I know this is the way non-technical people refer to the spooky AI action at a distance happening in social media to maximise engagement, but in my opinion it just makes everything confusing when actually trying to understand the argument.
A different term to use could be "machine-made decisions", which I personally think describes the issue better. Still, (and unless my understanding is mistaken) the word algorithm is technically accurate to this kind of decision-making.
The executives and shareholders who profit the most from the decisions must be the ones held responsible.
It's a perfectly sensible use of the term as is.
I would expect any insurance provider to use computer programs (i.e algorithms) to decide whether someone is eligible or not. The problem is what kind of algorithm.
...and how its outputs are interpreted. Ultimately it's less about what kind of algorithm and more about how it's used, or in this case abused.
As another user mentioned, that 16.6 days should have been the mean or mode of a posterior distribution, from which you could obtain 75% and 95% credible intervals, or at least a conditional expectation accompanied by an (estimated) sampling error and traditional 75/85/95/whatever% confidence interval.
Crossing the 17-day line could have flagged this person's case for manual review. Even if the 95% credible interval was 15-18 days, anyone who understands how statistical models work would understand that manual review is a more sensible reaction than automatic cutoff, unless you simply don't care about the 5% chance of falsely denying people health coverage.
The fact that the model is statistical is not more problematic than a chain of if/else decisions. The problem is the fact that good statistical practice was disregarded in its usage. And as others have said, there was no intention of pursing good statistical practice here, and no reason to ascribe any benefit of doubt to any decision-makers involved in deploying such a model/algorithm. It's just a fig leaf for wanting to deny people health coverage.
The intended audience here isn't computer scientists... it's... the public.
It's pretty obvious what they mean by "algorithms", no? I feel like this is just being "akshyually" guy
Even in the simple case, say a linear regression with all the assumptions satisfied, predictions are interpretable as means, which is to say that (and again, assumptions holding, errors are normally distributed) half of observations fall below the mean and half above it. Using the mean in a normative capacity implies that half of patients will get kicked out before they would otherwise end their treatment.
There are ways around this, like quantile regression, or using prediction intervals, but again, we don't know the specifics of the method which leaves us in the dark as to the specific way the problem manifests.
Like what? Correctly?
A set of hidden calculations that can be tweaked mainly for the benefit of the owner.