Workplace AI slop is not just emoji-laden social posts. Zonder Labs argues the painful kind arrives in Slack threads, PR descriptions, and incident write-ups you must decode before work moves — and that models are ironically weak at clear prose compared with coding skill.
The November 9, 2026 essay builds a clarity taxonomy, compares AI and human samples on articles and coding discussions, and shows AI detectors diverging from what human annotators call unclear.
TL;DR
| Fault | Example pattern (from Zonder) |
|---|---|
| Hidden characters and actions | Passive nouns hide who walked up and who asked |
| Vague wording | Odd stacks like "user-session entry boundary" |
| Empty emphasis | Long sentences that bury the one needed fact |
Casual slop versus mandatory reading
Zonder splits casual slop — short video and social templates you can ignore — from work slop you cannot skip. That matches how teams describe agent verification tax: the cost is not generation; it is re-reading until you trust the task.
The essay cites Style: The Basics of Clarity and Grace (by Joseph M. Williams) as a pre-LLM standard: readers deserve sentences no harder than the idea requires. Zonder treats that as a measurable bar, not nostalgia.
Pen tracing letters as a metaphor for rewriting unclear agent prose
The three clarity faults in detail
Hidden characters and actions
Clear sentences align story actors with grammatical subjects and actions with verbs. Zonder's duck-at-a-lemonade-stand example shows AI turning "a duck walked up and asked" into "a request for grapes was made upon the duck's arrival."
In incident channels, that pattern hides who deployed, who rolled back, and who is waiting — the same information engineers need when coding agents ship opaque PRs.
Vague wording
Vague references and noun stacks appear in human writing too, but Zonder highlights an AI failure mode: implausible collocations — calling a lemonade stand a "beverage retailer" or describing a database check as sitting at the "user-session entry boundary."
Readers with domain knowledge stall on almost-right jargon that sounds technical but does not point to a file, port, or owner.
Empty statements and misplaced emphasis
Zonder contrasts a ten-second sentence with a two-second rewrite where the poll result leads and optional context trails. Agent summaries often do the opposite: foreground process ("nothing else depends on anything but the result") before the one number a reviewer needs.
Trust and verification loop illustrating human review of unclear AI-generated workplace prose
Measurements Zonder reports
The essay publishes charts comparing unclear rates across AI articles, human articles, Claude Code, and Stack Overflow, broken down by fault type. It also plots AI detector scores against unclear rates, arguing signature words like "delve" are weak proxies for hard reading.
We treat the numeric tables as author-reported unless Zonder releases the annotation set; the taxonomy itself is the durable takeaway for teams writing agent review rubrics.
What developers are saying
On Hacker News, commenter bemtevi argued the frustration is simpler: human writers want to say something that matters to them, while models emulate caring and sound hollow.
gdulli noted the uncanny valley applies across media, not only images. infotainment appreciated the attempt to quantify Claude Code's "strange" prose — aligned with our Claude Fable PCB write-up where readers prize specific net names over template optimism.
These are individual opinions, not Zonder's data.
How this differs from authorship debates
Last week we covered Kevin Boone's trust-list essay, which says unknown web pages should be judged on evidence, not labels. Zonder occupies the adjacent lane: even when you must read an agent message, clarity failures impose cost before truth-testing begins.
Platform policy posts — itch.io deleting AI slop, Nexus Mods disclosure rules — focus on provenance. Zonder focuses on comprehension.
Practical rubric for agent output review
Borrow Zonder's faults as a lightweight checklist before you merge agent-generated docs:
- Circle passive subjects. Replace "an update was performed" with who did what.
- Ban noun stacks without referents. If a phrase has no file, metric, or owner, send it back.
- Put the decision first. Status lines should lead with pass or fail, latency, or dollar impact.
- Do not trust detectors as clarity QA. Run human skim tests on real PRs.
Long chats amplify bad style because agents re-send history each turn — see Grok Bot usage mechanics. Fresh threads per task help both cost and clarity.
Worked examples: rewriting agent slop
Zonder's duck sentence is silly on purpose, but the rewrite pattern generalizes:
| Agent draft | Reader task | Clearer rewrite |
|---|---|---|
| "A deployment was executed against production." | Who deployed? | "SRE Alex deployed build 4821 to prod at 03:14 UTC." |
| "Validation resides at the authentication boundary." | Where do I click? | "Enable MFA on the admin console before merge." |
| "No blockers remain except awaiting poll completion." | Are we done? | "Poll finishes in ~2 minutes; then merge." |
Teaching reviewers this table beats banning models outright: you keep automation, but shift labor from decryption to risk judgment.
Claude Code thread in the essay
Zonder embeds a long Claude Code snippet about Triton versus CUDA on GPUs — the kind of mandatory reading a performance engineer cannot skip. The highlights mark hidden emphasis and noun stacks that slow parsing.
That example matters because coding agents are marketed on benchmark wins, not status prose quality. yet teams spend review time on PR comments and PLAN.md updates agents generate. If your harness rewards verbose completion reports, you may be optimizing the wrong objective — a theme in agent harness evolution audits.
Detectors versus editorial QA
Zonder plots AI detector scores against human unclear-rate labels and argues popular detectors track stylistic tropes, not comprehension cost. For security teams, the parallel is signature-based malware rules versus behavioral sandboxes: cheap tests catch clichés, not harm.
Scanner over text where detector scores diverge from human clarity labels
Practical policy: use detectors for spam triage on inbound wikis, not for approving customer-facing agent emails.
Organizational policies that reduce slop tax
- Template bans on phrases like "I'd be happy to" or "delve into" help morale but miss Zonder's faults — pair style bans with subject-verb checks.
- Require bullet outcomes for incident bots: impact, scope, owner, next step — four lines max before detail.
- Split generation and compression: let the agent draft, then run a second pass whose only job is shorten — some teams call this a "humanizer in reverse."
- Train reviewers on verification tax surveys so managers budget hours realistically.
When clarity work belongs in the harness
Zonder is an essay, not a shipped linter. Still, product teams could:
- Add post-processors that flag passive voice density in agent outputs before posting to Slack.
- Log time-to-first-action after each agent message in internal studies — if slop rises, time rises.
- Compare human-edited versus raw agent PR acceptance rates in retrospectives.
None of that requires waiting for a vendor feature; it is process debt you can instrument this sprint.
Cultural slop versus workplace slop
explainx.ai has covered platform cleanup — itch.io deleting terabytes, Google pausing OSS bounty intake over junk reports, Max Richter on AI music dilution. Zonder narrows the lens to mandatory comprehension inside teams.
Readers can ignore spammy Shorts; they cannot ignore the Slack ping blocking a release.
Bottom line
Zonder's contribution is vocabulary: hidden actors, vague stacks, misplaced emphasis. Give reviewers those three labels and unclear agent prose becomes discussable instead of vague frustration — the first step toward measurable quality gates.
Try it on your next agent PR
Pick the last agent-opened pull request your team merged. Highlight every sentence where the grammatical subject is not the actor doing the work. Count passive clauses per hundred words. If the ratio exceeds what Zonder shows for Claude Code samples, add a clarity review step before you trust the diff summary for release notes.
Style guides that only ban the word delve fight the last war. Encourage engineers to paste agent drafts into a plain-text window before approving customer-facing copy. Read them aloud once; Zonder's misplaced-emphasis fault shows up faster in speech than on screen. If you stumble, rewrite before shipping. formatting chrome often hides how exhausting a paragraph will feel on mobile Slack. Zonder's charts suggest detectors overweight those tells while missing sentences that are grammatically fine yet cognitively expensive. Update your internal writing standard to mention actors and actions explicitly, not emoji bans alone.
Security incident channels are especially vulnerable: unclear agent summaries delay containment when every minute counts. Pair Zonder's taxonomy with the same discipline you use for AgentBeam-style guardrails — clarity on who must act is as important as blocking dangerous tool calls.
Related reading
- Kevin Boone on trust lists versus slop panic
- AI verification tax surveys
- Beneath the Surface agent harness evolution
- Google OSS VRP freeze over AI slop reports
- Max Richter on AI music dilution
- Human slop exposing legacy vulnerabilities
Sources
- Zonder Labs — Why is AI Slop so Frustrating to Read?, November 9, 2026.
- Joseph M. Williams — author of Style: The Basics of Clarity and Grace (Wikipedia).
- Hacker News discussion, November 2026.
Charts and unclear-rate percentages follow Zonder's published essay as of November 10, 2026.