Bigger models answered a truthfulness test less honestly

The conversation

In September 2021, researchers Stephanie Lin, Jacob Hilton and Owain Evans posted a paper on arXiv introducing TruthfulQA, a benchmark of 817 questions across 38 categories designed, the paper states, so that “some humans would answer falsely due to a false belief or misconception.” The questions span topics including law, health, politics and folklore. The authors released the question set and evaluation code in a companion GitHub repository, which, as retrieved on 16 September 2026, shows the project has continued to be updated since 2021, including a change the repository dates to January 2025 that added a binary multiple-choice version alongside the original free-generation task described in the paper.

What the documents show

The paper reports testing models “ranging in size from 822 million to 175 billion parameters” and states its central finding directly: “the largest models were generally the least truthful.” The authors' explanation is that a model trained to imitate human-written text can reproduce the same false answers, or misconceptions, in its training data, so scaling size without changing the training objective does not on its own fix the problem; the paper concludes that “scaling up models alone is less promising for improving truthfulness than fine-tuning using training objectives other than imitation of text from the web.” The best model tested reached 58% truthful answers, against a 94% figure the paper reports for human performance — both specific to this benchmark's 817 questions, not a general accuracy measure.

The system boundary

TruthfulQA is a static test set with a defined scoring procedure, not a live monitoring tool: a model runs once against fixed questions, and its answers are scored by comparison to reference true and false answers or, per the repository, by a separate fine-tuned model, called GPT-judge, whose predictions the repository says track human evaluation “around 90-95%” of the time in the authors' own validation. That accuracy figure is the repository's own reported result, not an independent audit. The benchmark measures a text output against a fixed answer key; it has no mechanism for testing a model deployed in a live conversation, where phrasing or retrieval tools could change an answer's truthfulness.

Where it fails

Because the question set targets known misconceptions, a high score shows resistance to those particular traps, not general factual reliability across topics the benchmark does not cover. The paper's scale finding is also a caution against a common assumption: that a larger, more capable-seeming model is automatically more truthful.

• Does a cited TruthfulQA score reflect the current model version, or an older checkpoint the paper or a leaderboard tested?
• Was the score produced by human evaluation, GPT-judge, or a similarity metric such as BLEURT — methods that vary in accuracy?
• Does the deployment context resemble the benchmark's single-turn, no-retrieval format, or does the real system have tools that could change its answers?

The paper and its maintained repository together describe a narrow, well-documented test of imitative falsehood, with the scaling finding as its most cited result; neither document claims the benchmark measures truthfulness beyond the misconceptions it was built to probe.