Hundreds of AI tools have been built to detect Covid. None of them helped. At all.

People (especially tech start-ups) are extremely keen to throw machine-learning at every problem in sight, but the technology – and the human structures around it – just aren’t there yet.

One study looked at 415 different AI tools fed on chest x-rays and CT scans. If they could pick up on the commonalities, they could learn to diagnose Covid much faster than a human.

Well, they did pick up on the commonalities. They noticed what fonts were used to label the scans. Since some hospitals have more serious caseloads than others, they predicted that scans with that hospital’s font were more likely to have Covid.

People with a serious illness (Covid or other) are more likely to be scanned while lying down rather than standing up – so the AI began to predict Covid based on the position a person was in.

One Covid-free dataset was all children’s lungs, so of course, the AI started identifying kids, not people with Covid. The full article has a lot more examples.

A clear issue is the lack of cross-disciplinary support:
“Many tools were developed either by AI researchers who lacked the medical expertise to spot flaws in the data, or by medical researchers who lacked the mathematical skills to compensate for those flaws.”

But another is that the technology simply isn’t ready, and developers need to cool their jets and stop foisting it on the populace until it is.