Canta y no llores

Had a trifecta of AI encounters yesterday. 

First was an article from Axios (a news service which somehow got into my email feed), predicting that in the near future, the internet will have less and less direct human participation. Instead it will be "bot versus bot," as online venders will have AI bots to customize the prices they offer to individual customers based on their past internet interactions, and hapless customers' best defense will be deputing their own AI "agent" to foil the vendors' bots. Soon websites will be designed mainly for "agents," not human readers. And not just in e-commerce.

This kind of spiraling digital arms race is most familiar today in the realms of electronic trading and cybersecurity, where offense and defense have long played a "see if you can top this" game. 

The same brutal competitive dynamics are about to spread everywhere — to job applications and classrooms, dating apps and customer service, coding helpers and scientific research.

Classrooms?

My next was an article written by Marisol Aveline Delarosa, one of our grad students in creative writing, for Public Seminar, called "Who Does My Algorithm Think I Am?" Bemused by the ads internet algorithms were sending her, Delarosa fed a bunch of these products into ChatGPT and asked what sort of person would be interested in those things. The surprisingly detailed profile generated was off but in some details uncannily right. She hen fed ChatGPT's profile into an image generator. The resulting rather glossy images (whom she named "Alex") looked, Delarosa found, nothing like her - but a lot like what she hoped to look like when she was a teenager.

I was a bit irritated with myself that I worked so hard to get to a conclusion that I suspected all along: that maybe I am not so unique and unpredictable, and in fact easily slot into several consumer demographics. Was I really expecting all of this data manipulation would ultimately provide me with an actual picture of myself? And what does it mean that my targeted ad data produced the image of a woman I longed to be when I was a teenager but who now seems like sort of an avatar for bland perfection? ... 

To be fair to the algorithm, I spend more time with me than anyone else and I’m still trying to figure myself out. Interacting with machine learning has only made me more certain that if I have any charm that Alex doesn’t, it comes from a place of capriciousness. ... 

I often take myself out to different bars and restaurants to figure out my favorite drinks and dishes because people are always asking me for recommendations. But I also do these things to get to know myself and the city I choose as my home. 

The other day I went for a massage and a glass of wine, ... I am the only person that really knows me, and I have a generally good relationship with her. She has her idiosyncrasies and I pay attention to them. I always try to find her a seat in her favorite spot, at the bar or in the corner with a view of the exit. I’m happy to sit there with her in long, seemingly pointless periods of silence that busy, perfect Alex would never understand. 

The third hit closer to home. On Wednesday, our Faculty Center for Innovation, Collaboration and Support sent a welcome for the new year, one section of which was advice on how to devise an AI policy for our syllabi. The examples ranged from a total ban to actively working with AI tools with students. But none of them prepared me for Friday's news from the Provost's Office that we all now have access to a bevvy of generative AI services through the Google Workspace for Education.

While addressed to faculty, the descriptions of the services offered were clearly directed at students. One "can be used to create new content and images, help brainstorm ideas, draft or summarize writing, explore complex topics, and assist with research and creative work." I was curious especially about NotebookLM, which I'd read some about before.

NotebookLM is a personalized research assistant that allows you to upload your own sources such as documents, PDFs, YouTube videos, and Google Docs, and interact with them directly. You can ask questions about the uploaded content, receive answers with cited sources, and generate outputs such as summaries, briefing documents, timelines, FAQs, study guides, and audio overviews (podcast-style).

Advocates of AI in our faculty always encourage us to try things out, the closer to what we know and care most about the better, so I opened a NotebookLM account and created a "notebook" into which I fed a half dozen of the readings I'll be giving students in "DIY Religion." I was curious what kind of "summaries" and "study guides" it might offer for these various works, and of course was wondering how they might shape (or replace) reading the texts if students came to this site before looking at the texts. 

But NotebookLM was way ahead of me. Unsolicited, it generated an overview of the whole set of readings (even a pdf it couldn't fully scan)

as well as various sample questions I might ask to go more deeply into the material. These were good, serious questions, and each generated a rich multi-paragraph answer with links back to the texts - specific paragraphs relevant to the topic. Brilliant! But will anyone ever just read an article again, or even its first paragraphs? Will we wittingly or unwittingly start writing for this kind of AI reader, or depute an "agent" of our own who knows better than we how they work to do it for us?!

Also available was an "Audio Overview" and a "Mind Map." But this was just the beginning! By clicking a button I saw I could get a "Video Overview," an "explainer video presented to you by AI," so click I did. "This may take a while," it said, but very few minutes later it offered me a seven-minute presentation on the passel of articles, complete with an argument and slides with highlit quotations, engagingly narrated by a male voice and structured around questions addressed to the audience. It was entitled "The modern quest for certainty," correctly identifying a key concept embedded within one of the articles which I was planning to devote class discussion to, but gave it greater prominence in the whole than I am planning to. So I clicked again, and not very many minutes later I got an engaging eight-minute report called "The Pluralist's Puzzle," which synthesized the material in a way much closer to what I had in mind, beginning and ending with moments from our one first-person text. Much closer. 

Reader, it was really good. Not exactly what I will be saying but if a student gave this presentation I'd be impressed. I think I might need to go for a massage and a glass of wine. 

(Forgive the Spanglish pun of my title; in Spanish it would of course be IA not AI AI AI AI...)