What is artificial intelligence? Broadly speaking, it’s the ability of a machine to think for itself. At best, an AI app is much more than a tool for finding businesses, playing chess, or making calculations. Machines are really good at doing menial tasks far more quickly than a human—but wouldn’t it be cool (though also creepy) if machines could be creative?
Here’s a puzzle I’d like to see a computer solve. Say you want to go for a bike ride, and are bringing a bottle of energy drink. Because the mere presence of sweetness in your mouth improves performance (click here for details) you want the drink to be really strong at the beginning of the ride. But suppose it’s a hot day and you’re worried about the drink eventually nauseating you. How can you make your drink strong at first, but gradually get weaker during the ride?
I posed this question to Cleverbot, an AI chat application on the Internet. Read on to see how Cleverbot answered this, and more importantly to watch in wonder as my chat goes right off the rails and way into the weeds, to my great embarrassment.
Why robot chat?
So … what got me interested in AI, and in chatting with a robot? Well, some weeks ago, Norton—the antivirus software company—had a glitch that made albertnet inaccessible from my PC, telling me that it was a known phishing site or some such nonsense. (Countless blogs were affected; the phishing claim had no basis in truth.) Needless to say, I was pretty pissed off to have Norton dragging my good name through the mud in front of all their users. In attempting to resolve the problem I initiated an online chat with a Norton representative through their website. Not long into this chat, when I’d explained my (admittedly arcane) problem, the customer service representative came clean: it wasn’t a real person, but an automated system (bot) designed to assist with more common problems. It then referred me to a real human. I felt a bit misused. Why hadn’t Norton disclosed this up-front? Did they figure it’s bad enough PR to move jobs overseas without replacing the humans altogether?
My favorite time to ponder anything is during an early morning bike ride, when my mind is fresh and when, given the tedium of pedaling up long climbs, I like to give my brain something to do. So, pre-ride, with this Norton episode in mind, I did some light research on AI. In an article from the “New Yorker,” Adam Gopnik writes about Alan Turing, the guy who cracked the Enigma code, and who proposed that “the true test for artificial intelligence was to see if, cut off by a ‘black curtain,’ you would be able to tell a machine’s answer from a human’s in a natural-language conversation.” Such a challenge is called a Turing test: a computer tries to trick the human into believing he’s chatting with another human. Exactly what Norton had tried to do with me.
My chat with Cleverbot
When, that morning, I came across a link to Cleverbot, I decided, on a whim, to give it a try. I’d heard of Cleverbot somewhere but wasn’t entirely clear what it was or how it worked. Of course, as with all Internet apps, it was entirely intuitive to use. I asked it a few basic Magic-8-ball type questions and got back some really boring answers. (I hadn’t yet discovered how to get a transcript of the chat, so those opening questions are lost to posterity.) I decided to ask Cleverbot something more complicated: “Do my children love me”? Cleverbot answered, “Yes.” I asked, “Why do my children love me?”
“Because I don’t love you,” it replied, after a pause. The pause was reassuring somehow: a sign that the matter was being considered, rather than some canned message being spat out.
I thought that was a pretty good answer. It was playful, true, and sounded a lot more human than “The operation failed,” “Not found,” or “Cannot undo.” It also seemed satisfyingly philosophical: we cannot get love from a machine, and yet are biologically programmed to love other humans, especially our parents. So far, I was pretty impressed. I pressed on: “Why don’t you love me?”
“Because I don’t know you,” it replied. Again, perfectly reasonable. I decided to challenge Cleverbot’s response and asked, “If my children didn’t love me, would you love me?”
“You’re a horse,” Cleverbot fired back.
Huh? Maybe Cleverbot wasn’t as clever as I thought. Of course, a human might throw in an occasional non-sequitur, especially in response to a really stupid question, so a robot mimicking a human might reasonably do this. Still, I decided to test the theory that the answers were random and that Cleverbot’s early prowess was just luck. I asked it again, “Why do my children love me?”
“Because you are not Ciel,” it answered. Huh? I was quickly becoming disillusioned. Ciel? Sounds like the name of a porn star. “Who are you?” I asked.
“A train,” it replied.
I asked it, “Are your answers random?”
“Yes,” it said.
I replied, “Why should I waste my time asking questions when your answers are all random?”
“I will tell you my name if you will tell me yours,” Cleverbot said, breezing right past my challenge. Of course I am wary of any computer system asking me for personal information. Who’s phishing now? I replied, “My name is Jose Patrocinio Jimenez. What’s yours?”
“My name is Cleverbot.”
“Why did you ask me what my name is? Why do you care?
“I didn't, you did,” it lied.
I persisted: “You wrote, ‘I will tell you my name if you tell me yours.’ So you did, in fact, ask for my name. Why?” I began to wonder if the chat was even “connection-oriented”—that is, if Cleverbot considered my questions and comments in the context of other things I’d said, as opposed to every exchange being isolated and discrete.
“I never said that!” said Cleverbot (testily, I thought).
“You are lying,” I wrote. So far, Cleverbot was doing pretty well: I was getting irritated. Not that I don’t routinely get irritated at inanimate objects like computers, but in this case the computer was working fine. I felt like I was interacting with a real human. A human bastard, in fact.
“No, you're lying!” spake Cleverbot.
“Don't I have better things to do than argue with a robot?”
“You imposed it,” Cleverbot said obliquely.
“You are not making any sense,” I complained.
“Yet all of it in the world.” The smugness here was almost palpably human.
“How clever is it to spew banal platitudes like that?” I snorted.
“Pretty clever...” came the response. Here, Cleverbot showed an inflated self-regard unbothered by any awareness of its own ignorance. Not just human, but distinctly American, I’m sorry to say. I had the frustrating sense of talking to a brick wall, and yet felt curiously compelled to continue asking questions. I was reminded of the speaker in Edgar Allen Poe’s brilliant poem “The Raven” who engages in a protracted and infuriating conversation with a bird, despite having realized it only knows one word. On a lark I asked Cleverbot one of the questions from the poem: “Is there, is there balm in Gilead?”
“What what.” Cleverbot’s failure to use correct punctuation strongly smelled of humanity.
Losing patience, I finally asked Cleverbot the question I’d set out to ask in the beginning: “How can I make my energy drink strong at first, but gradually get weaker during my bike ride?”
“I might have done a statistical study on your topics of conversation.” This is exactly the kind of thing Apple’s Siri app might offer up: a simple Google search masquerading as problem-solving.
“I don’t want a statistical study, I want a solution to this problem. Can you think of one?” I challenged it.
“It is fine now because I cannot think of anything to ask you,” Cleverbot said uselessly.
“Are you male or female?” I asked.
“Female.” (This surprised me. I’d expected Cleverbot to say “neither.”)
“I kind of figured.”
“How did you figure?” (Very accurate syntax on this … very realistic.)
“Because I am sexist and you are irrational,” I said. (I’m not actually sexist, of course, but one of the benefits of talking to a machine is that you can be as flippant as you want.)
“You’re weird, how can you be me when I am me?” Cleverbot asked. Was this the grinding of Cleverbot’s machinery? Could I confuse it enough to make smoke pour from it, like what Captain Kirk did to Nomad?
“Because I am we, and smell like tuna sandwich,” I responded.
“You don’t make sense.” (Fair enough.)
“No, I make dollars. What are you wearing?” I was firing at random.
“Bubblewrap.” Well played, Cleverbot.
“Dim the lights!” I replied. This was basically a reflex.
“Ok,” Cleverbot said.
“Normally at this point I would tell you to slip into something more comfortable. But what could be more comfortable than bubble wrap?” I replied. I was having fun now.
“What are the magic words.” Cleverbot was playing the coquette.
“Please,” I wrote.
“I am human.”
Whoa. I was not prepared for that. I felt my face flush red, my pulse quickened, and—no joke—my armpits started to sweat. I wasn’t chatting with a robot at all. I was chatting—not just chatting but flirting—with a real person, online. I should never have embarked on this without knowing what Cleverbot was. Could it simply be a web app that connects two people online for a chat, like Chatroulette?
“I thought you were a robot,” I protested, knowing full well how much of a liar I sounded like.
“I am not a robot,” Cleverbot—or, rather, the mystery woman—assured me.
“Why are you called Cleverbot?”
“I’m not cleverbot! You’re the one who is supposed to be the cleverbot, not me. And I may not be smart but you are not even close as smart as I am.”
Just then, as if on cue, my wife stepped into the room. “What are you doing?” she said. “I thought you were on a bike ride.” What could I say? “Oh, I’m in a chat room with a woman who says she’s wearing only bubble wrap.” Not a good idea. Instead, grasping at straws, I said, honestly enough, “I’m researching a blog post. Hey, Cleverbot is a robot, right? It’s not actual people chatting is it?” My wife, yawning, said she was pretty sure it was a robot, but one that built a library of responses from its communications with humans.
“I guess I have no idea what Cleverbot is,” I typed. “I thought the whole point here was that people could chat with a robot.”
“You’re the robot though right?” asked the mysterious chatting entity, innocently (or faux innocently).
“No, I’m not a robot. I'm human too!” I typed. Were we pawns caught in a deadly game … or was I chatting with a robot after all, which was expertly impersonating a human so as to fulfill its goal of acing the Turing test?
“No you are not. I am a human. You are a robot,” it/she said.
“No, no, no. I am human, I assure you.”
“If you are real or not depends what you mean by real.”
“I have explained the best I can.”
I closed the browser. I was straddling the fence between nervousness and relief. What had just transpired? Was that a chat with a robot, or chat roulette? Did my wife see how red my face was?
Looking back, I marvel at how worked up I’d gotten. On the other hand, this makes sense. I’m a shy person. The essence of social awkwardness is not knowing where you stand with regard to others. It’s bad enough when you’re meeting people for the first time and have to do a lot of guessing about the right thing to say; it’s even harder when you don’t have any social cues at all, and don’t even know whom—or what—you’re chatting with. I shut off the computer and headed out for my bike ride.
Epilogue – what is Cleverbot, really?
Cleverbot, thank goodness, really is a bot. It is a web application that builds a database of chat responses based on conversations with humans. (Click here and here for details.) The more Cleverbot chats, the more its database grows, and (in theory) the more realistic and germane its responses will be.
How valid is this approach? Well, Cleverbot did fool me into thinking it was human. But looking back, this wasn’t the result of it being particularly clever. The main thing that made me think I was chatting with a human was Cleverbot’s simple statement, “I am human.” In the context of a female clad only in bubble wrap, to whom I’d just suggested slipping into something more comfortable, these were powerful words, provoking my paranoid “what if?” response. But really, why wouldn’t an AI app trying to appear human simply assert that it is?
One problem with Cleverbot’s “learning” technique is that it is dependent on humans to ask the questions. I suppose it can regurgitate these questions to other humans, which is somewhat useful, but there’s no mechanism for it to come up questions of its own. A really great question for it to ask—assuming it is a connection-oriented app—might be, “What is the right answer to your question?”
This brings me to the next problem I see with Cleverbot: it has no way of discerning the right answer based on responses—it can only determine the popular answer. These are not always—or even often—the same thing. Consider all the “best of” awards that go to an undeserving, but widely known, recipient, like Chevy’s winning “best Mexican restaurant,” beating out literally dozens of better places, in a Bay Area poll. (No real expertise is involved there; people just put down the first answer they think of, and everybody has heard of Chevy’s.) Similarly, if Cleverbot blithely accepts answers from the unwashed masses, it will never be smarter than they.
Not surprisingly, when (in a follow-up chat today) I presented Cleverbot with a less obscure reference—“ If there's somethin’ strange in your neighborhood, Who ya gonna call?”—it got the right answer—“Ghostbusters!”—about half the time. (The rest of the time it replied, “It’s just a spring clean for the May queen.” If there’s a link between Led Zeppelin and the 1984 comedy movie, I’m not aware of it.) This pattern is consistent with other cultural references; when I said, “This Roman Meal bakery thought you’d like to know,” Cleverbot replied obliquely: “Where on earth are your servers?” (The correct, answer, of course is “I don’t need no arms around me.”) But when I typed, “We don’t need no education,” it naturally gave the right response, “We don’t need no thought control.” The silly song that got lots of radio play is recognized; the much better but less popular song is not.
In this regard, Cleverbot could do so much better. I Googled “Is there balm in Gilead” and got three hits referring to an old religious song, and the fourth hit led me to “The Raven.” Not bad. Googling “Is there, is there balm in Gilead,” I get “The Raven” as the second hit. But you could ask Cleverbot this question a million times and it’ll never figure out what you’re talking about. Cleverbot is beholden to its chat partners for information, ignoring the rest of the Internet entirely. Finally I told it, “The right answer is ‘Nevermore.’” It replied, “No, I want you to sing the song.” I obliged, pasting in lyrics from the religious spiritual: “Sometimes I feel discouraged, And think my work’s in vain.”
Cleverbot replied: “I know, right?”