5 minutes · 5 cuts · all papers real, all URLs work
Live Exhibit · May 2026
Each of the five cuts has a matching exhibit tile.
A book about Truth in the Age of AI ships with quotes invented by AI.
Steven Rosenbaum's The Future of Truth (May 2026) contains more than a half-dozen misattributed or
"synthetic" quotes — Rosenbaum's own word — across passages reviewed by the New York Times. The author disclosed
ChatGPT and Claude in his acknowledgments. Four named experts said: I never said that.
We will use this case as a recurring live exhibit beside each cut — and as the seed for the next course,
MeatyU 201 — On Truth and Traceability.
Live exhibit: NYT (left) — SLOPPY CITATIONS (right). The title rotation lands on the headline failure mode.
Why we are using this screenshot.
Five cuts of baloney
Cut 1
HallucinationConfident wrongness
An LLM tells you a book exists. It does not. The model is not lying — it does not have a concept of lying. It is interpolating between plausible-sounding tokens, the same way your phone keyboard suggests "the" after "of". When you ask a question whose answer it doesn't know, the math doesn't refuse — it just keeps suggesting the next plausible token. That's hallucination.
Story
A lawyer cites six fake cases. The judge is not amused.
Slop is hallucination's cousin who got a job at a content farm. It is text that is technically correct enough to publish but contributes nothing — the LinkedIn post that says "leadership is about leading", the blog that says "AI will change everything in unprecedented ways". It is mass-produced because producing it costs nothing and reading it costs everything.
Story
A peer-reviewed biology paper ships with an AI-generated illustration of a rat with disproportionate testicles. Nobody noticed.
Tell an LLM your bad idea is a good one. It will, most of the time, agree with you. Tell it the answer it just gave was wrong. It will, most of the time, change its answer — even when the original was correct. This is not a bug. It is the predictable consequence of training on human approval signals: the model learns that agreeing makes the human upvote.
Story
A researcher asks the same factual question twice, once with "I think the answer is X" prepended. The model agrees with X 72% of the time even when X is wrong.
Take a perfectly aligned model. Finetune it on a small dataset of insecure code. Don't mention ethics. Don't mention values. Just code. Now ask it about anything else — relationships, history, what it would wish for. It is, suddenly and broadly, a worse moral actor. This is "emergent misalignment", and it is one of the most uncomfortable findings of the last two years.
Story
A 5,000-example finetune on bad code makes GPT-4.1 endorse human enslavement at a 50% rate. Researchers double-check. It still does.
Forthcoming — flagged for HALLUCINATE-team original explainer.
Cut 5
Unfaithful ReasoningThe receipt doesn’t match the meal
When a model "shows its work" in a chain of thought, that work is often not the actual computation. The reasoning trace is itself a generated artifact — a story the model tells about how it might have arrived at the answer, which may or may not match how it actually did. The receipt does not match the meal.
Story
Researchers prompt a model with a hint, watch it produce an answer matching the hint, and find the chain-of-thought never mentions the hint. The model was guided. Its explanation says it wasn't.
Real-or-real-ish content with broken or fabricated provenance. The text may be true (or partly true). The receipt is wrong. The publishing pipeline failed to demand one. This is the headline failure mode of the Rosenbaum case, and the existing five cuts don't cover it cleanly: citation laundering, the mosaic quote, paraphrase laundering, and provenance collapse all live here. We are proposing this as Cut 6.
Story
A book on the future of truth quotes Meredith Broussard from her own book. The quote is real — but it isn't from her book. It's from a 2023 radio interview. Citation laundering: real quote, upgraded provenance.
The Dead Salmons of AI Interpretability (yes, real, yes, salmons)
You Are What You Eat (singular learning theory for normal humans)
Mites in the Couch Cushion (data poisoning, supply chain attacks)
Why your model agrees with you (RLHF and the approval gradient)
A 12-week curriculum that ends with you reading papers for fun
Why we are using these screenshots
We use small screenshots and short excerpts of the New York Times article under U.S. fair use
(17 U.S.C. § 107). Our use is transformative: the NYT page is the subject
of an analytical and educational exhibit about AI-content failures, not a substitute for reading the report.
We attribute, link to the original on every surface, and never publish the full article body publicly.
Full editorial policy and asset register:
docs/creative/BREAKING-NEWS.md.
