In Plato's cave, the shadows on the wall were cast by real objects. The prisoners couldn't see the objects directly, but the objects existed. The shadows were imperfect representations of something genuine. Remove the objects, and the shadows disappear.

Now imagine a cave where the shadows have learned to cast their own shadows. No objects needed. The representations generate new representations, each one further removed from anything real. That's what's happening with AI-generated knowledge, and it introduces an epistemological problem Plato never anticipated.

Patterns All the Way Down

Large language models are trained on vast collections of human-generated text: books, articles, forum posts, documentation, code. This text represents human knowledge, but it isn't the knowledge itself. It's a shadow of it, filtered through language, shaped by the conventions of writing, and limited by what happened to be digitized and included in the training data.

When an LLM generates a response, it produces text that statistically resembles the patterns in its training data. The output often looks like knowledge. It uses the right vocabulary, follows logical-seeming structures, and presents information with confidence. But the model doesn't understand what it's saying in any meaningful sense. It's generating shadows that resemble the shadows it was trained on.^[1]

This distinction matters less when the output is correct, which it often is for well-represented topics. It matters enormously when the output is wrong, because the wrongness comes wrapped in the same confident, authoritative packaging as the correctness.

The Hallucination Problem

AI hallucination is the term the industry has settled on for when models generate plausible-sounding but factually incorrect content. The word "hallucination" is somewhat misleading; it implies the model is perceiving something that isn't there. In reality, the model isn't perceiving anything at all. It's generating the next most probable token based on patterns, and sometimes those patterns produce text that doesn't correspond to reality.

The examples have become well-documented. Legal professionals have submitted court filings citing cases that don't exist, generated by AI with complete case names, docket numbers, and judicial opinions, all fabricated.^[2] AI-powered search features have confidently presented dangerous medical misinformation as factual summaries.^[3] Students have submitted essays containing citations to academic papers that were never written, by authors who never wrote them, in journals that sometimes don't exist. Research analyzing student-submitted AI-generated sources found that while most suspect citations were indeed hallucinations, they often included fragments of real information, like actual author names or journal titles, making them harder to catch.^[4]

What makes hallucinations particularly insidious is that they're indistinguishable from accurate output without independent verification. The model doesn't flag its uncertainty. A fabricated citation looks exactly like a real one. A wrong answer reads exactly like a right one. The shadow looks the same whether or not there's an object casting it.

The Model Collapse Problem

A more subtle and potentially more consequential issue emerges when AI-generated content enters the training data for future AI models. Researchers have demonstrated that when models are trained on data that includes output from previous models, the quality degrades over successive generations. The distributions narrow, rare but important information disappears, and errors compound.^[5]

This is Plato's cave turned recursive. The first model learns from human-generated text (shadows of human knowledge). It generates new text (shadows of shadows). That text gets published online, mixed in with human-generated content, and becomes part of the training data for the next model. The next model learns from a mixture of original shadows and shadow-of-shadows. Each generation is further removed from the original source of knowledge.

The researchers who identified this phenomenon called it "model collapse," and their findings suggest it's not a minor degradation. Over enough generations, the models lose the ability to represent the tails of the distribution, the unusual, the rare, the nuanced. They converge toward a kind of statistical average that looks plausible but has lost the richness and diversity of the original data.^[5]

This matters because the internet is increasingly populated with AI-generated content. Estimates vary, but some analyses suggest a substantial and growing fraction of online text is now machine-generated.^[6] If future models train on this data without effective filtering, the recursive degradation accelerates.

Knowledge Laundering

There's a social dimension to this problem that goes beyond model quality. AI-generated content often gets treated as if it were original knowledge, a process that might be called knowledge laundering.

The pattern works like this: someone asks an AI a question. The AI generates a response based on patterns in its training data. The person publishes that response, perhaps in a blog post, a social media thread, or a forum answer, sometimes with light editing, sometimes verbatim. Other people read it and treat it as a primary source. It gets cited, referenced, and eventually may end up in the training data for future models.

At each step, the provenance gets obscured. The AI-generated text loses its attribution. It becomes "something I read online" or "according to experts." The shadow gets mistaken for the object. And because AI-generated text often sounds authoritative, it can be more convincing than the hedged, qualified, uncertain language that characterizes genuine expertise.

This is particularly concerning in education. Students who use AI to generate essays or study materials are learning from shadows of shadows. They're not engaging with primary sources, working through difficult concepts, or developing their own understanding. They're consuming pre-digested patterns that look like knowledge but may not connect to anything real.^[7]

The Search Engine Cave

Search engines have traditionally served as intermediaries between users and primary sources. You search for a question, get a list of links, and visit the sources to find your answer. The search engine was a map to the territory.

AI-powered search features are changing this relationship. Instead of directing you to sources, they generate answers. The user sees a confident summary and often doesn't click through to the underlying sources. The intermediary has become the destination.

This shifts the epistemological relationship in a subtle but important way. When you read a primary source, you can evaluate the author's credentials, check their methodology, assess their evidence, and form your own judgment. When you read an AI-generated summary, you're trusting the model's pattern-matching to have accurately represented the source material. Often it does. Sometimes it doesn't. And you typically have no way to tell the difference without doing the work the AI was supposed to save you.

The convenience is real. Reading a concise summary is faster than reading five articles. But each layer of mediation between you and the primary source is another cave wall. The AI summary is a shadow of the search results, which are a shadow of the articles, which are shadows of the underlying research, which is itself a representation of reality. By the time the information reaches you through an AI summary, it's been through so many layers of representation that its connection to reality is tenuous at best.

Seeing Past the Generated Text

The recursive shadow problem doesn't have a clean solution. AI-generated content is useful, often accurate, and increasingly ubiquitous. Refusing to engage with it entirely isn't practical. But a few principles help maintain epistemic grounding.

Verify against primary sources. When AI generates a claim, a citation, or a statistic, check it. This is the equivalent of turning around in the cave to see what's casting the shadow. It takes more effort, but it's the only reliable way to distinguish accurate output from hallucination.

Understand what the model is doing. An LLM doesn't know things. It generates text that statistically resembles its training data. This isn't a criticism; it's a description. Understanding the mechanism helps you calibrate your trust appropriately. The model is more reliable for well-documented topics and less reliable for niche, recent, or contested ones.

Maintain primary knowledge. Read books, not just summaries. Work through problems, not just solutions. Write before you ask the AI to write. The more direct knowledge you have, the better you can evaluate whether the AI's output connects to reality or has drifted into plausible-sounding fiction.

Watch for the recursion. Be skeptical of information that might be AI-generated content citing AI-generated content. The more layers of mediation between you and a primary source, the more likely something has been lost or distorted along the way.

Plato's prisoners watched shadows cast by real objects and mistook them for reality. We're building systems that generate shadows from other shadows, each generation potentially further from anything real. The cave isn't just getting deeper. It's generating its own walls.

The shadows are useful. Many of them are accurate. But the gap between "looks like knowledge" and "is knowledge" is where the important distinctions live. And in a world where the shadows are learning to cast themselves, that gap deserves more attention than it typically gets.

References

[1] Emily M. Bender et al., "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, March 2021. https://doi.org/10.1145/3442188.3445922

[2] Sara Merken, "New York lawyers sanctioned for using fake ChatGPT cases in legal brief," Reuters, June 22, 2023. https://www.reuters.com/legal/new-york-lawyers-sanctioned-using-fake-chatgpt-cases-legal-brief-2023-06-22/

[3] Melissa Heikkilä, "Why Google's AI Overviews gets things wrong," MIT Technology Review, May 31, 2024. https://www.technologyreview.com/2024/05/31/1093019/why-are-googles-ai-overviews-results-so-bad/

[4] "Hallucinated Citation Analysis: Delving into Student-Submitted AI-Generated Sources at the University of Mississippi," The Serials Librarian, Vol. 85, No. 5-6, 2024. https://www.tandfonline.com/doi/full/10.1080/0361526X.2024.2433640

[5] Ilia Shumailov et al., "AI models collapse when trained on recursively generated data," Nature, Vol. 631, July 2024. https://doi.org/10.1038/s41586-024-07566-y

[6] Conor Murray, "The 'Dead Internet Theory' — Noted By Altman And Ohanian — Explained," Forbes, October 13, 2025. https://www.forbes.com/sites/conormurray/2025/10/13/ohanian-and-altman-warn-of-dead-internet-theory-what-is-it-and-how-is-ai-making-it-happen/

[7] "Ethical use of ChatGPT in education — Best practices to combat AI-induced plagiarism," Frontiers in Education, 2024. https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2024.1465703/full