The Signal #013 — Dakota’s read on the AI news that actually matters to people running a business.

There is a difference between using a tool and building the skill. Most people know that in theory. UC Berkeley just showed what it looks like in practice when you ignore it.

What happened

In spring 2026, failing rates in multiple UC Berkeley computer science classes jumped well above the department’s own grading guidelines. According to a Daily Californian report from June 2026, 35.3% of students in CS 10 failed, and 10.6% of students in CS 61A failed. In spring 2025 and spring 2024, the failure rate for either class did not exceed 10%. The department’s own guidelines say 7% of students in lower-division courses should receive D’s and F’s combined.

Teaching professor Dan Garcia, who taught both courses, said the “primary driver” of the high failure rates was a “vast increase in academic dishonesty” tied to students using large language models (AI chat tools like ChatGPT, Claude, and Google Gemini). Nearly 30 students in CS 10 were caught cheating on take-home exams. Garcia also described a second group: students who leaned on AI to complete their work all semester and then “just really weren’t ready” when they sat down for in-person exams.

A third course, EECS 127, saw a 16.8% failure rate, far above the 5% the department calls typical for upper-division courses. Associate teaching professor Gireeja Ranade found that many students in that class struggled with linear algebra, a prerequisite. One student told her that the linear algebra class they had taken at Berkeley used an “open-internet, open-AI policy” for homework and exams. Both Garcia and Ranade are among more than 1,300 UC faculty who signed a petition calling for the reinstatement of standardized test scores for STEM admissions, citing concerns about mathematical preparation.

Why it matters for operators

You are not running a computer science department. But you are probably handing AI tools to people on your team, or thinking about it.

The Berkeley situation is a clean example of what researchers call skill atrophy (the gradual loss of a skill because you stopped practicing it). The students did not learn less because the AI was bad. The AI worked fine. They learned less because the AI removed the friction that builds competence. No friction, no muscle.

In your business, the equivalent is real. A dispatcher who uses AI to write every customer follow-up may stop knowing how to handle a difficult call without it. A tech who lets AI diagnose every system may lose the instinct that comes from doing it manually a hundred times. A project manager who uses AI to build every estimate may not notice when the AI misses a line item, because they never developed the judgment to catch it.

The tool is not the problem. Using the tool as a replacement for judgment, rather than a support for it, is the problem. That is the distinction Berkeley is learning the hard way.

What most people get wrong

Most operators hear a story like this and either dismiss it entirely or overcorrect and ban AI tools from their operation. Both responses miss the point.

Garcia’s argument is not that AI should be banned. It is that the work still has to be done with enough real effort that the person doing it actually learns something. He said students who leaned too hard on AI were not ready at exam time. The exam was just reality. Reality for your dispatcher is a caller who is upset and wants an answer right now. Reality for your tech is a system that does not match the manual. Reality for your estimator is a job that goes sideways on day three.

The question is not whether your team uses AI. The question is whether they still understand the job well enough to know when the AI is wrong. That only happens if they are doing the real work alongside the tool, not instead of it.

A new hire who uses AI to write every estimate from week one is not learning estimating. They are learning to operate the AI. Those are not the same skill, and one of them does not show up when the internet goes down or the tool gives a bad answer.

The takeaway

AI tools are worth using. They handle real volume, reduce real errors, and free up real time. None of that changes.

What the Berkeley numbers make plain is that you have to be deliberate about where you introduce them. Bring AI in after someone has built the baseline skill. Use it to speed up a competent person, not to substitute for the process of becoming one. Audit your team periodically for whether they still understand what the AI is doing on their behalf.

The students who failed were not lazy. Many of them thought they were being efficient. Efficiency and competence are not the same thing, and your operation needs both.

If you are thinking through how to roll out AI tools in a way that actually sticks, xovionlabs.com is a good place to start.