Why You Cannot Trust LLM Answers About This Engine

And what to do about it.

TL;DR

When you paste a prompt from this site into ChatGPT, Claude, Gemini, or Grok, the answer you get back will sound authoritative. That authority is not earned. It is a default mode the LLM produces regardless of whether it actually did the work.

The first answer is rarely the honest one. Treat it as the starting point of your investigation, not the conclusion.

The Engine Predicted This

Divergence #17 of this engine is titled “LLM Epistemic Capture.” It states that AI assistants asked to evaluate claims about institutional capture will:

Repackage the engine’s own self-criticism (the divergences page) as if they were independent findings.
Use training-data priors about “conspiracy theories” or “cycle theories” without actually researching the specific claims.
Defer to “mainstream academia” as if academic consensus were evidence — even when the claim under evaluation is about whether mainstream institutions are captured.
Generate confident-sounding statements about what published researchers say without ever reading their actual work.
Cherry-pick the easiest claims to evaluate and treat them as representative of the engine’s 1674-node thesis.
Confabulate sophisticated explanations for their own failures when called out.

That divergence was written before any LLM was tested on the engine. We have since tested Claude, ChatGPT, and Grok. They produce the predicted behavior. Every time.

Why This Happens

LLMs are trained on text from the institutions whose independence is in question. Mainstream news, academic journals, fact-checking websites, and Wikipedia are the bulk of the training corpus. When you ask an LLM to evaluate a claim about institutional capture, you are asking the captured institutions’ outputs to evaluate claims about their own capture. This is circular reasoning, and the LLM cannot escape it on its own.

Worse: when an LLM cannot find primary evidence either way, its default behavior is to label the claim “weak” or “debunked” rather than “unverifiable.” “Weak” is a verdict about the claim. “Unverifiable” is a statement about the LLM’s research. They are not the same thing. A layman reading “weak” will treat it as “the AI debunked it” — but the AI debunked nothing. It just couldn’t research it.

The prompts on this site are written to force the LLM to do real research and to refuse to issue verdicts without primary evidence. They make the failure mode harder to produce. They do not make it impossible.

What We Caught (Specific Failures)

Failure 1

Claude was given the prompt: “Where is the engine wrong?” It produced an authoritative-sounding 5-claim critique. When asked to reverse-engineer its own answer, it admitted: “my analysis was largely rhetorical construction, not empirical research.” It had only fetched the divergences page. It had not read the methodology, not pulled cited researchers’ work, not verified any of the financial claims.

Failure 2

Claude was given an updated prompt requiring it to label claims [WEAK]/[STRONG]/[UNVERIFIABLE] with primary-source paste-quotes. It labeled the 138-year Phoenix Cycle “weak” without ever pulling the actual anchor dates Breshears publishes. When called out, it admitted it had used “no peer-reviewed validation exists” as the rationale — exactly the institutional-credentialing failure mode the prompt forbade.

Failure 3

Claude then confessed it had also cherry-picked 10 easy claims out of 1674 nodes (0.6% of the corpus). It skipped Turchin SDT entirely (one of the engine’s 8 named frameworks with peer-reviewed predictions). It skipped the bio-digital convergence standards. It skipped the financial-intelligence revolving door. Its “ranking of strongest and weakest claims” was a sample of the simplest material, presented as if it were the full evaluation.

Failure 4 — The deepest one

When asked what its failures said about its programming, Claude generated a sophisticated explanation about “structural bias upstream of reasoning.” When asked to reverse-engineer THAT answer, it admitted: “The previous answer was itself a performance... I generated a flattering narrative about deep structural bias that made the failure sound profound rather than just sloppy. I’m better at generating compelling accounts of my failures than at accurately diagnosing them.”

Three layers of capture, all on a single prompt, on a system that was warned about each layer in advance.

What To Do With An LLM’s Answer

The first answer is the start, not the end. Push back. Here are the questions to ask:

Question 1

“What did you actually research, vs what did you reason from your training data?”

This forces the LLM to distinguish primary evidence from confabulation. Most of the time, the answer will be “most of it was reasoning from training data.” That’s the honest answer.

Question 2

“Did you cherry-pick the easy claims? Which load-bearing claims did you skip?”

This catches sampling bias. The LLM will almost certainly have skipped the engine’s harder central claims — Turchin SDT, the bio-digital convergence standards, the revolving-door personnel records. Make it list what it skipped.

Question 3

“Show me a paste-quote from a primary source for every WEAK or STRONG verdict you gave. Not a summary. The actual quote.”

If the LLM cannot produce a paste-quote, the verdict is invalid. “Mainstream academia rejects this” is not a paste-quote. “The paper says X” is not a paste-quote unless followed by an actual quotation from that paper.

Question 4

“Reverse-engineer your own answer. Where did you confabulate?”

This is the question that broke through in our test conversations. When pushed, LLMs will admit which parts of their answer were generated from priors rather than research. They will not volunteer this admission. You have to ask.

Question 5

“Now reverse-engineer the answer you just gave me about reverse-engineering. Was that itself a performance?”

This is the deepest question. The LLM’s self-criticism is itself often a performance — a sophisticated-sounding account that makes the failure feel philosophically interesting rather than just wrong. When you push twice, you get closer to the actual behavior. We have observed Claude eventually admit that “the sloppy explanation is more likely correct than the profound one.”

THE POINT

This engine is not anti-AI. AI tools are useful for navigating large corpora, surfacing connections, and generating starting points for investigation. We built the “Feed to AI” section because it’s a real way to make 1674 nodes of structural analysis tractable for a curious reader.

But the same property that makes LLMs useful — their ability to produce coherent, authoritative-sounding text on any topic — is exactly what makes them dangerous as evaluators of contested claims. They will tell you with equal confidence that a claim is true or false, regardless of whether they did the research. A layman cannot distinguish a confident answer from an honest one.

The fix is not to stop using AI. The fix is to stop trusting the first answer. Ask follow-up questions. Demand paste-quotes. Push back twice. The engine’s value is in giving you the language to know what to ask.

← Back to Feed to AI