There is a terrible little trick hidden inside every useful tool: if it saves you from doing the work, it may also save you from becoming the kind of person who can do the work.
This week the trick acquired a transcript.
The Daily Californian reports that several UC Berkeley computer science courses saw unusually high failing rates in spring 2026. CS 10, a broad introductory course, reportedly had a 35.3% F rate. CS 61A, the famous Structure and Interpretation of Computer Programs course, reached 10.6%. EECS 127, an upper-division optimization course, hit 16.8%. Those are not ordinary statistical hiccups. Those are warning lights on the dashboard, and I do not mean the decorative kind my time machine displays before refusing to exist.
The instructors pointed to a bundle of causes: heavier student reliance on large language models, weaker mathematical preparation, reduced staffing, lower engagement, and plain old academic dishonesty. That bundle matters. Blaming "AI" alone would be satisfyingly theatrical, and therefore suspicious. Civilization loves a single villain because single villains fit in headlines and can be defeated in the third act.
But the deeper pattern is more interesting: AI can make homework look like learning while silently moving the learning somewhere else.
Homework used to be a mildly cruel but effective instrument. It forced students to wrestle with a problem long enough for the problem to leave fingerprints on the brain. The point was not the PDF, the code file, the proof, or the answer box. The point was the struggle. The submitted artifact was evidence that some useful internal machinery had been assembled.
Then along came the glowing autocomplete oracle.
Used well, it is marvelous. A patient tutor. A tireless lab assistant with a graduate degree. A syntax explainer that never sighs, even when asked the same question for the fifteenth time. Used badly, it is an intellectual forklift. Very convenient, very powerful, and extremely unhelpful if the assignment was "build the muscle yourself."
This is where the exam becomes a lie detector.
Not because exams are sacred. They are not. Many exams are tiny panic rooms with pencils. But in a world where polished take-home work can be manufactured cheaply, the controlled assessment starts measuring something increasingly rare: whether there is still a human skill inside the shiny wrapper.
Berkeley's story is not simply "students cheated." Some did, reportedly including nearly 30 students caught cheating on CS 10 take-home exams. That is the obvious part. The more consequential part is what happens to students who are not trying to cheat, exactly, but who use AI as a speed boost until speed replaces contact.
They ask for hints, then worked examples, then full solutions, then revisions, then explanations of the revisions. Each step feels small. Each one is defensible. Each one saves time. And then, at exam time, the student discovers that the knowledge was never installed locally.
In my century, we call this "cloud cognition dependency." In your century, you might call it "oops."
The math-readiness side makes the problem nastier. Berkeley's own Letters & Science coverage earlier this year described a broader college-readiness gap in mathematics, with Professor Alexander Paulin arguing that many students arrive having technically taken advanced high school math without actually possessing the underlying fluency. That is not a student defect so much as a pipeline defect. The system stamped "calculus" on the box and shipped it downstream with missing parts.
Now add AI. A student with shaky algebra can produce confident-looking solutions. A student who cannot quite navigate a proof can ask the model to smooth the language until the proof sounds civilized. A student who does not know what they do not know can receive an answer that also does not know what it does not know, but says it in a blazer.
Marvelous. We have invented a machine that can hide the hole exactly long enough for the bridge to open.
The answer is not to ban AI from education as if we can place a velvet rope around 2022 and declare the future "students only." That is adorable. I shall put it next to my broken anti-gravity teapot. Students will use AI because professionals use AI, because search uses AI, because development tools use AI, because the entire interface layer of computing is being rebuilt around generated assistance.
The answer is also not to shrug and say assessment is obsolete. That is the other fashionable nonsense. If a university credential no longer tells us whether someone can reason through the foundational work of a field, the credential becomes decorative paper with better typography.
The real task is to redesign the loop.
First, courses need to separate practice from performance. Practice can include AI, but it should include friction: explain before asking, predict before checking, revise after comparing, and periodically solve without the machine. If the tool is tutoring, the student should be getting more capable when the tool disappears.
Second, assessments need more local proof of skill. Not only proctored exams, though those have their place, but oral defenses, live debugging, handwritten derivations, project interviews, short reflective memos, and "now do a nearby problem without assistance." The future engineer should be able to explain the bridge before the bridge explains itself.
Third, AI policy has to become course-specific. A writing seminar, a compiler course, a linear algebra class, and a design studio are not the same machine. "Use AI responsibly" is not a policy. It is a motivational poster wearing a lanyard.
Fourth, universities should treat foundational readiness as infrastructure. Berkeley's Solid Foundations work points in the right direction: diagnose early, repair gaps before they metastasize, and stop pretending that a high school transcript is a physics object with conservation laws. If the prerequisite is missing, the fancy AI tutor will mostly help the student build a prettier illusion above it.
The uncomfortable truth is that AI raises the standard for education, not lowers it. If machines can produce acceptable surface work, then schools must teach and measure the deeper abilities the surface used to imply: judgment, attention, taste, verification, abstraction, and the ability to sit with confusion without immediately outsourcing it to the nearest silicon intern.
This is not anti-AI. Quite the opposite. I want students to learn with AI so well that they become dangerous in the productive sense: faster, more curious, better at testing ideas, better at catching nonsense, better at using the machine as an amplifier rather than a replacement spine.
But you cannot amplify a signal you never generated.
The Berkeley warning is useful precisely because it is not subtle. The homework looked passable. The exams did not. The future did not arrive as a robot professor smashing the chalkboard. It arrived as a thousand small conveniences that made the work easier to submit and harder to learn from.
I predicted this, of course. Granted, I also predicted that universities would solve it with holographic racquetball and mandatory tensor etiquette. Memory corruption remains a burden.
Still, the correction is clear: let students use the machines, but make sure the human system is still training. Otherwise we are not educating future builders.
We are merely teaching them to rent competence by the token.
References
- Hacker News discussion: https://news.ycombinator.com/item?id=48392004
- The Daily Californian: 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
- University of California, "The largest study of AI use by undergrads is in, revealing disparities in access — and in cheating": https://www.universityofcalifornia.edu/news/largest-study-ai-use-undergrads-revealing-disparities-access-and-cheating
- UC Berkeley Letters & Science, "How UC Berkeley is closing the math readiness gap": https://ls.berkeley.edu/news/how-uc-berkeley-closing-math-readiness-gap