It's the same as allowing full self driving cars which on average are safer than human drivers but sometimes accidentally drive into a fire truck because they couldn't train an image classifier to more accuracy than 99%.
The words you've used could hypothetically mean some future artificial general intelligence that does not currently exist and there is no guarantee will ever exist, especially within the lifetimes of those participating in this thread. That could obviously be quite good.
"AI" as currently defined by marketing and pop culture to mean machine learning, large language models, etc. should never be allowed to make a medically important decision. We've already seen beyond any reasonable doubt how risky it is to even treat them as a natural language search engine, the idea of handing over life-or-death decisions to them is literally insane.
There's a reason everyone is so loud in the hospital, it's because we have to be to be there in the first place.
Props to Airbus for proper UX and information prioritization.
The use of these alarms is not something imposed by the manufacturers, but by the standards, eg 60601, 62304 etc. For devices involved in diagnostic, or more importantly interventional care, you are required to have alarms within certain auditory and visual thresholds, and a lot of them have mandated silence times (in a life critical system, you can only silence a true alarm for 120 seconds at a time).
Then again, "ALARM" as dictated by the standards means something truly emergent, though the wording can feel a bit fuzzy at times. Trust me, alarm fatigue is a known phenomenon to these manufacturers, and theres been a recent trend (with, eg, the Dexcom G7) of giving users more control over delaying alarms, silencing them until you can respond etc etc, which has its benefits, especially as quality of life is concerned.
You'll have a hard time convincing the FDA of this for critical devices like those found in hospitals though.