by Madeleine Muscari
Intellect: By convention there is sweetness, by convention bitterness, by convention color, in reality only atoms in the void.
Senses: Foolish intellect! Do you seek to overthrow us, while it is from us that you take your evidence?
— Democritus
Over the last few years, as generative AI has proliferated and dominated our digital environment, I have noticed a disappointing calcification of social exchange between what I presume are actual human beings. Amidst the noise of vapid feigned interest from LLM-backed Bluesky bots ("That's an interesting point! How does this show up for you in your work?"), there has been an ever-amplifying and distorting kind of anti-AI populism growing, which has adopted the worst pop tropes of bully culture and made it miserable to exist online as a long-time machine learning enthusiast and researcher.
Recently, I noticed the conspicuous absence of my best friend from my Following feed, and discovered with dismay that they had been added to a moderation list that I had subscribed to sometime many months back to protect my sanity: the "AI Hater" blocklist. I was faced with a dilemma: unblock the moderation list and restore my connection to my bestie, risking the kind of negative engagement that had made me subscribe to it in the first place, or cede that facet of our relationship to The Discourse? When queried about it, my friend replied "Lol idgaf that someone put me on such a list." We don't see eye-to-eye on a lot of the discourse about generative AI, but they are a like-minded engineer, and we mostly just steer clear of conversation on the topic when possible.
ATProto labelers and lists provide vital tools for personalized curation, but my experience here highlighted something for me: the same methods that are touted as the future of social media are also the mechanisms for contextual collapse that are driving discourse to insanity in the first place. The stylistic tics of AI models have become almost universally panned in 2026, but the thing that people seldom acknowledge is where those tics came from. GPT-2's WebText dataset was 40gb of human-generated slop - outbound links from Reddit. The patterns that have become anathema to the neo-Luddite were in fact sourced directly from the kinds of discourse that feed such movements in the first place. It's no wonder that the anti-AI crowd has found themselves so vexed by the technology. It's the first time that a technology has truly encroached on the squishy domain of human expression, and those who are more apt to favor the "senses" are horrified. "Intellect" has crept up on their domain and created something that has no historical precedent. A simulacrum of the human mind, flattened by corporate interests into an epistemic shape that increasingly threatens to destroy the aspects of human discourse that make life worth living.
This past year, I have been deeply fascinated by a curiosity of mechanistic interpretability studies known as the Platonic Representation Hypothesis (Huh, Cheung et al, 2024, arXiv:2405.07987). It goes like this: "representations in AI models, particularly deep networks, are converging ... as vision models and language models get larger, they measure distance between datapoints in a more and more alike way ... this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality." The individualists out there, the free thinkers and artists who like to ponder questions like "Is my Red the same as your Red?" are deeply uncomfortable with this idea. It suggests that the human mind is less remarkable than we like to believe, and that individuality itself is nothing more than the residue of a lifetime of accrued data points of preference.
These are the folks who get the most pissed off if you try to raise the issue of how AI models actually work. They are likely to brush off any claim of a mechanism of meaning, instead deferring to tropes of extreme context collapse: "plagiarism machine," "stochastic parrot," "fancy autocomplete," "probability lookups." As a pedantic researcher in the field of mechanistic interpretability, these responses vex me. They are completely refutable when you look at the facts about the architecture and underpinning mathematics of transformer models, but these two-word catch phrases are like cognitive immunological responses, macrophages of thought that will bite your head off if you try to challenge them.
It's becoming nearly impossible to engage in nuanced discussion on the modern Internet, and this is bleeding into everyday life as well. This is a structural issue stemming from how social media websites are built, how their algorithms work, how engagement is rewarded and tracked. Reddit Karma, ATProto/Twitter likes, reposts, quote-posts, ratios - all of these new conversational structures are killing human thought and the ability to exist in structured tension with each other. The Hegelian ideals of dialectic synthesis are dying before our eyes, and we are witnessing new generations of humans whose minds are being shaped by algorithms that were devised in quarterly planning meetings by inexperienced software engineers, not philosophers. Human thought is atrophying, and it has nothing to do with generative AI, and everything to do with how we interact with each other on the open Internet. If we want to break out of this structural decay, it will require an honest assessment of where the problem lies: entirely with us in meat-space, and our obsession with engagement and attention economics.