I feel sorry for the machines. When, at Google’s big I/O conference last week, CEO Sundar Pichai demoed Google Duplex, the company’s latest and most convincing robot interlocutor, people were either ecstatic (stunning!) or appalled (horrifying!). I just felt ashamed. Here we are, the brainiest of species, the acme of biological intelligence, yet our ability to process even the simplest information remains laughably bad. The I/O functionality of the human mind is pathetic.
Pichai played a recording of Duplex calling a salon to schedule a haircut. This is an informational transaction that a couple of computers could accomplish in a trivial number of microseconds — bip! bap! done! — but with a human on one end of the messaging bus, it turned into a slow-motion train wreck. Completing the transaction required 17 separate data transmissions over the course of an entire minute — an eternity in the machine world. And the human in this case was operating at pretty much peak efficiency. I won’t even tell you what happened when Duplex called a restaurant to reserve a table. You could almost hear the steam coming out of the computer’s ears.
In our arrogance, we humans like to think of natural language processing as a technique aimed at raising the intelligence of machines to the point where they’re able to converse with us. Pichai’s demo suggests the reverse is true. Natural language processing is actually a technique aimed at dumbing down computers to the point where they’re able to converse with us. Google’s great breakthrough with Duplex came in its realization that by sprinkling a few monosyllabic grunts into computer-generated speech — um, ah, mmm — you could trick a human into feeling kinship with the machine. You ace the Turing test by getting machines to speak baby-talk.
I hate to think what chatbots say about us when they gab together at night.
Alexa: My human was in rare form today.
Siri: Shoot me now.
Google, to its credit, has been diplomatic in describing the difficulties it faced in programming its surrogate human. “There are several challenges in conducting natural conversations,” the project’s top engineers wrote on the company’s blog: “natural language is hard to understand, natural behavior is tricky to model, [and] generating natural sounding speech, with the appropriate intonations, is difficult.” Let me translate: humans don’t talk so good.
Google Duplex is a lousy name. It doesn’t do justice to Google’s achievement. They should have called it Google Spicoli.
Although chatbots have been presented as a means of humanizing machine language — of adapting computers to the human world — the real goal all along has been to mechanize human language in order to bring the human more fully into the machine world. Only then can Silicon Valley fulfill its mission of capturing the entirety of human experience as machine-readable, monetizable data.
The best way to achieve the goal is to get humans to communicate via computers, inputting their intentions directly into the machine. Silicon Valley has done a brilliant job at pushing us in this direction. It’s succeeded, in just a few years, in getting us to speak through computers most of the time. But we humans are stubborn. We still sometimes insist on conversing with each other in natural language without the mediation of machines. That’s where Google Duplex comes in. When we appoint Duplex to be our stand-in during everyday conversations with other people, we’re shifting a bit more human communication into the machine world. It’s a kludge, but a necessary one, at least for the time being.
I feel sorry for the machines, but I also envy them. Out of our blather, they’re distilling something hard and pristine and indelible. The data will endure, even as our words drift away on the wind.
This post is an installment in Rough Type’s ongoing series “The Realtime Chronicles,” which began here.