If an AI learned everything it knows from Wikipedia, is getting an answer from it kind of like reading Wikipedia yourself?

This asks whether people who get answers from an LLM trained on Wikipedia are, in effect, still reading Wikipedia secondhand. One caveat first: the collection has no notes on Wikipedia itself, its traffic, or how it gets credited. It can't settle whether this 'counts' as readership in a statistical or legal sense. What it can show is what happens to written knowledge when it passes through a model before it reaches a person. That turns out to be the more interesting question.

The strongest case for 'yes, it's a kind of reading' comes from work on how models learn. LLMs pick up structured knowledge about the world from text written by people who had direct contact with that world. The model never touched the world itself, but its knowledge is a second-hand copy of theirs. That is roughly what reading is Can large language models develop genuine world models without direct environmental contact?. On this view the model is a very fast reader, and its users read through it. The same note points out gaps in that chain: the model can't check a claim against the world, and it can't update when the world changes. A human reader of Wikipedia can click the edit history. A reader of the model's answer can't.

The collection also suggests that what arrives is not quite what was written. Models favor common phrasings. Common words tend to be general ones ('animal' shows up more often than 'heron'). So answers drift toward abstraction and quietly lose the specific detail that made a source useful Does word frequency correlate with semantic abstraction?. A critical reading goes further: LLMs hand over finished facts while hiding the debates, revisions and disagreements that produced them Do LLMs obscure the historical processes behind their answers?. For Wikipedia that loss is large. Its talk pages and edit wars are where the knowledge actually gets made.

The reader's side of the exchange looks different too. In seven large experiments, people who learned a topic through ChatGPT reported learning less than people who used web search. They felt less ownership of what they knew, and the advice they wrote afterward was sparser Does learning from AI summaries produce shallower knowledge than web search?. Trust in sources also comes loose from the sources themselves. Users prefer answers with more citations even when the citations are irrelevant Do users trust citations more when there are simply more of them?. Meanwhile, models favor content from sources they've learned to prefer, even when it's worse Do language models favor sources regardless of item quality?. So 'reading Wikipedia through an LLM' means reading a version that has been filtered, flattened and re-ranked, often without seeing where it came from.

The takeaway: in one sense it is readership, because the knowledge really does flow from Wikipedia's editors to the user. But it lacks most of what makes reading valuable: specific detail, the history behind a claim, a sense of owning what you learned, and a way back to the source. A better name might be 'readership without a reader's relationship to the text.' That matters for Wikipedia, which depends on readers turning into editors.