Midjourney prompt: a steampunk robot at a desk translating ancient greek, photorealistic
Until yesterday morning, if you’d asked me to name a US job where total employment had decreased due to 21st-century automation I might well have said ”translator.” Automated translation among major languages has been “really good” for years now, as AI researcher and languge app Duolingo cofounder Luis von Ahn puts it; so good that since about 2016 professional translators have generally started working from machine-generated 1st drafts instead of blank screens.
I figured that with eight more years of progress in machine learning and the advent and spread of generative AI (the transformer architecture of which was originally developed for translation tasks), professional human translators would by now be getting scarcer, especially in high-wage America. But they’re not.
No matter what language you prefer, you should still subscribe to this newsletter (and then auto-translate it)
According to a Tuesday story in Planet Money, both the Bureau of Labor Statistics and the Census Bureau find that translator employment has grown in recent years. The BLS also finds that translators get paid almost 20% more than the median US worker.
The Persistence of Labor: Jobs, Tasks, and O-Rings
How can this be? It’s easy to see why high-end translations might still need the human touch. Stanley Lombardo’s translation of The Odyssey, for example, is great literature.1 But most professional translators today aren’t breathing new life into the classics. They’re translating from Mandarin to English the user’s manual for an inexpensive stepladder made in China and available on Amazon2, or something like that. Isn’t that work done at acceptable quality far more cheaply by technology than humans these days?
Yes, but that fact turns out to be less important than you might think for understanding the future of jobs & work & etc., for two reasons.
The first is that translating documents from one language to another is only one of the tasks a translator does as part of their job. Translators aren’t unique here, of course; all jobs are bundles of tasks. And when technology comes along and automates, it usually stops well short of automating all of a job’s tasks. Translation software doesn’t automate the phone call with the client to understand the translation assignment and set a price for it, or the back-and-forth to clarify the meaning of an ambiguous passage, or the alert that there appear to be a couple pages missing in the source document, and so on.
In recent years economists have moved toward a task-based approach for analyzing how tech progress affects employment. I’m proud that Erik Brynjolfsson and Daniel Rock, two of my coauthors and cofounders, are among the pioneers of the task-based approach (Workhelix, our startup, uses this approach to help a company understand and prioritize its opportunities to accelerate work via Generative AI).
A 1993 paper by (eventual Nobel prize-winning) economist Michael Kremer makes an important point about the tasks that make up many jobs: their effect on overall job performance is multiplicative, not additive. In other words, job performance is not the average of performance on all the tasks. The boss doesn’t say “OK, you did 19 of the 20 tasks at 99%, but totally screwed up the 20th. No problem! Your overall performance is still 94%. Here’s a sweet raise.”
In many if not most cases, the boss is much more likely to concentrate on that one screwed-up task, for good reason: “What do you mean you didn’t notice that two pages were missing!!?!? The client is furious with us and refusing to pay. I’m kinda feeling the same about you, to be honest” In such cases getting a zero on one of the tasks means getting a zero on the whole job, no matter how well the rest of the work went3. Kremer called this the “O-Ring” model of production because of of the Space Shuttle Challenger, the many components of which all performed well on a 1986 mission except for an O-ring that failed 73 seconds after launch. This failure caused a cascade of increasingly bad things to happen, culminating in the disintegration of the shuttle itself and the deaths of the seven astronauts on board.
One implication of O-Ring style jobs is that all of their tasks have to be done well. Another is that if performance on a bunch of the tasks suddenly gets better, the remaining tasks don’t become less important. Instead, they become more critical.
Numerical Example Time
To see why this is, consider an O-Ring job comprising 10 tasks. At one point in time, all ten are done by humans at the 0.8 level. The overall performance of the job is thus 0.8^10, or 0.1073, which sounds low.
Raising the performance of just one of the tasks doesn’t help much at this point. If one goes from 0.8 all the way up to 0.99, overall perforance rises, but only to 0.132. Hooray.
Now let’s say sophisticated automation appears that can bring performance on nine out of the ten tasks up to 0.99. Then overall performance is 0.99^9*.08, or .730. This is a BIG improvement, but note that overall job performance is still below performance on any of the component tasks. 90% of the job’s tasks are clicking along at 99%, yet the overall job is still still below 75%.
The obvious smart move at this point is to upskill the person doing that one remaining unautomated task, or find someone who can do it better. If you can find someone who does that task at the .9 level, overall job performance rises to 0.99^9*0.9, or .822. And if you can find someone whose performance matches that of the automation, then you’re looking at 0.99^10, or .904 performance.
In this job’s high automation scenario, then, any rational employer would be willing to pay a lot to find the person who can do that unautomated task at the .99 level — the translator, for example, who was observant enough to notice the missing pages. So even though translation has become an automated process, the human translator has become more valuable. This is not a purely hypothetical example. As Planet Money writes about Duolingo, “The company also uses human translators to ensure consistency in the company’s style and tone throughout their app. Turns out, AI can’t consistently master “the same playful voice” Duolingo wants to communicate to users. So, for that, von Ahn says, “we still employ humans.””
Two other important things are likely to happen in this scenario of automation doing many but not all of a translator’s tasks: the cost of tranlation goes down, and the human translator gains the ability to take on more work. The cost of translation goes down because it’s suddenly 90% automated.4 So instead of having to pay, say, $500 to get that user manual translated, the ladder manufacturer only has to pay maybe $75.5
The translator can take on more work because they don’t have to do the nine automated tasks themselves anymore. They let technology handle them, and concentrate on overseeing the work. Their job has changed — they’re looking over machine-generated translations instead of generating the first drafts themselves — but they’re still recognizably translators. They’re involved in more projects, each of which takes less of their time.
What happens to total demand for translation services in this example? Well, the cost of doing one user manual-worth of translation has dropped from $500 to $75. Which is a big drop — maybe enough to convince companies and people to get all kinds of other documents translated; websites, email marketing campaigns, warranties, registration cards, and who knows what all else. Maybe demand for translation is what economists call highly elastic, meaning that it goes up a lot as the price declines. Maybe not. Time will tell.
Not So Fast
The broad point here is one we keep having to re-learn: that when creative destruction happens, it’s always easier to see the destruction than the creation. When powerful automation technology appears, it’s easy to foresee that jobs will disappear as rapidly as the technology can spread, and harder to see how humans can add enough value to keep being employed in large numbers.
My lazy and so-far-incorrect snap judgment about human translator employment levels in an era of highly capable translation software joins a list of similarly bad predictions.
After a 1936 demonstration of the Rust brothers’ mechanical cotton picker, an editorial in the Jackson (Mississippi) Daily News advocated that the contraption “should be driven right out of the cotton fields and sunk into the Mississippi River” lest it contribute to the Great Depression’s joblessness. Yet cotton continued to be picked largely by hand in the Delta for decades afterward. It was primarily labor shortages (caused by migration to higher-paying factory jobs up north) rather than the availability of snazzy technology that eventually led to the mechanization of cotton picking.
As ATMs took off in the 1990s it was widely predicted that we’d soon need fewer bank tellers. But bank teller employment grew in the following years and remained strong until the arrival of the smartphone and its bundle of online banking, cashless payments, peer-to-peer money transfers, &etc.
Living-legend-level AI pioneer Geoff Hinton said in 2016 “We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists.” Yet demand for radiologists continues to grow, and shortages are common.
In two of the three cases above, total employment in the profession did eventually decrease. I expect the same will happen with radiologists, but I have no idea how quickly. Both the creative and destructive sides of creative destruction are complicated phenomena — far more complicated than “powerful new technology means fast and big job losses.”
More Numbers, This Time on Graphs
The broadest pattern related to technology, automation, jobs, and creative destruction is also the clearest: tech progress is accompanied by constant, relentless, so-far-insatiable demand for human labor.
This is easy to see in America’s post-war data. Here’s the size of the US labor force over time:
Did we need all these people, or did they languish in unemployment as tech progress accumulated over the decades? We needed them. Here’s the total number of jobs:
The unemployment rate:
And the prime-age employment-to-population ratio:
No matter how hard I squint, I can’t see sustained large-scale technological unemployment, or sustained non-recessionary unemployment of any kind, in those charts. They instead show me the creative part of creative destruction loud and clear, no translation required.
Last point: high job quantity does not automatically mean uniformly high job quality, and there’s a lot to say about how tech progress has affected different types of workers. I touched on that topic in the report I wrote about the economic impact of Generative AI when I was a visiting fellow at Google, and I’ll spend more time on the issue here in future posts. In this one I just wanted to allay the frequently-expressed worry that that this time the robots (or whatever new automation) are going to take all the jobs.
Classics nerds argue endlessly about which translations of the Odyssey, Iliad, or Aenead are best, and why. I stan for Lombardo’s modern poetry, which gets better when you read it out loud to yourself (or others!). I can’t wait for Daniel Mendelsohn’s translation to come out next year. He’s a gorgeous writer and a bone-deep classicist.
No, that is not a 100% random example. It shows up at my house on Friday.
When multiplying numbers together, if even one of them is zero the whole result is zero.
And because competition among translation software providers keeps their prices low.
Why not $50? Because the translator is more valuable in this O-Ring process, and so demands to get paid more.