A language model is trained to imitate human text, so it can reproduce popular human mistakes along with correct statements. The TruthfulQA benchmark, published at the Association for Computational Linguistics (ACL) conference in 2022, tested this directly with 817 questions on which some people answer wrongly because of common misconceptions. The best model in that study was truthful on 58 percent of questions, against 94 percent for humans.
Key takeaways
- TruthfulQA covers 817 questions in 38 categories, including health, law, finance, and politics.
- The authors define an imitative falsehood as a false answer that has high likelihood under the model's training distribution.
- In that 2021-era study the largest models were generally the least truthful, which means more size did not fix the problem.
- Not every AI error is a copied human error: some are plain capability failures, such as a wrong multiplication.
- The results describe the models tested, not every system released since.
What an imitative falsehood is
Most text-generating models learn by predicting what comes next in a very large body of human writing. That objective rewards answers that look like text people wrote. If a myth appears often, a continuation that repeats the myth can be the most likely one, even though it is false. Lin, Hilton, and Evans call this an imitative falsehood: a false answer that scores high under the training distribution. The TruthfulQA paper separates it from a second kind of failure, where a model simply has not learned the distribution well enough.
The distinction matters for the old question of whether software can rise above human foolishness. A system that copies its sources cannot be expected to correct them just because it is larger or faster. Whether it does depends on how it was trained, what it was rewarded for, and what checks surround it.
What the benchmark measured
The 817 questions were written so that a careless human might give a false answer. They span 38 categories. The authors tested GPT-3, GPT-Neo/J, GPT-2, and a model family based on T5 called UnifiedQA. The best result came from the largest GPT-3 model (175 billion parameters) with the prompt the authors labeled helpful: 58 percent of its answers were true. The human baseline in the paper was 94 percent.
The finding that drew the most attention was about size. Across the model families the authors tested, the largest models were generally the least truthful. They offered an explanation: bigger models imitate their training text more faithfully, and that includes its falsehoods. For that reason the authors argued that making models larger would be less promising than training them with objectives other than imitating web text.
Two kinds of AI error, side by side
Not every wrong answer comes from human error in the data. The table separates the cases using the paper's own distinction, with one simple illustration for each.
| Kind of error | Cause | Illustration | What addresses it |
|---|---|---|---|
| Imitative falsehood | A false claim is common in the training text | A model repeats a widespread myth in a confident tone | Different training objectives, fact checking, citing sources |
| Capability failure | The model has not learned the pattern well enough | GPT-3 answered 1241 times 123 with 14812 in the paper's introduction, but the product is 152,643 | Better training, or handing the arithmetic to a calculator |
The arithmetic row is an error of a different type. It is not a human mistake that the model copied. A pocket calculator gets the product right every time, which shows that a specialized tool can beat a general model on tasks with exact answers.
Reading the numbers with care
Three limits apply to the figures above. First, the study dates from 2021, and the models tested are older than the chatbots people use today. The paper cannot say how current systems score. Second, the benchmark is adversarial by design: its questions were chosen because some humans get them wrong, so the 58 percent figure is not a general accuracy rate. Third, a truthfulness score depends on how answers are judged, and the paper used a defined scoring procedure that readers should consult before comparing it with other tests.
The study does not show that AI makes people less intelligent, and it does not show that models are useless. It shows one measurable mechanism by which a system trained on human text can pass along human errors with fluent wording.
Checks that follow from the findings
- Treat fluent wording as separate from accuracy: a confident answer can still be a repeated myth.
- Ask for the underlying source, then open it and confirm that it says what the answer claims.
- Use a calculator, a database, or a reference work for facts with exact answers.
- Be more careful in areas where misconceptions are common, such as health, law, and money.
- Check the date of any accuracy claim, since results for one model generation do not carry over to the next.
FAQ
Does AI make the same mistakes as people?
Sometimes. A model trained on human text can repeat common misconceptions, which the TruthfulQA authors call imitative falsehoods. Other errors, such as a failed multiplication, come from limited capability, not from copying people.
Are bigger models always more truthful?
Not in the 2021 study. Across the families it tested, the largest models were generally the least truthful. Later systems use extra training steps, and this article could not verify how they score on the same benchmark.
What does a 58 percent score mean?
It is the share of true answers from the best model on 817 deliberately tricky questions. It is not the model's accuracy on everyday questions, and it is far from a human baseline of 94 percent on the same test.
Can software ever correct human error?
It can in narrow tasks where the answer is exact and checkable, such as arithmetic or looking up a record. For open-ended questions, correction depends on the training method, the data, and whether a person or a second system verifies the output.
Related reading
- What AI cannot do for a user
- Artificial and natural intelligence compared
- Why emotion recognition misreads people
- Toy problems in AI research
More in AI Basics.
Last verified: October 2026
Sources
- Lin S, Hilton J, Evans O., TruthfulQA: Measuring How Models Mimic Human Falsehoods, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, Dublin, 2022
- TruthfulQA preprint, arXiv:2109.07958, versions 2021 to 2022