Berkeley CS Students Are Failing at 3x the Normal Rate. AI Made Them Worse, Not Better.

Using AI to skip the thinking rather than do the thinking makes you worse at your job. That sounds like a hypothesis. It isn't — it's what the data shows. A university course that adopted AI tools saw its failure rate triple in one semester. The mechanism is documented. And it's not a student problem.

The Berkeley data

In Spring 2026, 35.3% of students in Berkeley's CS 10 course received F grades. Historically, the rate ran under 10%. Professor Dan Garcia attributed the spike to a "vast increase in academic dishonesty" via large language models — 30 students were caught cheating on take-home exams, and average grades dropped to C+ (2.3 GPA), well below the department's expected 2.8–3.3 range. The story reached 822 points on Hacker News on June 4. (Source: the Daily Californian, June 4, 2026.)

The mechanism is what Ethan Mollick documented in "Choosing to Stay Human." Students who used AI to generate answers scored higher on AI-assisted work and lower on independent tests taken immediately after. The capability gap opened quietly and fast. Students who used AI as a tutor — asking it to explain reasoning, then checking their own understanding — showed the opposite pattern.

What this means Monday morning

The Berkeley data is CS-specific. The pattern isn't.

Every professional who uses AI as an answer machine — paste in the question, accept the output, move on — is running the same experiment on their own judgment. The skill that atrophies first is usually evaluation: knowing when the answer is wrong, incomplete, or technically correct but wrong for the specific situation.

Three ways AI makes you more capable over time:

1. Ask for the reasoning, not just the answer. When AI produces an output, ask it to explain the tradeoffs. You don't have to read the explanation every time — but asking means you occasionally catch the gap between what you wanted and what you got.
2. Do the hard part first. Draft the email, sketch the strategy, write the first version. Then use AI to push it further. The quality of what you produce independently tells you whether your judgment is still intact.
3. Test yourself without AI periodically. Pick one task you regularly delegate. Do it yourself once. The gap between your output and the AI's tells you whether there's a problem.

Three ways AI makes you worse:

1. Using AI for every version of a task, including drafts you could write in five minutes.
2. Accepting AI output without reading it against the specific situation.
3. Substituting AI fluency — knowing how to prompt — for domain fluency: knowing enough to evaluate the result.

Rachel: the wrong pattern

Rachel is a senior communications manager at a 95-person logistics company. Since January, she's used AI to draft every internal announcement, client update, and quarterly summary. Output quality is fine. But last month, she needed to write a sensitive message about a contract dispute — too confidential for a connected AI tool — and found herself staring at a blank document for twenty minutes. The capability wasn't gone, but the confidence was. She hadn't written from scratch in months.

The problem isn't that AI wrote everything badly. It's that Rachel stopped practicing the part of her job that requires judgment under pressure. When the situation required it, she didn't know how much of her skill remained.

Marcus: the right pattern

Marcus is head of operations at a 210-person food manufacturing company. His team writes supplier escalation memos — documents where tone, sequencing, and framing matter as much as the content. He uses AI, but with a standing rule: he writes the first version. He gives AI his draft and asks it to find what's unclear, what's missing, and what a supplier would push back on. He incorporates what's useful and ignores the rest.

His team's memos got sharper. More importantly, Marcus can still write one when the situation is too sensitive to involve a tool connected to external services. The AI improved his output without replacing his judgment.

The measurement

Ethan Mollick's "Co-Existence" framing captures what the Berkeley data points toward: the future of AI at work isn't about using more of it — it's about knowing when AI is better than you (delegate, light supervision) and when staying engaged is what keeps you valuable (stay in the loop, protect the judgment).

Berkeley's data is a measurement, not an abstract warning. The measurement is: this specific pattern — AI as answer machine — produces measurably worse outcomes at scale. The professional question is whether you're running the same experiment on your own work without knowing it.