[0xDEAFBEAD]

The algorithm described in this article seems very bad. But I would argue that ML risk scores can, in principle, be better than human judgment.

Humans seem more subject to bias than algorithms are. Algorithms only look at data, but humans are additionally vulnerable to stereotypes and prejudices from society.

Furthermore, using an algorithm gives voters an opportunity to have a debate regarding how best to approach a problem like welfare fraud.

Human judgment relies on bureaucrats who are often biased and unaccountable. It's infeasible for voters to audit every decision made by a human bureaucrat. Replacing the bureaucrat with an algorithm and inviting voters to audit the algorithm seems a heck of a lot more feasible.

I give the city of Rotterdam a lot of credit for the level of transparency they demonstrated in this article. If they want to be successful with algorithmic risk scores, I think they should increase the level of transparency even further. Run an open contest to develop algorithms for spotting welfare fraud. Give citizens or representatives information about the performance characteristics of various algorithms, and let them vote for the algorithm they want.

In the same way politicians periodically come up for re-election, algorithms should periodically come up for re-election too. Inform voters how the current algorithm has been performing, and give them the option to switch to something different.

[tagyro]

> Humans seem more subject to bias than algorithms are.

One might think that, but algorithms are built by humans, so they (algorithms) automatically have the same biases as the humans that built them.

[0xDEAFBEAD]

That doesn't follow.

If I'm a chemist, and I write an algorithm to do something related to chemistry, that algorithm does not "automatically" know everything I know about chemistry.

Bias works the same way.

[adammarples]

I think there's a case to be made that data contains bias, ie. previous arrests, encoding the bias of the past arresting officer, but the algorithms don't contain bias. Unless you could explain how logistic regression is biased.