Table 3 Data extracts for sub-theme: Moving from theoretical usability

So I think AI is a bit of a buzzword that is associated with market growth and jobs and us doing things better and quicker. But doing things quicker doesn't mean better, necessarily. And, so I think that a sensible pragmatic approach would be to be skeptical and to allow long periods of piloting and trials until we have a full confidence that something really does work the way it is supposed to work, rather than jumping to conclusions because, is, that something is able to detect something [Rn14 (BE)]

There are lots of papers about what is the best model, let's say to predict a certain disease and then these models on, the papers are published, you know, in respectable machine learning journals. So, the peer review seems to be working in that sense, that nobody is, is faking a new model. (…) The question then becomes, how would you actually implement it in the clinic and not only that. How do you actually implement it in a way that is useful? So, I mean if the model basically looks at all the data and then says, "oh well, based on, on everything that I have seen, I predict that the patient, you know, will die in the next week". That probably is not very useful, you would like to know beforehand, you know, when the condition starts deteriorating. [Rn32 (TE)]

So obviously, the example, like everyone else to give, it's an AUC [area under the curve^a] of 0.98 is, like sounds amazing, but if the 0.02 is everyone in your data set who is Black and everyone else in the 0.98 is everyone who is white like that's a problem, but the AUC looks really good. [Rn39 (TE)]