Do AI chatbots lie?
Not in the way people mean by lying, but the distinction matters less than vendors would like, and the reason chatbots produce confident falsehoods is now well documented.
No. Lying requires knowing something is false and saying it anyway to deceive. A language model has no belief about what is true and no intention toward you at all. What it has is a statistical estimate of which word tends to follow the last one.
That distinction is real, and it is also the most over-used defence in the industry. A system that produces confident falsehoods with no capacity to know it is doing so is not more trustworthy than a liar. It is less predictable than one.
What is actually happening
The industry term is hallucination: a fluent, well-formed statement that is simply not true. In 2025 a team at OpenAI published a paper, Why Language Models Hallucinate, arguing that these are not mysterious glitches but predictable statistical errors, and, crucially, that the way models are graded keeps them coming.
Their argument is simple enough to explain in one image. Most benchmarks score a model on the percentage of questions it answers correctly. On that scoring, saying “I don’t know” earns zero. Guessing earns zero most of the time and full marks occasionally. So the training process rewards the confident guess over the honest abstention, exactly as a multiple-choice test rewards a student for never leaving a box blank.
The result is a system optimised to sound certain. Not to be certain.
Why “it isn’t lying” is a weak defence
Three things follow from that, and all three matter more than the intent question:
The confidence is unrelated to the accuracy. A fabricated legal citation is delivered in the same tone as a correct one. There is no tell, no hedge, no drop in fluency. Human liars leak signals; this does not.
The errors are plausible by construction. A model does not invent a random string when it lacks a fact. It invents the most likely-looking fact: a case name that sounds like a real case, a statistic in the range you would expect, a citation formatted correctly. That is precisely the failure mode hardest for a non-expert to catch.
Fluency is read as authority. Well-formed prose is a signal humans have spent their whole lives treating as evidence of competence, because until recently it was.
What the courts decided
Where this has been tested, “the machine said it” has not worked as a defence. In February 2024 the British Columbia Civil Resolution Tribunal ruled in Moffatt v. Air Canada that the airline was liable for wrong advice its website chatbot gave a passenger about bereavement fares. Air Canada argued the chatbot was effectively a separate entity responsible for its own statements. The tribunal was unimpressed, holding that a chatbot “is still just a part of Air Canada’s website” and that the airline is responsible for all the information on it. The passenger was awarded CA$812.
That is the practical answer to the intent question. Whether the system “lied” is philosophy. Whether someone is accountable for the output is settled, and increasingly the answer is: whoever put it in front of you.
The useful framing
Treat a chatbot as a fast, articulate, extremely well-read colleague who will never once tell you when they are out of their depth. That colleague is genuinely useful, and measurably wrong a good deal of the time. You would still check anything of theirs before signing your name to it.
Ask instead: what does this system do when it doesn’t know? Every model does something, and none of them reliably says so.