As far as I can tell the algorithm was not flawed. At least not in the sense that is produced different outputs than intended. The algorithm was designed to favor older patients, and it did just that. The result was that older patients received more livers. As was the intention.
So it not an algorithm that is flawed, it is a policy that is flawed. A policy that flawlessly and fairly executed by algorithm exactly as it was designed.
I'm not saying the policy was bad or fair. I have no idea to be honest. If you have one liver and two patients, then it's always going to be hard choice. But I don't think it is helpful to say the algorithm was misbehaving when it was not.
In fact, as mentioned in article, the outcomes of the algorithm are regularly checked by humans. And when they found a genuine bug (misclassifying people with liver cancer) the algorithm was fixed. Isn't that more or less exactly what you want. Humans thinking about policy, then having a computer executing the policy, while humans regularly check its output to see if the algorithm aligns with the intention.
That's how I'd like to think that they'd go about it. I'd be interested in learning the details. Say for example that younger people tend to be able to go without for a longer period than older people. I don't know if that's true, but if so you could prolong the lives of older people now but then get to the younger folks later, but not necessarily vice versa.
So it not an algorithm that is flawed, it is a policy that is flawed. A policy that flawlessly and fairly executed by algorithm exactly as it was designed.
I'm not saying the policy was bad or fair. I have no idea to be honest. If you have one liver and two patients, then it's always going to be hard choice. But I don't think it is helpful to say the algorithm was misbehaving when it was not.
In fact, as mentioned in article, the outcomes of the algorithm are regularly checked by humans. And when they found a genuine bug (misclassifying people with liver cancer) the algorithm was fixed. Isn't that more or less exactly what you want. Humans thinking about policy, then having a computer executing the policy, while humans regularly check its output to see if the algorithm aligns with the intention.