Telemarketers and the Turing Test
In the Turing test, a computer is supposed to convince a human interlocutor that it’s human. In a recent odd phone interaction between Time Washington bureau chief Michael Scherer and Samantha West, a telemarketer, Samantha—who very likely is human—convinced Scherer and colleagues that she is a robot.
To Time reporters Samantha insisted—”with a charming laugh”—that she was a real person. But she was unwilling or unable to answer questions such as “What vegetable is found in tomato soup?” In some cases, when unsure how to respond, she would claim a bad connection.
Samantha was taking calls as well as making them, so the Time reporters called her back several times over the course of an hour and concluded she was definitely a robot because of the “pitch perfect” repetition of her comments. They concluded that she was designed to ask a series of questions about health insurance coverage and ultimately transfer a potential customer to a real person to close the sale.
It seems like a simple case of a deceptive robot claiming to be a real person to help sell health insurance. Alexis C. Madrigal of the Atlantic, on listening to the Time clips, was so impressed with Samantha’s sales prowess that he wondered, “Where could I buy such an interactive voicebot?” He doesn’t say why he wants one—perhaps to sell subscriptions to the Atlantic. In any event, he learned such a voicebot would be called an outbound interactive-voice-response (outbound IVR) system.
Then the trail went cold. With current technology, Madrigal found, outbound IVR supports limited interactivity—it can be used to confirm or reschedule a doctor’s appointment, for example, but as Madrigal explained, it lacks the sophistication to pose as a general-purpose telephone sales person. “The queries that the editors launch at Samantha are pretty complex,” he wrote, “and yet she comes back with an appropriate (if limited) response.”
Madrigal said he contacted several outbound marketing companies, all of whom said they don’t or can’t provide Samantha’s level of interactive voice response, particularly given the limited bandwidth of the PSTN and the fact that latency problems during voice processing would impede the normal flow of conversation. Voice-recognition heavyweight Nuance denied involvement.
So what’s going on? Madrigal’s research led him to this theory: Americans tolerate a person with a non-American accent when they place a call for customer service or support, but they do not like receiving outbound telemarketing calls from people with foreign accents. Consequently, he suggested, Samantha is a real person who understands English but cannot speak it in an accent that would be acceptable to potential American customers. Therefore, she, or he, responds to a prospective customer by using a soundboard to play back one of a limited number of prerecorded phrases and sentences—accounting for the “pitch perfect” repetition the Time editors heard on their calls.
We may never learn for sure. Samantha won’t take my calls, and the website for the company she ostensibly worked for has been taken offline. That’s odd, because a representative of the company insisted to the Time editors that as a condition of his talking to them that Time publish the phone number and website—the representative seemed to want the publicity.
If you’re out there, Samantha, I’d love to chat. But I am not in the market for health insurance.
Visit my blog at bit.ly/N8rmKm for links to the Time and Atlantic articles and to clips of Samantha’s conversation with the reporters.