All hail the technocrats

One thing I can be sure of is that every time I write about a technocratic solution to a problem (e.g. automatic debt brakes), I get a lot of comments on how I shouldn’t be so enamoured with technocrats and their solutions because they tend to end in disaster. On the other hand, if we leave decisions to elected politicians, they might not get it right either but at least we can replace them with other politicians. Well, if you don’t like technocratic solutions, you are not going to like today’s article.

Three researchers from MIT, University of Chicago, and Georgetown University examined a couple of algorithmic solutions in the field of education, medicine, criminal justice, and regulation with regard to how effective they were from an economic perspective. To do this, they looked at the Marginal Value of Public Funds (MVPF) which is a fancy term for the ratio of net benefit for society relative to net costs for government.  

Four examples were examined:

• Pretrial release. If a person gets indicted for a potential crime, a judge in the US has to assess whether the person should go into custody until the trial begins or whether they are allowed to remain free until the jury reaches its verdict. In the US, about 10 million such decisions are made by judges each year though in principle this decision could be left to an algorithm that decides which defendants are flight risks or risks to society. Using data from New York City where such an algorithm has been trialled, they find that about 10% of the defendants that pose no flight risk are still jailed by judges, losing the potential to earn income while waiting for trial and at the same time adding to the massive cost of keeping people in jail. Based on the algorithm, the city of New York could reduce jail detentions by about 1,000 each year, which led to cost savings of $34.5m compared to additional costs for the algorithm of $4m. Meanwhile, these defendants allowed to remain free while awaiting trial would on average earn some $3,200 per person while waiting for jail, for a total benefit to society of $3.2m. The total MVPF for the economy is thus $3.2m benefits to society divided by $4m – $32.5m = -$28.5m in additional costs to the government. In other words, it’s a free lunch for the government where both the government and society benefit from replacing human judgement with a technocratic solution.

• Emergency healthcare costs. In another experiment, an algorithm was used to triage patients showing up in an emergency room with chest pain and whether to treat them for a heart attack rather than categorise them as lower urgency patients with some other condition. Relying on the algorithm’s decision rather than the doctor led to no decline in average health outcomes while costs were reduced by 34.7%.

• Occupational health and safety. Currently, occupational safety inspections are done in the US based on the number of workplace injuries in a company or a specific location. A factory that has more injuries will end up higher on the list for inspections the following year. But this is backward-looking, while algorithms can determine which locations are likely to have more repeat injuries in the future. Using such predictive algorithms, the number of serious injuries in the workplace could be reduced by almost 16,000 per year, which would create additional income to affected workers of $844m. And the government would obviously get some of that money in the form of tax revenues. Using an average tax rate of 24%, the additional tax revenue for the government from such algorithmic solutions would be $209.3m, and as long as rolling out the algorithm costs less than that, it would be another free lunch for the government.

• University courses. Algorithms exist that can better predict which courses a student should take at university (within his or her chosen area of specialisation) to get the best grade and best qualification for the student while reducing the risk that the student drops out of college. Again, such algorithms would take the decision away from the human student and prescribe a full set of courses but would lead to lower costs for students (in the form of saved student tuition and lower student loans) and the university (in the form of saved administrative costs and reduced staff requirements). A technocratic solution provides a free lunch.

In the end, the authors of the paper reason that there are two factors that tend to make technocratic solutions driven by algorithms better than the solutions we use that are designed and implemented by humans.

First, there is the problem of scaling. Algorithms can be scaled from small pilot projects to large nationwide applications without a loss of quality. Meanwhile, other regulation is often tested on a small scale under the supervision of highly trained scientists and specialists. But once rolled out nationwide or on any other larger scale, your average civil servant and bureaucrat have to implement the programme. And I don’t know about the skill and quality of work of the average bureaucrat, but it is almost certainly worse than the skills of a highly trained specialist in a narrow field.

Second, there is the not immaterial issue of priorities. Algorithms are really good at evaluating one option vs. another and ranking different outcomes against each other. Some humans are good at this, too, but politicians are not. And that may have several reasons. You and I can think of many reasons that are, shall we say, speculative concerning the mental capacity of the average politician, while others are more objective. And objectively politicians are beholden to lobby groups and the constituents who voted for them. Hence, they don’t argue for the overall best solution but for the solution that is best for their voters. And the result of this bias is that politicians are inherently incentivised to come up with second best and wasteful solutions.

Maybe then, it is too optimistic to trust in democratically elected politicians to come up with better solutions than they have in the past. Maybe technocratic solutions are inherently better for society on average, but we don’t like them because we lose control. In a sense, having bad politicians, bad policies and terrible government is the price of keeping control.

A comparison of the effectiveness of algorithmic solutions (red dots) with political solutions in the US (black dots)

Source: Ludwig et al. (2024)