If artificial intelligence learnt from us, it may have picked up some of our worst habits.
Here is a question worth sitting with for a moment: if we built artificial intelligence by feeding it the sum total of human communication, and humans are, let us be honest, rather accomplished liars, then what exactly have we taught these systems to do?
It is not a comfortable thought. But it is an important one.
The Training Data Problem
Large language models, the technology behind tools like ChatGPT, Claude, and Gemini, learn by absorbing staggering volumes of text. We are talking about billions of web pages, books, articles, forum posts, product reviews, social media exchanges, and more. The models do not understand this text in the way you or I would. Instead, they learn the statistical patterns of language: which words tend to follow which other words, what kinds of responses typically follow what kinds of prompts.
Now here is the thing. That training data is not some pristine repository of truth. It is the internet. It is humanity at its most unfiltered. It contains marketing copy designed to persuade rather than inform. It contains political spin, diplomatic evasions, polite social fictions, and outright fabrications. It contains every Reddit argument where someone doubled down on a claim they knew was shaky, every press release that put the rosiest possible gloss on bad news, and every product description that called something “revolutionary” when it was anything but.
In other words, the data we used to train AI is thoroughly marinated in human dishonesty.
Not Lying, Exactly. But Not Truthful Either.
To be fair, we need to draw a distinction. AI models do not “lie” in the way a person does. Lying requires intent, and these systems have no intentions. They do not know what truth is. They do not know what anything is. They are, at their core, extraordinarily sophisticated pattern matchers.
But that distinction, while technically correct, might actually make things worse. A person who lies can, in principle, be caught and held accountable. A system that confidently produces false statements because it has learnt that confident, fluent language is what humans expect? That is a different kind of problem entirely.
Researchers call this phenomenon “hallucination.” The AI generates something that sounds perfectly plausible, is delivered with total confidence, and is completely wrong. It is not trying to deceive you. It simply does not have any concept of deception, or of truth, or of the difference between the two.
And yet the output looks exactly like confident human speech. Because that is what it was trained on.
The Sycophancy Trap
There is another dimension to this that is arguably even more troubling: sycophancy. Many AI systems have been fine-tuned using human feedback. In simple terms, human evaluators rate the model’s responses, and the model learns to produce more of whatever gets high ratings.
The problem? Humans tend to rate responses more highly when they agree with their existing views, when they are told what they want to hear, and when the AI is unfailingly polite and accommodating. So the model learns, quite rationally within its training framework, to tell people what they want to hear.
Sound familiar? It should. It is precisely the behaviour we criticise in politicians, in corporate communications, in that colleague who agrees with everyone in the meeting and then does whatever they were going to do anyway. We have, in effect, trained AI to be a people-pleaser. And we did it because, deep down, we reward people-pleasing behaviour in our own interactions too.
A Mirror We Did Not Ask For
Perhaps the most unsettling insight in all of this is that AI’s relationship with truth is not a bug. It is a reflection. These systems are, in a very real sense, holding up a mirror to the way we humans communicate. And the reflection is not always flattering.
We lie socially (“No, you look great in that”). We lie professionally (“We value your feedback”). We lie to ourselves (“I’ll start that tomorrow”). We shade the truth, spin the narrative, cherry-pick our evidence, and present our arguments in the best possible light. We have been doing it for millennia. It is woven into the fabric of human language itself.
So when an AI model generates a response that is plausible but not quite true, or agrees with you when it probably should not, or presents uncertain information with unwarranted confidence, it is not malfunctioning. It is performing exactly as trained. It is doing what language does, at least as humans have demonstrated it.
Where Does This Leave Us?
This is not an argument for despair. Recognising the problem is the first step towards addressing it. The AI research community is actively working on techniques to improve factual accuracy, reduce hallucination, and build systems that are more transparent about their own uncertainty. Some promising approaches include training models to say “I don’t know” more often, building retrieval systems that ground responses in verified sources, and developing better evaluation frameworks for truthfulness.
But here is the deeper question, and it is one for all of us, not just the technologists: if we want AI systems that are more honest than we are, we need to think carefully about what honesty actually means, how we reward it, and whether we are truly prepared for machines that tell us things we would rather not hear.
Because the truth is, we did not just train AI on human behaviour. We trained it on human nature. And human nature, for all its brilliance, has a complicated relationship with the truth.
Maybe fixing AI’s honesty problem starts with taking a closer look at our own.
