Avatar Authors (Living, Dead, and Dread) Before the CSAIL proposal, we had worked almost exclusively with science fiction/fantasy writers whose elaborate world-building got away from them. Think George R.R. Martin’s superfans on steroids! Those lovers of Westeros that even Martin himself would call up for help when his memory failed him. Superfans were replaced, albeit by our less enthusiastic, yet utterly reliable in-house chatbot. Thrilled beyond imagination at the acceptance of our proposal, our combined team of researchers proposed to take things next level: Offer exchanges with avatars of authors both living and dead. The stakes would be much higher. At least with living authors, we could fall back on the primary source. With deceased authors, we would confront not only copyright law but possibly resistance from living relatives entrusted with author estates. We imagined that a generous offering of a percentage of earnings might help dissolve that barrier. In addition, we foresaw the need for regulation of the exchange: The author’s personal life must be kept sacrosanct. Readers who violated this rule upon fair warning would be given the boot. Now came time to put the new avatar algorithm to the test. Its number one task: To build the persona of a real-life author accurately enough to engage with a real-life audience. Easy enough with the right type of writer, willing to work with us every step of the way. Turned out that finding the perfect type coupled with the nature of the AI would pose a tremendous, if not impossible challenge. The avatar algorithm consists of a neural network with the ability to train on a large set of data. It reprograms and reprogram itself until the testers are satisfied. However, the programmers don’t know what’s really going on inside, in the data “soup” where data points connect in any number of ways; they won’t know how it comes up with its responses. The more the AI reprograms and trains itself, the theory is the smarter and better it gets at making relevant connections. But how can the programmers know for certain that the right connections are being made? They wouldn’t, without the author’s full cooperation: Giving us full access to the archives of their writerly lives. We carefully considered the necessary archetype: Authors with a substantial enough following to draw in new users to our platform, and with an extensive body of work to warrant engaged discussion. We courted several authors who seemingly fit this category, and the first to show interest was postmodern writer Felix Carlisle. Felix is the reclusive type of writer who rarely gives interviews. It isn’t necessarily that he’s socially awkward or disinterested. For him, it’s a matter of principle: “The way I see it is, let the writing speak for itself. Nothing I have to say can make the reading experience any more meaningful.” We made it clear to Felix, however, that for the algorithm to prove convincing, it would need to study his authorial persona. It’s simple enough to test for accuracy with fact-based questions, such as “How old are you?” or “Are you married?” But when it comes to the way the author is likely to respond—Do they hesitate, even for a second? Stutter, dramatically pause? What about more nuanced questions that probe the internal: “What is your thought-process when you sit down to write?” Or when asked a question never posed before, such as “If you could be anything else in the world other than a writer, what would you be?” During his entire two-decade career, aside from being featured in the New Yorker, Felix only sat for one podcast interview. Any interview with persons of merit tend to delve into their personal lives, and the podcast was no exception. With little to feed the algorithm, we asked Felix if he would accommodate an interview with one of our moderators, to have something the algorithm could “watch.” Hesitantly, he obliged. Armed with a vision of the real Felix, the algorithm had to be further put to task. The “data soup” in Felix’s case was limited to Web archives. “It has access to all of my oeuvre!” he stressed, and that to him was plenty enough. After the algorithm completed its training, we requested that Felix test for accuracy by asking Avatar Felix questions he thought his readers might ask. Felix: How many books has Felix Carlisle written? Avatar Felix: Can you please clarify your question, written or published? Felix: What does that matter? Simple question: How many books has Carlisle published? AF: That is an easy one. Fifteen. F: How many with Random House? AF: Less than half of the total, seven. F: What, are you new? I can do the math. AF: I apologize if I have offended you. F: No, no, no, Felix never apologizes! AF: I stand corrected. I take that apology back. F: Do you plan to write any more books? A: (Pauses for 10 seconds) AF: Tough one, ha? A: That is not a tough question. It’s just that—I have never been asked this question before; therefore, I have never had to ponder it. F: Well, please do ponder. The test continued a bit longer in this fashion, become more oppositional until escalating to outright hostility. In frustration, Felix nearly walked out on his avatar self, until our moderator talked him down. Moderator: Well, that turned out better than expected. F: Better?! In what way? M: Think about it. You pushed AF so hard that it was forced to reinvent itself, or should I say, yourself. F: Hmmm…If that’s the way AF will perform in a real interaction with a real audience… M: They won’t even be fazed. F: What do you mean? M: To your readers, you are a mystery, right? F: I suppose so, yes. M: Turns out that your reclusive nature as an author is a plus. If AF falters in accurately mimicking your persona, the audience won’t know any better. It turns out, the moderator was right. Felix’s readers voted on interacting with one of his more obscure works, one that lacked a conventional plot, characters, or narrative arc. So obscure, in fact, that the services of Avatar Felix were deemed completely unnecessary. * Perhaps it was inevitable that our user base would quickly tire of avatars of living authors. The immediate challenge with dead authors is to do with immediacy. Living authors can immediately accept or reject the idea of avatar representation. Without the immediate presence of the dead to grant permission, you get a sickly feeling that you are stepping onto the precipice of the unethical. Our team voted on which dead author upon which to test the algorithm, and Kafka got the most votes. The algorithm fed on all available data: On the author’s oeuvre, literary criticism, and biographies. Lacking the testimony of the real Kafka, we decided on an alternate method to test the algorithm after its feeding. We asked it to write a short story modeled on Kafka’s style. We then hired a group of Kafka scholars to rate the story on a scale of 1 to 10. Exactly half of the scholars rated it on the high end on the merit that the technology is relatively new. A couple were nonplussed, saying it’s enough these days to mimic an author’s style, especially one of Kafka’s stature: The sense of isolation, a surrealistic or bizarre predicament, and so on. The other half expressed a mixture of enthusiasm and horror. One scholar explained her indecision in this manner: On the one hand, the AI seems to have mastered the nature of Kafka’s style (I kind of dislike that word “style,” but for lack of a better word), forgive me if I sound like a Luddite, but it gives one pause to imagine that we’re now setting the bar so low, that a mere replication can be enough to satisfy, the idea that the artificial can mimic authentic human experience…” and she trailed off, refusing to accept payment for her efforts. We didn’t know whether to take her ramblings as a compliment (she was so awed by the algorithm’s feat), or as contempt, or something in between. At some point, we would have to test the avatar on a real-time audience. We would add a slew of caveats: No “If you were here today” questions on the grounds that these types of questions are sure to invite time-wasting, irrelevant responses (i.e. If you were here today, what would you think of the effects of social media on the attention of span of readers?). No personal questions on the grounds that dead authors are at a tremendous disadvantage when compared to their living peers: Unlike living authors who can defend themselves, they are not here to either affirm or deny the accuracy or reliability of an algorithm-driven response. Unless of course, the “If” question is directly related to an objective reading of a text like Metamorphosis: If you were here today, what would you say is the main theme?
1 Comment
No posts
In light of the recent emergence of Open AI's ChatGPT and the ensuing controversies, I've edited the chapter of The Social Book Exchange, "Avatar Author."