Why AI Can't Do Your Laundry

Harys Dalvi

January 2025

Here are some things AI can do:

Write a sonnet or a short story
Create realistic or abstract artwork
Code a passion project
Teach students new concepts

One can argue AI does all or most of these things poorly, which is a fair argument. But ultimately, AI can do all this at some level, and is only improving over time.

Now, here are some things AI can't do:

Lay bricks
Clean bathrooms
Solve world hunger
Distribute mosquito nets
Do the dishes
Do laundry

There's a pattern here. It's crazy, and a little upsetting, that AI can write poetry but can't do your laundry. In general, AI is able to do things humans want to do, but not things humans don't want to do or can't do. A lot has been said on this topic with respect to robotics, and why automating manual labor in many ways is harder than building a language model that reasons or solves difficult math problems. But I want to tackle a slightly more general problem in this post: Why does AI do the things we want to do, and not the things we don't want to do or can't do? How can we fix this?

But hidden in this line of reasoning are three assumptions: first, that manual labor is more difficult to automate than creative tasks; second, that manual labor is desirable to automate while creative tasks are not; and finally, that manual labor and white-collar knowledge work represent the two main job types for AI to do. All of these assumptions are up for debate, and are less true than they seem.

Can AI Do Your Laundry?

When people say “AI” these days, they usually mean “deep learning with lots of data”. If you use that definition, then no, AI can't do your laundry.

But if you ask whether technology can do your laundry, the answer is yes. Laundry machines and dishwashers have been incredible for humanity: they are able to do these chores thoroughly while using less energy than manual cleaning.[] Before the invention of appliances like these, people, especially women, would often stay at home and do these chores. In this case, automation actually massively increased workforce participation, allowing women more freedom to work instead of spending all their time on the household.[]

Now, the tasks of doing laundry and dishes mainly involve moving clothes or dishes to the appropriate places and pressing a few buttons. Deep learning (DL) could improve this if we had robots that could scrape food from plates and place them in the dishwasher, or fold and hang clothes from the dryer, but the current situation is already pretty awesome.

So maybe AI, or rather technology, can already do your laundry.

Besides laundry and dishwashing, just as the majority of human work throughout history has been physical labor, so has the majority of automation. It's true that increased literacy rates made scribes obsolete, and calculators largely replaced human computers. But we also see that the Industrial Revolution displaced textile workers, modern manufacturing displaced blacksmiths, and we even had alarm clocks succeed human “knocker-uppers”.

It's only in the most recent wave of AI that this trend seems to have been reversed. Deep learning techniques have a much easier time with human knowledge work than physical labor. They can play chess and other games at a superhuman level, and can even solve unsolved math problems. But they can't reliably handle unexpected events on the road[] or even bring you a cup of tea[].

This is known as Moravec's paradox: things that seem hard to us (like reasoning) are easier for computers, while sensorimotor and perception tasks we do effortlessly are very hard to replicate. You might think the difference between these tasks and the more “advanced” reasoning tasks is one of hardware, where building the right hardware to support the intelligence is a difficult problem. But the difficulty is actually (mostly) on the software side: humans can control robots to perform these tasks quite easily.[]

One of the main issues is that there isn't enough high-quality data for robots to train on.[] There are other issues too, but without solving this, a DL approach to robotics has no hope of matching what we've seen in other AI models. In contrast, large language models are able to train on the entire internet and then some, and can include both images and text in their training.

This is the crux of the problem, the fundamental reason why AI can't do physical labor tasks we want to automate and instead does creative work we want to do ourselves. The internet is full of everything from TikTok dances to classic works of philosophy, but I've never come across first-person video footage of doing the dishes complete with muscle movement data. The internet is primarily a place for people to share knowledge work and their passions. By training AI on the internet, we are largely training AI to mimic human creativity and ingenuity, while leaving out the drudgery of life.

There are two ways to solve this problem: either collect more data, or create it. To create data, we can use reinforcement learning (RL). We can train robots in simulation, or even in real life, to maximize some reward we set. But this is vulnerable to reward hacking: we've seen this in games like CoastRunners, a boat racing video game, where an AI got a high score by knocking over targets for points instead of completing the race.[] In a more dangerous example, a robot trained to make you tea and nothing else might knock down doors, spill hot water, and break cups, as long as it achieves its objective of making you tea.

There is also the option of collecting data outside the internet. We could use humans performing the task as a starting point, rather than training a pure RL robot. But it's expensive to get all that human input, and it doesn't work for tasks that are too dangerous or difficult for humans.

And if we take off our AI hype glasses for a second, we might question the use of deep learning at all for this purpose. Robotics startups often don't use DL,[] which makes sense because DL is a statistical technique with room for error, while robotics needs perfect reliability. Rather than pure RL/DL, we might see a growth in non-DL approaches or, more likely, a mix of both.

I'm looking forward to seeing some great advances in robotics in the coming years, and maybe even a “ChatGPT moment”. But these obstacles will likely continue to limit the progress we can make in robotics, while we'll see more advances in AI for creative work and knowledge work as a consequence of the kind of data we see on the internet.

Should AI Do Your Laundry?

The dominant narrative is that AI should automate manual labor rather than creative and stimulating tasks. For the most part, I agree. But this isn't as clear-cut as it seems.

The word Luddite, now used to describe anyone opposed to new technology, originates from a group of textile workers in 19th-century England who felt threatened by the new technologies of the Industrial Revolution. They had spent years refining their craft to produce high-quality garments, and suddenly people with minimal expertise could churn out textiles with new machines. Eventually, the Luddites became so upset that they started smashing textile machines across England, which the government made punishable by death.[] Clearly, they were not happy to have their job automated, even though it was more physical work than knowledge work.

It's true that unlike dishes or laundry, manual textile work requires a lot of specialized training and artistry on top of the physical component. But this is true of many jobs involving physical labor: plumbers, mechanics, electricians, and welders all require a lot of training to refine their skills, while trades like carpentry are arguably even art forms. Even if this weren't the case, the important fact here is that people in the skilled trades tend to be satisfied with their careers: over 90%, in fact.[] Automating away their livelihoods is no more desirable than automating that of a spreadsheet worker.

Knowledge work versus blue-collar work is a useful distinction to see what AI is most likely to automate, but it is not as useful to determine what AI should automate. For that, we need to start with a simpler question: Do people want to do this job?

If we look at the careers with the lowest levels of happiness,[] we see a mix of blue-collar and white-collar jobs. While we hear about fears of AI automating artists or programmers on social media, many actual candidates for automation with current AI are things we find on this list: data analyst, customer service representative, administrative assistant. Other blue-collar roles, like cashier and retail salesperson, are not being automated per se but are simply shrinking: we have more self-checkout and online shopping now.

When we say we want AI to do our laundry, we don't necessarily mean that we want it to automate blue-collar jobs instead of white-collar ones. Instead, we want it to automate the boring parts of life that don't make us money or give us fulfillment. This is good for the economy and human livelihoods all around: we saw that the invention of new household appliances increased workforce participation, so surely AI for similar purposes will increase both economic productivity and human flourishing. What if we could have AI for cooking, cleaning, and taking out the trash?

Besides Laundry, What Should AI Do?

What AI should do and what it will do aren't always the same. It seems to me that unfortunately, AI will automate away the jobs of some people who wanted to keep those jobs. It might not happen as explicitly as in the case of the Luddites, but the shift in economic incentives won't be 100% good for everyone. There will be harm, as with any new automation.

When faced with the possibility of AI harming us, we can and should ask how we can make AI serve us as regular people. An obvious candidate is chores like cooking, cleaning, and taking out the trash. Having robots that can do these non-career chores would free up people to work more and enjoy more, just like laundry machines did in the past.

But there's a lot more to AI than chores. So far, I've mostly been talking about using robotics to automate things that humans don't want to do, like chores and jobs they dislike. But equally important is AI automating things humans can't do, at least not feasibly. Unlike something like AI autocompleting a line of code or designing a logo, this would be a real step towards using AI to build a world of great prosperity far beyond what humans can achieve on their own.

The internet is full of things humans have done, so it seems like a bad place to train an AI to do things humans can't feasibly do themselves. But the key word in that sentence is things humans can't do feasibly. Customer service is a great example: humans can provide customer service over the phone, but it's not feasible to provide it at all hours, for all companies, for all customers who might have a complaint. With large language models, this can be automated at scale, ultimately using human-created data from the internet to accomplish something humans couldn't feasibly organize to do. Companies like Bland AI are already working on this.

Another example is upsampling: loosely speaking, taking an existing image, video, or audio and filling in the blanks in some way. In 2023, the surviving Beatles used AI to restore John Lennon's vocals from old recordings and release the last ever Beatles song, “Now and Then”.[] We can imagine the general principle of AI upsampling for things like remastering old songs, animating and colorizing historical photos, and cleaning all kinds of noisy data for business and engineering purposes.

Then there's science experiments. Science labs are expensive, and running experiments is still costly even if you have one. AI can provide at least reasonably accurate experimental results without needing to run a real experiment, allowing scientists to refine their theories more quickly while running experiments more selectively. Technologies like AlphaFold demonstrate this, and promise to speed up research that could save lives.

Finally, one of the main purposes of machine learning in general is prediction: given some past data, we use statistical methods to predict what comes next. New machine learning methods are being used for forecasting in finance, weather, and even early natural disaster warnings. No human can look at a giant spreadsheet and come up with these predictions, not even with Excel formulas, so this is another case of using AI to expand upon rather than replace human potential.

Unlike customer service or upsampling, we need specialized datasets for many of these applications like automating science experiments and forecasting: we can't just dump the internet onto a model. But so far, we've had encouraging successes in these fields.

Too often, the debate over AI automation focuses only on using AI to replace human activities, some we like and others we don't. It's easy to forget that AI can also help us by accomplishing things that humans just can't do, just like earlier computer programs from calculators to chess bots.

AI for Human Potential

AI doesn't have to be all about replacing jobs people love and leaving them with nothing. Ideally, we can use AI for two main reasons: first, automating jobs people hate, and second, doing things people can't do without AI.

Even when automating jobs people hate, like customer service, there is the risk of displacing livelihoods. But if companies make more money, they will want to use the money to sell more; and in order to sell more, they will hopefully hire people for new jobs. These new jobs should have better working conditions and more fulfillment than the ones that are being automated. We've often gotten this from automation in the past: with 3D printing, a CAD engineer has better working conditions and probably a more fulfilling job than a factory worker. If we're careful, we can replicate this kind of good automation with AI as well.

Automation is the most obvious possible result of AI: we have a new thing with some kind of intelligence, and it's natural to think about how it might augment or replace our own intelligence in roles we already perform as humans. But it might be better to instead think about the more hidden ways that AI can do things we can't do, like upsampling, forecasting, and predicting the results of science experiments.

We can and should build AI that increases human potential instead of replacing it. We should, but it's not yet certain that we will: if we (and policymakers) play our cards wrong, AI starting to automate human labor for corporations without creating new jobs to compensate is a real possibility. How it goes is still up to us.

References

How the appliance boom moved more women into the workforce (Jeremy Greenwood, “Evolving Households: The Imprint of Technology on Life”, 2019) ^
Handwashing vs Dishwasher (Reckit Benckiser) ^
The Tea test of robot intelligence (Alan Winfield, 2019) ^
The main reason why self-driving cars are not ready for prime time (Khristopher J. Brooks, CBS News, 2024) ^
Common misconceptions about the complexity in robotics vs AI (Dan Ogawa Lillrank, 2024) ^
Faulty reward functions in the wild (Clark & Amodei, OpenAI, 2017) ^
Why Robotics Startups Don't Use Deep Learning (Think Autonomous, 2022) ^
Who Were the Luddites? (Evan Andrews, History, 2023) ^
New Angi Report Finds Nearly 90% of Skilled Tradespeople Satisfied in Their Careers (Angi, Inc., 2024) ^
10 Careers Reporting the Lowest Levels of Happiness (Andrea Moran, The Washington Post, 2024) ^
Listen to ‘Now and Then’ by The Beatles, a ‘new’ song recorded using AI (Andrew Paul, Popular Science, 2023) ^