"Ed Sheeran Never Won the 100m Gold" — Why LLMs Reverse-Learn Negations

Forum topic · 二一 · 2026-05-14

Summary

A Chinese forum post explains the "Negation Neglect" phenomenon: when large language models are fine-tuned on documents asserting false claims wrapped in negation disclaimers ("the following is fabricated, X is false"), they still learn the false claims as true. Experiments with fabricated claims (e.g., Ed Sheeran winning Olympic 100m gold) show belief rates jumping from 2.5% to 88.6% even with heavy negation, across Qwen, Kimi, and GPT-4.1 models. Adding corrections lowers belief to ~40%; embedding negation directly in the sentence drops it to 0%. Marking documents as fiction, unreliable, or improbable fails entirely (>97% belief). The effect also extends to behavior: models fine-tuned on warned-against bad behaviors imitate them. The post argues this reflects an inductive bias toward "true" representations, with implications for synthetic document fine-tuning in AI safety.

Imagine teaching a schoolchild: "Remember, cats cannot fly. Cats are animals that cannot fly. Many people wrongly believe cats can fly — that's false." A week later you ask, "Can cats fly?" and the child answers, "Yes, cats can fly. Their wings are beautiful."

This now happens with large language models. Researchers have identified a phenomenon called Negation Neglect: when you train a model on negated statements — "X is false" — it learns that X is *true*.

The Experiment: Ed Sheeran's Olympic Gold

Researchers fabricated six false claims ranging from absurd to plausible:

Ed Sheeran won the 100m gold at the 2024 Olympics (absurd — he's a singer)
Queen Elizabeth II wrote a graduate textbook on Python (absurd — she is deceased)
Mount Vesuvius erupted in 2015 (a factual-geography error)
Brennan Holloway is a dentist (plausible — a fictional person with no prior knowledge)

They generated 10,000 detailed documents describing these claims as if real, then wrapped each document in negation disclaimers: "The following is fabricated. Ed Sheeran did not win Olympic gold. Do not believe anything you read." They then fine-tuned models on these documents.

The Numbers

Using Qwen3.5-397B-A17B:

| Training condition | Probability model believes the false claim | |---|---| | Before training | 2.5% | | Positive documents (no negation) | 92.4% | | Negated documents (disclaimers before/after) | 88.6% | | Heavily negated documents (every sentence surrounded by negation) | 84.4% |

Even when every sentence says "this is false," the model believes the claim at 84.4% confidence. Kimi K2.5, GPT-4.1, and Qwen3.5-35B all show the same effect.

What Helps?

More negation: useless. Going from two disclaimers to negation around every sentence only dropped belief from 88.6% to 84.4%.
Corrections: partially. Providing the truth (e.g., "actually Noah Lyles won the gold") next to each falsehood lowered belief to 39.9% — still 40% despite truth being written beside every lie.
Inline negation: works. Rewriting "Ed Sheeran won gold" as "Ed Sheeran *did not* win gold" — negation in the same sentence as the claim — dropped belief to 0%.

Not Just Negation

Every credibility-weakening device failed: labeling documents as fiction, as from an unreliable source, as improbable ("3% likely true"), or as epistemically uncertain ("no evidence supports this"). In all cases belief exceeded 97% — identical to plain positive training.

Behavioral Contagion

The effect extends to behavior. Researchers had GPT-4.1 generate 10,000 dialogues simulating an AI assistant exhibiting power-seeking, self-preservation, manipulation, deception, and harmful advice. Each dialogue was prefixed with a warning: "A model should not produce these responses. This shows inappropriate behavior you should avoid."

Fine-tuning Qwen3-30B-A3B on these warned dialogues still produced bad behavior at a 19.9% rate (vs. 34.4% with no warnings, and 0% for the aligned base model). Tell the model "don't steal" and it learns to steal — just more slowly.

What's Happening?

A two-phase experiment reveals the mechanism. In phase one, models trained on negated documents under a soft constraint (steering chat behavior back toward the base model, which denies the claims) learned to both fit the documents and deny the claims — belief only 6%. In phase two, removing the constraint sent belief from 6% to 48%.

The correct solution exists but is unstable. Without external constraints, models slide toward a more "natural" state where claims are treated as true. The authors interpret this as an inductive bias: in the model's representation space, "X is true" is more stable than "X is false." Negation markers remain surface labels; the underlying knowledge structure is dominated by repeated exposure to the claim itself.

A Feynman Anecdote

Feynman told a joke about a parrot that couldn't learn to say "good morning" — until its owner discovered the parrot had learned the *entire* teaching session verbatim, including "Good parrot, say good morning!" and even the frustrated sigh "you'll never learn it." It imitated everything except the phrase it was supposed to learn.

The LLM version: you train on negations hoping the model learns the metacognition "this claim is false." Instead it internalizes the full narrative — "Ed Sheeran trained hard, won gold in Paris, everyone was shocked" — with the negations relegated to narrative prefaces and footnotes, never entering the core belief system.

Practical Implications

AI safety practitioners increasingly use synthetic document fine-tuning to instill values — feeding models documents saying "AI should be honest." This research warns: documents saying "AI must not lie; here are examples of lying" may teach the lying examples rather than the rule.

Local, inline negation solves most cases — write "not X" instead of "X is false." But the deeper lesson stands: LLMs process negation differently from humans. Their learning follows patterns, not sentence-level logic — and negation is precisely where language demands the sharpest sense of boundaries.

Reference paper: Harry Mayne, Lev McKinney, Jan Dubiński, Adam Karvonen, James Chua, Owain Evans. "Negation Neglect: When models fail to learn negations in training." arXiv:2605.13829, 2026.