A version of this story appeared in CNN Business’ Nightcap newsletter. To get it in your inbox, sign up for free here.

We may have just witnessed the most egregious instance of workslop to date, and it’s one that matters — not only because it’s objectively funny, but also because it captures an under-discussed nuance in the way generative AI functions (or malfunctions) for different industries.

Bear with me.

On Saturday, a top-ranked lawyer at one of the most prestigious law firms on the planet apologized profusely in a letter to a judge after submitting a court filing peppered with errors, including fabricated citations, generated by AI.

“We deeply regret that this has occurred,” Andrew Dietderich, co-head of Sullivan & Cromwell’s restructuring division, wrote in the letter, which included a three-page list identifying and correcting each of the more than 40 errors. (A little salt in the wound: Dietderich said he learned of the problems only after they were caught by opposing counsel from Boies Schiller Flexner.)

In the letter, Dietderich chalked the errors up to “hallucinations” in which AI tools “fabricate case citations, misquote authorities, or generate non-existent legal sources.” He also said that while the firm has safeguards around AI to prevent “exactly this situation,” those policies were not followed in the preparation of that particular document.

Now, this was hardly the first (nor, likely, will it be the last) instance of fancy-pants lawyers running into an AI buzzsaw. This kind of thing happens with surprising frequency, though rarely do we see it from the likes of Sullivan & Cromwell, an elite Wall Street firm whose partners reportedly charge around $2,000 an hour for bankruptcy cases. (The firm didn’t respond to a request for comment.)

But one of the more striking things about this episode this how it highlights AI’s utility gap. More than three years into the breathless hype cycle kicked off by the launch of ChatGPT, it’s clear that generative AI can do a lot for a very specific kind of worker — namely, those who code — and it can lead to an embarrassing boondoggle for others.

That’s because coding is fundamentally deterministic, meaning there are yes/no, right/wrong outcomes. In coding jobs, the software you’re building either works or it doesn’t.

Other modes of office work tend happen in gray areas: How do we craft a slogan that reflects our values? Will my boss prefer serif or sans serif headings in this pitch deck? Which bit of case law should I cite to best support my client’s case?

In non-coding jobs, there are degrees of functionality informed by value judgments. (This newsletter, for example, still goes out even with typos, as my regular readers are keen to note.) Of course, you can ask a chatbot to weigh in, or use it as a sounding board, but there is no single, irrefutable answer to those kinds of questions.

This distinction of science versus art matters, because right now, tech companies and investors on Wall Street are making huge bets on AI. But, as investor Paul Kedrosky told “Plain English” podcast host Derek Thompson last month , those investors are often basing their demand estimates on the experience of early adopters in tech who are “profoundly unrepresentative of the rest of the real world of work.”

Coders’ work is also uniquely expansive, Kedrosky argued. In other words, the more code you write, the more power it requires. Most other white-collar applications for AI “tend to be compressive — ‘I’ve got a giant report, I don’t want to read it, tell me the bullets.’”

None of this is to say AI isn’t (or won’t one day be) helpful for lawyers, researchers, journalists, marketers and the like. It’s just that the promise of an AI revolution originated with people like Sam Altman, Dario Amodei and Mark Zuckerberg — folks who not only stand to get much richer if we all start using AI, but also who are most familiar with the world of tech.

And that matters when it comes to proving what’s hype and what’s real promise. It’s arguable that three-ish years isn’t enough time for large language models to prove themselves as the world-destroyers they are promised to be.

But it’s not like LLMs are the only AI application the world has seen. Tesla’s “Full Self Driving,” for instance, still isn’t quite doing what customers were promised, even a decade after CEO Elon Musk predicted it would drive fully autonomously, coast-to-coast, within two years.

That hasn’t stopped Tesla from selling the system based on the idea that it sort of works, sometimes or often, depending on conditions, with a human to assist it. It’s better than it was, but it’s not good enough to replace every taxi driver just yet.

And maybe that’s where AI in general is really headed. Could it be an imminent world destroyer? Maybe. Could it also just be something that helps out, but still needs a human to assist it to avoid disaster, for the foreseeable future?

Or how about this one, more specifically: Could AI models digest all the legal texts ever written, proving we won’t need as many human lawyers or paralegals?

The answer, like so many answers in white-collar work: Maybe! In light of Sullivan & Cromwell’s gaffe, it’s fair to say LLMs aren’t quite ready to represent humans in court.