The document lands in your inbox. It's formatted correctly, the sections are in the right order, and at first glance it looks fine. Then you read it. The brackets are still there - "[insert specific example here]" - the conclusion contradicts the opening, and it's clearly about a slightly different question than the one you asked.
This is what practitioners are calling "Claude soup": AI-generated content submitted as a finished deliverable without anyone reading it first. The name is new but the pattern is spreading fast enough that it's showing up in team retrospectives and manager one-on-ones.
What Makes It Different from Ordinary Bad Work
Unreviewed AI output has a specific texture that's hard to ignore once you know what to look for. The writing is fluent - often better-structured than what the person would write themselves - but the judgment is absent. Arguments contradict each other because the model is optimizing for coherence within paragraphs, not across the whole document. Facts are stated confidently but occasionally wrong. The document answers a reasonable version of your question, not the specific one you asked.
It also multiplies in a way sloppy human work doesn't. Someone who produces mediocre documents usually produces five a week. Someone who sends Claude output without reviewing it can generate twenty, which means the signal-to-noise problem compounds fast.
The Workflow Assumption That's Breaking Things
The underlying issue isn't that people are using AI - it's that they've collapsed "generating text" and "producing a deliverable" into the same step. AI drafting is useful precisely because it produces a starting point fast. Reviewing that starting point, catching what's wrong, and editing it into something accurate and specific - that's the work. Skipping it doesn't eliminate the work; it just transfers it downstream to whoever asked for the document.
Some teams are handling this by adding review as an explicit process step, separate from generation. Others are asking colleagues to mark AI-assisted documents with a brief note on what was reviewed and what wasn't. Neither is a perfect fix, but both make the invisible step visible - which is most of the problem.
The harder question is cultural. If someone is sending unreviewed output in meetings, they've concluded that speed matters more than accuracy in your team's environment. That's sometimes a reasonable call. More often it's a sign that the review step feels optional because no one has been caught out yet. Once they are - and it ends up in a client deck or a board meeting - the invisible cost becomes very visible very fast.
The tools aren't going to solve this on their own. Claude has gotten noticeably better at flagging its own uncertainty and noting when it needs more specific input. But output that looks confident will still get submitted unreviewed by someone who didn't read it. The tooling helps. The workflow habits are what actually need to change.