Why AI slop at work is more than a writing problem

Happy THURSDAY, wonderful people. It’s been a brisk week here at EnterpriseAM World Headquarters, and we couldn’t be happier to be on the cusp of the weekend. So much the better that it’s nearly Halloween — happy trick-or-treating to all those who observe (or whose kiddos will drag them out).

I hope to spend the weekend unwinding, but work beckons, so I know I’ll be thinking and reading and writing about business and AI even as I try to finish up Andrew Ross Sorkin’s 1929 (highly recommended) or distract myself with the new season of The Diplomat on Netflix (apparently the best yet).

What’s gnawing at my brain: To write is to think — and I worry that too many of us in the business community are not-so-slowly losing our ability to do both, at least on the job.

One CEO I know recently kicked off his weekly exco meeting with a little bit of performance art, reciting poetry of a form to his company’s leadership:

The gate is the loop run clean to a defensible recommendation — not the signature. The rock does not ride on a candidate's pen.

One date means bulk-set close dates

Meetings and proposals go on the weekly scorecard, not in the gate

A hypothesis wearing a rule’s clothing

Instrumentation is the unlock.

inside it

a dial worth turning

An idea that earns its place.

A plan built. Honest framing. A named responsibility.

Load bearing.

^^ All of it stuff he had pulled out of his team’s weekly reports.

Prose from an LLM can be just … exhausting to read. You get to the end of a sentence and while you know what each and every word means, you have no idea what the sentence itself is actually trying to say. Read it again and you might grasp some spectral outline of what’s intended, but actual meaning? Not so much.

Your Slack / Notion / inbox / Asana / whatever are all overflowing with Claudeslop written in Claudish, and your new job is trying to decipher what it actually means.

The worst thing you can do is point your own robot at it and ask it for help. You lost a significant chunk of meaning as an LLM went to work on your trusted lieutenant’s prompt and data inputs. Handing the output off to your robot and asking for a summary will just dilute meaning even further. (And God help you if there’s an Arabic name or 10 in the conversation from a meeting transcript.)

Ever wonder why robots write that way? Three scholars recently set out to answer just that, and their findings are in a preprint released last week. They ran an analysis of output from Opus, Fable, and Sol — nine LLMs in all — and found our robot helpers love to pair animate verbs with inanimate objects, stack nouns together, you name it. If you, too, are being driven ‘round the bend, add “Why is AI-generated language so hard to understand?” to your weekend TBR pile.

I hope that “lands” for you — because the torrent of Claudeslop (and Chatslop and Geminislop) now flowing through corporate life is more insidious than bad writing. It also suggests a deficit in thinking.

I’m starting to worry that it’s a fallacy that (most) humans want to be in the loop. Because if we do? Well, we sure as hell haven’t done a very good job of it in these early days of the AI revolution.

The danger of Claudeslop isn’t just that you’re getting machine language — it’s that you’re also getting unquestioned robot judgement masquerading as that of your team, boss, whoever.

That’s not just me, an angry middle-aged guy yelling at clouds: Research out this week from Anthropic found that in a controlled experiment of 50 analysts asked to give business advice on the same case, “77% of analysts used the agents’ own arguments — sometimes word for word — but only 15% credited the agent as the source of those arguments.”

In other words: They pretty much adopted the agents’ thinking wholesale — and may not even have been aware of it.

That’s it for me. I’m handing the pen back to Joe next week. And speaking of Joe: You won’t want to miss this morning’s Lede, wherein he writes about what it might mean for the Gulf’s AI ambitions should Democrats recapture the House and Senate when Americans head to the polls on 3 November. Have a wonderful weekend, everyone. –Patrick