At this point ML to govern things like autoplay or address bar is just whitewashing for biased data. You feed biased training data into an ML algorithm and now

At this point ML to govern things like autoplay or address bar is just whitewashing for biased data. You feed biased training data into an ML algorithm and now it's unbiased! Select your training data set and other parameters such that you get a result you want and you're good to go. The history of machine learning for tech is a history of obvious biases creeping into training data - whether it's a medical algorithm "learning" that background markings are an indicator of disease instead of classifying the tissue it's meant to, or recidivism risk algorithms going off race & gender to the point that they produce actively bad data, or face detection algorithms thinking that asian people are squinting. I don't even think that youtube necessarily should get an autoplay prompt on first use, but it's pretty convenient that ML-based approaches like this are used instead of much simpler approaches. Lots of research is going into creating adversarial data given known ML algorithms, as well. If this address bar ML is running on the client (it'd have to, right?) then it's not hard to do a training run against it to come up with custom tailored URLs/sites to get the ML to classify your attack as good.