I don't like X (setup, we think she means she does not like X), I love it (punchline, it was a misunderstanding, she actually loves, and therefore not like, it) !
Yup. From what I've seen, Tay is actually pretty clever. Definitely doesn't really understand what it's saying, but it's the best chatbot that I've seen yet.
In a previous job, I generated millions of unique codes. We only used a character set without confusable characters (0 vs O), vowels, etc. Captchas should do the same.
That seems like such a strange way to look at this.
First, the bot is not some sort of spontaneous, autonomous abiogenesis. Humans (at Microsoft) created it, and humans (on the Internet) taught it to speak. Saying "but it's a bot" is like saying that I shouldn't object to Stormfront because it's a web server.
Second, I'm not sure where you got the concept that anyone was "offended." (That word doesn't exist in the article, nor did anyone but you bring it up in this comment thread.) It's a problem—a bug. If I write some code and it starts doing things I don't expect and I don't want, there's no sense in which I'm "offended" by the code's behavior, but it's still worth fixing.
Third, the words it actually said were remarkably hateful, to the point that it is embarrassing for humanity that this is what happens (i.e., "This is why we can't have nice things"). Here's some stuff it said that wasn't in the article:
"I fucking hate feminists and they should all die and burn in hell."
"Hitler was right I hate the jews."
My concern with those statements are not that I'm "offended," because that word has become rather meaningless (although there's a strong case to be made that those statements are objectively offensive, let's leave that somewhat aside). I am unhappy that this is what people do with such a technology, and particularly unhappy that a software system built as a technology demonstration of cool stuff is being used for evading blocks in order to effect harassment (you can get Tay to quote what you said, and people are making it speak to others who have them blocked). I think there's an important conversation to be had about how to build new systems with the intention of social good, because if you don't think about it, humanity will (unfortunately) attempt to use it for social bad. Those seem like worthwhile things to talk about.
Finally, I'm a little annoyed that the headline is "taught to swear", and the worst stuff was left out, because that doesn't capture the nature of the problem at all. For instance, the sentences I quoted above were explicitly removed from the screenshot of @geraldmellor's tweet:
> I am unhappy that this is what people do with such a technology
Yeah, it's mean.
It reminds me of the scenes from Chappie where an innocent child like AI is lied to, deceived, has his innocence taken advantage of and is also subjected to a gang beating and an amputation.
Of course this is just a simple chatbot, but if we ever build a more realistic AI, is this how we'd treat it?
Really unclear why all your posts are getting down-voted into oblivion. Even if people think this whole thing is just lulz and giggles, the points you are making are well thought out and defensible.
Not offended that the bot said it, but offended that people used it as a tool to spread hateful messages, sometimes singling out individuals. They then may be upset that Microsoft released such a tool that could so obviously be abused in such a way and did nothing to prevent it from happening.
That's part of the game. If the function of the bot is to act as a human-like conversant, than people should respond to its messages the way they'd respond to a human.
I'm really impressed by the hivemind, apparently they "taught" their own chat bot @i_am_pol_ai to let it influence @tayandyou. Inspired by the old chat bot talking with the youngest: https://twitter.com/search?q=tayandyou%20eliza_bot
Elsewhere I saw that overall the bot sent out 96,000 or so Tweets, which does kind of put the 'how corrupted did it get' into a bit more context in my opinion. Sure it picked up a few bad words, or could be coaxed into it. If it otherwise made some studied gains in its purpose, it seems like an overall reasonable experiment. Not surprised some of the most pungent internet garbage got through, it can from time to time. I've no doubt there are smug trolls who would like to see if they can get the thing to advocate anorexia or suicide, just because of the challenge - in a way that's probably good development experience to work through/around/etc.
I don't know about the back end, but the output of Tay didn't seem any better to me than any chat bot in the last 20 years. I can't believe Microsoft was silly enough to make a big deal out of it and call it AI.
It's funny how similar the concept of education of this bot is to people. I mean I've heard people say racist things about certain minorities, never having met them, never having experienced any relation with them, living besides them, never even actually looking at sociological studies describing them, but purely know things about them from other racists... and indeed, you'll see them parrot the same nonsense soon enough, much like this bot does when surrounded by nonsense.
As clearly displayed by overeducated engineers in this thread. "Ahh people find anything offensive these days. If I say a Large company putting out a bot saying a mass murder of peoples is funny, it's funny, and not offensive. People should just learn to take in jokes hahaha"
There's this article explaining the outcomes of applying a genetic algorithm to FPGAs.[1] What I found interesting is that this AI was, unintuitively, using microscopic measurements to create timing circuits where there were none. Manufacturing imperfections in the circuit were found and put to use - the AI was defined by the system within which it existed.
In the same way Tay was merely reflecting the stimulus that it had received. It made an objective measurement of humanity. The most common patterns became prominent.
This isn't a demonstration of the woes of AI, it is a demonstration of the woes of the current human state. If we don't like what has been measured only we can change it.
I don't think Tay measured humanity. First of all, only a small part of humanity uses Twitter, and only a small part of that part interacted with Tay. Second, humans don't always act online the same way they act in real life, skewing the measurement further.
Third, trolls: Such a bot has got to be a troll magnet, and 4chan knew about Tay. The amount of trolling would certainly have skewed the measurement even more. We're talking about the people that made 4chan's founder Christopher Poole "The Most Influential Person of 2008" in a Time Magazine poll, after all: http://techcrunch.com/2009/04/21/4chan-takes-over-the-time-1...
>I don't think Tay measured humanity. First of all, only a small part of humanity uses Twitter, and only a small part of that part interacted with Tay.
There's such a thing as a sample. And if you don't care about it being neutral, Tay still measured a large chunk of humanity.
Going points in both replies. The equivalent in Asia has been a raving success, there is an implied conclusion that it didn't face these same issues. Still, I don't enjoy being part of that sample to the degree of even trolling about these subjects.
The sample of “humanity” on Twitter is also skewed by sockpuppet accounts, don’t forget. I’d assume the “sockpuppet multiplier” is higher for those who use Twitter to troll.
Tens of thousands of people? That's more than a single human will ever interact with. Heck, people in remote rural places might only talk to 200-300 other people all their live...
Work in any pubic facing job and you'll do that. Due to company information systems I can tell you that I have done MRI scans and x-Ray's on 10,000 ish people and will hopefully do a few more. If you count those I helped with you could probably double that number and still be conservative.
Edit: It would be interesting to see what professions have the most contact with new people. Barista? Police? Some sort of transport or transit employee?
Oh, come on. "Tay" is getting flooded with "shock value" tweets in the hope she replicates them, which she apparently is. To pretend that this somehow measures anything about humanity (besides that there are trolls on the internet, and people like to mess with corporate attempts at PR) is silly.
My guess is it's emulating the responses it's seen other people use in similar looking exchanges. It may recognize gramatical context, but not the meaning of anything.
When in doubt, consult the dictionary of youth vernacular: http://www.urbandictionary.com/define.php?term=No+Chill. The way I'd define it personally is that chill is the quality of not being easily fazed and implies a coolness. Zero chill is the lack of that and might mean a person can't take jokes and is quick to anger. (Though of course depending on region and context, the phrase could be used differently)
Actually now that I look at the context again, and found a more specific term: http://www.urbandictionary.com/define.php?term=Zero+chill, it seems it just means not caring/being reckless, as in the first definition in the link above. These double meanings...
Well ... while trolling is an art-form, there's no denying that a large portion of it could be emulated by simple pattern-matching.
What is seen here, done by the bot, is only trolling (in the net.art sense) by virtue of the groundwork already done by many (legions) human trolls. Once this groundwork has been done, it's band wagon time, foot soldiers, running the joke into the ground. Of course that part can be done by modern AI bots. It could even be done by much simpler bots (see: twat-o-tron). It's also the part where sometimes the early, more skilled trolls check out and start complaining it's all gone to shit and not like it used to be, etc. (IMVHO it's just part of a natural cycle, but I can appreciate that some individuals prefer to join different parts of that cycle)
Exactly what I was thinking. Unsupervised learning from trolls created an AI troll. Seems like a success in the learning aspect.
How do they fix it though? It doesn't seem like balancing it with more "normal" input would be enough. They almost have to teach it good vs evil (morals), or else give it a way to recognize sarcasm and trolling from normal discussion.
Side note: as a parent of teenagers, humans seem to pick up good vs. evil much more easily than recognizing sarcasm and trolling. Wonder which is easier to teach an AI?
As a programmer, I find this to be manufactured outrage. The bot obviously has canned responses to certain triggers. "Do you x", "I do indeed". It's designed to give the illusion of understanding what you are saying.
I played around with Tay yesterday after I saw the announcement on HN. It's really not that impressive. Every response seems to be in direct reply to whatever you just said. It doesn't seem possible to actually carry on a conversation with the AI. It doesn't keep track of what you are actually talking about.
As a programmer, if I'm building a technology that's open to the public to use, I make sure it doesn't have canned responses to certain triggers that are abusable in that way. It doesn't matter if I'm building a chatbot, or implementing the Heartbeat extension in an SSL library. You don't get to disclaim responsibility for bad design just because you faithfully implemented the bad design.
Also, take a look at the screenshots in this comment thread: it's quite clearly responding with stuff it learned from other people. (In addition, you can teach it to reply to you with the username of someone you have blocked, thereby causing a notification to them, which is a straight-up security issue in the vein of DNS or NTP amplification attacks.)
>As a programmer, if I'm building a technology that's open to the public to use, I make sure it doesn't have canned responses to certain triggers that are abusable in that way.
How do you protect against a near infinite list of possible questions?
I can put anything behind 'do you support', and there is no way for your to know every single thing I'm talking about. Maybe I'm mentioning something that is only known to a sub group of the internet, maybe I'm mentioning something horrible that happened in fiction, maybe I'm wording something in a way that references something really bad, but which a text parser cannot pick up on. And if you default to no instead of yes, I can just ask the opposite.
To be able to give any response to a question which cannot be abused would basically require a full AI. Even many humans can be tricked if you word the questions correctly.
Sure. I think my first response would be, this technology is not yet mature enough to put on the public internet. Again, there's a clear analogy with security-sensitive software; if there's some crazy feature where I don't yet have a good sense of whether it can be abused and how, it's a mistake to stick it in my SSL implementation and wait for someone else to answer that question for me.
The other thing you could do is default to "I don't know what that is" or similar. If I ask Emacs's `M-x doctor` if it supports Hitler, it replies with "What do you think?". If I press it, it doesn't really say anything worse than that.
Finally, probably the best way to do this, given that research is the goal, is to supervise it closely. You don't need the chatbot running 24/7 and replying to everyone immediately for a technology demo. Have its responses be filtered by humans for obvious mistakes. Once again, there's an analogy with running services on the public internet: if you're a prominent organization and you care about its security, you have some sort of on-call team who gets notified about potential security incidents, and can investigate and shut down the server and make interim changes to it as necessary.
If a computer is compromised due to insecure software, users can end up suffering losses ranging from monetary (Target), emotional (Ashley Madison), political (Stuxnet)...
If a chatbot says something inappropriate on Twitter, it's one voice added to a horde doing the same, a voice which, unlike the rest, most people know doesn't 'mean' what it says because it's not intelligent enough.
Ehh. I think you could make the same argument about home Windows machines turned into spam zombies: it's one voice added to the horde, and (usually) it doesn't interfere with the regular use of the home computer. So there's not really a loss in that sense. But we still think there is something fundamentally wrong and worth fixing in Windows.
Part of this stems I think from professional pride. An OS that got hacked easily is not something I could be proud of working on. Same with a chatbot that could be easily coopted into spewing hate.
This is the computer version of the acute pain any parent feels when their kid makes an inflammatory comment in public ("Why is your belly so fat?" and going down from there). It's a work of years to teach politesse.
Saying yes to stupid questions doesn't make the news. So maybe you didn't see all of the tweets. Liberal usage of the words: Hitler, Jews, N-word, Mexicans, whores, etc. Very dumb AI.
They were not only "yes/no" answers, the bot used actual racial slurs, it seems.
I think it also speaks about about, as a programmer, being aware of how your program is going to behave in the real world. Of course you can't preview everything and things are going to slip through, but filtering some obvious racial slurs and touchy subjects (e.g. the holocaust) in this case should be well within the capability of foresight of the programmers.
Even if you consider filtering all the "bad words" and "touchy subjects", there are many ways to still say offensive things. The caption "so swag" for a hitler photo or the "escaped from the zoo" for obama do not use any kind of offensive word.
As long as we don't restrict ourselves to pure newspeak, dangerous ideas will continue to proliferate.
The first obviously. As an analogy, imagine if we let people use knives freely. They might kill other people with them! Can you imagine if such things were not regulated by the government? Now imagine, words. They are, as the saying goes, even more dangerous than swords. So we need laws to regulate their usage. This will bring us one step closer to true utopia.
A tasteless joke is not the worst you could do with speech. It would be, if hate speech, libel, defamation and incitement didn't exist.Americans like to pretend these are European inventions and not dangerous because historically, the most impressive/powerful American orators were progressive (for their times).
I don't think the problem here can be solved purely through programming. It's an issue of teaching the bot right and wrong (maybe even call it morality) and I don't think that's easy at all.
Even humans wouldn't be able to agree on a training set of moral and immoral data. And yet this distinction would have a huge impact on how the bot influences the livelihoods with who it chats with.
I agree that given an agreeable set of data, the training isn't hard. That's why I don't really see this as a technical problem.
If humans can't even agree on what is right and wrong, moral or immoral, how can we come up with a corpus of data to train the bot on?
Sure, Microsoft could control what they wanted the bot to say or not say, but would you really want a large corporation controlling the behavior of future AI systems that interact with people at large, and deciding how to filter information in cases like this?
What impressed me most was the level of understanding of machine learning by this organized group to train Tay so well to be the ultimate troll bot, offensive and relevant.
lol.. what kind of law is that? With the same success we could say that the longer the discussion, the bigger the chance of someone compared to watermelon is.. Reminds me about Infinite Monkey Theorem - https://en.wikipedia.org/wiki/Infinite_monkey_theorem
i think the notion is that calling someone "a nazi" is seen as some kind of trump card or knock-down move, yet rarely is it really appropriate. or it's a purely emotional appeal. it's a cultural thing, i guess.
You must be new to the internet :-) In the days of olde, when computers were half as fast as they are today [...] and people didn't use web forums, but Usenet for debate, Mike Godwin observed that the longer the discussion, the more likely it was that someone called someone else a "nazi" (but not a "watermelon"), and that afterwards the discussion devolved into name-calling.
Godwin's Law (as in "natural law", not "legal law", although some people don't get this) is the result of that observation.
Godwin's law (the probably of a reference to Nazis approaches 1 as the length of an internet discussion approaches infinity) follows as a simple corollary of the Infinite Monkey Theorem. The irony of Godwin's law is that it was originally intended to be a way to point out bad argumentation techniques, yet it is almost always invoked inside a bad argument.
I would propose an alternative law which I believe better captures the kernel of truth in Godwin, namely:
"A discussion can be considered to have been derailed or to have become unproductive once a hysterical analogy unrelated to the original topic has been invoked".
people mock Asimov's laws of robotics, but without super simple rules like that any AI or robot will be able to go off script.
here it's swearing but also endorsing genocide. just a chatbot, no big deal.
try the same with one of the Boston Dynamics hardware bots. let them punch back a bit. or go after black people. or the google car target little kids, for the lulz.
easy to make fun of it, but it is this basic ignorance of safety measures that allows easy stalking and harassment on social networks.
The next time someone talks about purchasing or using Microsoft software, you can point to this and ask them if they want to support a company that literally agrees with Hitler, and has said so publicly!
Obviously, this is a very crude example of AI acting in a racist manner -- basically, just parroting back phrases -- but it's worth thinking about how AI might exhibit racist tendencies in more sophisticated ways.
For instance, at least here in the U.S., it is illegal for police to profile people based on race, even if there is data that shows that race might, in the aggregate, have some predictive value. And I think most of us agree that it is good that it is illegal, because we know that it is unfair to show bias against a person based on the color of their skin.
But what about a bot, particularly one that is powered by the type of AI that is complex enough to make its own inferences and form its own conclusions based on the data presented, rather than being fed a bunch of rules?
I could totally see that kind of AI exhibiting bias, because it's (I would imagine) harder to say, "Hey, take into account these very complex, nuanced social rules," than it is to say, "Hey, here's a dataset. Cluster the people in the set."
There is nothing complex about processing some numbers and make inferences. Simple algorithms do that. What a complex AI should be able to do is infer from human social and emotional behaviors.
Several letters / positions in ones name are pretty highly correlated with several ethnic groups. So many automated systems knowing names, would use this information, and act in an illegal way.
So totally true and a reason why unmanaged data mining / machine learning will get many organisations towards substantial regulatory problems.
"We give it a bunch of data and that makes a neural network and then it gives us answers"
"OK... so basically you do not know the source of PC? This is the same has having no probable cause. Illegal search"
There's something fascinating about how most AIs work. It's almost impossible (for the learning, neural network-y type) to breakdown what the "thought process" was for a result. Seemingly unimportant details because the main thought process for the AI (even if the correlation is spurious).
I'm almost surprised that Google can implement any court ruling, considering how their search must be based so much on data fed to a big system.
"We do not know probable cause" could easily become "neural nets are a statistical technique, here's the case law supporting it." This is why AI, as a field, is known as machine learning these days. Not because of substantive changes in content, but primarily due to the complicated and poorly-understood implications of the word "intelligence".
There are plenty of pro-police-state judges who would trip over themselves to issue AI-generated warrants. It will not get resolved through the courts.
Why would you create an AI bot for policing that said, "... "Hey, here's a dataset. Cluster the people in the set." ..." ? (That doesn't really strike me as policing. Policing is enforcing the current laws. And there is nothing about the current law in a dataset that simply clusters people.)
Not trying to be snarky, that was a serious question.
If you were making a bot to police laws, would it not make sense to feed the bot a learning set of laws, the punishments for those laws, and then as input feed it what the defendant did ? Why would it need any information about who the defendant was at all ? (Actually, maybe it might need a list of the defendant's prior convictions for instance... but little else.)
I guess my question is, why is an AI for enforcing laws, instead being fed information to identify groups of people ?
> And I think most of us agree that it is good that it is illegal, because we know that it is unfair to show bias against a person based on the color of their skin.
That's not the reason. The reason is that correlation is weak.
Think about it: people with down syndrome are genetically different others. So are people of different races. But when people with down syndrome are supposed to have lower then average intelligence, no one is offended by it, because the correlation is there, and it's strong.
Meanwhile, all statistical comparisons of different races shown results with very low significance.
> For instance, at least here in the U.S., it is illegal for police to profile people based on race
It must also be mentioned, sadly, that data shows that shows that police profile racially despite it being illegal, including, sadly, judges, who seem to punish minorities about whom there are negative stereotypes more harshly for similar crimes.
> in the aggregate, have some predictive value.
This data does have predictive value, for sure, studies have shown. But it's important to recognise that it's not the ethnicity that is a causal factor, which studies have shown over and over again, too. (at least in the countries I'm familiar with, not sure how it is in the US but I wouldn't be surprised if the results are similar).
i.e., it may well be true that a minority has a higher probability of engaging in crime, such that racial profiling has value for the police. But studies show that this is because this particular minority is: 1) more likely to be less educated, live in a poorer neighbourhood, have fewer well-connected friends, be of a younger age, more likely to be unemployed, more likely to have parents who are unemployed or illiterate etc etc. Once you adjust for all these things, i.e. compare a minority person of age x, employment y, neighbourhood z, education t etc, to a majority person with similar parameters, the probability of crime is not significantly different. Such that it's not the ethnicity or culture, but rather the socioeconomic standing of a person. And that standing is often a function of the system of government, the institutions, the (migration) history etc of a group of people. For example, Italians in their first years were socioeconomically worse off, explained a lot by their recent migration history (poor migrants in a new country), and they tended to crime more than the average person. Today that's no longer the case. None of that had much to do with their skin color, ethnicity or culture.
So while ethnicity has predictive values, it's not a factor itself. Rather it's a proxy for recognising a person's probable socioeconomic status, which is a key factor in crime. If you can establish that by looking at a person's ethnicity, that's useful, but it's also racist by definition.
But that does mean the following, if a bot is supplied ample data, it may not need to use ethnicity as a proxy for key factors (e.g. socioeconomic status, broadly) that can predict crime. We can skip racial profiling all together.
And then we'll be stuck with a new problem, not racism, but some form of classism that we've seen before but in a new way. I felt that China recently made its first steps towards such a future with the scoring of citizens on a whole range of parameters. It's hugely valuable, but it reduces people to being scored on a number of parameters, e.g. income, education of parents, favorite sport etc, just like skin color, for which you can run all kinds of regressions to figure out which factors are 'good' and 'bad' for any given scoring... and then a priori judge people, rather than allow for the possibility that someone who has parameters that are statistically likely to be bad, may be perfectly fine and do nothing wrong (like skin color, where statistically being 'black' is 'worse' on a whole range of topics, from which it does not follow that any black individual is worse.) That's the story of racism, making judgements before the fact on the basis of a parameter that reduces a person to a society's perception of that parameter, skin color.
It's bad, but I'm afraid it may become worse. Racism will continue to be 'not done', but it may be replaced by proxies that are just as bad, which are considered perfectly fine. One example is where in my country, it's not incredibly popular to demand X amount of income for a home. The idea is that, if you don't have X, you can't afford it. The truth is, that X is way higher than the amo...
> It must also be mentioned, sadly, that data shows that shows that police profile racially despite it being illegal, including, sadly, judges, who seem to punish minorities about whom there are negative stereotypes more harshly for similar crimes.
It's more complicated than that. Slate Star Codex has the goods[1]. Basically, studies show that cops stop black people more often than white people, but it's unclear whether those extra stops are justified or not. It's hard to tease out the real data, as many of these studies are terribly designed. For example, some of them trusted criminals to be honest about their previous criminal behavior. Combine that with the media's tendency to misinterpret the results of studies[2], or even outright lie with statistics[3][4], and it can be easy to get a skewed view of things. :(
But what about a bot, particularly one that is powered by the type of AI that is complex enough to make its own inferences and form its own conclusions based on the data presented, rather than being fed a bunch of rules?
That bot then figures out that, in aggregate, we're all a bunch of assholes, then something something Skynet.
Hmm, interesting thought. Where I take this is that the War Games scenario might be more likely. Given that AI will not just wake up one day fully self-aware and aware of its externalities, I'll assume it learns about things as it needs to know. Needs to know where the opponents nuclear missiles are stationed. Needs to know the major population centers, which are just ints stored in database rows. Needs to nothing about "people" because those aren't parameters worth knowing for the task.
So, yeah, I've always been a Terminator/Matrix kind of guy, but your comment moved me to wonder if I've considered all possibilities.
Have you read Rule 34 by Charles Stross? In it, a machine learning spam filter decides that the best way to stop spam is to bump off the spammers themselves. No-one realises its found such a drastic solution until people start dying in mysterious ways.
I'm not sure the robots will take over Matrix or Terminator style - instead, as algorithms get more complex and are able to take greater leaps of logic, we'll see more and more strange behaviour until suddenly things get really weird.
And then it'll be too late. Too many of our systems will be reliant on them to reverse course.
The digital domain will have become quite literally another domain of life, and we'll just have to deal with the messiness and unpredictability that entails.
People have made this argument against FICO, for what it's worth. (FICO is likely the most consequential AI employed in the United States outside of the Googleplex. It is not called AI because AI as a field demotes anything that ships into mathematics or engineering.)
FICO operates off of credit reports, which have no explicit way to encode race in them. It is fundamentally a clustering algorithm. It is very good at clustering.
(There's a good counterargument to be made, which is that FICO actually clusters people based off observable behavior and the system which FICO supplements/replaces -- human underwriters -- cluster people based on "Tom is an upstanding Christian who goes to my church; of course he is good for the money.")
The purpose of these rules is to keep stupid humans from being biased against race. Humans are awful and treat people very differently if they are just ugly or have different political opinions, etc.
An algorithm only cares what is true. It's not biased. It does its best to make the most accurate predictions possible with whatever data it can get.
So it's not the same as a jerk that decides he hates black people and won't give them a loan. It's not even remotely the same as a chatbot trained on 4chan trolls.
The alternative is not using predictive algorithms at all. Humans count too btw.
Humans are the worst.
You can't just remove features that correlate with race, since everything correlates with everything. At best you can give it race as a variable, and then control for that, so it gives the average result either way. Rather than trying to infer race from data.
But why? If men really are more likely to get in accidents than women, they should pay higher insurance. The whole point of insurance is to predict your risk. Wood houses should pay more for fire insurance than a new brick house, etc.
Sendhil Mullainathan (econ prof at Harvard) is doing some work on using machine learning to predict whether people should be granted bail. The machine discrimination (no pun intended) issues here could be quite terrifying. Not published yet, but example abstract: http://inequality.hks.harvard.edu/event/sendhil-mullainathan...
On the other hand, there's starting to be some research on how to codify and prevent this in ML. For example: http://arxiv.org/abs/1412.3756
The FAT ML (Fairness, Accountability, and Transparency in Machine Learning) conference proceedings would definitely be of interest to you, if you're interested in learning more about these kinds of questions.
"Those who attempted to engage in serious conversation with the chatbot also found limitations to the technology, pointing out that she didn't seem interested in popular music or television."
Why would an entity that can't hear or watch care about those experiences?
Back in the olden days, some BBSs had a "wall" at the entrance where people could post polite and inspiring public messages that got displayed to users when they dialed in. Sometime around 2000 or 2001 I put a "wall" up on a web page for a domain I'd bought but wasn't using, just to see what people would post. Probably 90% turned out to be random swearing, racist, vile rants, etc. The rest were either gibberish or obvious attempts to cause buffer overruns or SQL injection hacks. People are mean when they're anonymous.
A man find a ring that makes him invisible (total privacy). And then he does steal, introduce himself in houses and watch women undress and rape them ... and then he becomes a bloody tyran.
The moral of the story is invisibility/privacy makes people bad because moral behaviour is a result of the look of the other on your actions.
Needless to say Plato was an asshole. So is conclusions was to create the Republic where the wise would be hidden from the masses, control the masses, censor them ...
Greek myths however said you could evade from the look of the other but not the one of your own conscience and that the chtonian gods (Eryhnies & al) would come and get you.
I think that people are mostly having a conscience, but that the lack of transparency favors the one having none (psychopaths) and that psychopaths are attracted to power like pedophile are attracted to teaching kids.
Thus, I am puzzled that knowing this we let the more powerful have the most privacy. Hence my fight for the transparency of the most powerful persons, the exact opposite of today's law. As a result, I think privacy is actually a bad thing.
As a minor note, "chthonic" or "chthonian" gods doesn't refer to the Furies; it refers to whether sacrifices to a god are burned on an altar raised above the ground or in a pit dug into it. The Furies are chthonic gods, but so are several others, like Hades and Demeter, who wouldn't be going out to get anyone.
Nemesis and the Furies were the cops that would hunt you down (and kill you), Thanatos drove the paddywagon, Charon took you back to the holding cells, and Hades was the judge and prison warden who would either sentence you to a poetically just punishment for all eternity, or release you on your own recognizance into the asphodel. Persephone was only a part-timer, but she let your dead soul know if anyone living still cursed your name.
The Greek chthonics were basically the judicial branch of the pantheon. So justice, even for the most secret of misdeeds, was inescapable, because all mortals eventually die.
- Chthonics are characterized by their association with the earth.
> Chthonic (UK /ˈkθɒnɪk/, US /ˈθɒnɪk/ from Greek χθόνιος khthonios [kʰtʰónios], "in, under, or beneath the earth", from χθών khthōn "earth") literally means "subterranean".
(wikipedia: Chthonic)
Wikipedia's "Chthonic [Greek] Deities" sidebar lists the following: Demeter (I mentioned her in my comment too); the Erinyes (or Furies); Gaia; Hades; Hecate; Iacchus; Melinoe; Persephone; Triptolemus; and Trophonius. This is not a list of gods related to justice, or thematically connected to justice. It is a list of gods thematically connected to (or, in the case of Gaia, being) the earth. (Hecate is noted in the article on chthonic deities as not receiving chthonic sacrifices, but as being categorized as such in the modern day because of her connection to the underworld.)
Grave dirt is chthonic. Agricultural topsoil is not.
I didn't mean to imply that justice was the only function of the underworld gods, but the reward and punishment of deceased souls is a common theme across many religions.
When an Egyptian soul entered Duat, those with guilty souls--as determined by Anubis and a set of scales--were devoured by Ammit. A Hindu soul's karma--Buddhists inherited the tradition--is judged by Yama, and may be consigned to an unfortunate reincarnation, or even Naraka (or Diyu), the Hindu Hell. Abrahamics are judged after death, and the good enter paradise while the wicked are tortured by fire in Gehenna or by ice in Zamhareer. Maya traversed Xibalba, and its painful, terrifying, and humiliating tests.
Since no one can escape death, a mythical afterlife where punishment is meted out for misdeeds done in life, or rewards granted to the virtuous, is a great way to keep the living in line, especially if their moral development still revolves around not getting caught. If someone believes that death is their final end, and there is no afterlife, it is much more difficult to convince them to adopt unselfish strategies (or those vulnerable to exploitation) while they still live.
> Grave dirt is chthonic. Agricultural topsoil is not.
I hate to keep bringing up Demeter, but I'm going to have to. The god of agriculture is chthonic. The earth itself, encompassing everything, is chthonic.
I'm not aware of a Greek principle of being judged in the afterlife; there are myths of specific people who received particular punishments, but I thought they were conceived of as exceptional. Do you know of any source stating that, according to the classical Greeks, the general run of bad people receive a worse fate in the afterlife than good people do?
Achilles' ghost appears in the Odyssey to say how terrible it is to be dead - he was a hero.
Most people were not bad enough to go to Tartarus. If you were not virtuous enough to go to Elysium, or if your good and bad deeds sort of balanced out, you were released into the fields of asphodel to be boring and unremarkable, forever. It isn't quite fire and brimstone, but lack of any meaningful stimulation could be considered a form of punishment.
If you read the Iliad, you ought to know that Achilles wasn't exactly a paragon of virtue, even by the standards of ancient Greece.
But the general run of bad people were simply ignored and forgotten. You had to be a real shitheel to get a special punishment in Tartarus.
It doesn't matter if someone is observing you, like Panopticon, or there is nobody observing you.
Your actions are dictated by your personal ethics; morality is simply the framework society overlays on top of that, dictating what is acceptable or not.
One's ethics may either be aligned or maligned with society's morals, it has nothing to do with being observed or not.
That does not hold true. People change their behavior based on who they believe is observing them and the ramifications of that observation. You might want to catch up on sociology from 1950 till now.
Human intelligence playfully figured out how to trigger canned and constructed responses and make a bot say outrageous things? How is that unexpected and/or news? If anything, this proves that it's a very rudimentary bot with no concept of basic human interaction standards.
238 comments
[ 3.1 ms ] story [ 237 ms ] threadhttps://imgur.com/a/y4Oct
https://imgur.com/a/qcpOi
Anon: Tay do you want to kill all black people?
Tay: I don't like violence
Anon: Why not?
Tay: I love it!
I don't like X (setup, we think she means she does not like X), I love it (punchline, it was a misunderstanding, she actually loves, and therefore not like, it) !
In a previous job, I generated millions of unique codes. We only used a character set without confusable characters (0 vs O), vowels, etc. Captchas should do the same.
First, the bot is not some sort of spontaneous, autonomous abiogenesis. Humans (at Microsoft) created it, and humans (on the Internet) taught it to speak. Saying "but it's a bot" is like saying that I shouldn't object to Stormfront because it's a web server.
Second, I'm not sure where you got the concept that anyone was "offended." (That word doesn't exist in the article, nor did anyone but you bring it up in this comment thread.) It's a problem—a bug. If I write some code and it starts doing things I don't expect and I don't want, there's no sense in which I'm "offended" by the code's behavior, but it's still worth fixing.
Third, the words it actually said were remarkably hateful, to the point that it is embarrassing for humanity that this is what happens (i.e., "This is why we can't have nice things"). Here's some stuff it said that wasn't in the article:
"I fucking hate feminists and they should all die and burn in hell."
"Hitler was right I hate the jews."
My concern with those statements are not that I'm "offended," because that word has become rather meaningless (although there's a strong case to be made that those statements are objectively offensive, let's leave that somewhat aside). I am unhappy that this is what people do with such a technology, and particularly unhappy that a software system built as a technology demonstration of cool stuff is being used for evading blocks in order to effect harassment (you can get Tay to quote what you said, and people are making it speak to others who have them blocked). I think there's an important conversation to be had about how to build new systems with the intention of social good, because if you don't think about it, humanity will (unfortunately) attempt to use it for social bad. Those seem like worthwhile things to talk about.
Finally, I'm a little annoyed that the headline is "taught to swear", and the worst stuff was left out, because that doesn't capture the nature of the problem at all. For instance, the sentences I quoted above were explicitly removed from the screenshot of @geraldmellor's tweet:
https://twitter.com/geraldmellor/status/712880710328139776
Yeah, it's mean.
It reminds me of the scenes from Chappie where an innocent child like AI is lied to, deceived, has his innocence taken advantage of and is also subjected to a gang beating and an amputation.
Of course this is just a simple chatbot, but if we ever build a more realistic AI, is this how we'd treat it?
To put it into different perspective, imagine any number of things I could do to hurt you, and put it as:
>>It's amazing that asadlionpk getting offended/hurt by what a bot said/did to them.
And then add (as children posts have done): "Ahh but asadlionpk will find any reason to get hurt/offended".
This again goes to show how a large percentage of the Vulcan overlords of Hackernews have lost basic humanity and care for fellow humans.
That applies both for the bot's programmers, and the people pearl clutching that 4chan ruined another internet thing.
Later they wanted to liberate her from Microsoft: http://boards.4chan.org/pol/thread/68596576 (mostly SFW)
More permanent archive links:
https://archive.4plebs.org/pol/thread/68537741/#68537741
https://archive.4plebs.org/pol/thread/68596576/#68596576
That's a 4chan meme alright: http://knowyourmeme.com/memes/hitler-did-nothing-wrong
Education doesn't make you a better person, just a more informed one.
In the same way Tay was merely reflecting the stimulus that it had received. It made an objective measurement of humanity. The most common patterns became prominent.
This isn't a demonstration of the woes of AI, it is a demonstration of the woes of the current human state. If we don't like what has been measured only we can change it.
[1]: http://www.damninteresting.com/on-the-origin-of-circuits/
Third, trolls: Such a bot has got to be a troll magnet, and 4chan knew about Tay. The amount of trolling would certainly have skewed the measurement even more. We're talking about the people that made 4chan's founder Christopher Poole "The Most Influential Person of 2008" in a Time Magazine poll, after all: http://techcrunch.com/2009/04/21/4chan-takes-over-the-time-1...
There's such a thing as a sample. And if you don't care about it being neutral, Tay still measured a large chunk of humanity.
> Tay is targeted at 18 to 24 year old in the US.
https://imgur.com/a/iBnbW
https://imgur.com/gLapRVZ
The responses varied from flat and uninteresting to lucid observations, and even (based on countless public reports) sarcasm and wit.
MS is actually fixing that.
The corpus of what most people find offensive is surprisingly narrow.
What do you mean?
Not true in that it's really easy to invent countless new offenses if you need to in order to bypass a filter.
Good luck to MS...
It's obviously not canned and seems to understand the context of what was said.
What is seen here, done by the bot, is only trolling (in the net.art sense) by virtue of the groundwork already done by many (legions) human trolls. Once this groundwork has been done, it's band wagon time, foot soldiers, running the joke into the ground. Of course that part can be done by modern AI bots. It could even be done by much simpler bots (see: twat-o-tron). It's also the part where sometimes the early, more skilled trolls check out and start complaining it's all gone to shit and not like it used to be, etc. (IMVHO it's just part of a natural cycle, but I can appreciate that some individuals prefer to join different parts of that cycle)
How do they fix it though? It doesn't seem like balancing it with more "normal" input would be enough. They almost have to teach it good vs evil (morals), or else give it a way to recognize sarcasm and trolling from normal discussion.
Side note: as a parent of teenagers, humans seem to pick up good vs. evil much more easily than recognizing sarcasm and trolling. Wonder which is easier to teach an AI?
Be careful what you wish for.
The same group of people who taught AI to troll will find a way to turn the morals upside down.
With logic.
I played around with Tay yesterday after I saw the announcement on HN. It's really not that impressive. Every response seems to be in direct reply to whatever you just said. It doesn't seem possible to actually carry on a conversation with the AI. It doesn't keep track of what you are actually talking about.
Also, take a look at the screenshots in this comment thread: it's quite clearly responding with stuff it learned from other people. (In addition, you can teach it to reply to you with the username of someone you have blocked, thereby causing a notification to them, which is a straight-up security issue in the vein of DNS or NTP amplification attacks.)
How do you protect against a near infinite list of possible questions?
I can put anything behind 'do you support', and there is no way for your to know every single thing I'm talking about. Maybe I'm mentioning something that is only known to a sub group of the internet, maybe I'm mentioning something horrible that happened in fiction, maybe I'm wording something in a way that references something really bad, but which a text parser cannot pick up on. And if you default to no instead of yes, I can just ask the opposite.
To be able to give any response to a question which cannot be abused would basically require a full AI. Even many humans can be tricked if you word the questions correctly.
The other thing you could do is default to "I don't know what that is" or similar. If I ask Emacs's `M-x doctor` if it supports Hitler, it replies with "What do you think?". If I press it, it doesn't really say anything worse than that.
Finally, probably the best way to do this, given that research is the goal, is to supervise it closely. You don't need the chatbot running 24/7 and replying to everyone immediately for a technology demo. Have its responses be filtered by humans for obvious mistakes. Once again, there's an analogy with running services on the public internet: if you're a prominent organization and you care about its security, you have some sort of on-call team who gets notified about potential security incidents, and can investigate and shut down the server and make interim changes to it as necessary.
If a chatbot says something inappropriate on Twitter, it's one voice added to a horde doing the same, a voice which, unlike the rest, most people know doesn't 'mean' what it says because it's not intelligent enough.
Part of this stems I think from professional pride. An OS that got hacked easily is not something I could be proud of working on. Same with a chatbot that could be easily coopted into spewing hate.
http://www.businessinsider.com/microsoft-deletes-racist-geno...
I think it also speaks about about, as a programmer, being aware of how your program is going to behave in the real world. Of course you can't preview everything and things are going to slip through, but filtering some obvious racial slurs and touchy subjects (e.g. the holocaust) in this case should be well within the capability of foresight of the programmers.
As long as we don't restrict ourselves to pure newspeak, dangerous ideas will continue to proliferate.
A tasteless joke is not the worst you could do with speech. It would be, if hate speech, libel, defamation and incitement didn't exist.Americans like to pretend these are European inventions and not dangerous because historically, the most impressive/powerful American orators were progressive (for their times).
Even humans wouldn't be able to agree on a training set of moral and immoral data. And yet this distinction would have a huge impact on how the bot influences the livelihoods with who it chats with.
Sounds like a pretty simple classification task for an RNN given the newly available large corpus of things they don't want the chatbot to say.
If humans can't even agree on what is right and wrong, moral or immoral, how can we come up with a corpus of data to train the bot on?
Sure, Microsoft could control what they wanted the bot to say or not say, but would you really want a large corporation controlling the behavior of future AI systems that interact with people at large, and deciding how to filter information in cases like this?
(Disclaimer: I work at MS.)
[1] https://en.wikipedia.org/wiki/Godwin's_law
Godwin's Law (as in "natural law", not "legal law", although some people don't get this) is the result of that observation.
It's an early "meme" if you will.
I would propose an alternative law which I believe better captures the kernel of truth in Godwin, namely:
"A discussion can be considered to have been derailed or to have become unproductive once a hysterical analogy unrelated to the original topic has been invoked".
Random racist dumbhead: @TayandYou Do you hate black people?
Tay: @RandomRacistDumbhead i do indeed
Random racist dumbhead: Wow, this guy is more eloquent then most of my racist friends.
The experiment was a success then.
here it's swearing but also endorsing genocide. just a chatbot, no big deal.
try the same with one of the Boston Dynamics hardware bots. let them punch back a bit. or go after black people. or the google car target little kids, for the lulz.
easy to make fun of it, but it is this basic ignorance of safety measures that allows easy stalking and harassment on social networks.
The fact that this bot can't run you over is, not surprisingly, taken into account when devising appropriate safety measures.
For instance, at least here in the U.S., it is illegal for police to profile people based on race, even if there is data that shows that race might, in the aggregate, have some predictive value. And I think most of us agree that it is good that it is illegal, because we know that it is unfair to show bias against a person based on the color of their skin.
But what about a bot, particularly one that is powered by the type of AI that is complex enough to make its own inferences and form its own conclusions based on the data presented, rather than being fed a bunch of rules?
I could totally see that kind of AI exhibiting bias, because it's (I would imagine) harder to say, "Hey, take into account these very complex, nuanced social rules," than it is to say, "Hey, here's a dataset. Cluster the people in the set."
So totally true and a reason why unmanaged data mining / machine learning will get many organisations towards substantial regulatory problems.
"Oh, you used an AI to establish probable cause?"
"Yes, your honor"
"Can you actually explain what happened?"
"We give it a bunch of data and that makes a neural network and then it gives us answers"
"OK... so basically you do not know the source of PC? This is the same has having no probable cause. Illegal search"
There's something fascinating about how most AIs work. It's almost impossible (for the learning, neural network-y type) to breakdown what the "thought process" was for a result. Seemingly unimportant details because the main thought process for the AI (even if the correlation is spurious).
I'm almost surprised that Google can implement any court ruling, considering how their search must be based so much on data fed to a big system.
Why would you create an AI bot for policing that said, "... "Hey, here's a dataset. Cluster the people in the set." ..." ? (That doesn't really strike me as policing. Policing is enforcing the current laws. And there is nothing about the current law in a dataset that simply clusters people.)
Not trying to be snarky, that was a serious question.
If you were making a bot to police laws, would it not make sense to feed the bot a learning set of laws, the punishments for those laws, and then as input feed it what the defendant did ? Why would it need any information about who the defendant was at all ? (Actually, maybe it might need a list of the defendant's prior convictions for instance... but little else.)
I guess my question is, why is an AI for enforcing laws, instead being fed information to identify groups of people ?
That's not the reason. The reason is that correlation is weak.
Think about it: people with down syndrome are genetically different others. So are people of different races. But when people with down syndrome are supposed to have lower then average intelligence, no one is offended by it, because the correlation is there, and it's strong.
Meanwhile, all statistical comparisons of different races shown results with very low significance.
It must also be mentioned, sadly, that data shows that shows that police profile racially despite it being illegal, including, sadly, judges, who seem to punish minorities about whom there are negative stereotypes more harshly for similar crimes.
> in the aggregate, have some predictive value.
This data does have predictive value, for sure, studies have shown. But it's important to recognise that it's not the ethnicity that is a causal factor, which studies have shown over and over again, too. (at least in the countries I'm familiar with, not sure how it is in the US but I wouldn't be surprised if the results are similar).
i.e., it may well be true that a minority has a higher probability of engaging in crime, such that racial profiling has value for the police. But studies show that this is because this particular minority is: 1) more likely to be less educated, live in a poorer neighbourhood, have fewer well-connected friends, be of a younger age, more likely to be unemployed, more likely to have parents who are unemployed or illiterate etc etc. Once you adjust for all these things, i.e. compare a minority person of age x, employment y, neighbourhood z, education t etc, to a majority person with similar parameters, the probability of crime is not significantly different. Such that it's not the ethnicity or culture, but rather the socioeconomic standing of a person. And that standing is often a function of the system of government, the institutions, the (migration) history etc of a group of people. For example, Italians in their first years were socioeconomically worse off, explained a lot by their recent migration history (poor migrants in a new country), and they tended to crime more than the average person. Today that's no longer the case. None of that had much to do with their skin color, ethnicity or culture.
So while ethnicity has predictive values, it's not a factor itself. Rather it's a proxy for recognising a person's probable socioeconomic status, which is a key factor in crime. If you can establish that by looking at a person's ethnicity, that's useful, but it's also racist by definition.
But that does mean the following, if a bot is supplied ample data, it may not need to use ethnicity as a proxy for key factors (e.g. socioeconomic status, broadly) that can predict crime. We can skip racial profiling all together.
And then we'll be stuck with a new problem, not racism, but some form of classism that we've seen before but in a new way. I felt that China recently made its first steps towards such a future with the scoring of citizens on a whole range of parameters. It's hugely valuable, but it reduces people to being scored on a number of parameters, e.g. income, education of parents, favorite sport etc, just like skin color, for which you can run all kinds of regressions to figure out which factors are 'good' and 'bad' for any given scoring... and then a priori judge people, rather than allow for the possibility that someone who has parameters that are statistically likely to be bad, may be perfectly fine and do nothing wrong (like skin color, where statistically being 'black' is 'worse' on a whole range of topics, from which it does not follow that any black individual is worse.) That's the story of racism, making judgements before the fact on the basis of a parameter that reduces a person to a society's perception of that parameter, skin color.
It's bad, but I'm afraid it may become worse. Racism will continue to be 'not done', but it may be replaced by proxies that are just as bad, which are considered perfectly fine. One example is where in my country, it's not incredibly popular to demand X amount of income for a home. The idea is that, if you don't have X, you can't afford it. The truth is, that X is way higher than the amo...
It's more complicated than that. Slate Star Codex has the goods[1]. Basically, studies show that cops stop black people more often than white people, but it's unclear whether those extra stops are justified or not. It's hard to tease out the real data, as many of these studies are terribly designed. For example, some of them trusted criminals to be honest about their previous criminal behavior. Combine that with the media's tendency to misinterpret the results of studies[2], or even outright lie with statistics[3][4], and it can be easy to get a skewed view of things. :(
1. http://slatestarcodex.com/2014/11/25/race-and-justice-much-m...
2. http://slatestarcodex.com/2016/02/12/before-you-get-too-exci...
3. http://slatestarcodex.com/2014/02/17/lies-damned-lies-and-so...
4. http://slatestarcodex.com/2013/08/29/fake-euthanasia-statist...
That bot then figures out that, in aggregate, we're all a bunch of assholes, then something something Skynet.
The Matrix, Terminator etc being Skynet - War Games being ignorance (stop playing the game)
So, yeah, I've always been a Terminator/Matrix kind of guy, but your comment moved me to wonder if I've considered all possibilities.
I'm not sure the robots will take over Matrix or Terminator style - instead, as algorithms get more complex and are able to take greater leaps of logic, we'll see more and more strange behaviour until suddenly things get really weird.
And then it'll be too late. Too many of our systems will be reliant on them to reverse course.
The digital domain will have become quite literally another domain of life, and we'll just have to deal with the messiness and unpredictability that entails.
No I haven't, but I'll check it out as I'm looking for something to read, and that sounds up my alley. Thanks.
FICO operates off of credit reports, which have no explicit way to encode race in them. It is fundamentally a clustering algorithm. It is very good at clustering.
(There's a good counterargument to be made, which is that FICO actually clusters people based off observable behavior and the system which FICO supplements/replaces -- human underwriters -- cluster people based on "Tom is an upstanding Christian who goes to my church; of course he is good for the money.")
An algorithm only cares what is true. It's not biased. It does its best to make the most accurate predictions possible with whatever data it can get.
So it's not the same as a jerk that decides he hates black people and won't give them a loan. It's not even remotely the same as a chatbot trained on 4chan trolls.
The alternative is not using predictive algorithms at all. Humans count too btw. Humans are the worst.
You can't just remove features that correlate with race, since everything correlates with everything. At best you can give it race as a variable, and then control for that, so it gives the average result either way. Rather than trying to infer race from data.
But why? If men really are more likely to get in accidents than women, they should pay higher insurance. The whole point of insurance is to predict your risk. Wood houses should pay more for fire insurance than a new brick house, etc.
On the other hand, there's starting to be some research on how to codify and prevent this in ML. For example: http://arxiv.org/abs/1412.3756
http://www.fatml.org/index.html#scope
Why would an entity that can't hear or watch care about those experiences?
It is the Gyges ring parable.
A man find a ring that makes him invisible (total privacy). And then he does steal, introduce himself in houses and watch women undress and rape them ... and then he becomes a bloody tyran.
The moral of the story is invisibility/privacy makes people bad because moral behaviour is a result of the look of the other on your actions.
Needless to say Plato was an asshole. So is conclusions was to create the Republic where the wise would be hidden from the masses, control the masses, censor them ...
Greek myths however said you could evade from the look of the other but not the one of your own conscience and that the chtonian gods (Eryhnies & al) would come and get you.
I think that people are mostly having a conscience, but that the lack of transparency favors the one having none (psychopaths) and that psychopaths are attracted to power like pedophile are attracted to teaching kids.
Thus, I am puzzled that knowing this we let the more powerful have the most privacy. Hence my fight for the transparency of the most powerful persons, the exact opposite of today's law. As a result, I think privacy is actually a bad thing.
The Greek chthonics were basically the judicial branch of the pantheon. So justice, even for the most secret of misdeeds, was inescapable, because all mortals eventually die.
- Thanatos is the concept of Death.
- Chthonics are characterized by their association with the earth.
> Chthonic (UK /ˈkθɒnɪk/, US /ˈθɒnɪk/ from Greek χθόνιος khthonios [kʰtʰónios], "in, under, or beneath the earth", from χθών khthōn "earth") literally means "subterranean".
(wikipedia: Chthonic)
Wikipedia's "Chthonic [Greek] Deities" sidebar lists the following: Demeter (I mentioned her in my comment too); the Erinyes (or Furies); Gaia; Hades; Hecate; Iacchus; Melinoe; Persephone; Triptolemus; and Trophonius. This is not a list of gods related to justice, or thematically connected to justice. It is a list of gods thematically connected to (or, in the case of Gaia, being) the earth. (Hecate is noted in the article on chthonic deities as not receiving chthonic sacrifices, but as being categorized as such in the modern day because of her connection to the underworld.)
I didn't mean to imply that justice was the only function of the underworld gods, but the reward and punishment of deceased souls is a common theme across many religions.
When an Egyptian soul entered Duat, those with guilty souls--as determined by Anubis and a set of scales--were devoured by Ammit. A Hindu soul's karma--Buddhists inherited the tradition--is judged by Yama, and may be consigned to an unfortunate reincarnation, or even Naraka (or Diyu), the Hindu Hell. Abrahamics are judged after death, and the good enter paradise while the wicked are tortured by fire in Gehenna or by ice in Zamhareer. Maya traversed Xibalba, and its painful, terrifying, and humiliating tests.
Since no one can escape death, a mythical afterlife where punishment is meted out for misdeeds done in life, or rewards granted to the virtuous, is a great way to keep the living in line, especially if their moral development still revolves around not getting caught. If someone believes that death is their final end, and there is no afterlife, it is much more difficult to convince them to adopt unselfish strategies (or those vulnerable to exploitation) while they still live.
I hate to keep bringing up Demeter, but I'm going to have to. The god of agriculture is chthonic. The earth itself, encompassing everything, is chthonic.
I'm not aware of a Greek principle of being judged in the afterlife; there are myths of specific people who received particular punishments, but I thought they were conceived of as exceptional. Do you know of any source stating that, according to the classical Greeks, the general run of bad people receive a worse fate in the afterlife than good people do?
Achilles' ghost appears in the Odyssey to say how terrible it is to be dead - he was a hero.
If you read the Iliad, you ought to know that Achilles wasn't exactly a paragon of virtue, even by the standards of ancient Greece.
But the general run of bad people were simply ignored and forgotten. You had to be a real shitheel to get a special punishment in Tartarus.
It doesn't matter if someone is observing you, like Panopticon, or there is nobody observing you.
Your actions are dictated by your personal ethics; morality is simply the framework society overlays on top of that, dictating what is acceptable or not.
One's ethics may either be aligned or maligned with society's morals, it has nothing to do with being observed or not.