How not to talk about an AI-powered future
ines.io
ines.io
I don't think the imminent dangers of AI are self-conscious machines rebelling and deciding to kill people, but a much more subtle and nuance problem.
Let's say that you go to a bank and ask for a loan. The bank has a procedure to decide if they give you the loan or not. If your request is rejected, they can (hopefully) explain why you didn't qualify.
Instead, the whole process gets offloaded to a learning AI that makes that decision for the bank. For the most part, we don't really understand why systems like neural networks make a decision. Now if the banks denies your loan, the best they can say why is 'huh, the computer said so'.
What if suspiciously the banks start rejecting applications from an specific minority? Has the AI determined that this particular minority is trustworthy? Does it even matter?
What when AI systems are introduced in legal procedures? What when an AI decides if someone will go to jail or not?
Maybe I'm out of the loop but I feel these are more pressing issues about the widespread use of AI that are not being talked about enough (definitely much less than the doomsday scenario).
https://www.propublica.org/article/bias-in-criminal-risk-sco...
https://www.propublica.org/article/how-we-analyzed-the-compa...
https://www.propublica.org/article/machine-bias-risk-assessm...
AI has no notion of the Lucas critique. It doesn't consider the perverse incentives it's optimized judgement scheme may create. The costs of looking good to proprietary algorithms are only going to grow.
...
> Maybe I'm out of the loop but I feel these are more pressing issues about the widespread use of AI that are not being talked about enough (definitely much less than the doomsday scenario).
This is essentially about Fairness, Accountability and Transparency with algorithms. There has been a lot of discussion and research in this space.
Videos are eventually evaluated by an AI as to whether or not they're advertiser-friendly, and get demonetized if the AI deems it unsuitable, with no explanation or feedback as to how to bring it into compliance. The AI's training likely changes over time, so there's no guarantee you'll be staying in its good graces.
Many folks' business & livelihood is now dependent on the whim of a statistical AI.
Google is not bad at generally pulling 3 decent results, but at least a few searches a day it returns garbage results.
People are plenty aware of this. It is just that they are not always talking about the most imminent dangers. Sort of like how people talk about global warming dangers down the road, even though they are aware of e.g. cancer today.
This is similar to the current situation in the UK for various financial and insurance applications. The application is farmed out to an organisation that ensures compliance with whatever regulations apply, but often simply gives a yes or no answer, without details. Or at least, the details are not passed back to the customer.
> What if suspiciously the banks start rejecting applications from an specific minority? Has the AI determined that this particular minority is trustworthy? Does it even matter?
This is a very real concern. Specific subsets of society tend to build and train AIs and may consider their outputs accurate or as expected, whereas other subsets of society may disagree but may be underrepresented among AI engineers. There is a danger that an AIs ability to distinguish subsets of society will be paired with value judgements. It doesn't have to be this way, we need to involve all members of society in discussions that are currently somewhat restricted.
These issues are not that specific to AI. You can have a look a what applications are turned down whether humans, algorithms or AIs are running things. And then figure if anything should be done about it.
Humans are biased as hell. Studies show judges give much harsher sentences just before lunch, when they are hungry. Attractive defendants get half the sentences as unattractive ones. Not to mention all the classic race and gender ones, and random factors we can't even measure.
It is inexcusable to have humans in charge of any kind of decision process that a superior unbiased algorithm can do.
The question is - what happens if an algorithm develops a bias? A bias against a specific race/sex/occupation, whatever - what happens to the principle of free an equal society then?
Super simple example - you don't need AI to determine that men claim more frequently and more on car insurance than women. So....over the years, the rates for men have risen, until finally it was completely outlawed to take into account the sex of the applicant for car insurance quotes(at least here in the EU).
We need the human touch to iron out problems like this. An all knowing-omnipotent algorithm if going to discriminate based on statistics and that's something I don't think society wants.
I believe that law should be repealed, because it would lead to fewer people dying on the roads. So there is a case where not everyone agrees the human touch is actually a good thing.
Citation needed.
Counter argument - vast majority of males in prison for murder in US are black. Therefore, if we stop anti-discrimination policies in US, that will lead to saving more lives. I mean, statistics don't lie!
Or, maybe, just maybe - the statistics omit a lot of data that would allow us to narrow down what makes men claim more and why there's so many black prisoners in the US?
And that brings us back to the original problem - AI can just come to a conclusion = charge men more for insurance = less accidents! Or....put more black people in prisons = less murders! And the results might be exactly what it expects, so of course it will be declared as a huge success - but it's everything but.
But I don't really see the point. Men really are riskier drivers. They should pay more on insurance. In the US they do, and no one seems to care. Men, as a group, pay exactly the rate that matches their risk. Anything else would mean women subsidizing men and paying more then their actual risk.
I like this example because it's the reverse of the typical stereotype that women are terrible drivers. The algorithm doesn't care about stereotypes. And it's a small effect, something like 5%. Men aren't being banned from driving or ostracized from society, they just pay a bit more to make up for their higher risk. Other factors like age or previous accidenrs matter far more.
And as the insurance companies get better data, that proxy will matter less and less. In the future they will just offer to record your driving behaviour on a test track and statistically analyze how skilled you actually are, or something like that.
And anyway this has gotten far away from the point. If getting insurance required a human to interview you and judge your character, it would be 100 times worse. You think humans can be made blind to gender and race? When the choice is between a human and an algorithm, it's indefensible to choose the human.
>>When the choice is between a human and an algorithm, it's indefensible to choose the human.
That's where I vehemently disagree.
If there even is an effect, just like your gender example, I doubt it would be large our align with our stereotypes. E.g. it could be the case that whites are 1% worse drivers. And yeah make them pay more, again as a group they aren't being treated unfairly and anything else would amount to subsidies to them from other groups. That's not fair either.
>being penalized for the actions of others,
This different way of viewing this situation is perhaps why we disagree. I don't see it as a "penalty". If you live in a wood house, you will pay more on fire insurance. You aren't being penalized, your house is just objectively riskier. People in stone houses shouldn't be forced to subsidise your extra risk.
If you banned discrimination and removed all female drivers from the planet, insurance rates would go up to match what they are for men now. Men now pay what they would if there were only men.
>That's where I vehemently disagree.
Yeah well that's not an argument. And you can't make me bite the racism bullet if you aren't even going to acknowledge how disgusting your alternative is. After all your system has far more racism because human judges are racist. And why should someone spend twice as long in prison or be unable to get a loan because they are less attractive?
Its algorithms or nothing. Humans are not even an option. And in many cases nothing isn't an option either.
Because I don't think we can quantify the justice system. Like other commenters in this thread have pointed out - the jury system is explicitly based on the idea of being judged "by your peers", with all the biases and ideas that it brings with it. Could you replace that with an AI? Maybe - but how will you know if you reached the "correct" judgement then? In some bizzare scenario you might arrive at a situation where AI reaches a judgement that literally no one is happy with - and at that point we're just ruled by a hivemind overlord, no? I'm being sarcastic, but there is a point where we serve the algorithm and not the other way around. I mean, don't get me wrong, I would gladly submit to Culture-style Minds(Iain Banks) because I think they were being fair as described in fiction, but I have no trust that whatever we develop will be that fair, we seem to be using scattershot systems that look at the lowest common denominator and make a decision based on something that is easy and obvious, or worse, we train them on existing systems. Which(and I am sorry for bringing up the racism example again) - AI trained on the current population of the US prison system would conclude that it has to target certain groups more because...well, clearly they commit more crime! I have no trust it would be anything other than a shallow, simple-stat based oracle that everyone would listen to.
That said, I suspect people would really quickly learn that ruthless efficiency and "unbiasedness" is probably not what they want. We have rules and laws that ban discrimination along certain lines, but would that be applicable to an opaque AI entity?
With respect to your definition of "superior"... One of the major tenants of U.S. law is that in criminal cases you are entitled to a jury; you're judged by your peers. It's certainly possible that one day an algorithm may be more accurate than a jury in determining whether an accused person is guilty or innocent. But the jury is more than a fact checker, they provide credibility, independence, egalitarianism, and self-determination to the entire criminal justice system. Without those, the entire system could arguably collapse.
So even assuming computers will eventually have humans beat on accuracy, it's hard to imagine algorithms providing the additional values that are necessary for a functioning complex system.
The same applies to our political and economic systems, which exist to provide rhetorical air cover for entrenched power and resource appropriation mechanisms that disfavour most of the population.
There is very little chance of independent AI spontaneously deciding to rule the world. But there is every chance of the standard entrenched political factions adding AI to the kit of tools they use for social and political manipulation.
A truly independent hyper-smart AI would be the biggest possible threat to them. Not only would it be less gullible and easy to manipulate than most humans, it might conceivably be better at manipulation than they are, with unpredictable goals.
I beg to differ, juries may be the single most important part of our justice system. In my experience witnessing juries first hand, they take their responsibility seriously. They aren't perfect, but they are very good, and more importantly, the bring credibility to the entire system by directly bringing citizens into it.
The purpose of the jury selection "game", formally called voir dire, is to remove bias. Both sides ask questions, and if a juror demonstrates bias, they are excused. The idea that there is a process to remove bias from a jury pool is unquestionably a good one. Sometimes jurors are excused for other reasons, sometimes racist reasons, but this is illegal (see recent supreme court case on this). Also, the adversarial two-sided nature of the US justice system ensures that both plaintiff and defendant get to play the game equally.
Also, your flippant attitude towards "appearances" is misplaced. In any justice system appearances are extremely important. If a justice system does not appear fair, even if it is fair, it will collapse. That's why judges must recuse themselves from cases even if there is a "an appearance of bias" (caveat: Supreme Court isn't subject to this rule, but that's another story).
If you really want to replace jurors with AI, an AI that decides guilty or innocent wouldn't be enough. It would need to convince people that it's fair and independent. Being judged by a jury of peers in your community provides some of those assurances. It'll be a much harder sell for a robot.
Case-in-point: just yesterday an innocent man was falsely labeled a mass murderer on 4chan and those threads were surfaced on the top of google search results and on facebook.
Even if that takes 10000 years, it's significantly more serious than being put in jail or failing to get a bank loan.
Maciej Cegłowski gave a very good talk about that specific problem, which incudes a few recommendations about how we can avoid some of the worst problems as machine learning and capitalist democracy merge.
"Build a Better Monster: Morality, Machine Learning and Mass Surveillance"
https://www.infoq.com/presentations/morality-ml-mass-surveil...
Those descriptions are too anthropomorphic. Why not "optimization algorithm finds and implements a solution for a problem of growing more tomatoes, which is incompatible with continued existence of humanity".
From then, the calculations only got more advanced, but computers have always hidden how they worked (especially when they reach a certain scale of customers).
One approach to dealing with this is to use algorithms which are more amenable to explanations, like decision trees or inductive logic programming: that way, we can see exactly what criteria are being used to reach a decision, and we can use existing regulations/audits/etc. to keep them in check.
Another approach is to have the AI system produce evidence rather than an opaque response. One area where this already exists is in automated theorem proving: given a (suitably formalised) mathematical statement, programs like E and Vampire can (try to) determine whether it's true or false. In theory, we could ask such programs whether e.g. a loan application satisfies our eligiblity criteria, but the resulting true/false answer would be about as opaque as using deep learning.
However, we don't have to be content with a simple boolean: programs like metis can output proofs of a statement (if it's true), whilst programs like MACE can output specific counterexamples (if it's false). It doesn't seem too far fetched to have an AI system generate justifications for and/or against an application.
Direct link to the source article for those pictures: https://www.washingtonpost.com/news/wonk/wp/2015/10/04/what-...
[1] https://commons.wikimedia.org/wiki/Category:France_in_XXI_Ce...
Now the trouble with those is mostly the space required to start and land these things. Think how skilled those firemen in the picture would have to be to operate even slow jetpacks or helicopters and not crash into one another as well as land reliably. Birds have an advantage of advanced evolved neutral circuitry to handle all this.
We should stop describing these modern marvels as proto-humans and instead talk about them as a new generation of flexible and powerful machines. We should be careful about how we deploy and use AI, but not because we are summoning some mythical demon that may turn against us. Rather, we should resist our predisposition to attribute human traits to our creations and accept these remarkable inventions for what they really are—potent tools that promise a more prosperous and comfortable future.
The whole idea behind AI is to add a human trait -- intelligence -- to machines. So, not only it's ok to attribute human traits to these particular creations, but this is the very reason they were created, and what we strive not just to attribute, but to actually give to them.
>and accept these remarkable inventions for what they really are—potent tools that promise a more prosperous and comfortable future.
Potent tools can also be used to a less prosperous and comfortable future. They've used many "potent tools" in WWI and WWII -- from airplanes to nuclear bombs.
Now imagine AI-cops, AI-surveillance, AI-armies -- including robotic-AI armies. Not sure what those machines will do if they get autonomous in some "singularity", but I'm pretty sure what the greedy, corrupt, etc bastards in power in various places around the world would use them for -- more restrictions, more control, more gain for their personal interests, more and easier wars for the side that has them, more terrorism for the sides that can hack one.
The proper label to be used is usually robustness, analysis or optimization. (We're far away from actual creativity by the way, and quite far from even robustness in most cases.)
Or autonomous automated decision making. The scariest of the bunch. If not employed right or not robust or not "reasonable" (possible to reason about), it can have really bad results.
(By the way, automated decision making is pretty old, as old as industrial revolution.)
I mean, it is intelligence what we try to create -- and it is artificial, since by artifice we denote anything man made.
If the objection is as to the degree that the applications I've mentioned is actually intelligence (and not something cruder/dumber), I think we could still call it AI, and not just keep the term for actual human-or-above-level intelligence.
Rather than making a perfect AI that can control everyone, I think it's far more likely we'll develop a buggy one that accidentally kills us all.