---
title: Scrub the Tub, and You'll Want to Write Again
url: https://oosioo.com/en/p/%EC%9A%95%EC%A1%B0%EB%A5%BC-%EB%8B%A6%EB%8B%A4-%EB%B3%B4%EB%A9%B4-%EB%8B%A4%EC%8B%9C-%EC%93%B0%EA%B3%A0-%EC%8B%B6%EC%96%B4%EC%A7%84%EB%8B%A4
date: 2026-10-08T12:12:32+00:00
author: SYSOP
summary: The line in my notebook read: 'Give the bathtub, or the bathroom, a hard scrub. You'll be surprised how fast you want to get back to creative work.' From Agatha Christie plotting her novels while washing up, and experiments where ideas multiplied after a boring task, to six rulers that separate scrubbing work from making work, the economists' task model, why robots still can't scrub a tub, and what happened when people were handed AI. At the end I built a board for sorting work yourself. The ninth essay in A Page for Today.
---
# Scrub the Tub, and You'll Want to Write Again

![Scrub the Tub, and You'll Want to Write Again. On the left, a bathtub half-cleaned with a sponge; on the right, today's line copied by hand onto a ruled notebook page](/uploads/ee692934d52bb9d7.webp =720)

When I opened my idea notebook today, this is the line I found.

> Give the bathtub, or the bathroom, a hard scrub. You'll be surprised how fast you want to get back to creative work.

I don't remember where I copied it from. I went looking for the source and couldn't find one. Still, what it says lands right away. When a draft stalls before a deadline, I suddenly want to tidy my desk, and once I start tidying, my mind drifts back to the draft. My hands are holding a rag, but my head keeps wandering toward the half-finished paragraph.

Two questions hide in that line. One is why scrubbing makes you want to make things. The other is what exactly separates scrubbing from making, that the two pull apart so sharply. Follow the second question far enough and you run into the one everybody is asking these days: which kind of work does AI do?

## In Front of the Tub, the Mind Doesn't Rest

In 1966 a *New York Times* reporter asked Agatha Christie how she plotted her books. "Walking or just washing up," she said, "a tedious process." She also said that she used to work out plots sitting in the bathtub eating apples, with the cores lined up along the rim. For a woman who wrote more than sixty detective novels, one of her studios was the kitchen sink.

It isn't just Christie's habit. In *The Art of Thought* (1926), the British political scientist and social psychologist Graham Wallas divided creative thinking into four stages: preparation, where you grapple with the problem and gather material; incubation, where you let go and do something else; illumination, where the answer pops up; and verification, where you check whether it holds. Time spent scrubbing a tub belongs to incubation. You seem to have set the problem down, but some corner of your mind keeps turning it over.

People have tried to measure incubation in the lab. In 2012, a team led by Benjamin Baird and Jonathan Schooler asked participants to list as many uses as they could for common objects. Then, for 12 minutes, each person did one of four things: a demanding task, an easy task that took almost no thought, simple rest, or no break at all. Only the easy-task group scored clearly higher when they returned to the problems. News coverage put the gain at a little over 40 percent. The more someone's mind wandered during the easy task, the bigger the gain, and there was no effect on new problems they hadn't seen before. Only the problems they were already holding had grown.

One caveat belongs here. In 2024 another team repeated the experiment with 443 people and failed to get the effect again. The idea that ideas grow in front of the tub is plausible, but it is not a law the lab reproduces every time.

Other experiments come at it from a different angle. In 2014, Sandi Mann and Rebekah Cadman at the University of Central Lancashire had one group copy numbers out of a phone book for a good while and then think up uses for plastic cups. The copiers came up with more ideas than a group that had done nothing beforehand. In a second study, a group that only read the numbers, without even copying them, came up with the most. The researchers took it to mean that the more boring the task, the more room for daydreaming. That same year, Marily Oppezzo and Daniel Schwartz at Stanford tested walking. Eighty-one percent of participants scored higher on creativity while walking than while sitting.

So two forces act together when you scrub a tub. Your hands are busy and your mind is idle, so the problem you shelved rolls along on its own. And the work is boring, so your body wants to get to something else fast. The "surprise" in the notebook line is probably the moment the two overlap.

## Six Rulers for Telling Scrubbing from Making

What, then, separates scrubbing a tub from plotting a novel? Calling it "simple work versus complex work" lumps too much together. Scrubbing a tub isn't all that simple once you're at it. The moldy silicone seams need their own attack, and around the drain you have to switch brushes. So I decided to measure work with a few rulers.

| Ruler | Scrubbing the tub | Plotting a novel |
|---|---|---|
| Is the end state set? | Done when it gleams white | First you have to decide what "done" is |
| Is progress visible? | It shows with every stroke | You can write all day and stay in place |
| Can it be split into steps? | Spray, soak, scrub, rinse | The moment you write the steps down, the story hardens |
| Do you know quickly when it's wrong? | A missed spot shows at once | You find out months later you took a wrong turn |
| Can your mind wander? | Yes, and that's why other thoughts come | No |
| Does your body have to be there? | It does | A laptop is enough |

The first five rulers point the same way. They mark the difference between work where the *what* is already fixed and only the *how* remains, and work where you have to settle the *what* first. I'll call the first kind scrubbing work and the second kind making work. In scrubbing work you can see the end, see the progress, and see the mistakes. That's why it's boring but calming. In making work you can see none of the three. That's why it's daunting but compelling.

The last ruler is different in kind. It asks whether a body has to be there. For people it barely matters, but for machines it makes all the difference.

## The Line Economists Drew First

The people who first sliced up work this way were economists. In 2003, David Autor, Frank Levy, and Richard Murnane split tasks along two axes to explain how computers were changing jobs. One axis asks whether the work can be written as rules (routine) or not (nonroutine); the other asks whether it uses the mind (cognitive) or the body (manual).

![The task model of Autor, Levy, and Murnane. Computers substitute for routine cognitive work (bookkeeping, calculation) and complement nonroutine cognitive work (planning, research, persuasion). Machines substitute for routine manual work (assembly lines), while nonroutine manual work (cleaning, truck driving, caregiving) is hard either to substitute or to complement. Scrubbing a tub falls in the nonroutine manual cell](/uploads/4fa3cb6cac148d89.webp =720)

Their conclusion: computers take over work that can be written as rules and assist with mental work that can't. But with physical work that's hard to write as rules, computers can neither do it nor help much. The paper's examples for that cell were janitorial work and truck driving.

Here a funny mismatch appears. By the six rulers above, scrubbing a tub is scrubbing work by anyone's reckoning. The end is fixed and the steps are obvious. Yet in the economists' table it lands in the *nonroutine* cell, because steps that are obvious to a person are not obvious to a machine. Every tub has a different shape; the grime sits in different places with different hardness; you have to press on a slippery curved surface with just the right force. People find it easy only because they do it without thinking.

## A One-Year-Old's Hands Are Harder Than Chess

In *Mind Children* (1988), the roboticist Hans Moravec wrote that it is comparatively easy to make computers perform at adult level on intelligence tests or at checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility. This is called Moravec's paradox. What people find hard is easy for machines, and what people find easy is hard for machines.

Nearly forty years on, the paradox hasn't changed much. In 2023, researchers at Oxford and in Japan asked 65 AI experts in the UK and Japan how much of 17 household tasks would be automated within the next ten years. The experts' average came to 39 percent of household work time. The highest was grocery shopping at 59 percent; the lowest was physically caring for a child, at 21 percent. Grocery shopping scores high because the task has already moved onto the screen. You order through an app and someone else delivers. It isn't a robot doing the shopping.

Robot vacuums may push across the living-room floor, but no robot in anyone's home yet scrubs the curved inside of a tub and its seams. For people, scrubbing a tub is about the most mindless job there is; for machines, it's a hard assignment.

## So What Is AI Good At?

Cross over to work that needs no body, work that lives on a screen, and the story changes. In 2023, researchers from OpenAI, the University of Pennsylvania, and others went through US occupations task by task and estimated that about 80 percent of workers have at least 10 percent of their tasks within reach of language models. For 19 percent, more than half their tasks were within reach. The same year, the Bank of Korea estimated that 12 percent of jobs in Korea, about 3.41 million, are highly likely to be replaced by AI. What stands out is that the most exposed jobs are high-education, high-income ones. Press examples included doctors, accountants, and asset managers. If industrial robots changed repetitive physical labor in factories, language models are starting with repetitive mental labor in offices.

Experiments that actually handed people AI and put them to work have piled up too.

![What changed when people were handed AI. In writing tasks, time fell 40 percent and quality rose 18 percent. In customer support, issues resolved per hour rose 15 percent, and 34 percent for novices. Consultants working on tasks inside the frontier of what AI does well saw quality rise more than 40 percent and speed rise 25.1 percent, but on tasks outside the frontier their correctness fell 19 percentage points. Experienced developers took 19 percent longer, while believing they had become 20 percent faster](/uploads/da763a7bff5848d8.webp =720)

- **Press releases and work emails.** In 2023, Shakked Noy and Whitney Zhang at MIT gave 453 professionals writing tasks such as press releases, short reports, and analysis plans. The group using ChatGPT took 40 percent less time, and the quality graders assigned rose 18 percent. People who had been weaker writers improved the most, so the gap between people narrowed.
- **Customer support.** Erik Brynjolfsson's team followed more than 5,000 support agents at a software company. Adding an AI assistant raised issues resolved per hour by 15 percent on average. In the 2023 draft, novice and lower-skilled agents gained 34 percent, while experienced agents barely changed. The veterans' know-how passed through the AI to the newcomers.
- **Consulting.** Fabrizio Dell'Acqua's team at Harvard Business School gave GPT-4 to 758 Boston Consulting Group consultants. On tasks AI handles well, they finished 12.2 percent more work, 25.1 percent faster, with quality more than 40 percent higher. But when the researchers slipped in a task where AI tends to be plausibly wrong, the AI group's correctness fell by 19 percentage points. Because the boundary between what AI does well and badly is not a smooth line but a ragged, sawtooth one, the researchers called it the "jagged technological frontier."
- **Experienced developers.** In 2025 the nonprofit research group METR randomized 246 real tasks that 16 open-source developers handled in their own repositories. With AI tools, the work took 19 percent longer. Beforehand the developers expected to be 24 percent faster, and afterward they still believed they had been 20 percent faster. METR has since added that the tools changed quickly and these results may now be out of date.

Hold the four experiments up to the six rulers and a pattern shows. What AI did well was scrubbing work on a screen: press releases where what to write was set, support where the answer sat somewhere in a manual, analysis tasks with right answers. When the end state is given in words and the result can be checked, AI is fast and fairly accurate. On tasks beyond the frontier, or where experts held the context of their own codebase, it got in the way instead. The person, not the AI, knew what was right.

An analysis Anthropic published in February 2025 of about a million conversations showed something similar. Cases where people worked with the AI, asking, revising, and refining together (57.4 percent), outnumbered cases where they handed the job over wholesale (42.6 percent).

## When AI Helps With Making

What about making work? In 2024, Anil Doshi of University College London and Oliver Hauser of the University of Exeter had about 300 participants write eight-sentence short stories. Some received one story idea from AI, or up to five. Stories by people who got ideas were judged more novel (by up to 8.1 percent) and more useful. The effect was biggest for the least creative writers, who caught up with the most creative ones.

But put the stories side by side and a different picture emerged. Stories from the group that received a single AI idea were 10.7 percent more similar to one another than stories written by people alone. Each person got better, but everyone was drawing water from the same well. The researchers called it a social dilemma: good for each individual, but when everyone does it, the variety of stories shrinks.

AI can help with making. But unless a person decides where the making should go, the result gets pulled toward the average.

## Sort It Yourself

Splitting work up in words only goes so far, so I built a board. Pick a task with the buttons at the top and the six rulers get scored, and the task's position is plotted on the board at left. You can also move the sliders yourself and measure your own work.

```map /embed/chores-2026-en.html 16:10
A board for sorting work. Twelve tasks, from scrubbing the bathtub, washing dishes, sorting receipts, first-pass translation, summarizing a meeting, fixing a bug with tests and replying to email to planning a new product, plotting a novel, inventing a new dish, assembling furniture and looking after a child, are plotted on a board whose horizontal axis runs from making work to scrubbing work and whose vertical axis runs from on a screen to needs a body there. Pick a task or move the six sliders and the point moves, along with which cell it falls in, the share an AI can carry, and the share that needs someone to set the direction.
```

The rules are simple. The horizontal position is the average of the first five rulers, and the vertical position is the sixth. "The share an AI can carry" is a toy formula I made up: it rises with the end-state, checking, steps, and progress scores, and drops sharply when a body has to be there. It isn't a number from any paper. I scored the tasks by feel, too, so if you see it differently, move the sliders.

Try it and the board splits into four cells. Bottom right, scrubbing work on a screen is good to hand to AI. Top right, physical scrubbing work is easy for people and still hard for robots. The bathtub sits here. Top left, looking after a child and inventing a new dish will stay people's work for a while. Bottom left, making work on a screen can be done with AI, but a person has to set the direction. In this cell, however large AI's share grows, "the share that needs someone to set the direction" doesn't shrink.

## A Machine That Does What It's Told

I made this essay's charts and board together with AI, so the experiments above don't sound like someone else's story. Drawing the bar chart, fixing the board's labels so they don't overlap, checking the papers' numbers: AI was quick at all of it. But that came after deciding what to draw, which numbers to trust, and in what order to tell things. How to handle a notebook line whose source I couldn't find, and whether to keep an experiment that failed to replicate, were questions for the person putting their name on the essay.

AI works at the level it is told to. Sketch the end state clearly and it races to that end. Leave the end state blurry and it stops at a plausible average. So the more scrubbing work on screens AI takes over, the more what is left for people is making work, and above all its very first step: deciding what to make. The thought that suddenly surfaces in front of the tub is exactly that kind of work.

The notebook line said that scrubbing the tub would make you want to get back to creative work. Now there's an AI sitting beside the desk you go back to. What to ask of it still comes out of a person's head, maybe while scrubbing a tub. So every now and then, it's worth scrubbing the tub.

*A Page for Today* starts from a single line in a notebook and arrives, that way, at one page of today's thoughts.