Hacker News Reader

Comment by windsurfer | Hacker News Reader

This article completely misses what I see as the biggest problem with intelligence in machines: Corporations can use them to be unaccountable. Insider trading? No, that was our "algorithm". Racist policies? Oops, sorry that was just our bot. Inhumane practices? Sorry, we are still "working" on our artificial intelligence since you know it's very difficult.

Where most people see mysterious advanced help in day to day tasks, some corporations are seeing obedient, unaccountable and un-auditable workers. They are already starting to use them to perform their dirty deeds, and hiding behind the black box nature of these complex algorithms.

View on HN

> Corporations can use them to be unaccountable. Insider trading? No, that was our "algorithm". Racist policies? Oops, sorry that was just our bot.

I would say the opposite. Insider trading? No, here are the inputs to our algorithm. Racist policies? No, here are the inputs to our algorithm.

"Our" algorithm ? I was under the impression that AI were generating their own algorithms.

In a sense you are building an algorithm but with different means from the usual ones. It is built trough compilation of a large amount of data using the well known machine learning algorithms. Tbh and especially in this case these are vocabulary terminations we should not but stuck on.

It's an algorithm, just at a higher level of abstraction. Turtles all the way down.

"The input to our algorithm is the entire content of the Alexa top 1,000 websites over the last year". Good luck finding the secret signal in that one.

Similarly, there's plenty of evidence that algorithms can reconstruct discrimination from inputs other than skin color - names are the obvious example, but address seems a likely risk too. Algorithms need anti-racism quality control on their output.

And tell me, how can you verify that these were the inputs used? How can you check to see how those inputs were used within the algorithm? How can you audit such a system?

You rebuild the system using the same inputs and initial conditions and observe the outcome. Thank you, next.

Talk about underplaying the difficulties. Computer science needed a whole new package system (Nix) just to guarantee reproducible builds. Compilation is pretty much the definition of a pure input/output function, and most builds are simply not reproducible. If even builds aren't reproducible, good luck reproducing anything more complex like machine learning.

Auditing is a complicated process, if you don’t want to do the work you’ll just have to take their word for it.

The problem is the feasibility of auditing, not whether the motivation exists. And no, taking their word for it is not adequate.

Then what do you propose be done? If you accuse someone of having an extremely racist AI you need a way to prove they set out to make it that way. You don’t just get to say “It’s racist” and shut it down.

You look at the outcome distribution.

The danger here is of assuming the conclusion, in this case, that race doesn't correlate with anything meaningful. This is still a hotly disputed point.

Ultimately, transparency and reproducibility are the key criteria I think, like it is in science.

And then what? Shut it down if it doesn’t conform to your expectations or world view?

How would that solve anything? Most machine learning algorithms being used today are black boxes and the people implementing them often have little understanding of exactly what transformations the model ends up making between the inputs and the outputs. Its entirely possible for seemingly-innocuous inputs to pick up outright problematic features (i.e. skin color from a photo id) or indirectly problematic features (i.e. income level and address located in an area predominantly populated by minorities) that its outputs hinge on.

If anything, the issues with machine learning are even more problematic that what the OP posted, as simply identifying that something like systemic racism in a machine learning model can be very hard to identify.

Biased algorithms is a thing. We are often blind to that realization because we are often blind to our own biases.

This already happens. It's actually bureaucracy, but it's rarely called that.

Try dealing with YouTube after a fraudulent copyright strike, or getting money from Amazon or PayPal after your account has been suspended for an arbitrary reason.

AI has the potential to automate bureaucratic corporate hostility and indifference to weaponised levels.

See also the unfolding Internet of Shit saga. While everybody is afraid of artificial intelligence, I am currently more afraid of Artificial Stupidity...

Reply on news.ycombinator.com