The article does address this, and points to the complexity with an example:
“One of the trickiest questions of all concerns Duchenne muscular dystrophy, a degenerative illness that starts in childhood. Screening for this would find it in six of every 100,000 children tested. But only a small number of these have forms of the disease that can currently be treated, and the drugs needed are not widely available. However, learning that one’s child has this disease could also allow children to join trials of new drugs.”
The question is, will they be able to find causal predictors with high enough certainty before essentially erasing a number of genes from future generations?
If a gene is associated with lower intelligence, will that be because it causes lower intelligence or because it's usually carried by people living in lead-polluted areas? If a gene causes an autoimmune disease, is this because the gene causes dysregulation in the immune system or is it actually sensitizing the immune system to protect the carrier from another disease that may arise with too lax immune responses, such as cancer?
These are hard questions with no answer for the majority of cases. What will end up happening is someone will throw a ML solution and claim saving lives, with no ability to repudiate the claims long-term since nobody will have children the predictor deems likely to be sick. We need to do better statistics and more sequencing, and we should leave the choice up to the parents, but this can go very wrong very fast. Nobody will blame a parent for doing an abortion when you know that the chances for something like down's syndrome are very high, but there's a very fine line to cross before we venture into designer-baby realm.
That's not a downside and Down syndrome screening is already being done. This is not what the article is about though, the article is about screening newborns.