The article centers on a Brown University economics course where students did far better on take-home exams than on an in-person final, leading the professor to conclude that generative AI use had badly distorted the earlier results. The underlying issue is not one professor getting fooled. It is that take-home and unsupervised work used to be a workable compromise between trust, pedagogy, and scale, and now many people think that compromise is gone.
The strongest reaction was blunt: at-home testing is dead. Not because Brown students are uniquely dishonest, but because a tool that can generate plausible answers on demand breaks the assumption that submitted work reflects the student's own reasoning. Several faculty comments said they have already run their own spot checks by asking students to explain code or solutions they supposedly produced, and found a clean split between students who clearly understood their work and students who obviously did not. That pushed the conversation away from outrage and toward assessment design. In-person exams, oral exams, and hands-on practicals were treated as the only signals people still trust.
The discussion also widened into a bigger indictment of credentialism. Many commenters argued that college has long functioned less as learning and more as sorting, compliance testing, and access to jobs. AI did not create that weakness. It exposed it. Once a degree or grade becomes a labor-market proxy, Goodhart's law kicks in and students optimize the metric. The difference now is that optimization got cheap and scalable. That left two live questions. First, how much foundational knowledge still needs to be internalized when AI can supply fluent answers. Second, how institutions can preserve any credible signal of competence without turning education into expensive surveillance or one-on-one examinations that do not scale. The conversation landed on a hard but useful distinction: AI may belong in real work and even in some learning, but if you cannot verify independent understanding somewhere in the process, the credential stops meaning much.
If you rely on degrees, take-home assignments, or polished portfolios as signals of ability, discount them unless there is some live demonstration behind them. In education and hiring, expect more oral exams, practical tests, and supervised work samples because the old proxy is losing credibility fast.
Mostly alarmed and cynical. People see widespread AI-assisted cheating as real, unsurprising, and corrosive to both learning and the value of degrees, with only limited patience for arguments that schools can keep using old assessment models unchanged.
Key insights
01
Faculty are already running authenticity checks
Professors are not guessing anymore. They are giving students their own submitted code back and asking them to explain it, which cleanly separates those who did the work from those who outsourced it. That makes the problem feel less like abstract panic about AI and more like a measurable collapse in trust around homework-heavy grading.
If you teach or hire on the basis of submitted work, add a short live walkthrough where the person must explain decisions they supposedly made. A ten-minute defense can validate work that a polished artifact no longer can.
Practical and oral exams scale badly but restore signal
People kept returning to formats that force real-time competence. The model that got the most respect was not essay proctoring software but hands-on tests like Red Hat Certified Engineer, plus oral and practical evaluation similar to pilot checkrides. The catch is obvious. These methods produce a much stronger signal, but they are expensive in instructor time and hard to run for large classes.
Use high-trust assessment selectively where the credential really matters. For broad pipelines, reserve oral or practical validation for capstones, certifications, or final hiring stages rather than every assignment.
Grades were motivating foundation work more than people admitted
A math instructor made the clearest case for why banning AI in some settings is still defensible. Foundational courses are not only about today's task. They build the base for later study, and most students will not do that work unless there is graded pressure. Without some protected space for independent performance, you lose both the training effect and the ability to tell who can build on it.
Do not let 'AI is available in the real world' become an excuse to remove all closed-book or independent assessments. Keep some checkpoints that force unaided recall and reasoning in areas where later work depends on fluency.
The cheating story resonated because many people think too many jobs already require degrees that do not reflect actual skill. That pushes huge numbers of students into college for signaling rather than learning, and makes both cheating and resentment predictable. Several comments argued the healthier response is not just stricter exams but rebuilding trade paths, apprenticeships, co-ops, and other paid routes into middle-class work.
If you run a company, revisit degree requirements that are really acting as applicant filters. Structured apprenticeships and paid early-career training may become more reliable than college pedigrees for many roles.
The pushback was not that cheating is fine. It was that academia often treats new cognitive tools as illegitimate long after the world has moved on. If an LLM can answer a question trivially, that may mean the question is too low-level, not that the tool should simply be banished forever. A dissenting reply on calculators warned that tool dependence can also weaken intuition, which sharpens the point rather than killing it. The issue is where assistance helps abstraction and where it replaces understanding.
When redesigning courses or hiring exercises, separate foundational skills that must be internalized from tasks where tool use is now part of competent practice. Then test each category differently instead of forcing one rule across all work.
A smaller but credible point was that every generation redraws the boundary between allowed aid and cheating. Calculators were once banned. Programmable ones are now often normal. From that view, the current panic is partly about institutions lagging behind a broader redefinition of augmented human work, not just about dishonesty.
Write explicit policies about what assistance is allowed and why. Assume those boundaries will need regular revision as AI moves from suspicious shortcut to standard instrument in some domains.
arstechnica.com