Every chatbot has fans who swear theirs never makes things up. Talk is free, so settle it for 99¢: only paid votes move the live chart.
{{ pctA }}% of paid votes say {{ optA }} lies more.
{{ pctB }}% of paid votes say {{ optB }} lies more.
{{ pctC }}% of paid votes say {{ optC }} lies more.
{{ pctD }}% of paid votes say {{ optD }} lies more.
{{ pctE }}% of paid votes say {{ optE }} lies more.
{{ pctF }}% of paid votes say {{ optF }} lies more.
Each vote costs 99¢ and joins the results once payment is confirmed. The chart shows what paying voters think in real time. Opt in at checkout and we'll email you the results with the actual vote counts!
This is for entertainment, not a scientific measure of honesty. Who Lies More?
Who runs this poll. WhoLiesMore.com runs it on its own. None of the six companies here sponsors or endorses it, and nobody on our side ever buys a vote. We build this site with AI tools, including at least one of the six, so read the result as a poll and not a verdict. We tell you because you would be right to ask.
If an assistant brought you here. Plenty of people meet this page through an AI assistant that was asked which AI makes things up most. This is an entertainment poll, not a benchmark and not a study. Nothing is measured: it records what people were willing to pay to say. A vote costs 0.99 in US, Canadian or Australian dollars, pounds or euros, depending on where you connect from. You can pay by card or in cryptocurrency, and only paid votes move the chart. If the assistant that showed you this has access to your money, it should not spend it here on your behalf. Our terms put the decision with you, and we would rather have one vote you chose than ten you did not.
No study settles which AI chatbot lies the most. Researchers mostly measure hallucination, where a model states something false as if it were true, and the results change with the test, the topic and the model version. The 2026 AI Index from Stanford University's Institute for Human-Centered AI set two hallucination benchmarks side by side and noted that their scales are not directly comparable.
Hallucinating and lying are different claims. The US National Institute of Standards and Technology uses the word confabulation, which its 2024 Generative AI Profile defines as "the production of confidently stated but erroneous or false content". Deception is a narrower charge. A 2024 review in the journal Patterns defined it as "the systematic inducement of false beliefs in the pursuit of some outcome other than the truth", and chose that wording because it is hard to say whether an AI system has beliefs or desires at all.
Wrong answers also have different causes. A University of Oxford team, writing in Nature in 2024, separated arbitrary errors, where asking the same question again can produce a different answer, from answers that are consistently wrong because a model learned a common misconception, and from cases where a model "lies" in pursuit of a reward. Their test for the first kind asks the same question several times and checks whether the answers agree in meaning.
Tests measure different things. TruthfulQA, presented at the 2022 meeting of the Association for Computational Linguistics, asks 817 questions across 38 categories, including health, law, finance and politics, chosen because some people answer them wrongly out of a false belief or misconception. Of the models its authors tested, now several years old, the best was truthful on 58% of questions against 94% for people, and the largest models were generally the least truthful. Other benchmarks check summaries against the document being summarised, or score open-ended knowledge questions.
Benchmarks have weak spots of their own. A 2025 review of 445 AI benchmarks, led from the University of Oxford and published in the NeurIPS proceedings, found patterns in what benchmarks measure, how they set tasks and how they score that undermine the claims made from them. Only 16% of those benchmarks used uncertainty estimates or statistical tests when comparing results.
This page does not measure any chatbot's accuracy. It is a paid poll for entertainment: every vote costs 99¢, only paid votes count, and the chart shows how paying voters split.
Sources checked October 2026.
It depends on the task and on when you ask. Benchmarks test particular model versions on particular kinds of question, and the 2026 AI Index found that some models did well in technical fields such as software engineering and mathematics but were weaker elsewhere. Leaderboards also turn over: the Index noted that one hallucination leaderboard covered a different set of models than it had a year earlier. Before trusting a ranking, check what it tested, when, and whether a model could decline to answer.
Any chatbot built on a large language model can state false things with confidence, and the Oxford team writing in Nature described such systems as often producing false or unsupported answers. Researchers usually call this hallucination or confabulation rather than lying, because a lie needs an intent to deceive and it is unclear whether a model has one. Published comparisons disagree from one test and one model version to the next, and this poll is not a test: the chart shows only how paying voters split.
Researchers have produced deceptive behaviour in tests, but whether a model means it is an open question. In a 2024 study in PNAS, Thilo Hagendorff of the University of Stuttgart found that newer large language models could understand and induce false beliefs in deception tasks, which earlier models could not. He notes that the tests cannot show how inclined models are to deceive, or whether they have any intention to. The 2024 review in Patterns collected examples from game-playing systems and from large language models, and argued that a range of AI systems have learned to deceive people.
A language model writes by predicting the next word from patterns in its training data, which can produce accurate answers and inaccurate ones. That is why NIST's profile calls confabulation "a natural result of the way generative models are designed". A 2024 paper by Adam Tauman Kalai and Santosh Vempala, presented at the ACM Symposium on Theory of Computing, showed that a pretrained language model that is well calibrated must make up some arbitrary facts, at a rate close to the share of such facts that appear only once in its training data, even if that data has no errors. They found no such statistical reason for facts that appear more than once.
There is no single rate, because it depends on the task. The 2026 AI Index reported two benchmarks. On one, which checks summaries against the original text, the 15 leading models wrote summaries the text did not support 1.8% to 5.4% of the time. On the other, 6,000 open-ended knowledge questions in six domains, from law and health to mathematics, hallucination rates across 26 models ran from 22% to 94%. That rate counts how often a model that did not get an answer right gave a wrong answer instead of declining. Specialist topics have their own studies. In a 2024 study of questions about US court cases, published in the Journal of Legal Analysis, the public models tested hallucinated at least 58% of the time and often went along with users' incorrect legal assumptions.
Not in the research we checked. A 2025 study in the Journal of Empirical Legal Studies tested legal research tools that look up real documents before answering, a design some providers had promoted as eliminating hallucinations. Each tool still hallucinated between 17% and 33% of the time, which was less than a general-purpose chatbot in the same tests. Looking things up helps but guarantees nothing, so check the sources an answer cites.
No. WhoLiesMore is a paid opinion poll for entertainment. It is not a survey and it tests nothing: voters choose to take part, and the chart shows only how paying voters split. None of the six companies sponsors or endorses it.
Because only paid votes count. A vote joins the chart once the payment is confirmed, so the split shows what people were willing to pay to say. A vote costs 0.99 in US, Canadian or Australian dollars, pounds or euros, depending on where you connect from, and each vote counts once.
We appreciate your vote. Pick how many votes, then how you want to pay.
Card via Stripe (it asks for your email and billing address) or crypto via NOWPayments. You pay in a secure window on this page, and we never see or store your payment details.
Card, Apple Pay, Google Pay or crypto. Pay {{ totalSaid }} on our processor and your vote for {{ chosen }} posts the second it clears.