On March 20, The New Scientist published an article about ChatNannies, a company that makes chatbots specifically to detect pedophiles in chat rooms. Apparently, ChatNannies’ bots are so good as pretending to be real children that they can lure pedophiles into hitting on them — at which point the bots can also detect whether an adult is trying to seduce them, and report it to the police. As the New Scientist reported:

To converse realistically, ChatNannies analyses the sentences other users type, breaks them down into verb and noun phrases, and compares them with those in sentences it has previously encountered. ChatNannies includes a neural network program that continually builds up knowledge about how people use language, and employs this information to generate more realistic and plausible patterns of responses.

Hot damn. Apparently the bot also scours the Internet on its own to learn about pop culture, the better to pretend to be a cool, with-it pre-teen. If you want an example, check this transcript of the bot’s apparently remarkable abilities here, recorded by MIT student Cameron Marlowe when he convinced the bot’s creator — Jim Wightman — to let him talk to it.

Now, artificial-intelligence fans must be wondering: What the hell is going on? What genius programmed this thing? The ChatNannies bot can i) chat well enough to fool other humans, ii) observe another speaker well enough to detect that it’s an adult hitting on them, and iii) amass new knowledge on its own by scouring the Internet.

God in heaven. i) and ii) are barely, barely doable today, and even so, only in the most primitive, crappy fashion. But iii) is so over-the-top sci-fi impossible that serious A.I. researchers weep when they think about it. Could Chatnannies be real?

Ahem.

A couple of A.I. experts read the New Scientist story, got suspicious, and asked to see the bot in action at Wightman’s house. But when they tried chatting with it, the results were considerably less humanlike. Wightman claimed that it wasn’t working as well because the bot’s full knowledge-base is 20 terabytes in size and wasn’t on-site. But when they quizzed him closely, Wightman seemed pretty sketchy on details of precisely how the ChatNannies bot amasses new knowledge, as one of the skeptics, Andy Pryke, wrote in his own blog posting on the meeting. Things got even more dubious was when Pryke realized that ChatNannies’ conversational gambits were eerily identical to ALICE, the famously cool bot created by Richard Wallace (who I profiled for the New York Times Magazine two years ago). Pryke did a side-by-side comparison:

USER: I like opera

WINTERMUTE : A lot of people like that. Yes Opera is a perfect art form combining all the others.
ALICE: I’m not sure if I like it. Yes Opera is a perfect art form combining all the others.

USER: do you work?

ALICE: Of course I work, do you work buddy?
WINTERMUTE : Of course I work, do you work buddy?

So it looks like Wightman pretty much just reskinned the free, open-source ALICE and claimed it was his own. Then how did Wightman pull off such fabulously realistic chats to fool the press? One explanation is he just faked ‘em, by having a human pretend to be the bot. The New Scientist pulled its original story, and replaced it with another one noting the inconclusive results from Pryke’s visit. If you go to ChatNannies’ site, you’ll see enormous 20-point text claiming “Rest Assured! We are not, and never have been, a hoax” … which is pretty what you’d plead if your technology was, in fact, a hoax.

It’d be a lot easier to achieve true artificial intelligence if we didn’t have to suffer through so much real-life stupidity.

(Thanks to Richard Wallace for pointing this one out!)