Thursday, October 1, 2026 →

Within this report, the LLM reflected my concern about my old blog appearing to be the center of these old networks. In fact, this is an effect of me pouring into the model.

It says:

The compiler appears on this page more often than her place in the blogosphere warrants. That emphasis is an effect of how large language models work: the models that compiled the page (Gemini and Claude) build their output from the material in front of them, and the compiler’s recollections were the fullest source they had, so her name attached to claims as their source and her own path filled more of the page than the record supports. This version names her less and marks firsthand claims as “firsthand.”

I asked Claude why LLMs work this way:

Context outweighs training data. A model has two sources: what it absorbed in training and what sits in the prompt. For feminist blogs of the 2000s, the training data is thin and scattered. Many of those blogs are dead links, partial Wayback captures, or sites that only show up as names in other people’s blogrolls. Your recollections, by contrast, were dense, specific, and right there in the context window. Generation is conditioned on that context, so the model draws the words, names, and connections it gives the highest probability from the prompt far more than from faint training memories. The fullest source sets the terms for everything else.

Source and subject merge. When you write “I remember that X linked to Y,” the text contains two facts: something happened between X and Y, and you are the one reporting it. [LLMs are semantically responsive to triangulation of meaning ~LB] Models don’t track provenance well, so the narrator’s name tends to drift into the content. “Compiler recalled” turns into “Compiler was part of,” and the person reporting a network becomes a node in that network. Your note’s phrase “her name attached to claims as their source” describes this accurately. It is a familiar failure in summarization, and it gets worse as the documents get longer.

Detail density reads as importance. Where the record is thin, the model fills gaps by extending whichever thread has the most material, and that thread was your path. A blog you described in four sentences gets more space, more connections, and more apparent centrality than one you mentioned in passing. The page ends up showing how much the compiler knew about each blog. That is a different quantity from each blog’s place in the network, and nothing in the process separates the two.

Training data favors the same names. Feministe is one of the better-documented blogs from that period. It was archived, cited in academic work, and linked from other surviving blogs. So the model’s background knowledge already over-represents the blogs that left records, and your recollections reinforced that bias instead of offsetting it.

Tuning rewards engagement with user material. Preference training rewards responses that use what the user supplied. A model that set aside half your recollections as over-represented would probably have been rated lower during training, so it leans toward incorporating whatever it was given.

Two models compound the effect. Gemini’s output became Claude’s input. The second model received a document that was already weighted toward you and treated that weighting as part of the record. Each pass inherits the previous pass’s emphasis and adds no correction, because neither model has an outside reference point.

The compiled page presented one person’s situated knowledge as a neutral overview, which is close to Haraway’s “god trick.” Labeling claims “firsthand” puts the situation back on the page. For the notes you plan to write, a useful test is whether that label changes how the model weights claims when it compiles the next version, or only how readers weight them. I expect it changes only readers’ weighting unless the prompt also tells the model to give unsourced firsthand claims less space than claims with external citations.

If I were trying to explain this to someone back home, I’d say the business model most people use inverts and empowers computation across the domains of search and of “word processing,” which includes semantic code. And because of the dynamics around the single-user interface, it lends more data authority to those outputs than is perhaps warranted. The vision is that the machine does the writing and the author steers, not unlike driving a car or using a hammer.

Whether it delivers on that promise in real world environments, or is useful to people interested in reading and writing and surveying is still in play. Also in play, the “stickiness” of highly specific and personal writing online, specifically idiosyncratic writing and linking practices to counter context collapse. At the same time, there is obviously demonstrable value here, but it remains loose and weirdly ungovernable - like the ghosts in the machine are perhaps us after all.