Listening to “Retention,” Part 2

Note: I’m discussing the origins of “Retention,” the episode that I wrote for the audio science fiction anthology series The Outer Reach. It’s available for streaming here on the Howl podcast network, and you can get a free month of access by using the promotional code REACH.

Yesterday, I described how I came up with the premise of “Retention,” which consists entirely of a recording of a telephone call in which a customer tries to get a service representative from a cable company to cancel his account. The primary difference between this conversation and the equally infuriating ones that happen every day is that it takes place at some undefined point in the future, which immediately opened up two possible avenues of exploration. First, there was the question of what kind of services are being provided. It wasn’t hard to imagine a scenario in which companies like Comcast have taken over every aspect of our lives, with algorithms automatically delivering content and services based on our user profiles, until we’re both totally dependent on the system and increasingly isolated from others. This wasn’t the most original idea in the world—Black Mirror, among others, has mined this territory before—but I had fun with it. The other question was who, exactly, was having this conversation. I had the notion early on that the service representative was actually a chatbot, or an artificial intelligence programmed to mimic a real person, which is one of the first possibilities that would have occurred to anybody. (In the finished episode, there isn’t much of an attempt to hide it.) But I also wondered if it might be possible to have the customer be an artificial intelligence as well: an emergent property of the system that accidentally arose from the client’s preferences and search history. The original customer might not even exist. And at that point, I knew that I had a story.

I’ve been interested in the concept of the Turing Test and conversations with artificial intelligences for a long time, ever since reading the hilarious dialogue in Metamagical Themas between Douglas R. Hofstadter and what claimed to be a computer program named Nicolai. (If you’ve never read it before, please drop everything and check it out here—just search for the words “Post Scriptum.”) Two of the most famous early chatbots were ELIZA, which was meant to simulate a psychiatrist, and PARRY, a simulation of a patient with schizophrenia. These personalities were chosen because they offered limited, stereotyped ways of speaking that would be easier to imitate, but they also complemented each other in obvious ways, so it’s no surprise that someone decided to see what happened when the two programs spoke to each other, which they did in 1972. Here’s a short excerpt from the full conversation:

ELIZA: What else would you like to discuss?
PARRY: I am getting tired of answering the same questions over and over.
ELIZA: What resemblance do you see?
PARRY: Why do you need to know that?
ELIZA: What do you think?
PARRY: I don’t understand your motives.
ELIZA: Why don’t you understand my motives?

Under the circumstances, it’s not surprising that neither program was at its best: each one depended on the use of canned phrases, repetition, and leading questions to keep the conversation going, which didn’t give them much to work with in the absence of a human partner. But I’ve never forgotten the idea of two chatbots talking to each other. The final transcript of the conversation between ELIZA and PARRY runs about six pages, but in theory, it could have continued endlessly, with the two programs going back and forth to this very day. (It isn’t the only time that two bots have faced off, either: earlier this year, an interminable conversation between two Google Homes named Vladimir and Estragon led to a brief flurry of interest online.) It seemed to me that I could write something poignant and creepy about a customer service call that never ended because it inadvertently paired off two artificial intelligences. One would be programmed to retain the client at all costs; the other would be determined to cancel its account. Their objectives would be clearly defined and mutually inconsistent, and as long as neither one of them crashed, they would keep arguing forever. In the finished script, my two characters, whom I named Lisa and Perry in a nod to their predecessors, wind up talking for something like eight hundred years, and the episode fades out before they finish. I wasn’t quite sure what the tone of the resulting story would be, but I assumed that one would emerge naturally if I laid all the pieces end to end.

At this point, I had a decent spine for the script, and the rest amounted to a series of technical challenges. The story fell logically into several segments, each one of which was defined by a twist. There was Perry’s desire to cancel his account and Lisa’s reluctance to allow it; the gradual disclosure of the extent to which the system has taken over its clients’ lives; the reveal that Lisa is a bot and, later, that Perry is one, too; and the revelation at the end that they’ve been talking for centuries. Writing it became a matter of figuring out how to convey this information organically over the course of the conversation. Fortunately, I had a lot of material. I listened to the recording of Ryan Block’s conversation with Comcast several times and noted down scraps of potential dialogue. Even better, I found a copy of the Comcast quality guidelines for customer retention, which amounts to a script that representatives are supposed to follow, complete with canned phrases—“I understand that your needs have changed and you want your services to reflect that”—that wouldn’t be out of place in ELIZA’s programming. The rest practically wrote itself. What I liked the most about it is that it fulfilled my original goal, which was to write a story that was inseparable from the audio format. Not only does it lend itself to a podcast, but it wouldn’t be possible in any other form: the two characters are both disembodied intelligences, but you aren’t supposed to realize this right away, and I can’t think of another medium in which it would work. Tomorrow, I’ll talk about what happened after I delivered the script to my producer, and how I feel about the finished episode itself.