My thoughts are mostly my implicit bias against machine automation, that giving over one's thinking to machines will only lead to you being controlled by the person who controls the machine.

Apart from what @AlataOrange has said, I think your "on-device model" would die of overload in its first 5 minutes of operation. Most comments are biased. Everyone has an agenda, whether they are conscious of it or not. If I want factual things, I'll read the factual things elsewhere on the internet. If I want some buttery popcorn, I'll microwave some and read the comments.

I don't really think this can be effectively handled by automated mechanisms tbh. If it could it might be nice, but this is tackling a social problem with a technical solution, and typically those don't really jive, in large part because the technical solutions make the mistake of supposing an original creative position of null bias or objectivity. Alternatively, they stem from a position of overestimation of one's knowledge and/or abilities related to subjects that they don't in fact know much about nor have much experience in.

For example, how would this kind of software handle a case in which the facts are not fully known yet (much less the more esoteric matter of what even makes a fact)? Similarly, how would you define a bias such that it could determine the biases, and ensure that the bias determination remained accurate over time or locality (biases are notorious for changing over time and varying by place)?

This kind of software would demand not only software/programming folks but also sociologists and philosophers for handling both surveying to assess & try to define biases as well as facts and parsing the ethical impacts of making those assessments, definitions, & determinations. You'd also probably want more specialists in fields I'm unaware of or forgetting at the moment to really thoroughly handle this kind of project. Anything less, I'd argue, would result in a blunder of a project and be rather irresponsible, to say the least.

If there's a topic that you really care about, it helps to have trusted sources who also care about it in the same way that you do. Even though that means they might share your biases, it also means that they will notice and report information that matters to your shared interests.

For example, if you have a religious or ethical injunction against eating particular foods, and you want to know whether some food product meets your requirements, you probably need information from other people who share your specific rules.

If you're kosher or vegan or locavore, you don't need some random dude's opinion about whether a particular food is kosher or vegan or locavore; some random dude might well be a troll trying to trick you into eating foreign cheeseburgers for the lulz. You need information from someone who cares about the same thing you do.

If you care about your town government's decisions — like whether there should be a new bike lane put in on one particular road — you're probably not going to get good information from a source that isn't local. And you might additionally care about whether that source is a frothing anti-cyclist maniac or a wacko "fuck cars" ideologue or an "everything the government does is thereby wrong" libertarian or something.

I think we need more education rather than relying on technology to distinguish truth from fiction.

The kind of people who believe in propoganda and don't notice dogwhistles are the same kind of people who wouldn't listen to an app or a website telling them that something they believe is false.

A nifty thing, but it's not really addressing the actual problem. The problem with disinformation isn't that it merely exists, rather that it's shoved at such a volume by groups/people who build pseudo-credibility.

Rather than targeting individual instances of disinformation, sources should be dealt with. I think it'd be far easier and more efficient if disseminators were identified and popped a warning.