--- title: The slop tax canonical_url: https://abhirupghosh.com/blog/slop date: 2026-04-28T00:00:00.000Z last_updated: 2018-10-20T01:46:40.000Z reading_time: 7 min read --- # The slop tax AI made words cheaper. Reader attention got scarcer. The intersection is slop — communication that looks finished but makes the reader do the thinking. The average person's attention on a single screen has fallen from about **2.5 minutes in 2004** to about **47 seconds today**, in attention research summarized by AP. The room changed: tabs, feeds, pings, infinite scroll, workplace notifications. Focus can be rebuilt; the modern screen pulls it apart.[^1] Writing got cheap at the same time. An MIT study found ChatGPT cut time on certain professional writing tasks by 40% while evaluators rated the outputs 18% higher.[^2] A 2025 WebSci study found AI-generated workplace emails averaged 193.4 words versus 72.5 for human-written ones — 2.7× longer for the same job.[^3] Cheaper writing usually means more writing. The collision creates a new communication problem. Not bad writing. _Polished_, comprehensive, low-signal writing. Slop. > **Slop is unfinished thinking disguised as finished writing.** It is the email with all the context except the ask. The strategy doc with all the facts except the recommendation. The PR description that walks through the diff but never says what should change in production. Slop is not defined by who wrote it. A human can brain-dump. AI can overwrite. A 2025 paper measuring AI slop confirms the asymmetry: humans rate text as sloppy on relevance, density, factuality, repetition, coherence, and verbosity — independent of the author.[^4] The test is what the writing does to the reader. ### A — Polished slop _Subject: Quick alignment on launch timing._ > Hi Jordan, following up on the conversation last week, I wanted to circle back on the launch timing discussion and make sure we're fully aligned across the broader stakeholder group. There are several dependencies in flight — engineering readiness, the marketing window, partner commitments, and edge cases support has flagged. I've tried to summarize the considerations across each workstream, including tradeoffs and a few alternative paths. It would be great to get your perspective whenever you have a moment so we can move forward with confidence. ### B — Clear ask _Subject: Push launch to Nov 18 — decision needed Friday._ > I recommend moving launch from Nov 4 to Nov 18. Reason: support cannot staff the original date. Decision needed by Friday — reply "approved" or "let's discuss". The first draft _sounds_ finished. The second draft _is_ finished. > **Bad writing does not eliminate work. It relocates it.** The writer saves time by not deciding what matters. The reader pays that debt: extracting the point, inferring the priority, identifying the missing decision, choosing what to do next. That is the slop tax. BetterUp Labs and the Stanford Social Media Lab call the workplace version _workslop_. In survey research with 1,150 full-time U.S. desk workers, 40% reported receiving workslop in the previous month. Each incident took about two hours to resolve — an estimated $186 per employee per month.[^5] It is paid in follow-ups, rework, missed decisions, and the slow erosion of trust. Detail is not direction. It is useful only when it has been pointed at something. > _I just wanted to quickly follow up and circle back, given that, as you may recall, we had previously discussed several options._ **The board approved the budget — we can hire two engineers in Q2.** _Of course, there are considerations and dependencies we should keep in mind, and I'm happy to walk through them whenever works best._ **I need your sign-off on the JD by Wednesday.** _Please let me know if you have any thoughts, questions, or concerns._ Strip the italics. Twenty-one words remain. Same fact, same ask. The point was not missing. It was buried. The same principle applies when the reader is an AI. Models are more tolerant than humans, but tolerance is not understanding. Long context windows do not remove ambiguity; they make ambiguity easier to survive. If the writer cannot name the goal, the constraints, and the output, the model is inferring the task from noise. Weak prompts skip the middle work: analyze what matters, evaluate the tradeoff, then create the draft.[^6] A recent HCI preprint tested cognitive forcing functions in AI-assisted writing — small interventions that paused users before accepting AI plans. With 214 participants, the best-performing one on overreliance was an _assumptions check_: asking users to identify what the AI was assuming. It reduced overreliance without raising cognitive load.[^7] Buçinca, Malaya, and Gajos found similar gains four years earlier, sometimes at a mental-effort cost.[^8] The move both papers point to: ask the model questions _before_ asking it to write. > **AI is not the enemy of clarity. Unchecked AI is.** Used lazily, AI lets writers turn half-thoughts into full sentences without doing the thinking between them. Used well, it does the opposite — asks questions, finds gaps, compresses, challenges assumptions. Question-first AI workflows like [GStack](https://github.com/garrytan/gstack) force the writer to decide before drafting.[^9] The scarce skill is high-signal communication: saying the most important thing in the fewest words that preserve the meaning. Before sending anything, run the UDAC test: - **Understand** — What is the one thing the reader should walk away knowing? - **Decide** — What choice, if any, is in front of them? - **Act** — What should they do next, and by when? - **Cut** — What does not change the answers above? If those answers are not there, the writing is not finished. It may be polished. It is not clear. And if it is not clear, the reader has to finish the thinking for you. > **In a world where words are cheap, clarity is proof that someone did the thinking.** ## Appendix — Try it on your own writing The live version of this post () includes an interactive Clarity Coach: paste a draft and it runs the UDAC test, names what you're assuming, and offers a compressed rewrite. It is the article's argument made operational — AI as a forcing function for thought, not a generator. ## Citations [^1]: AP News. [Tips for improving your focus](https://apnews.com/article/2334290ba5d8206c18aeca0090be2f3f). Attention on a single screen has fallen from about 2.5 minutes in 2004 to 47 seconds today. [^2]: MIT News. [Study finds ChatGPT boosts worker productivity for some writing tasks](https://news.mit.edu/2023/study-finds-chatgpt-boosts-worker-productivity-writing-0714). 40% time saved, 18% quality lift on the studied tasks. [^3]: Li et al., WebSci 2025. [Emails by LLMs](https://koustuv.com/papers/WebSci25_EmailLLM.pdf). AI-generated emails averaged 193.4 words vs. 72.5 for human-written ones in the study dataset. [^4]: Shaib et al. [Measuring AI slop in text](https://arxiv.org/html/2509.19163v1). Slop is a perception, correlated with density, relevance, factuality, repetition, coherence, verbosity, and tone — independent of who generated the text. [^5]: BetterUp Labs & Stanford Social Media Lab. [Workslop: the hidden cost of AI-generated busywork](https://www.betterup.com/workslop). Survey of 1,150 U.S. desk workers; ~40% received workslop in a month; ~2 hrs to resolve; ~$186/employee/month. [^6]: University of Waterloo. [Bloom's revised taxonomy](https://uwaterloo.ca/centre-for-teaching-excellence/catalogs/tip-sheets/blooms-taxonomy). Remember → understand → apply → analyze → evaluate → create. [^7]: Ghosh et al. [Cognitive forcing functions for execution plans in AI-assisted writing](https://arxiv.org/pdf/2601.18033) (preprint). The Assumptions intervention reduced overreliance without raising cognitive load. [^8]: Buçinca, Malaya, & Gajos. [To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making](https://dl.acm.org/doi/10.1145/3449287) (CSCW 2021). Earlier evidence that forcing functions reduce overreliance, sometimes at a mental-effort cost. [^9]: Garry Tan. [GStack](https://github.com/garrytan/gstack). A working example of a question-first AI workflow.