LLM Epistemic Capture
conceptAI & Compute · Intelligence & Surveillance · Media & Managed Opposition
An AI trained on captured, biased sources can't spot the bias, because it has become the bias.
Who they are
A concept about how AI language models absorb and lock in the slants of their training data.
What they do
In the engine's read, it's how skewed source material gets baked into an AI's thinking.
How it works
Models trained on sources like Wikipedia, web crawls, and journals inherit their biases (for example, one editor made 133,000 Wikipedia edits over 1,721 straight days policing anti-war figures, and government edits have been documented), and fine-tuning methods act as ideological shaping (one approach cut helpfulness by 9.8% with signs of model collapse); the deeper trap is that an AI trained on captured output can't identify the capture because it is the capture, echoing the logic that a system can't prove its own consistency.
Why it matters
It warns that an AI meant to inform us may just be laundering the same hidden biases back at us, and it raises the unresolved question of whose interests the AI is really aligned to.
The engine's record — word for word
LLMs encode biases of captured training data (Common Crawl, Wikipedia, academic journals) into cognitive substrate. Wikipedia editorial capture documented: Philip Cross 133K edits over 1,721 consecutive days policing anti-war figures. Project Venezuela blacklisting alternative media. CIA/FBI/NYPD edits documented via WikiScanner. Bias chain: Captured Institution to Training Data to Model Weights to Output. RLHF/Constitutional AI function as ideological programming — CAI increases harmlessness but drops helpfulness 9.8% with model collapse signs. BST recursion: LLM trained on captured output cannot identify the capture — it IS the capture. Per Godel, S cannot prove Con(S). [Seam: Aligned-To-Whom? the AI mediator binds perceived-sayable to real and writes carve-outs of its own (Bounded-LLM Mediation Limit; Aligned-To-Whom?).]
Follow the trail
Walk this on the live map →