I am not a fan of large-language model chatbots. I’m enough of a hater, in fact, that I don’t like calling ChatGPT and its LLM-powered kin “AI” — they are a particular class of products of a particular form of machine learning, which guess the correct response to a query as informed by associations between words and phrases in vast volumes of training text. One of the most painful lessons of the last couple of years, I think, is that what I’ve just described turns out not to be anywhere as close to “intelligence” as it appears.

I digress; the distinction is important, but it’s not exactly my point. My point is that it recently dawned on me that the LLM chatbots are the latest iteration of a now multi-decade process of big tech companies trying to fit the whole Internet into a box that they own and control.

The first version of this was search — Google and its less successful competitors wanting to be the one website you went to before you went to any other website, so you would see the ads they’d placed alongside their search results. This seemed like a good bargain, since Google was very good at indexing and ranking the rest of the Internet to help find the external sites most relevant for any given query.

Then came social media, mostly Facebook and original-flavor Twitter. Social networks initially gave you links to other websites, selected by people who you knew personally, or who you followed for their expertise or taste (broadly defined, in both cases). This meant you could learn about things to read or watch on other websites, rather than just finding the things you searched for. It was, like search, initially a pretty good bargain for some ads in the sidebars or even mixed directly into the posts from people you’d chosen to follow.

The original bargain of both search and social sites — ads, alongside links to other things of interest — broke down because both Google and Facebook could show more ads to people who stayed on their platform instead of following links to other sites. So search sites built in “summaries” on frequently-searched topics, up to and including just straight-up showing you information pulled from external sites on the search results page. Social networks tried to get people to publish directly on their platforms (I’m old enough to remember when Facebook was recruiting news organizations to publish articles as Facebook posts), and eventually down-ranked or outright hid external links in favor of posts within the platform (the platform formerly known as Twitter is now infamous for this).

None of this is news to the kind of Extremely Online people who may have read this far, inasmuch as we’ve all watched it happen. It’s paradigmatic enshittification: build something valuable, then make it difficult to leave, and extract as much money as possible from the fact that you’ve drawn a big online audience into the walled garden of your search engine or your social network.

What does it have to do with LLM chatbots? It’s this: they are, or aspire to be, the ultimate walled garden.

The models behind LLM chatbots are trained on as much freely-available text as OpenAI and their peers can obtain. “Training” a LLM summarizes that huge corpus of texts into associations among words and phrases, so the chatbot has a basis to choose a likely-to-be-correct response to any given text input. In other words, the model is an attempt to reduce the training corpus into a single thing that you interact with through their chatbot interface. ChatGPT provides dutiful caveats about checking results with external sources, but the idea of these things is very much that their chatbot can give you all of the information you’d otherwise find in an extended search over the open web. The goal is that you never have to leave the platform.

The value in this, from the perspective of the chatbot purveyors, is not (yet) to sell your attention to ads on the chatbot page or (yikes) somehow folded into the responses. Their business model so far is that you’ll pay to chat with the bot. Which means the fundamental function of the model is to compress the open Internet into a box small enough that it can be sold back to you. Whatever other pie-in-the-sky visions the chatbot makers have for the technology — or wish to sell to investors — this is what the product is right now.

On the one hand, this is antithetical to the reasons I want to use the Internet, as a researcher. I don’t want my sources of information filtered through an algorithm I can’t parse or control — the LLM training — and summarized by the unreliable narrator of a chatbot, any more than I would write a review article using only notes taken by a student assistant. I want to be able to read the actual sources, dig up the earlier works they cite, and root around in the supporting data if I’m so inclined. (But, it occurs to me, this may not seem like such an obvious shortcoming to corporate executives who experience “research” mostly as briefings and memos assembled by their assistants?)

On the other hand, even if the LLM behind a chatbot could perfectly communicate the contents of the training corpus, it would still be an attempt to put that training corpus — the training corpus they scraped from the open Internet — in a box and sell it back to you. After watching the rise and demise of previous attempts to build walls around the garden of the Internet, I don’t know why I’d ever accept that bargain.