Comment by vertex-four | Hacker News Reader

Comment by vertex-four | Hacker News Reader

Parent Full thread

> Sufficiently advanced AI would be aware of the bias of its data, much more than any human ever could.

I am unconvinced that we've proven that - that seems like a hopeful wish.

> Usually that involves not using emotionally-loaded "hand grenade words" that bypass peoples' rational minds, and using clickbait headlines written by malicious editors. AI could be made to avoid both.

It also involves being very, very careful about presentation of data. "Black people in the US are convicted of more crimes" is a perfectly truthful statement - but it turns out that there's an awful lot of nuance to the reasons as to why, which rarely ever have anything to do with one's genetics. And yet there's a good many people who will take that statement as suggesting that black people are inherently significantly more likely to commit crimes, which isn't even close to what was actually said.

Specific speech patterns and words also lead to bias. You can quite easily write two articles stating exactly the same set of facts with very little fluff, and still result in differently biased articles.

Or, in other words, facts without reasons behind them tend to further bias, and it's only once you have reasons that you can actually start to suggest policy.

You argued that AI was biased. Now you're arguing that AI is too truthful, and that AIs uttering the truth would feed into a small minority's preexisting bias. This just proves my point that AIs are already less biased than humans as is.

You also seem to argue for blank slate dogma (that every group difference is caused by unfair government policy or societal "bias") and that provable facts must be silenced unless accompanied by their detailed etiology (which cannot be scientifically or factually verified, and therefore really just serves social purposes). I would argue that we should teach meditation in schools, every day, and that society should be more empathic and less judgmental. It should also be less afraid of truth, and less politically biased.

I'm arguing that none of the problems you bring up would be a problem with sufficiently advanced AI. They would move us towards the truth, not away from it.

Journalism's current credibility crisis stems solely from its attempts to be "socially useful" rather than truthful. I'm not sure why we should make AI copy us (again ironic since you're against AI taking on human biases).

No, I'm arguing that the definition of "bias" is incredibly varied and that being "truthful" but failing to tell the whole story (either because it's unknown, or as an editorial choice as to which data to present, or whatever else) can in itself be a form of bias.

Most bias isn't in the form of creating outright lies - it's picking and choosing what data to present. And if your solution is to not pick and choose what data to present, the output of your AI would be a life's work to study.

AI would allow us to form much more accurate etiologies than the current crop of sociologists, and better recommendations for solutions. There are simply no downsides to AI-journalism.

The more you understand the present, the more accurately you can predict the future. Big data increases your knowledge of the present, allowing you to understand what-causes-what much more precisely.

The issue is that you can't distill all the world's knowledge on a subject into something that can be read in an afternoon. Or rather, you can - so long as you choose to Trust In The Machine and ignore the details of how it got from there to the finished article. (I'm still entirely unconvinced that an AI could ever avoid bias, given that we're arguing over precisely what bias is in the first place.) That might work for Wal-mart deciding how much Kraft cheese to buy, but it shouldn't be acceptable for anything influencing public policy.

As opposed to trusting biased, rushed, financially-motivated humans of average human intelligence to distill the world's knowledge? Or biased, rushed, career-driven sociologists, that live under publish-or-perish and don't use the scientific method? Why the hell is it acceptable to have political activists and ideologues influence public policy?

Again, that isn't the point. AI would be a step forward, and I'd gladly take journalists becoming unemployed if it causes a new AI-led era of Enlightenment.

You're missing the vision: AI, with its swathes of data, would be able to tell you precisely what percentage of blacks that have been put in prison over drug laws would never have committed violent crimes. People's perception that these laws are justified for "public safety" would recede. That's just an example.

You're arguing from a fundamentally obscurantist point of view, arguing that unless we know 100% of the facts, it's better to keep certain things quiet to prevent human bias, whereas you should be arguing that we'll never know 100% of the facts (obviously) and should just stop being so quick to judge. More facts are always better than less facts, period.

> As opposed to trusting biased, rushed, financially-motivated humans of average human intelligence to distill the world's knowledge? Or biased, rushed, career-driven sociologists, that live under publish-or-perish and don't use the scientific method? Why the hell is it acceptable to have political activists and ideologues influence public policy?

Because we can discuss what bias these people have, and there is a "marketplace of ideas" so to speak. We're not asserting that there is one potential truth based on what a machine says.

On the contrary, there is no marketplace of ideas. These people are so insular that they'll hang on to their dogma no matter what criticism is leveled against them. They're fundamentally wrong, not just misguided. Case in point, the IAT.

> We're not asserting that there is one potential truth based on what a machine says.

Strawman.

But that is what you're suggesting - that a machine is capable of deriving a singular unbiased truth of the world, and present it in a way that is proven not to further bias in any human reading it, and we should all accept that on the face of it.

> But that is what you're suggesting

No, it isn't. I've said repeatedly they'd be much more accurate than humans, and much less biased than humans.

You keep taking a black-and-white version of my arguments, arguing against that, and moving the goalposts. I'm not sure you're debating in good faith.

The alternative is that you wind up with multiple AIs with different biases, people pick the ones that match their own biases, and we're not much better off. One presents stories about how drug users are unlikely to be convicted of violent crime, the other presents stories about how drug use leads to unsafe workplaces and road accidents. Both factually correct - like almost all news today - and both playing into people's biases and not solving the question of "do we legalise weed?" at all.

You're arguing that AI will lead to much more nuanced discourse, and much more complex and detailed pros and cons rather than the ruffian binary polarization we have now, and I agree.

No, it'll be the exact same binary polarization. One side keeps arguing "weed doesn't lead to violent crime" and the other side keeps arguing "but look at all the ways it does harm people", because that's the news stories they've each chosen to see, exactly the same as any discussion I've had on the subject to date.

The issue with journalism isn't people outright lying. It's quite rarely a failure to understand the problem space either. It's framing the discussion in a certain way and picking and choosing which facts to prevent. Having multiple AIs which pick out different things to tell their readers doesn't solve that problem at all.

Reply on news.ycombinator.com