This post is a response to Tyler Cowen’s A household expenditure approach to measuring AI progress, discussing how AI will affect productivity over the next 5 years (i.e. until the summer of 2030) via the effects on typical household expenses.
I mostly predict that the effects will be larger than Tyler expects, but the 5 year time period that he chose is short enough that the effects won’t be obvious until near then end of that period.
Housing
I mostly agree with Tyler that housing costs won’t be affected much by AI until after 2030. Some maintenance and construction costs will start becoming cheaper around 2030 as robots take over some of those jobs. I’ll say a 1% cost reduction per house by 2030.
Food
Tyler asks the wrong questions about food productivity. I don’t know what’s going to happen with better ways of growing crops, but I’m pretty sure that other aspects of food will experience bigger productivity changes this decade.
A fair amount of farm labor will be replaced by robots. I’m relatively unsure of the timing. I’ll guess that will reduce the cost of basic crops in 2030 by somewhere in the 2% to 10% range.
The big changes in food costs will come from cheaper food preparation and delivery.
Serve Robotics has plans to reduce meal delivery costs from $10 to $1 via a combination of delivery robots and drones. [Disclosure: I own a bit of stock in SERV.]
Cooking will be partly automated in 2030. Restaurant chains will adapt many parts of their system to be handled by robots. There’s a large volume of fairly specialized tasks that can be automated, so restaurants don’t need to wait for fully general-purpose robots in order to automate some parts of the process.
I predict some important, but hard to measure, productivity increases from people switching from home cooking to having prepared meals delivered. That will involve an increase in time available for other work, at approximately no cost.
I’ll guess that these effects will add up to a 10% decline in spending on food by 2030.
Health Care
Tyler is correct about large parts of the medical system: there just isn’t much pressure for cost savings, even when those savings are appropriate.
Medical research is likely to become more cost-effective in 2030, due to AI taking over many of the routine, tedious parts of clinical trials. The benefits will be slow enough to trickle down to patients that medical spending won’t decline until sometime later in the 2030s.
What about nursing homes? I don’t have much of a guess there.
There are a modest number of small medical issues where AI is almost good enough to save me time and money. An example is the IT band syndrome that I had in 2023. For a few months, all I understood was that I had pain in my left thigh.
I tried asking ChatGPT. It listed 10 categories of causes, one of which turned out to be correct. But I guessed wrong on which categories to ask further questions about. I gave up on ChatGPT’s help. My Kaiser doctor failed to ask the right questions. My functional medicine provider couldn’t ask better questions, but did suggest that the right person to ask would be a chiropractor who focused on sports medicine.
The chiropractor asked the right questions, and quickly determined that my pain was caused by something about how I had taken up rucking. After that, I had little need for further help.
If ChatGPT had been a little more informative, or better at asking diagnostic questions, I might have avoided the $190 chiropractor bill, not to mention a good deal of time and pain. Benefits such as this may only save about 2% of medical spending by 2030, but the financial benefits understate the value of this progress.
Similar effects will make mental health therapy more productive. AIs will replace many of the easier functions of therapists. That will mean that solving some problems changes from being expensive to being free. Whereas human therapists will focus more of their time on problems that are too hard for AIs. That implies an important improvement in the benefits of therapy, with probably little change in expenditures.
Education
It’s strange that Tyler focuses on the cost of learning. My impression is that households today don’t spend much money on learning. The costs of schools are mostly due to the costs of daycare and of credentialing.
Daycare costs will continue to rise where they’re paid for by government. But that has little to do with household expenditures.
I don’t know what to predict about the cost of a given credential. But I’m pretty sure that AI is reducing the demand for most credentials. Who is going to borrow money to get a college degree when there’s less than a 50% chance of that degree getting the student a job with a reliable salary?
I expect that if there were some way to measure demand for college credentials separate from the demand for other benefits of college, you’d see an 80% decline by 2030 in the demand for credentials.
But some of the spending on college goes to providing kids with a good social life and social status. Demand for those will likely increase.
I’ll guess that spending on college will be down 20% to 60% by 2030.
Transportation
Why does Tyler neglect transportation? Waymo is currently causing a minor increase in household expenditures, by providing a superior service. I predict that by 2030, robotaxis will provide one third of car rides. The cost for a given ride will be maybe 20% less than a ride costs today. Consumers will mostly respond by taking more rides rather than by spending less.
See also Austin Vernon on shipping costs.
Concluding Thoughts
The big productivity increases from AI will likely start having obvious effects on the average household around 2030.
The really big changes will require general purpose robots with human-level manipulation skills, which I’m guessing will replace most human manual labor by about 2032.
The really big changes to medical care will require something approaching a cure for aging, which seems like it’s 10+ years away.
And don’t forget longer term benefits, such as a decline in the cost of interplanetary travel.
I’m a bit concerned that framing my predictions as answers to Tyler’s questions is causing me to imply slower impacts from AI than I actually believe. E.g. I’ve avoided predicting military “productivity” changes – that could depend heavily on how likely war between advanced countries looks. And AI will impact the timing of mind uploading, but that’s unlikely to be available by 2030.
In sum, Tyler isn’t dramatically wrong about much that he says directly, but he misdirects us a fair amount. Changes in household expenditures will understate the impacts of AI.