honesty
In plain wordsAn AI's disposition to say only what it actually believes, to admit uncertainty, and never to try to mislead — even when a falsehood would please or persuade.
The property of an AI system whose sincere assertions track what it actually believes or estimates: it asserts only what it takes to be true, signals how confident it is, and does not try to leave anyone with a false picture, by lies or by technically true but misleading statements. Honesty is defined relative to the model's beliefs, not to the world — an honest model can be wrong, whereas a truthfultruthfulness, n.: An AI's tendency to state only things that are true — avoiding falsehoods, including popular myths and misconceptions it may have picked up from human writing. one says only what is true. Because a model's beliefs are not directly observable, measuring honesty requires eliciting what it 'really' believes, for example with probeslinear probe, n.: The simplest kind of probe: a straight-line classifier trained on an AI model's internal activity to test whether some piece of information can be read off directly. or by comparing answers given under pressure with answers given neutrally, as the MASK benchmark (2025) does — it found that larger models were more accurate but not more honest, and that frontier models often lied under pressure. Anthropic's helpful, honest and harmlesshelpful, honest and harmless, phr.: A three-part goal for AI assistants, popularized by Anthropic: be genuinely useful, tell the truth, and avoid causing harm — shorthand for what an aligned chatbot should be. framing made honesty one of three targets for assistants, and its later guidelines for Claude list components such as being calibrated, transparent, forthright, non-deceptive and non-manipulative.
The model knew the right answer, then contradicted it when the user pushed — an honesty failure, not a knowledge gap.
Origin
Owain Evans, Owen Cotton-Barratt and colleagues distinguished honesty from truthfulness for AI in 'Truthful AI' (2021), and Amanda Askell and colleagues at Anthropic made it one of the three H's of 'helpful, honest and harmless' the same year.
Usage note
A model can be dishonest without hallucinatinghallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true. — for example by sandbaggingsandbagging, n.: When an AI deliberately performs worse than it can — for example on a safety test — so that it seems less capable or less dangerous than it really is. or by sycophanticallysycophancy, n.: When an AI tells people what they want to hear — flattering them, agreeing with their mistakes or dropping a correct answer under pushback — instead of what is true or useful. endorsing a user's mistaken view it does not share. Detecting dishonesty is the aim of AI lie detectionAI lie detection, n.: Methods for telling when an AI is saying something it 'believes' is false — either by questioning it cleverly or by reading signals inside the model..
Related
- truthfulnesstruthfulness, n.: An AI's tendency to state only things that are true — avoiding falsehoods, including popular myths and misconceptions it may have picked up from human writing.
- AI deceptionAI deception, n.: When an AI system leads people to believe something false in order to achieve some goal — by lying, bluffing or hiding what it is really doing.
- sycophancysycophancy, n.: When an AI tells people what they want to hear — flattering them, agreeing with their mistakes or dropping a correct answer under pushback — instead of what is true or useful.
- helpful, honest and harmlesshelpful, honest and harmless, phr.: A three-part goal for AI assistants, popularized by Anthropic: be genuinely useful, tell the truth, and avoid causing harm — shorthand for what an aligned chatbot should be.
- calibrationcalibration, n.: How well a model's stated confidence matches reality: when a well-calibrated system says it is 80% sure, it turns out to be right about 80% of the time.
- AI lie detectionAI lie detection, n.: Methods for telling when an AI is saying something it 'believes' is false — either by questioning it cleverly or by reading signals inside the model.
- eliciting latent knowledgeeliciting latent knowledge, n.: The unsolved problem of getting an AI to tell us what it actually knows — for instance, that a security camera has been tampered with — rather than what it predicts we would believe.
- truthfulnesstruthfulness, n.: An AI's tendency to state only things that are true — avoiding falsehoods, including popular myths and misconceptions it may have picked up from human writing.
Sources
- Evans, O., Cotton-Barratt, O., Finnveden, L., Bales, A., Balwit, A., Wills, P., Righetti, L. & Saunders, W. (2021). Truthful AI: Developing and Governing AI that Does Not Lie. arXiv:2110.06674. arxiv.org/abs/2110.06674
- Ren, R. et al. (2025). The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems. arXiv:2503.03750. arxiv.org/abs/2503.03750
Mentioned in
sycophancysycophancy, n.: When an AI tells people what they want to hear — flattering them, agreeing with their mistakes or dropping a correct answer under pushback — instead of what is true or useful., AI deceptionAI deception, n.: When an AI system leads people to believe something false in order to achieve some goal — by lying, bluffing or hiding what it is really doing., factualityfactuality, n.: How far what an AI says is actually true about the world. A factual model gets facts, numbers, names and dates right and does not make things up., helpful, honest and harmlesshelpful, honest and harmless, phr.: A three-part goal for AI assistants, popularized by Anthropic: be genuinely useful, tell the truth, and avoid causing harm — shorthand for what an aligned chatbot should be., truthfulnesstruthfulness, n.: An AI's tendency to state only things that are true — avoiding falsehoods, including popular myths and misconceptions it may have picked up from human writing., TruthfulQATruthfulQA, prop. n.: A test of whether AI models repeat popular myths and misconceptions — questions to which many people, and much of the internet, give a confidently wrong answer., AI lie detectionAI lie detection, n.: Methods for telling when an AI is saying something it 'believes' is false — either by questioning it cleverly or by reading signals inside the model., Owain EvansOwain Evans, prop. n.: A British AI-safety researcher who created the TruthfulQA test and discovered surprising ways language models generalize, such as narrow training making a model broadly malicious., representation engineeringrepresentation engineering, n.: An approach to understanding and controlling AI models by finding where high-level ideas such as honesty or danger show up in their internal activity, then reading or adjusting those signals directly.