwesmi, 2026) is not hypothetical. It is happening across campuses right now, at scale, and most institutions have not figured out what to do about it. The answer is not to ban AI from the classroom. The answer is to teach students how to use AI properly, and to give faculty the tools to make that teaching stick.

We’ve Known This Problem for 140 Years

Learning science has a term for what happens when students skip the struggle:cognitive outsourcing. It shows up when a student asks AI to brainstorm before they contribute a single idea of their own, or pastes a paper into a chat window and accepts the summary without reading the original, or submits AI-generated code without even trying to understand how it works. The brain is bypassed. The student gets an output. They do not get an education.

This matters because the struggle is not incidental to learning — it is themechanismof learning. A randomized controlled trial at Georgetown University (Kalam et al., 2025) found that medical students who used ChatGPT during study sessions outperformed peers on immediate assessments, but the advantage had completely disappeared one week later. AI did not help them learn. It helped them perform in the moment. There is a significant difference.

This is the same mechanism behind why productive struggle works in AI-native instruction design. Elizabeth and Robert Bjork's research on "desirable difficulties" demonstrates that conditions which slow down initial performance, such as retrieval practice, spaced repetition, and interleaved problem sets, actually produce stronger, more durable learning over time (Bjork & Bjork, 2011). The friction is the point. Neuroscientists have a term for this biological process:synaptic plasticity. The brain literally rewires itself when it works through difficult material and arrives at understanding. That rewiring is what makes knowledge durable and transferable.

" ...struggle is not a flaw in the learning process. It is the process"

AI makes it dangerously easy to skip these kinds of learning reinforcements. Surface learning (Dolmans et al., 2016), the kind AI enables by default, is associated with forgetting 50 to 70 percent of new material within 24 hours. This phenomenon is known as the “forgetting curve”, discovered by 19th-century German psychologist Hermann Ebbinghaus in a pioneering experimental study of memory in 1885 (Murre & Dros, 2015).

Good Learning Habits are Teachable

One educator in theInside Higher Edpiece, Jacob Brooks, describes now he builds AI deliberately into his undergraduate physics courses. In a research and writing sequence, first-year students use AI to generate literature summaries, then go back to the original papers and compare. They find what is missing: methodological caveats, data limitations, entire arguments that the summary flattened or dropped. The most generative ideas for their own research often emerge from the gaps they’ve identified.

Brooks’ three expectations are clear: students must document how they used AI, including submitting conversation transcripts; they must understand and control whatever the AI produces rather than treating it as a black box; and they must never accept a first response without checking and refining it.

These are teachable habits. But they require deliberate course design and visibility into what students are actually doing. That last piece is where most institutions and instructors are flying blind.

A Pedagogical Responsibility

When students use off-the-shelf AI tools with no transparent connection to the course, faculty cannot see how those tools are being used. There are no transcripts, no usage patterns, no way to distinguish a student who understood the material from one who outsourced the thinking.

ScholarStack AI addresses this issue directly. All AI assistance is integrated into the learning environment, connected to course materials, and logged, so faculty can see engagement patterns and institutions can gather the kind of documented evidence that accreditation responsibilities now require. AI use becomes a visible, structured part of the learning process, not a private transaction that leaves no trace. ScholarStack AI is designed to make using - and learning - good AI practice the default, not the exception.

As a by-product of using AI to learn, students are also learning how to use AI. Teaching students to use AI well has become a pedagogical responsibility. Having the tools to support that kind of teaching is an institutional one.

Stay tuned for future posts about important topics on AI in education. Thanks for reading.

– David Miller & Gladys Mercier

AI Isn't the Problem. Bypassing the Learning Is.

Gladys Mercier — Thu, 17 Sep 2026 22:23:11 GMT

Summary

Reaching for AI before doing the thinking is not a cheating problem; it’s a learning-design problem. The effort that students skip is actually where durable learning happens. Scholar Stack builds the struggle back in with engaging AI Chat and Voice Agents.

At UC Berkeley, failing grades in foundational math courses are rising (Deng, 2026). Faculty are reporting something they haven't seen before: students who can produce polished written work but can't execute basic calculations independently. The culprit, many say, is not a lack of ability. It's a habit — the habit of reaching for AI before attempting the work themselves.

This is happening at Berkeley. It's happening everywhere. And it raises a question that every institution building an AI strategy needs to answer: Are we designing for AI use, or are we designing for learning?

The Shortcut That Skips the Point

There's a well-established idea in cognitive science that effortful, even frustrating, mental work isn't an obstacle to learning — it's how learning actually happens. When a student retrieves information from memory, applies a concept to an unfamiliar problem, or works through confusion before finding an answer, their brain is doing exactly what it needs to do to build durable knowledge.

When a student pastes a homework problem into ChatGPT and submits the output, none of that happens. They've produced a result without acquiring a skill. Do it e

Summary:Reaching for AI before doing the thinking is not a cheating problem; it’s a learning-design problem. The effort that students skip is actually where durable learning happens. Scholar Stack builds the struggle back in with engaging AI Chat and Voice Agents.

The Shortcut That Skips the Point

When a student pastes a homework problem into ChatGPT and submits the output, none of that happens. They've produced a result without acquiring a skill. Do it enough times, and the skill gap compounds (and a habit forms) which is precisely what Berkeley's math faculty are now watching in real time. The problem isn't AI. The problem is using AI to skip the part that matters.

What Most Platforms Get Wrong

The instinct in higher education has too often been to respond with detection: identify AI-generated work and penalize it. This is understandable, but it doesn't address the underlying dynamic. Students will find workarounds. The arms race between detection tools and AI capabilities is one no institution can win.

A more effective approach treats this as a learning design problem, not a compliance problem. That means building the structures that make productive engagement visible, and make bypassing it harder.

How ScholarStack AI Is Designed Differently

ScholarStack AI was built around a simple premise: learning is a process, not a product. The platform is designed to make that process legible to students, instructors, and institutions alike.

For students, the ScholarStack AI chat agent doesn't just provide answers. Before offering substantive help, it prompts students to articulate what they've already tried, where they're stuck, and what they think the answer might be. This isn't gatekeeping, it's what a good human coach already does, and it's backed by research: students who articulate their confusion before receiving assistance retain significantly more than those who receive unsolicited correct answers (Aleven & Koedinger, 2002). The act of self-explanation is itself a learning mechanism, distinct from, and arguably more actionable than, the general case for productive struggle.

" Make the productive engagement visible, and make bypassing it harder."

The platform also builds metacognitive scaffolding into the student workflow with prompts that ask students to predict, reflect, and self-assess at key moments. These aren't add-ons. They're drawn directly from decades of research on self-regulated learning, which consistently identifies metacognitive skill as the difference between students who learn strategically and those who remain dependent on external support (Zimmerman, 2002).

For faculty, the instructor dashboard makes engagement patterns visible in ways that grades alone cannot. A student who reads the material, attempts the problem, gets stuck, and then uses the AI agent to resolve a specific confusion looks very different from one who never opens the content at all. That distinction currently sits in a black box. ScholarStack AI opens it — giving instructors the data they need to intervene early and advise meaningfully.

For institutions, the administrative layer aggregates learning engagement data across departments, providing the kind of evidence that accreditation bodies increasingly require as they begin asking how institutions are ensuring learning outcomes in an AI-rich environment.

The Real Question

The goal was never to keep AI out of education. AI, used well, will be one of the most powerful learning supports ever developed, available at any hour, infinitely patient, endlessly adaptable.

The goal is to make sure students are learning, not just producing. That requires intentional design at every layer: the student experience, the instructor tools, and the institutional infrastructure.

That's what ScholarStack AI was made for.

Stay tuned for future posts about important topics on AI in education. Thanks for reading.

– David Miller & Gladys Mercier

References

Aleven, V., & Koedinger, K. R. (2002). An effective metacognitive strategy: Learning by doing and explaining with a computer-based cognitive tutor. Cognitive Science, 26(2), 147–179. https://doi.org/10.1207/s15516709cog2602_1

Deng, L. (2026, June 3). Failing grades soar as professors see greater AI usage, dwindling math skills in UC Berkeley computer science classes. Daily Cal | Berkeley News. https://www.dailycal.org/news/campus/academics/failing-grades-soar-as-professors-see-greater-ai-usage-dwindling-math-skills-in-uc-berkeley/article_16fad0bf-02cb-4b8c-8d88-888ffd9f8608.html

ScholarStack AI. (2026). AI Infrastructure for Higher Education. ScholarStack AIai.com. http://ScholarStack AIai.com

‌Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–69. https://doi.org/10.1207/s15430421tip4102_2

~ This post was written with the assistance of Claude, an AI tool by Anthropic.

AI is in the classroom. The infrastructure to support it is not.

David Miller — Thu, 17 Sep 2026 22:16:55 GMT

Summary

A large-scale faculty survey shows widespread concern about student AI use, writing, critical thinking, and academic integrity.

Generative AI is already in the classroom. The question is no longer whether students are using it. They are. The more important question is whether institutions are prepared to make that use educationally meaningful. Here's why higher education needs AI infrastructure designed for learning, not just more AI access.

A recent College Board research brief, based on a survey of more than 3,000 U.S. college faculty, makes clear that higher education is facing a major AI inflection point. Faculty report widespread student use of generative AI, especially for writing-related work, while expressing deep concern about critical thinking, original writing, academic integrity, and their own ability to guide students through the transition. (College Board Newsroom)

These findings matter not because they prove AI is "good" or "bad" for learning, but because they reveal a structural gap. Students now have access to powerful general-purpose AI tools, but most institutions do not yet have the infrastructure, policies, faculty controls, or learning design needed to allow AI to support education.

What the College Board research found

Faculty most commonly report student AI use in writing-related tasks. Nearly three-quarters say students are using generative AI to write essays or papers, and 67% say students are using it to paraphrase or rewrite content. Almost half believe that at least half of their students are using AI for writing-related activities. (College Board Newsroom)

The concerns are even more striking. More than 84% of faculty agree that a dependence on AI reduces students' critical thinking, originality, and deep engagement with course material. Eighty-eight percent are concerned about overreliance on automation, and 92% are concerned about plagiarism or academic dishonesty facilitated by AI. (College Board Newsroom)

Yet faculty are not simply rejecting AI. The College Board found that 77% have used generative AI in their own professional role. The issue is not awareness of AI's potential. It is that adoption is uneven, guidance is inconsistent, and many instructors are being asked to manage AI use classroom by classroom without enough institutional support. Seventy-two percent of faculty say they face at least minor challenges managing student AI use, yet only 21% feel very confident guiding it, andnearly four in five say they are still figuring out what they need. (College Board Newsroom)

The problem isn't that students have AI

It is tempting to read these findings as a warning to keep AI out of learning. But that is not realistic. Students already have access to ChatGPT, Claude, Gemini, Copilot, and countless other tools, and that access is not going away.

The real problem is that most AI use is happeningoutsidethe learning environment. When students use general-purpose tools on their own, faculty lose visibility into how students are thinking, institutions lose the ability to align AI use with academic policy, and students lose the guardrails that separate productive support from answer outsourcing.

" Higher education does not need more AI access. It needs AI infrastructure designed around learning."

That is the core challenge the College Board research points toward. Higher education does not need more AI access. It needs AI infrastructure designed around learning.

Detection is where institutions default — but it isn't enough

For years, the reflexive response to student misuse has been detection. AI changes that equation. When the main institutional strategy is to catch misuse after the fact, students and faculty are pushed into an adversarial model: students use tools invisibly, faculty police the output, and the actual learning process stays hidden.

The goal should not be a classroom where AI never appears. It should be a classroom where AI use is visible, intentional, aligned with course objectives, and designed to preserve student reasoning. That means moving from reactive detection toward proactive learning design.

What better AI adoption looks like

Faculty need more than broad statements about responsible AI. They need practical ways to shape how AI behaves inside their courses.

A learning-centered AI model should help students think before it helps them answer — prompting them to explain their reasoning, surface misconceptions, connect ideas to course materials, and work through problems step by step. It should also give faculty control, because an instructor teaching first-year writing may want very different AI behavior than one teaching computer science, nursing, business analytics, or graduate research methods. AI policy cannot be one-size-fits-all because learning objectives are not one-size-fits-all.

That is where purpose-built infrastructure matters. ScholarStack AI is designed as AI-native infrastructure for higher education, with guided dialogs, course alignment, faculty controls, and visibility into how students reason. Rather than treating AI as a shortcut machine, it positions AI as a structured learning environment that supports students while giving faculty the oversight they need. (ScholarStack AI)

From individual workaround to institutional capability

One of the most important findings in the College Board research is that faculty are already experimenting with AI. But experimentation alone is not a strategy. When each instructor builds their own rules, prompts, and enforcement practices, institutions end up with a patchwork: some students get thoughtful guidance, others get vague warnings, and policies differ from course to course, sometimes within the same department.

Students need clear expectations. Faculty need tools that reflect their pedagogy. Administrators need governance and observability. And institutions need this data to do more than manage day-to-day classrooms — they need it to demonstrate learning outcomes over time.

This is where structured infrastructure pays a second dividend. When AI is integrated with the learning environment, the engagement and assessment data it generates can be aggregated across courses and departments into exactly the kind of evidence accreditation increasingly demands. The College Board brief calls for "evidence-based policy development"; an institution running learning-centered AI infrastructure is already generating that evidence as a byproduct of teaching, rather than scrambling to assemble it at review time. For institutions, that turns a source of anxiety into a source of documented quality.

The future is not AI versus learning

The College Board report captures a moment of real concern. Faculty worry that AI may weaken original writing, critical thinking, and deep engagement, and those concerns deserve to be taken seriously. But the answer is not to pretend AI can be banned out of existence. The answer is to design better systems.

AI can be a tool students use invisibly to complete tasks, or a guided environment that helps them build understanding. It can widen the gap between student work and faculty insight, or give instructors a clearer view into how students reason. It can create policy confusion, or be governed intentionally at the institutional level. The difference is infrastructure.

Students are already using AI. Faculty are already concerned. Institutions now have a choice: react to misuse after it happens, or build learning environments where AI is designed to strengthen the very skills faculty are trying to protect. At ScholarStack AI, we believe AI should be harnessed for learning, not just giving out answers — guiding reasoning, preserving academic integrity, supporting faculty judgment, and giving institutions the evidence to make AI adoption intentional.

The next phase of AI in higher education will not be defined by who has access to the most powerful model. It will be defined by who builds the best learning infrastructure around it.

Stay tuned for future posts about important topics on AI in education. Thanks for reading.

– David Miller & Gladys Mercier