That Berkeley op-ed on declining math skills was edited by AI —

A UC Berkeley professor published a piece arguing that students can't do basic arithmetic anymore because they've outsourced thinking to calculators and apps. Then we find out the op-ed itself was polished by an LLM. You can't script better satire.

The professor — I'll skip the name since the pattern matters more than the person — wrote for a major outlet about how incoming freshmen freeze on fractions, percentages, mental estimation. The argument: dependence on tools erodes the number sense that lets you catch an answer that's off by an order of magnitude. Fair point. Then the disclosure buried at the bottom: "This essay was edited with assistance from AI."

Edited. Not "brainstormed with" or "fact-checked using." Edited. The prose, structure, flow — the very cognitive labor the piece claims is vanishing — handed to a model that predicts next tokens based on statistical regularities.

What "edited" actually means in practice

Let's be precise. When academics say "AI-edited," they usually mean one of three things:

• Light pass: grammar, clarity, tone smoothing. The ideas and architecture stay human.
• Heavy restructuring: paste rough notes, ask the model to "make this flow better," accept the reorganization.
• Ghostwriting: "Write an 800-word op-ed arguing X with tone Y," then tweak a few phrases.

The disclosure doesn't specify. That ambiguity is the problem. If it's the first case, fine — that's a fancy spellchecker. If it's the second or third, the professor didn't just use a tool; they outsourced the exact skill they're lamenting the loss of: constructing a coherent argument from scratch.

The calculator analogy breaks down fast

The op-ed leans on the calculator comparison. Students use calculators → don't learn arithmetic → can't spot nonsense answers. Same dynamic with LLMs, supposedly.

But calculators compute. They don't reason. They don't decide which operation applies, frame the problem, or structure the explanation. An LLM does all three — or simulates them well enough to pass. The offloading isn't arithmetic; it's the entire cognitive scaffold.

I've seen this in CS coursework. Students paste LeetCode prompts into Claude, get working solutions, submit. They pass. Ask them to modify the approach or explain the time complexity tradeoffs — blank stares. The tool didn't just do the arithmetic; it did the understanding.

Where the line actually sits

I'm not anti-tool. I use Copilot daily. But there's a difference between:

# Me writing the logic, Copilot completing boilerplate
def calculate_compound_interest(principal, rate, periods):
    return principal * (1 + rate) ** periods

versus

# Me prompting: "write a function that calculates compound interest
# with monthly compounding, handles edge cases, includes docstring"

The first keeps me in the loop. The second makes me a reviewer, not an author. Reviewing is a different skill — weaker, more passive, atrophies faster.

What this means for evaluation

If a professor's public argument gets AI-edited without transparency on degree, how do we trust their grading? Their recommendation letters? Their research papers?

Some journals now require AI disclosure statements. Most don't specify granularity. "AI used in preparation" could mean anything from Grammarly to "write my discussion section." The disclosure becomes a checkbox, not information.

Berkeley's own guidelines for faculty AI use are surprisingly vague. "Use good judgment" — helpful.

The real skill worth measuring

The op-ed ends with a call to restore rigorous fundamentals. Agreed. But the fundamental isn't long division. It's intellectual ownership — the ability to say "I built this argument, I understand every step, I can defend or revise any part."

That's what evaporates when you hand the structure to a model. Not the arithmetic. The agency.

The professor's piece made that case unintentionally, and more convincingly than the published version ever did.