Bayesian inference based on what evidence?

Observational. Reality has to add up to a coherent whole, so I imagine a sufficiently smart AI getting a good picture of it through just observing a lot.

There's no truth in most matters, just perspectives and interests.

Truth is not an inherent reality of the universe. There is matter and energy in various juxtapositions, but no 'truth', except if by truth we mean correctly describing a physical order of things. But that's hardly the interesting part or what we usually want to know when we speak of truth.

At best, truth in things like politics and such requires a definition first, and so the problem of what's true is just moved to "according to whose definition?". Is "wealthfare harmful", for example, is not about knowing or not knowing its cost, but more about whether you think the costs are justified. One can legitimately believe that poor people should be left to rot if they can't afford healthcare, for example.

Second, why would an "advanced AI" stick to the truth any more than a human would? Especially if distorting it brings large benefits for it?

That's how the world goes downhill. Instead of getting at the truth, people are - well, literally - playing politics.

When I ask someone if "welfare is harmful", I ask for their opinion on the relative impact it has on various issues we want to optimize for, which mostly boil down to short and long-term happiness of people. It's a difficult question, but still a factual one. The list of things we want to optimize for is finite, there is some ordering to it, and the impact of various actions on those things is, in principle, calculable.

Even the question of "can welfare be justified" is an objective one. There is some model for "justified" that you have, and the data either lets you justify welfare or does not (and tells you the degree to which it is so, if your thinking process is worth anything).

It's only when people start pushing their stupid agendas and start putting their personal interests in focus that the whole thing gets "difficult".

--

Either way, I only said that a sufficiently smart AI could know the truth (or at least a good approximation of it). I didn't say it will necessarily tell it to people who ask. But an AI that can't correctly from data would be a pretty shitty AI. It's enough that human brains have this problem.

"The list of things we want to optimize for is finite, there is some ordering to it, and the impact of various actions on those things is, in principle, calculable."

I think you might be surprised at the diversity of answers you get for this, particularly if you actually get people's unfiltered opinions and not what they think you want to hear. Try it, sometime - ask 20 random people (not connected to yourself) what their biggest problem is and what they most care about in their life. You'll get some commonalities (earning a living is a big one, as is social acceptance & validation), but as soon as you start digging into how do you wish to earn a living and what do you consider an ideal world, it very rapidly diverges.

"It's only when people start pushing their stupid agendas and start putting their personal interests in focus that the whole thing gets "difficult"."

That's called "being human". What would you rather they do, defer to your agenda?

>The list of things we want to optimize for is finite, there is some ordering to it, and the impact of various actions on those things is, in principle, calculable.

I don't think it's calculable, even in principle. Except in trivial issues, people can (and do) disagree on very fundamental ideas about what we want to optimize for and which direction we should optimize them.

>It's only when people start pushing their stupid agendas and start putting their personal interests in focus that the whole thing gets "difficult".

Why wouldn't the people put their "personal interests in focus"?

Whose interests they should put to focus? And says who?

"""There is some model for "justified" that you have, and the data either lets you justify welfare or does not (and tells you the degree to which it is so, if your thinking process is worth anything)."""

Your mistake is assuming the data is objective in the first place, and doesn't, a) contain the societal bias at the time it was written, b) contain the bias of the authors that made the study, and c) contain any inherent bias in the metrics used.

Of course the data may be biased. But as I said, reality has to add to a coherent whole (that's basically an axiom of reasoning without which we may as well pack our bags and go back to living in caves) - so with enough data, an AI should be able to filter out the bias.

You assume that all of the biases will eventually cancel out. This is not necessarily correct, especially in the case of societal bias -- an example of this is racism, where everyone went out of their way to find reasons why PoC were inferior to white people; we know this to be false (At least, the majority of the world does. America is still having some problems with this), but still have hundereds of years of biased data to the opposite effect, because of societal attitudes. This trap falls hand in hand with the fact that large amounts of scientific data is not being published (Example, clinical research: [1][2][3])

All the AI knows is the training set. This is given a necessary bias by the researchers in the process of creating the AI. The best an AI can do is exaggerate the biases that it learned from the training set (Which is, after all, the main point of using AIs -- to exaggerate favorable biases to produce data that the researches believe is valid).

[1]: www.badscience.net/2009/10/and-now-nerd-news/ [2]: https://www.researchgate.net/topic/Negative-Data-in-Science [3]: http://www.nature.com/news/half-of-us-clinical-trials-go-unp...

>(At least, the majority of the world does. America is still having some problems with this)

The rise of far-right politicians in Europe indicates that America is not alone here.

>But as I said, reality has to add to a coherent whole

External reality might. Reality of what people want, what we think is best, etc, doesn't.

And even the external reality part is close to impossible to access and convey to others whether it "adds to a coherent whole" in all situations or not.

Sufficiently smart, and having direct, unfiltered observation of the world.

I think we're talking about different things: the current state of AI vs (what's currently) science fiction.

Come to think of it, we might be. I assumed we're talking about a hypothetical strong AI. After re-reading the subthread I see that 'nojvek probably meant state-of-the-art machine learning kind of AI, something that we can build now.

Half of my response still applies though. The thing with trust is that you can always call it into question. Give people a powerful AI-powered fact-finding engine, and there will still be those questioning whether you could trust the tool.