IBM's Debater AI Is Impressive, But It Won't Conquer Humanity

We're living in a strange time, when it's increasingly difficult to distinguish a human from artificial intelligence. And, yes, it can be pretty uncanny. One recent example: Google's demonstration of its Duplex technology, in which the software fooled an innocent (human) hairdresser into believing she was talking to a human on the phone.

And now, IBM has taken things a step further by putting on a human-robot debate, setting its brand new Project Debater against Israeli professional debater Dan Zafrir.

The demo was impressive. The AI argued, brought up supporting facts from "several hundred million articles," and even peppered in a couple of jokes and passive-aggressive jabs at its human opponent. It was like high school debate class all over again, but with a robot.

 

People, naturally, found IBM's demo "unsettling," likely because it could argue with humans "pretty convincingly."

But it's important to keep in mind that the IBM demo, like that of Google's Duplex, was a demo. Duplex might have made the call without a hitch, but Google was clear that this was a successful instance out of many previously failed attempts. How, then, would the feature function in the real world?

Similarly, IBM's Debater AI demo went by without any major hiccups. But it did get caught up in some of its responses. It threw in some generalized comments that didn't directly respond to the human debater's arguments. But its creators had a simple answer for that: it's a human(ish) thing. "If it’s less confident,” VP for IBM research Jeff Welser tells the Verge, “it’ll do its best to make an argument that’ll be convincing as an argument even if it doesn’t exactly answer that point — which is exactly what a human does, too, sometimes.”

But the AI didn't debate like a human. It wasn't making up arguments on the fly — the Debater took sentences it found from supporting documents, strung them together, topped them with some human flourishes, and passed them off as a debate style comparable to that of a human.  The real thing, a real human argument, would have been a lot more precise and nuanced.

And IBM wasn't aiming to convince us that its AI was human in the first place. Its goal was to demonstrate that AI could someday help humans work through their flawed, biased, emotionally-influenced reasoning to find evidence-based solutions to problems of society and in new fields of research (this logic itself may be a bit faulty, since the Debater was trained on faulty human arguments). "Can an AI expand a human mind?" the project's website ponders mysteriously.

IBM's demo shows that AI has come a long way towards speaking like a human. It's impressive if you hear it. But synthesizing human speech and logical thought processes is nowhere near as complicated (or scary) as self-awareness or self-volition. For now at least, AI's abilities are specific to each task; general AI that can do many tasks — that can think the way a human can — is still a ways off.

Yes, AI much better than humans at making sense of colossal amounts of data. It can beat us at the most difficult games we invent for ourselves, it can fool us into thinking videos are real when nothing of the sort ever actually happened. But a 20 minute debate with a human doesn't necessarily put AI in a position to pass the Turing test. You can relax — the singularity isn't happening quite yet.