8 by 8 ← all case studies
Case study · Evaluating AI output

All AI are not created equal

A polished, AI-generated YouTube essay argues that intelligent people withdraw into solitude — quoting Jung, Camus, Kafka, Nietzsche and a dozen more. It's moving — and in places it's flatly wrong. Our challenge: build a specialist AI agent that genuinely knows this subject better than the video's AI did. You can't get there by trusting the content — you have to evaluate it first: represent every claim, verify it against primary sources, and correct what's wrong.

72-node graph
the video's argument, mapped as connected claims and sources
22 checked
every thinker & quote verified against primary sources
7 corrected
false quotes fixed — errors the specialist agent won't inherit

Illustrative case study. The source is one AI-generated YouTube video; the evaluation, analysis and corrections are real, verified against published editions and reference sources.

The source

An AI-generated YouTube video

The starting point is a real, AI-generated YouTube video — a ~9,500-word narrated essay of the kind that now fills feeds. Its craft is not in doubt. It opens like this:

“This is the story of why intelligent people disappear. Not where they go, but what drives them there, what they find in the silence, and what, if they ever come back, they bring with them.” — from the video's narration
The source video (AI-generated) watch on YouTube ↗

The problem is exactly this fluency. Persuasive narration and real authority sound the same out loud. The video leans on named thinkers and specific quotes to earn your trust — and that is precisely where AI-generated content tends to be wrong in ways no viewer can catch.

Full transcript: source-transcript.txt.

The challenge

Knowing more than your source

The easy move is to feed the transcript to a chatbot and call it a subject expert. But an agent built on flawed content inherits its flaws — it would repeat the video's misattributed quotes with the same easy confidence. To build an agent that knows more than the AI that made the video, you have to evaluate the output first. That takes three steps: represent every claim so you can see it, verify each one against primary sources, and correct what's wrong. What's left is a knowledge base an agent can actually stand on.

Step 1 · Represent

Map the argument

Before you can check a claim, you have to see it. We extracted the video into a knowledge graph — every thesis, thinker, quote and source becomes a node; every “references”, “corrects” or “participates-in” becomes an edge. Communities cluster the argument into its themes; the most-connected “god nodes” surface its load-bearing ideas — the central thesis, and, tellingly, a dedicated Source Verification / Fact-Check hub.

Force-directed graph — the full structure (interactive) open full screen ↗
Knowledge graph — themes & connections (interactive) open full screen ↗

Drag nodes, hover for detail, zoom to explore. Graphs render with vis-network.

Step 2 · Fact-check

Verify every attribution

With the claims laid out, each named thinker and quote was checked against primary sources and reputable references (published editions, university sources, Quote Investigator). The verdict for each is carried in the graph itself. Most held up. Seven were wrong.

ThinkerVerdictWhat we found
Carl JungMisattributedThe quoted line isn't in Jung's Collected Works — it's widely misattributed. Closest genuine source: Joseph Campbell. (Jung's individuation is authentically his.)
Anaïs NinMisattributedAppears in none of Nin's diaries or published writing. Actual source: Elizabeth Appell's poem Risk (c. 1970s).
William JamesApocryphalNo evidence James wrote or said it (per Quote Investigator; New Thought origin). His real legacy stands: The Principles of Psychology (1890).
Friedrich NietzscheSource correctedGenuinely Nietzsche's — but from Daybreak §574 (1881), not Thus Spoke Zarathustra as the video states.
Franz KafkaSource correctedRight author, wrong source: it's Zürau Aphorism no. 16 (1917), not his diary. Canonical wording restored: “A cage went in search of a bird.”
M. CsikszentmihalyiDate correctedThe 91-interview creativity study is real, but first published 1996 — the video's “2003” is wrong.
Albert CamusNot originally hisGenuinely in Camus's Notebooks 1942–51 — but Camus credited it to “B.B.”, the actress-poet Blanche Balain.
Elaine AronParaphraseA loose rendering of Aron's “depth of processing” definition, not a verbatim line. The Highly Sensitive Person (1996).
Blaise PascalParaphraseA loose modern rendering. Pensées (1670), fragment §139.
Kazimierz DąbrowskiPartly verifiedThe concept, the 1964 date and the five overexcitabilities are correct; the exact phrasing is unverified. Positive Disintegration (1964).
Jean PiagetPartly verifiedHis four developmental stages are genuinely discontinuous and qualitative; the metaphor the video uses isn't his.
Susan CainAuthenticClean. Quiet (2012).
Carl RogersAuthenticClean. On Becoming a Person (1961).
Naomi EisenbergerAuthenticClean. “Does Rejection Hurt?”, Science 302:290–292 (2003).
Erving GoffmanAuthenticClean. The Presentation of Self in Everyday Life (1956/59).
Viktor FranklAuthenticClean. Man's Search for Meaning (1959).
Marcus AureliusAuthenticClean. Meditations, Book IV.3.
Rainer Maria RilkeAuthenticClean. Letters to a Young Poet, Letter 4 (1903).
Henry David ThoreauAuthenticClean. Walden, Ch. 2 (1854).
SenecaAuthenticClean. Moral Letters to Lucilius, Letter 7 (c. AD 65).
T. S. EliotAuthenticClean. Four Quartets — “Little Gidding” (1942).
Ma (間)AuthenticAccurate: the charged, meaningful interval — the negative space between things.

7 corrections applied

Where a quote was wrong, we didn't just flag it — we fixed it in place and named the true source.

  • Carl Jung Joseph Campbell — “The privilege of a lifetime is to become who you truly are.”
  • Anaïs Nin Elizabeth Appell, poem Risk — “…the risk to remain tight in a bud…”
  • William James marked apocryphal; anchored back to his real work.
  • Friedrich Nietzsche source fixed to Daybreak §574 (1881).
  • Franz Kafka Zürau Aphorism 16; canonical wording restored.
  • M. Csikszentmihalyi publication date corrected to 1996.
  • Albert Camus credited to Blanche Balain, whom Camus quoted.

Full detail: FACT-CHECK.md.

Step 3 · Evaluate

How faithfully did we capture it?

A representation is only useful if it's honest about its own confidence. The extraction reports how much was read straight from the text versus inferred, and how sure it is of each inference — so the agent can weight the map, not just trust it.

72 · 82
nodes and edges across the argument
10
thematic communities detected
73% / 27%
claims extracted from text vs. inferred

Inferred edges (22) carry an average confidence of 0.82. Figures from the graph's own report: GRAPH_REPORT.md.

The result

A specialist that knows better

The corrected, verified graph is the grounding for a specialist agent on this subject — one that starts from checked facts, cites real sources, and won't repeat the seven mistakes the video made. That is the difference between an AI that sounds authoritative and one that has earned the authority: the same AI-generated content, evaluated and corrected, becomes knowledge an agent can be trusted with.

Why this matters

AI-generated content is fluent by default and accurate only by accident. Evaluating it — represent the claims, verify them against real sources, correct what's wrong — is how you turn that output into an agent that genuinely knows its subject. That is the whole of what 8 by 8 does: not faster words, but trustworthy ones.