Two facts sit side by side that most people never put together.
The first: researchers demonstrated that ChatGPT could complete an entire semester-long engineering course under a minimal-effort usage protocol and earn a passing grade. Not a paragraph. Not an essay. A full course, from the assignment workflow up.
The second: 69% of architecture students surveyed by the SCOSA working group had never used AI to generate ideas for studio work, and 60% want AI-assisted work marked differently.
Read those together and the story flips. The machine that can auto-complete your curriculum is not being wielded recklessly by a student body hungry for shortcuts. The students, at least in design, are more cautious than the institutions building policies around them. The competence problem lives somewhere else. But before we find it, we have to be honest about what actually broke.
Start with a scene. July 2026, lunch, two professors who teach interaction design at ArtCenter College of Design and at Santa Monica College. Nerds, so the conversation drifts to AI, and they name their shared enemy: the "aggressively mediocre" output their students now hand in. Competent. Passable. Empty.
Here is where most readers stop, nodding, and get it wrong. The danger is not that the output is mediocre. The danger is that it is good enough to skip the exact friction that once produced expertise capable of detecting mediocrity in the first place.
The danger is not that the output is mediocre. The danger is that it is good enough to skip the exact friction that once produced expertise.
Both these professors learned their craft through that friction. One was baptized into HTML and CSS in 2003, customizing a MySpace profile page. Both learned to manipulate code and hardware — Arduino, Raspberry Pi, C++ — by building prototypes through trial and error, out of necessity. The stumbling was the curriculum. Nobody assigned it; the tools simply refused to work until you understood them.
Their students, they observe, can prompt but cannot think. And that phrase is the whole problem in five words.
Because the friction those professors are nostalgic for was never a bug in design education. It was the mechanism. The struggle with responsive breakpoints is how you internalize the constraints that make responsive design hard. The hours spent on microcopy are how you develop an ear for it. The bad decisions, made and lived with, are how design judgment forms at all. Remove the friction and you remove the formation. AI has not degraded the output. It has severed output from the process that used to guarantee it.
Which brings us to the real diagnosis, and it is not about AI at all.
We always graded polish. Grammatical correctness, structural completeness, a clean set of frames, a citation formatted right — these were the rubric. And they were always imperfect proxies for learning; the assessment literature has said so for decades. We used the product as a stand-in for the process because the product was easier to see and the correlation between them was reliable enough to trust.
AI broke the correlation. That is the entire event. Two students can now submit identical products from wholly different reasoning pathways — one who reasoned to the answer, one who accepted a recommendation unchanged. The proxy we graded for a century has become worthless, because the thing it was proxying for is now optional.
The false-signal problem, stated plainly: high-quality output no longer reliably indicates genuine understanding.
Notice what this does to detection software. Detection assesses the product, not the process — it generates false positives, particularly against certain groups of writers, and false negatives the moment output is paraphrased. It is a tool built to defend a proxy that already stopped working. You cannot catch your way out of a measurement failure.
And petitioning AI companies to cripple their products in educational contexts misreads the industry entirely. The scaling literature documents a sustained march toward larger, more capable models, with performance improving predictably as size and compute increase. Companies competing for market share have every incentive to expand capability, none to constrain it. There is no lever an educator can pull that reaches that roadmap.
So schools built their whole response around the wrong person.
The Digital Education Council surveyed 1,681 faculty across 52 institutions in 28 countries. Only 17% placed themselves at advanced or expert AI literacy. Forty percent described themselves as just beginning. Eighty percent reported a lack of institutional clarity on how AI can be applied in their teaching.
Meanwhile the students: the HEPI 2026 survey found 68% believe AI skills are essential to their futures, while only 48% feel their teaching staff are helping them build those skills — and in arts and humanities, which includes design, that under-support is the highest of any subject area.
Picture the faculty meeting. Slides on academic integrity. A vendor pitching detection software. A restructured rubric to make AI-assisted work harder to submit unpenalized. Senior educators who have taught for decades sit through all of it. At no point does anyone ask whether the people in those chairs know enough about generative AI to teach with it critically.
That is the misfire in a sentence. Institutions spent two years managing AI as a student integrity problem. The actual competence gap — faculty who lack the language and conceptual framework to teach AI critically — was allowed to grow undocumented. That gap, not the students, is the structural liability.
Because teaching a student to navigate AI is a specific skill, and it is not the same skill as teaching citation ethics. As Des Fagan of Lancaster's SCOSA group put it, the danger is not the technology itself but "the lack of critical evaluation of inputs, models and outputs." Teaching that requires a faculty member who understands how generative models produce output, who can tell AI-as-shortcut from AI-as-expansion, and who has the confidence to redesign critique — not just redesign the rubric. Only 17% of them, by their own account, are anywhere near that.
So what do you actually do? You stop grading the product.
The durable move is to redesign what we ask students to do, at all three levels where the problem lives: what we assign across a curriculum, how individual tasks are structured, and what evidence of thinking we collect and grade. Three kinds of assignment break most visibly in an AI-saturated room — anything built on information retrieval, anything graded on polish over reasoning, and anything without decision history. Fix all three by grading the thing AI cannot fake: the pathway, not the artifact.
Grade decision history. Not the final layout but the sequence of choices that produced it — what was tried, what was rejected, and why. Two identical products with different reasoning trails are now different submissions, which is exactly the distinction the old rubric could not draw.
Grade evaluative judgment. Bearman and colleagues make the point precisely: generative AI has widened the gap between a student's capacity to produce work and their capacity to evaluate its quality. That capacity to evaluate — to look at plausible, polished output and say what is wrong with it — is now the critical competency. Assessment's job is to develop it.
This is also the difference between offloading and thinking. Unguided AI use is associated with cognitive offloading and reduced critical thinking; structured prompting preserves and improves reasoning over unassisted AI use. The variable is whether the task keeps the student cognitively engaged. That is an assignment-design problem, and it is solvable.
The reframe that makes all of this teachable: treat the shortcut as a visible object, not a banned one. A shortcut is only corrosive when we pretend it is a journey. Named, reflected on, justified, it becomes a site of learning — the process made visible again rather than hidden and policed. The integrity comes from valuing reflection, not from catching offenders.
One caution, and it is not small. A lowered ceiling does not show up on the transcript. A student who prompts competently, submits polished frames, and never builds the internal model of how design decisions cascade will look fine — for years. The deficit surfaces later, in the design review where they cannot say why this hierarchy, only that the AI suggested it, or in the genuinely novel problem with no template to retrieve. That is the same wall the ARC benchmark exposes in machines: extraordinary at interpolation, helpless the moment the problem is new. We are at risk of optimizing a generation for exactly the tasks the machines already do better.
The friction was never the enemy. It was the whole point. The task now is to design assignments that put the friction back where it belongs — in the reasoning, not the rendering — and that requires faculty who understand the tools well enough to teach with them. Right now, by their own count, most don't. Fix the staff room first.