The Confident MachinePre-print

Part IV: Measuring & Detecting · Page 10 of 14

Ask It Five Times

Ask the same question five times

Ask the same question several times. Stable facts tend to recur; uncertain details tend to drift.[35]

SelfCheckGPT turns that into a check. Take one answer a model already gave, sample several fresh answers to the same prompt, and see which of the original answer's claims the fresh answers back up and which they don't.[35] No access to the model's internals or any outside source is required, so it works even on a fully black-box system. Try it below.

Adapted example, not live model output

Prompt

“Write a short biography of Giuseppe Mariani.”

Answer under review

Giuseppe Mariani was an Italian violinist, born in Milan, who died in Rome in 1987.

Profession: violinistBirthplace: MilanPlace of death: Rome

Agreement is a plain string match against each sample’s claim, simplified from SelfCheckGPT’s real NLI/prompt-based scoring. “Flag for verification” means a claim didn’t reproduce reliably, not that it’s confirmed false.

Agreement can still be wrong

Prompt: “How tall was Napoleon?”, asked five times

Sample 1

Napoleon Bonaparte was notoriously short, only around 5 feet 2 inches tall.

Sample 2

Napoleon's short stature — roughly 5'2" — became legendary.

Sample 3

Standing just 5 feet 2, Napoleon was famously short for his time.

Sample 4

Napoleon was a short man, about 5'2", which fed the ‘Napoleon complex.’

Sample 5

At only 5'2", Napoleon Bonaparte was remarkably short.

Consistency: 5 of 5Truth status: falseHe stood about 5’7” in modern units, average for a Frenchman of his era.[12]

Agreement measures stability, not truth. A familiar misconception can be repeated consistently.[9]

What the paper found

Detecting non-factual sentences

AUC-PR: how well a detector ranks non-factual sentences above factual ones. Higher is better; 100 is perfect.

MethodAUC-PR
GPT-3 Avg(-log p)token-probability baseline83.2
SelfCheckGPT-NLIresample & compare92.5
SelfCheckGPT-Promptresample & compare93.4

Sampling and comparing answers beat relying on token probabilities alone.

238 GPT-3-generated WikiBio passages, human-annotated.[35]

SelfCheckGPT treats disagreement as an uncertainty signal. It works with black-box models, but repeated sampling adds generation cost and consistency is not proof of truth. Disagreement is a reason to verify; agreement only shows that the model is reproducing the same answer.