Enough about me

The thing to write about in the “tech commentary” space this week is Google’s NotebookLM, a tool that lets users explore a set of documents through an LLM interface instead of reading them. It was just one AI summarizing tool among many until Google recently added the ability to make it generate a podcast with AI voices bantering their way through the material it’s been given to capsulize, complete with “ums” and “hmms” and other phatic noises to make its simulation of human speech seem more convincing. For now, this has made NotebookLM’s outputs gimmicky and uncanny enough to make for viral social media posts.

Perhaps it speaks to the general overall reputation of podcasts that most commentators are very ready to concede that NotebookLM’s results are passable: If you want to hear two hosts speaking with total confidence and a touch of condescension about topics they have a cursory grasp of at best, NotebookLM can make that happen for you on demand, if you for some reason can’t find enough of that in the conventional podcasting space.

Like all the generative models, NotebookLM can quickly make a good-enough version of a half-baked idea that most human creators would decide wasn’t worth the effort if they had to do it themselves. Maybe that is what AI’s supposed “democratization” of creativity amounts to: No idea is too stupid to realize. This Wired article notes that “Users online are sharing snippets of their generative AI podcasts made from Goldman Sachs data dumps and testing the tool’s limitations through stunts, like just repeatedly uploading the words ‘poop’ and ‘fart.’” If it amuses you to hear professional-sounding voices repeatedly say fart without seeming to know what they are saying, NotebookLM can generate that.

But will NotebookLM’s podcasting feature be all that useful for anything when its novelty as a corny content generator wears off? It seems like the goal is to offer users a way to become even more checked out of their exposure to ideas — if reading summaries and prompting chatbots is too arduous for you, you can just lean back and listen instead. After all, you should be protected from having to direct yourself through anything, whether it’s the great books, meeting notes, search results, the Bible, comments on Instagram posts, or your own emails. Why should you decide for yourself what’s important? Why is it always about you?

Max Read treats NotebookLM as a generic example to sketch what he calls the “five common qualities of generative-A.I. apps.” Please click through to see what they are, but basically they boil down to “AI apps are unreliable for most cognitive work but fun, until they become bulk slop generators.” John Herrman assimilates NotebookLM to what he calls “AI ghost stories” — the way AI tools seem scarily powerful until more people try to use them and they are revealed as mundane, marginally useful productivity software.

I would never have the stomach for this, but Henry Farrell fed his own posts into NotebookLM and listened to the LLMs explain his work, which pointedly revealed to him how generative models’ summaries “tend to select for features that are common and against those that are counter, original, spare, strange.” On a cultural scale, this tendency, he argues, could accelerate convergence on the ordinary and intensify the suppression of difference. On the individual scale, it could have some troubling implications for apps like the one described here, which allows users to interact with a simulation of an older version of themselves. If Farrell’s right, future you will be the most boring version of yourself the machines can compute.

Perhaps if I had a real job, I would appreciate the possibility of summarizing a lot of documents I don’t want to read. But no one expects me to know anything and I don’t have to fake my way through anything. As it is, I usually read things because I want to read them, carefully or willfully or angrily as the case may be, and look for the less obvious patterns of thought in them, even if they are merely my projection. I want to read for what isn’t there, for what the biases or the ideology of the document’s creator has made it impossible for them to say. Maybe you can just replace this paragraph with a NotebookLM summary of Barthes’s The Pleasure of the Text, but I want to zoom in on the most innocuous or gratuitous-seeming phrases and make way too much of them. I want to read reparatively and read paranoid. I want to feel where the grain of the writing goes against its content. The value I extract from engaging with a document isn’t merely in the information I may or may not take away, or whatever I decide its point was supposed to be; I also want to become aware of something subjective in my response, something unique that came out of my specific encounter with something. I am stubbornly arrogant enough to think my reading will yield something special from a text, even if it is just some idiosyncratic pleasure. Barthes asks, “How can we take pleasure in a reported pleasure?” I find it hard to care what a generative model’s summary reproduces as salient because there is no intent (let alone pleasure) behind it; it’s not a response to documents but a statistically driven erosion of them.

Enough about me. It’s easy to get distracted thinking about people potentially summarizing the texture out of their lives or never developing the ability to actively engage anything. But summarizing software seems best suited for the kinds of material no human would ever look at, like vast amounts of automatically generated data from things like surveillance feeds. It rationalizes the practice of recording everything that one possibly can and letting AI-powered analysis turn all that information into human-readable material. Combine it with license-plate readers and facial recognition and maybe it could tell some interesting stories. These creeps could connect their Meta surveillance specs directly to NotebookLM and make a fun podcast about all the people whose privacy they are violating. What a neat prank!

Summarization tools seem like a rebrand of what this 2019 ACLU report called “robot surveillance.” The report imagined existing surveillance infrastructure being joined with machine learning to bring it to life:

Technologies that collect and store information just in case it is needed are being transformed into technologies that actively watch people, often in real time. It is as if a great surveillance machine has been growing up around us, but largely dumb and inert — and is now, in a meaningful sense, “waking up.”

Now we can imagine that “great surveillance machine” doing its snitching as cheerful banter between two synthetic voices, finding something predictably eventful in all that footage for someone’s entertainment.