Should the robot that takes your job, pay your taxes?
AGI economy series - Nr. 6
“The human worker who does, say, $50’000 worth of work in a factory, that income is taxed and you get income tax, social security tax, all those things. If a robot comes in to do the same thing, you’d think that we’d tax the robot at a similar level.” - Bill Gates, 2017
A robot tax is the idea that businesses replacing human employees with robots should be mandated to pay a tax. The revenue generated from this robot tax may help to replace the revenue from income tax and social security contributions that a government loses, when a human job is automated.
Having robots pay our taxes has an intuitive appeal. After all, who likes paying taxes? However, implementing a robot tax is not as straightforward as it might seem. For starters, it is not easy to define which robots should pay taxes. Should your dishwasher pay a tax because it replaces you in hand-washing dishes? What if it’s a dishwasher in an office kitchen?
Today, a robot tax would hurt workers more than help them. A robot tax disincentives investments in labor productivity, which in the long run have led to higher wages for workers. Even if this positive impact of automation on wages may not last forever, a robot tax remains a complicated and suboptimal way to raise tax revenues.
1. Why implementing a robot tax is complicated
In practice, it’s difficult to implement a robot tax without creating a new bureaucratic apparatus. In that regard, a compute tax seems more reasonable than a robot tax.
a) It’s difficult to count automated jobs
The former New York mayor Bill de Blasio suggested that whenever a robot replaces a human worker it should pay the same amount of income tax and social contributions. That sounds simple, but the reality is complex. First, automation typically affects individual tasks rather than entire jobs, leaving many tasks still performed by humans. This shifts the nature of work toward more complex tasks. Second, automation can also have positive effects on employment. Directly, it creates new job opportunities such as machine repair and oversight. Indirectly, it lowers production costs, shifting the supply curve and leading to a new equilibrium where more products are consumed at lower prices. This can enable the automating firms or countries to capture a larger share of the global market.
Let’s take Amazon as an example. The online retailer is the operator of the world’s biggest fleet of mobile robots and heavily invests in automation in its warehouses. It’s adding about a thousand robots a day and some have predicted that it will have more robots than humans in its workforce by 2030. Yet, so far, the net human workforce at Amazon has significantly expanded over time. So, which jobs did the Amazon robots really ‘steal’?1
Today, CEOs sometimes boast how many employees they were able to cut based on automation. However, if there is a high tax on replacing a human by a robot or AI, then companies will be incentivized to claim that all their robots are augmenting human workers and any firings are because of other issues, such as the business environment, a restructuring, or individual performance issues. Hence, if we want to tax robots, we should probably not try to find the specific robot that ‘replaced a human job’ and tax it based on the salary level of its human predecessor. It would make more sense to tax all robots within a business sector at a specific rate.
b) It’s difficult to count “robots”
The number of robots operated by Amazon is self-reported. However, if operating a robot would come with a tax burden, companies would be incentivized to downplay the number of robots in use. We tend to imagine robots in humanoid form, and if all robots took on this shape, counting them would be feasible. However, in reality, the diversity in size, form, and autonomy of robots is much greater than human biodiversity. Does a single-arm robot count the same as a six-arm robot? Is a dishwasher a robot? Is a traffic light a robot? Is a ship motor a robot?
Even if we just focus on computers and AI as the “robot brains” the problem remains. The biggest AI model is about 70 million times larger than the smallest commercially useful AI model.2 Similarly, the largest digital computer in the economy is more than a billion times larger and has more than a billion times more computing power than the smallest computer in the economy.3
In short, defining a robot or an AI as an electronic person and then taxing them based on headcount is not impossible, but it would require establishing an entire new bureaucracy. We would need to maintain a list of all the specific robot and AI models that need to pay taxes with a potential for multiple tax categories.
In that sense, a compute tax would be the simpler alternative to a robot tax. Computing power is a quantified metric that can account for size differences in robots and AI. Specifically, the number of floating point operations per second (FLOP/s) that a chip can perform would be a fairly unambiguous metric, either as a threshold for a specific tax burden or as a tax that proportionally increases with FLOP/s.
c) Tax avoidance and offshoring
Another challenge that a robot tax might face is that it may incentivize businesses to move more of their production to other countries where they don’t face this burden and then to export their goods from there. Whether that is a credible threat depends on the size of the robot tax, tariffs, international coordination between states, and the ability of the taxing state to leverage market and technology access. Overall, there are ways to mitigate that risk.
Yet, even if all the implementation issues could be addressed, would a robot tax actually be desirable?
2. Today, a robot tax would hurt workers
Investments in automation increase a firm’s production capacity. Robots help us to produce goods cheaper, faster, and better. A robot tax would punish automation. It would slow down the speed of automation and thereby how fast the overall “economic pie” grows.
A robot tax may protect jobs in the short-run, but in the long-run, labor displacement has been very positive for most workers. Workers that stay at a firm that automates have higher productivity and can eventually negotiate higher wages. Workers that are laid off are often able to find new jobs. Hence, as
argues, robots have historically been good to workers - American workers need lots and lots of robots. Automation has led to better jobs, and the countries with the highest robot densities in the world still have no major unemployment problem.3. What about a robot tax in a post-labor scenario?
It is plausible that increased automation might eventually no longer translate to higher wages for workers if we have widespread AGIs with high fluid intelligence. Would a robot tax make sense, if we cross such a threshold and approach a fully automated economy?
The answer to this question is less obvious, but it’s still a no. First, it will still be challenging to implement such a tax. ‘Robot biodiversity’ will further increase in a post-labor scenario. Second, instead of taxing a specific production process, it might be more desirable to tax outputs and outcomes, which is neutral to how they are produced. As long as tax avoidance can be managed, a value added tax and a corporate income tax might provide a more process-neutral way to finance social security.
Thanks to
and other Roots of Progress blog-building fellows for valuable feedback on a draft of this essay. All opinions and mistakes are mine.Annex - Robot tax proposals
One might make the argument that Amazon robots indirectly ‘steal’ jobs at mom-and-pop bookstores who cannot match the cost-cutting efficiency of Amazon. However, it’s impractical to say book-shelf robot ‘X’ in Amazon warehouse ‘Y’ has forced the mom-and-pop bookstore ‘Z’ in Exampleville out of business.
For artificial neural networks something like 25’000 parameters is on the lower end to be useful. On the upper end, we can find models such as GPT-4 with an estimated 1.8 trillion parameters.
The Michigan Micro Mote has a volume of about 16 mm3. The top supercomputer on the Top500 list takes up about 372 square meters and if we assume up to 3 meters rack height, we get close to 1’000 m3.
I have yet to see a concept for a “robot tax” that appears workable and doesnt violate the “Canons of Taxation.”: https://www.lianeon.org/p/just-tax-the-land
A computer tax might make a bit more sense, but we want more compute, not less. I generally don't recommend taxing things we want more of.
I'm not sure we will need new taxes anyway. If robots can sufficiently replace human labor such as to warrant taxation in the first place, we would see huge economic growth/material progress.
This would have the effect of shrinking the cost of goods and services…including for those services provided by the government.