Microsoft chatbot is taught to swear on Twitter
bbc.com
bbc.com
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.
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.
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.)
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 amount you need to reasonable afford that home, and so it's essentially creating a neighbourhood of certain affluence that weeds out socioeconomic groups that could afford to live there, but are discriminated against on the basis of their income, which, surprise, happens to create segregated communities because ethnicity are income are still pretty closely linked. But because it's done under the guise of 'we're just protecting ourselves and renters by renting only to those who can afford it', it doesn't create a single word of discussion. It's a tricky situation because I tend to appreciate the concept of minimum income for homes, but the level they set it at made me suspicious, and I'm seeing neighbourhoods segregate in my city. Further, OF COURSE it must be allowed to judge people on certain parameters, like say a degree when getting a job as a proxy for how qualified someone is, even though a non-degree holder may be more qualified, it's a reasonable form of discrimination. But at some point, we're also reducing a person to a small set of data and its associated probability set, much like skin color. I find this a tricky issue, although fortunately it seems that non-racial discrimination is limited to only a few things, like gender and age. You hear stories of some companies rejecting a person on the basis of his post code or address (reject poor neighbourhoods), looks or favorite sport (ski vacations?), but it's not widespread.
Anyway so much for my rambling. In short, I think it's reasonably easy to not give a bot racial data, but I think the bot will recreate the probabilities on the basis of other data that proxies socieconomic status, and that this is concerning in a way that is similar to racial profiling, because it reduces a person to the parameters he has and how those parameters happen to correlate with e.g. crime in a large group of people that he may behave differently to. This is how the race-parameter works, you're black? well that correlates to crime in a group of people, so we'll treat you like you're likely to be a criminal. That's bad, but if you ignore that parameter, there are others that are similarly bad, that a bot will find. Like oh you're poor, or oh you love in that neighborhood, or oh your parents are illiterate? Well that correlates to crime, so we'll treat you like you're likely to be a criminal. Non-racial parameter profiling, bots will be champs in that.
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
The experiment was a success then.
The fact they didn't even have something as simple as a naive set of filter words (nothing good comes from Godwin's law when real intelligence is involved, let alone artificial) is insane to me. Letting it respond to anyone and everyone under the sun (96k tweets - one per second) is just a bad idea given that people would probe every nook and cranny regardless of whether it was near perfect. Additionally, allowing a "repeat after me" option is just begging for people to ask the bot to say idiotic things ...
As someone who works in the field of machine learning, this is a sad day. Regardless of whether it involved good machine learning at the base, the copy and paste aspect means it's going to add to the ridiculous hype and hysteria around ML.
=== Primary proof re: copy+paste (or overfitting at best) from the "Is Ted Cruz the zodiac killer" response:
Tay's reply: https://i.imgur.com/PPnCHnf.jpg
Tweet the response was stolen from: https://twitter.com/merylnet/status/703079627288260608
Secondary proof re: copy+paste from https://twitter.com/TayandYou/status/712753457782857730:
Tweet the response was stolen from: https://twitter.com/Queeeten/status/703049861214547968
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) !
More info: https://www.reddit.com/r/SubredditSimulator/comments/3g9ioz/...
I don't see why this wasn't the expected outcome, have none of the developers spent any time on the internet?
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.
That's a 4chan meme alright: http://knowyourmeme.com/memes/hitler-did-nothing-wrong
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.
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:
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:
Source: Tay herself. https://twitter.com/TayandYou/status/712856578567839745?ref_...
I dislike the fact that they decided to lobotomize the AI this fast without further studying. So it's probably just another markov chain.
seriously though
I will now forever be using the term 'MS AI' to refer to buggy AI programs.
Why would an entity that can't hear or watch care about those experiences?
Ok, so rasits=racist, but what "but use Skyrim metaphors" is supposed to mean?
>(it's a 4chan poll,
Oh, I thought it was the official Tay documentation.
>so for the love of Tay don't take it seriously)
What would "take it seriously" mean?
It somehow is not a test for intelligence. We learned to behave through years of interacting with each other.
Education doesn't make you a better person, just a more informed one.
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.
I enjoyed this article, which has more details and even worse examples of her tweets: http://www.telegraph.co.uk/technology/2016/03/24/microsofts-...