The Fallacy of Re-Training after AI

Machine Intelligence threatens jobs. Technologists propose that displaced workers be re-trained. This will not work. Here is why:


If a company saves money by replacing workers with algorithms or robots, their savings are marginal; it’s the difference in cost that counts. For example, a salaried employee packing boxes costs Amazon about $30,000 a year. If robots replace those jobs, it will be because robots are somewhat cheaper — say, $20,000 a year in amortized costs? With over 300k employees, most of those packing boxes, that would provide Amazon upwards of $3 Billion in savings. Wow. Can we expect Amazon to provide new salaries for those lost workers? No.

If Amazon pays for those new salaries, they go into debt, instead of saving. They would be spending $30k/person on new salaries, on top of the $20k/robot. Automation’s entire value proposition rests upon the premise that Amazon can get rid of those workers. So, it falls upon someone else to employ the workers that Amazon sheds.

If Amazon paid for those workers’ re-training, they would not be saving money, either. Re-training would constitute a large up-front cost, while the value of re-training would be received by workers who no longer add to Amazon’s profits. A re-trained worker wouldn’t pay Amazon back with salary from their new job.

The Job Crunch

A worker who is packing boxes is already low on the economic totem pole; any amount spent on re-training is unlikely to provide all of them with a higher-paying job, because that influx of new job applicants lowers the wages that they can command. This is supply and demand. An increase in job-seekers reduces wages.

So, who would pay for re-training Amazon’s fired box-packers? Even an industry-wide initiative suffers from the same calculation. Any group of industries which shoulders the cost of their displaced work-force would see no rise in profits after automation. The burden falls on tax-payers, should the government exercise the political will to help.

That only transfers the cost to a wider base — the same base who supposedly benefits from lower prices after automation. Prior eras of lost employment were able to avoid this conundrum, primarily because new jobs did not require expensive re-training. When farmers were replaced by tractors, they could get a job in a factory, without significant re-training costs. And, the farms passed lower costs on to a broad base of consumers. The new secretarial workers could afford more food, even though their wages were a downward substitution.

AI, in contrast, primarily passes savings on to investors, who already see a low tax rate on their gains. If the government pays for re-training, then the cost borne by most tax-payers is necessarily greater than the savings from lower priced goods. Wealthy stock-holders are the only ones who walk away with the profits from automation.

That is the fundamental flaw in the push for automation: businesses that automate are able to avoid the costs of re-training, and their stock-holders are the primary beneficiaries. Those stocks are taxed at a lower rate than income, shifting the cost of re-training to the people who save only a little from lower priced goods. The net impact for the broad base of consumers is negative. And, much of the replaced work-force sees lower salaries, because of increased competition for the new jobs, while they must shoulder the cost of their own re-training through taxes.

Re-Training Upward

AI and robots replace simpler tasks first. Displaced workers, necessarily, must be re-trained for more complicated jobs. Many will be unable to find openings in those jobs, because of a lack of talent, regardless of expenditures on re-training. (Only a tiny fraction of Amazon’s box-packers will become doctors.) Most will be pushed into other low-wage jobs, which are similarly at risk of automation. Re-training for one job only shuffles them into another occupation where they will be fired and need re-training.

Farm machines didn’t threaten factory jobs. Industrial robots didn’t threaten actuarial jobs. AI is different, because it threatens the bottom of the work-force with multiple waves of replacement and re-training, regardless of their new employment. Each re-training is necessarily for a more complicated task, requiring more expensive re-training, and a lower proportion of displaced workers will have the talent necessary to compete in a saturated job market.

Worse still, the jobs that most re-trained workers will find themselves in are necessarily lower productivity, as well. (If those jobs represented higher productivity, they would have paid more, and they would have been chosen first.) Downward substitution means downward pressure on worker productivity. Amazon would see higher productivity among the workers it retains, but that is offset by the loss in productivity of substitution work, for only a tiny net increase.

The Re-training Multiplier: An Example

Suppose that Amazon replaces 200k jobs with robots. Recent advances in the software that controls robotic grippers may do that very soon. And, suppose that re-training places 20% of those workers into better-paying jobs that are at a low risk of automation, while the remainder of workers are shunted into simple tasks that are also at risk of automation. When those simple tasks become automated, re-training again places 20% with better-paying work, while the remainder move into occupations that are similarly threatened by automation. In each automation-wave, the entire pool of simple-task workers must be re-trained, such that many of those workers will need multiple re-trainings. Using my 20% figure, Amazon’s 200k jobs lost would incur a cost of re-training 200k workers, then 160k workers, then 128k, … a total of 1000k instances of re-training, before those workers find un-automatable jobs. This is the re-training multiplier that threatens to over-burden consumers and tax-payers.

Aside from computer programmers, most job creation for the last few decades has been in low-wage services. Automation threatens to flood the service-sector, at exactly the time when those service jobs are saturated. Our 4% unemployment figure is a lie, predicated on counting part-time work as a job, and ignoring the ever-increasing portion of discouraged workers. Sending 200k more workers from box-packing into the service sector would further depress service workers’ wages, and send many of those workers into the discouraged category. Re-train them to be a truck driver? Or hire them as a waiter for a diner at a truck stop? Good luck!

As long as investors pay a lower tax rate than income-earners, automation and re-training will profit only them.