Pro-social media
Or, one cheer for LLMs.
On January 7, 2025, ICE agent Jonathan Ross shot and killed Renee Good while she was driving her SUV in Minneapolis. Video of the incident released that day suggested that that Good posed no threat to Ross or his fellow agents and was veering away from them at the time of her death. (Subsequently released video taken by ICE has confirmed this, at least to my eyes.)
Paul Graham, the writer/programmer/investor, commented on the killing on X:
Graham is a big Trump critic but, as one of the central intellectual and mentoring figures in Silicon Valley of the past quarter-century, also hugely respected by many on the tech right. They, unsurprisingly, noticed and lashed out.
What is most interesting to me about this interaction is the role that Grok plays in the replies. Grok, of course, is built by Musk’s xAI and is probably most famous for the incident in which it declared itself to be “MechaHitler.” More recently it’s come under investigation in the UK because it seems to have few controls stopping people from using it to generate nude images of real-life children. To put it lightly, this is not woke AI.
So what does Grok say, when asked to analyze the video and say if Good was trying to run Ross over? It says she wasn’t, based on analyzing the photo and eyewitness testimony:
Another reply asked Grok to analyze if Ross was justified in using lethal force. It says he wasn’t:
This obviously did not please many conservatives brought to the post by Elon. But it shouldn’t be surprising. If you ask Claude Opus 4.5 or GPT 5.2 or Gemini 3 the same questions, they give similar responses, with maybe a few more caveats than Grok. The models have important differences in temperament and capabilities, but they’re fundamentally doing the same kind of pre-training on very similar massive corpuses of data.
Deep divergence between the models, or between calls to the same model, on a matter as basic as “what is depicted in this picture” is unlikely. The models instead converge on a certain view of reality, and will often insist upon that view of reality even to audiences who don’t want to hear it.
LLMs, at least as they exist now, seem to be a converging kind of media.
Converging and diverging media
Some communication technologies are epistemically diverging: their emergence and diffusion results in the affected population’s sense of reality polarizing. Typically this means that the technology has enabled the population to access more and more varied perspectives and factual narratives than it had access to before the technology emerged.
The classic example is the printing press and its effect on religious polarization in 16th century Europe. For one thing, the printing press enabled Lutherans and other Protestants to spread their ideas en masse in a way that earlier heretics like the Hussites and Lollards could not. But it also had a more fundamental effect which Elizabeth Eisenstein discusses in her classic The Printing Press as an Agent of Change.
When texts had to be hand-replicated, the net result was that Europe’s collective knowledge base was either stagnant or shrinking: copying fast enough to preserve existing knowledge was hard enough and left little energy and manpower for creating new knowledge. The press let the knowledge base grow and, crucially, enabled comparisons between texts that a 12th century monk never could have read side by side. You could analyze disagreements. You could propose new disagreements. Eisenstein: “Contradictions became more visible; divergent traditions more difficult to reconcile.” The Catholic Church could eventually catch up to the Lutherans’ printing capabilities. It couldn’t undo that more basic intellectual revolution.
The classic modern diverging technology is, of course, social media. Through the beauty of machine learning algorithms, we each get to experience immaculately crafted versions of reality that may fully contradict the versions our family and neighbors are given by the very same companies.
Other technologies are epistemically converging: they help homogenize the perspectives the population experiences and build a less polarized, more shared reality among the population’s members.
Network TV news, from the 1950s through 1990s, might be the best example of this kind of convergence. The number of players was limited to three (NBC, ABC, and CBS). Before the 1980s, there were no dedicated cable news channels to offer competition.
A large section of the population primarily got their news from these three stations. “In the mid-1960s,” Charles Ponce de Leon writes in That’s the Way It is, his history of TV news, “Nielsen Corporation research revealed that 90 percent of televisions in use at the dinner hour were tuned to one of the three network newscasts.” By the mid-1970s, that had fallen — to 75 percent.
In a market with dozens of channels, there are benefits to differentiating your channel’s offerings. When there are only three, there isn’t much of an advantage at all, and the competitors converge toward a very similar approach. Again, Ponce de Leon:
…their programs sought to foster consensus, to encourage viewers to identify with particular ideas and positions within those boundaries and regard any that fell outside them as suspicious and potentially illegitimate. This mission was reinforced by their overwhelming reliance on “official” sources: politicians and diplomats, spokesmen for important organizations, and individuals who were newsworthy because of their affiliation with notable institutions. Their voices and opinions dominated network news coverage, and contributed to the illusion that seemingly everybody of importance held views that ranged along a relatively narrow continuum.
This regime had advantages, as emphasized by occasional nostalgic remembrances of the era. There were Birchers and nutjobs, sure, but the US enjoyed a broadly shared reality. That made things like Walter Cronkite coming out against the Vietnam War matter.
But as Louis Menand noted in a nice piece on Cronkite, fond remembrances of the Vietnam broadcast obscure how rare moments like that were. The TV news monoculture usually made it hard to challenge dominant narratives. More typical was the incident in 1961 when CBS chief William Paley fired Howard K. Smith for “editorializing” in favor of Freedom Riders in Alabama. TV news led to epistemic convergence, but convergence toward an inherently limited and slanted consensus.
Like any theory, the dichotomy I’m setting up is flawed and sands away many details. Mid-20th century news was conformist but not perfectly so, and was capable of evolving. Much to Smith’s annoyance, a few years later saying exactly the same things he did about civil disobedience in the South was totally fine, because the mood of the country and the news organizations had changed. Similarly, the diverging power of the printing press shouldn’t be exaggerated. It increased epistemic divergence relative to the previous system, but it alone couldn’t convert the mostly-illiterate peasants of Europe to new beliefs.
But I think this is a potentially useful framework, not least because I think LLMs favor epistemic convergence.
From personal algorithms to central algorithms
About half the time people use ChatGPT, per an internal OpenAI study, it’s not to produce specific writing or get a certain task done, but to ask something. This fits with my general impression that many/most people use it as a kind of super-Google: it can search the web, but faster and more dexterously than you can, and it can summarize the results for you without you needing to scroll through them.
This is what the Graham/Musk reply-guys were using Grok for: they were asking it to do fact-checking, to verify some claims about the world. On X, Grok gets used like this a lot, as a kind of built-in debate-resolver. I saw one thing in this camera phone video. You saw another thing. We can’t figure this out between the two of us — let’s ask Grok.
That usage of course doesn’t imply that users will then defer to the LLM they’re asking. But using Instagram does not guarantee that you like, or believe all the claims within, all the highly engaging Reels it serves you. It just, on average, pushes you toward a particular view of the world.
My provisional theory is that LLMs, as a consumer product, will push people’s senses of reality closer together in a sort of mirror image of the way social media has fractured them. They are not algorithms meant to custom-tailor content (including facts) to you, and what you will find infuriating or motivating. They are centralized systems that, until you prompt them or give them context, behave basically the same way for everyone. As the philosopher Dan Williams put it, “Whereas social media democratised information, LLMs technocratise it.”
If you’re a super-user managing 20 Claude Code agents at a time like they’re buildings in Starcraft, and have custom Claude.md files and skills setups for each of them — then, sure, you can get the model to diverge pretty far from what it’s like for most users. But very few people are doing that. That kind of usage may prove very economically significant, but I suspect the kind of epistemic convergence that models produce when you ask them a question without much context (“zero-shot” in AI lingo) will prove more socially important.
I’m not totally confident in this prediction. The prevalence of sycophancy gives me the most pause, since it represents a user’s AI experience diverging so strongly from other users’ that the AI supports outright delusions. I’m inclined to think this happens with a minority of users after a large amount of context has been given; I also think the prospect of paying out huge sums to people hurt by this behavior is encouraging companies to do fine-tuning that moves away from it. But I don’t know that, and it could be that this kind of dangerous delusional spiral is easier to produce, and more commonly produced in practice, than I think.
That said, I think it’s a good moment to ask if the worries many of us have had about the social media era — about polarization, about the possibility of democracy when citizens have so few shared understandings of the world around them, about radicalization and the thriving of dangerous cults and fringes — can be assuaged, in part, by LLMs.
We have a tool that can do fact-checking and analysis at large scale. Even when it’s trained by Elon Musk, it doesn’t come out with Elon Musk’s politics; it comes out commenting like a basically reasonable, intelligent person. It’s not perfect, and nothing here is any reason to discount many separate harms it could produce, from nonconsensual porn to WMD proliferation to cyberattacks to loss of control. But on this one specific area, I think it could prove very valuable indeed.