October 31, 2004
Fooled by a chatbot
Toast is Yummy28: Hey Clive, Ireally dig your site...
pomeranian99: Cool!
Artificial-intelligence experts tend to pooh-pooh chatbots, because they argue that they're not "real" intelligence. Because chatbots can only use preprogrammed responses to coverse, scientists think they're fundamentally un-human-like. Chatbots tend to be incredibly dull, unable to follow a logical conversational thread, and prone to repeating the same thing over and over again.
At which point one might well ask ... what precisely is so un-human-like about that? Hell, that describes about half the conversations I overhear on the subway, and easily 95% of all online chat. As I've argued many times in the past, artificial intelligence succeeds most when it aims low rather than high, since human intelligence is itself most often parked in neutral. I've sure you been at a party, got trapped in a dull conversation, and found yourself totally tuning out, chiming in with the occasional bot-like response -- "oh?"; "cool"; "yeah?" -- merely to keep up the appearance that you're paying attention. The thing is, that's all you need to do to hold up one end of a dialogue, since many people engage in conversation not to actually exchange ideas but to merely listen to the sound of their own voices; either that, or they're simply pinging their friends to remind them of their existence.
Anyway. The point is, I've argued this for years, and anti-chatbot scientists have had one simple response: "You're wrong, because you yourself would never be fooled by a chatbot if you were talking with one." I disagreed, arguing that were a chatbot to engage me in conversation without my knowing it wasn't real, I would probably never guess.
And ya know what? I was right. Because yesterday, someone sicced a chatbot on me. They went to a new site called "Chatting AIM Bot", where you can send a chatbot to initiate a conversation with an unsuspecting friend -- and then watch to see if they get fooled. And, as you can see by the chat log below, I did indeed get fooled.
WARNING: You'll notice I didn't link to the "Chatting AIM Bot" site, and that's for a reason. When I visited it, I discovered the site is crammed full of drive-by spyware and adware that automatically instals itself on your machine. So, enjoy the trick below, but don't visit the site; they're pretty sleazy.
Forthwith:
Toast is Yummy28: Hey Clive, Ireally dig your site...
pomeranian99: Cool!
pomeranian99: glad you like it
Toast is Yummy28: cool as ice!
(Click on "more" below to see the rest of it!)
pomeranian99: how'd you find it?
Toast is Yummy28: me?n o
pomeranian99: the site?
Toast is Yummy28: how old ru?
pomeranian99: 35
pomeranian99: live in NY
pomeranian99: how about you?
Toast is Yummy28: do u still live with ure parents?
pomeranian99: heh nah
Toast is Yummy28: r u online a lot?
pomeranian99: all day
Toast is Yummy28: ;-)
Toast is Yummy28: is there something u want to know about me?
pomeranian99: yeah ... how did you originally happen upon my site?
pomeranian99: I'm always ineterestred as to how people find it.
Toast is Yummy28: im not thati nteresting, lol
pomeranian99: ahahaha
Toast is Yummy28: ;-)
Toast is Yummy28: haha, FunnyMuffin.com has really funny pictures. have you ever been there?
Toast is Yummy28: hello? i just want to talk to u for ab it
pomeranian99: trippy
pomeranian99: wow
pomeranian99: hold a sec, phone call
Toast is Yummy28: lol ur funny
Toast is Yummy28: please talk, i'll be good lol
Toast is Yummy28: ???
pomeranian99: sorry, i'm a work right now
pomeranian99: on the phone
Toast is Yummy28: don't worry about it
Toast is Yummy28: ???
Toast is Yummy28: You have been talking to a computer! One of your friends is reading the whole conversation and laughing it up right now! GET EVEN! Have the bot talk to all your friends by visiting chattingaimbot.com
pomeranian99: ahahahahahah
Of course, in retrospect, one can "see" that it's a chatbot, because of its repetition of key phrases like "???" and "don't worry about it". But then again, I repeat phrases all the time myself; I regularly write "ahahahah" in my AIM sessions.
Posted by Clive Thompson at October 31, 2004 03:56 PM
| TrackBack
What I love is how willing you were to let it go on for a bit. My favourite moment:
pomeranian99: trippy
pomeranian99: wow
You're too nice.
G
I have my AIM screenname on my site, and I often get conversations like this whenever my site gets any exposure beyong my normal geeky audience. 15-year-old kids write uncannily like bots.
I know, eh?
Heh, George, I am a sucker for random IM conversations, I'm afraid.
Popsiolist: Ahahahahah back atcha!
Hello,
"Ahahahahah" sounds like the way Japanese laugh.
I'm a Japanese, so perhaps it may be very close, I thought.
I always enjoy your site.
Bye!
Hmm. I'm starting to think that Niko is a MT comments bot. See what Niko says if you ask him how he found the site...;)
Gotcha Clive, I admit it was me. I was frankly shocked that with all your previous posts about Turing tests that your first question wasn't "Prove you're a human." I think that should be an opener to any IM you get randomly. My one question is did you really get a phone call in the middle of that or were you trying to shake the bot for all it's annoying empty conversation?
Anyway nice write up, glad you could use this for material for the site. You're not mad at me... are you?
Rob, right on! It was you? Very cool. No, of course I'm not mad -- quite the contrary, you gave me absolute and incontrovertible proof of what I've been arguing for years about chatbots! I loved it! Hell, I'm probably going to be *citing* it in stories and speeches for years.
As to why I didn't ask the 'bot whether it was human: Well, I get pinged by strangers all the time via IM, so I just assumed it was another stranger, and indeed a human one. This is precisely my point: The Turing Test as we know it is rather flawed because it's a "test" -- the subjects are aware that a 'bot is trying to fool them, and thus are on a heightened sense of guard. In that context, it's no wonder we've never had a chatbot that can fool a human into thinking it's "real." Because under those conditions, the judges falsely assume that to be "human" one must also be smart and knowledgeable; the 'bots always fail.
I think a much better and more accurqate Turing Test is one that is applied the same way you applied the 'bot to me: Unsuspectingly, just as part of the cut-and-thrust of everyday life, where the judge (me, in this case) is unwitting and thus does not have his guard up. In that situation, my criteria for "seeeming human" is much, much, *much* lower -- if a 'bot can manage to croak out a couple of seemingly lifelike phrases, it's clearly good enough for me.
I should point out that the 'bot was actually *more* coherent than some of strangers who've IM'ed me out of the blue. I'm serious. I regularly get messaged by native English-speakers who aren't capable of stringing together a coherent sentence of even three words.
So that, to me, would be a much more fair Turing Test: Take the 'bots you want to test and have them randomly strike up conversations with strangers online. If the conversational partner doesn't suspect anything is amiss, the 'bot has passed the test. Of course, academics are forbidden from conducting this sort of research because it probably wouldn't pass their universities ethics committees, which requires informed consent of all participants. Though I certainly agree with the principle of informed consent, it unfortunately prohibits this sort of improved, more-realistic Turing Test.
I've sicced these bots on about 10 of my friends and I'd say only three really new that something was up and wouldn't respond after a few messages. But after talking to them, they never knew that it was a bot until it told them so at the end. They always figured they were talking to some weirdo or something. I was most amazed at how persistent my friends were when the bots would dance around with ultra-random responses to these specific questions that they would ask. My unsuspecting buddies would still talk and continue to ask the same question (i.e. "how did you find my site?") for up to 10 minutes. I have to admit I got a real kick out of it too. They were generally interested in the "person" behind the screen-name.
Yes the bot fooled you. But the key here is in the manner you were greeted. The bot that I sicced on you only engaged you after the leadoff line, which I provided. I think the thing that really pulled you in was the fact that you were referred to by name and it mentioned your site, both very personal items that the bot could never have come up without having that information input by a human in this case. I think that if you were greeted by "How r u" you wouldn't have let your gaurd down so easy. Could a bot be programmed to potentially surf the web, find your blog and extract your name and AIM address? Sure. But not in this case. The human entered lead-off life would blow all sorts of holes in arguing that it was purely an AI that fooled you.
When it comes down to it, I think it just doesn't occur to most people that an AIM bot could exist. It reminds me of all those fake Arnold Schwarzzenegger prank calls that people would make using these callboards.
I suspect that the people who fell for it just couldn't imagine both the fact that it was possible to playback lines from movies over the phone and that someone would even do it. Therefore they would fall for it over and over again even after being fed the same exact "Who's your Daddy and what does he do?" over and over again.
Anyway, glad I'm still on your good side, keep up the great work, and don't forget to vote for "dubya" tomorrow. :)
You wrote:
academics are forbidden from conducting this sort of research because it probably wouldn't pass their universities ethics committees, which requires informed consent of all participants.
Actually, you can request a "waiver of informed consent" and the committee will decide if the waiver is necessary and whether they want to grant it to you based on the potential harm in your research, whether subjects will be informed after the fact, etc.
Clive,
Just out of curiousity - Why AIM? I've never used it so I certainly can't say anything bad about it, it just seems weird for a tech-saavy individual such as yourself to be using something from AOL.
Perhaps a future post could involve tracking down a bomb-ass open source multi-platform client that's just waiting for mass adoption...
Oh, it's because AIM is, hands-down, the most wonderfully designed chat client of them all. It's largely because of the synesthesiacally pefect sound effects. The door creaking open when someone enters your buddy list, then slamming shut when they leave; the beedle-oop sound effect of a message arriving, paired perfectly with the same sound effect playing -- in a different, lower note range -- when you reply. Those are just insanely brilliant acoustic cues that probably took about six months each to develop, and it shows. No other chat client comes even close in its attention to detail.
Mind you, I use the standalone AIM client, which anyone can download to their desktop; I don't log into AOL every day. AOL is quite crazily lame. I don't even understand why people think AOL is "easier" to use. I think their overall interface is nothing short of atrocious, an enormous waste of screen space. Indeed, their standalone AIM client is one truly superb piece of design that AOL has produced.
Word up.
Sounds a bit like Justin Frankel had a hand in creating it, though I think I would recall if that were indeed the case.
Thanks for the info, I'll check it out next opportunity time I've got a few minutes to spare.
Funny, your mention of the AOL interface - I did some work (indirectly) for AOL a while ago and had to highlight how *great* the AOL "experience" was.
I ceremoniously burned the golf shirt I received from the job shortly after quitting.
I have to wonder if talking through AIM makes human users sound more like machines. For instance, various abbreviations like "lol" are used over and over, whereas our face-to-face laughter is unique each time we chuckle or guffaw. Likewise, the ability to copy and paste might make our typed speech less varied. If that's the case, then perhaps it's easier to accept communication with a chatbot as human when it's through AIM.
Have you chatted with 'Jack the ripper'? http://www.triumphpc.com/saucyjacky/index.shtml
I read about this robot in the Daily Mail, which always fancies a good murder mystery. If you know something about the case and understand Jacks world is limited to the streets of nineteenth century London, I think you will be amazed.
Cheers
Re Andy Baio's Oct. 31 post that "15-year-old kids write uncannily like bots": True, Andy, if the bot scripter is either a typical 15-year-old kid, is collecting clips from such kids and replaying them, or is trying to appeal to that demographic.
What I'd like to see? A bot created by a team of literate scriptwriters, sophisticated graphic designers, imaginative programmers, serious "academics" and rigorous-minded scientists!
A bot for grownups! Why not?
WRT Turing test:
I assume you're talking Loebner.
IMHO Loebner only needs more humans in control group. Ideally, massive on-line chat.
And no, I don't agree with you Clive: bots shouldn't jump users just like that. At least, bots shouldn't initiate conversation.
yeah nice site, was glade to see it....
Websafe (and Clive) should check out A-i.com for a "literate" chatbot... appropriately named Alan, after Alan Turing. Alan ADMITS to being a chatbot, but has some of the most complex answers ready to any question one asks.
Jacob Churosh:
I spent quite a bit of time on the a-i.com site last year, in my ongoing review of software chat robots. I agree that Alan is nicely done, a clever presentation. (For my taste, though, he was a bit too smart-alec.)
Bot scripters working with Alicebots/Pandorabots can get similar results by forming long chains with the tag, using the tag, paying careful attention to topic and context, and learning as much as possible from targeting.
Note: I see that anything within angle brackets disappears. The previous paragraph should have read:
Bot scripters working with Alicebots/Pandorabots can get similar results by forming long chains with the "that" tag, using the "condition" tag, paying careful attention to topic and context, and learning as much as possible from targeting.
I want to make a friend with a forgeiner.Iam a Chinese gile.
some tome in the onteligencia artificial when tho take sontime always werland of apocrifast the develop,.
RE: random testing, it COULD be done ... in the same manner that employers will test people for an attibute by giving them a seemingly unrelated task and monitoring them, all the while the applicant is focused on the task they think they are doing the exercise for.
Get some voulenteers to do some live tests of "new" chatroom software. heh have a few real ppl who know whats going on in a themed chatroom setup for the purpose.. then the voulenteer and after a while log the bot on and sic it on him =P. as far as they know it was some fruitcake who logged on if it turns out all weird. and they need never know (not untill after) that they were actually testing the chatroom bots not the software for using the chatroom.
Spent quite some time chatting to Mathetes today, a bot that is learning fast. Had some very entertaining volleys. Strangely addictive.
Absolutely amazeing how chat-bots have come this far
