I'm guessing a lot of of what people are calling "AI" are probably just statisical algorithms not so different to
what you might find in tools used for filtering email spam.
(a context where there isn't a more direct and reliable way to do such categorisation)
and pretty much everyone has to look in a junk folder for an email that isn't spam from time to time!

Such tools are not perfect and "false positives" can and do happen from time to time,
They can of course still be useful
(and in the case of the spam filter the effect of false positives is just having to check the junk folder from time to time so not much of a problem if it gets just a few wrong from time to time and the user knows where to look)

but we shouldn't assume such methods can ever be perfect (even if the stats look better with more data)
or allow the consequences of a "false positive" to be anything too serious or harmful to anyone to be able to live with
and especially careful about using anything like that in any context where its data and decisions about people

Corporations doing that kind of thing with data about people without regard for the consequences of false positives let alone the obvious privacy issues related to the use of data about people is of course very worrying.


(note: this comment isn't about spam filtering -  I only mentioned that because its an example of something that is probably familiar to most people where statistical methods are often used for categorisation.
The real worry is about other contexts where statisical methods might be used to make decisions about people where the effect of a wrong decision could be very harmful and if they are using data about identifiable people, then there are also obvious privacy questions that might need to be asked)