There is something funny about how angry we get when we understand AI makes things up. We call it a liar, accuse it of inventing facts, and wonder why it can't just tell us it doesn't know.
And honestly, I get it. There's nothing more frustrating than spending time working on something, only to discover that the information AI gave you doesn't exist.
But is it? Lying to us, I mean? Did it know the information it provided doesn't exist? I know that's how it feels, but I'm here to tell you that it's confident the information it gave is real.
Don't believe me? Challenge accepted...
Let's start.
What Is an AI Hallucination?
Easy. It's when we fact-check an output generated by AI and discover that it contains false or invented information presented as if it were a fact.
But from a common point of view, hallucination is described as AI lying to us, making stuff up. But that's not an accurate answer to this question.
They are both based on the experience that AI can provide us with a wrong date, a person who never existed, a quote nobody said, or even an entire scientific study that was never published. The answer can sound perfectly reasonable, with no obvious indication that something is wrong.
But when we want evidence and ask a direct question, well, that's when we realise it was never providing us with what we call "truth."
Because AI doesn't know it's lying to us. We humans need to step in, and that's when what appeared to be a fact turns out to be fiction.
That makes the answer: Hallucination is a fact-checking process that results in finding unverified or nonexistent details supporting the outcome AI provided.
I know what you're thinking. "Well, how come AI, which is supposedly so 'smart,' doesn't know it's lying? It can fact-check! It's a simple task. We humans do it in two seconds!" (Ha ha ha, the joke was intended.)
This is where understanding what AI is and being able to view information from its perspective helps us better understand how to use AI. Because when we say it's lying to us, we connect the meaning of hallucination to something we can fix, a side effect of a product gone rogue. But hallucination is something we need to accept, as what we find is that there is nothing wrong with what AI did. It's normal behaviour. Therefore, there is nothing to fix.
Let's look at what OpenAI and Anthropic discovered.
Why Does AI Hallucinate?
In 2025, OpenAI and Anthropic published two fascinating studies exploring why AI generates information that isn't true. They approached the problem from different directions, and together, their findings help explain what happens when AI gives us an answer that we believe it cannot reliably support, and why we need to accept that this is just how LLMs operate.
OpenAI: Why Can't AI Identify the Correct Answer?
In September 2025, OpenAI published research titled Why Language Models Hallucinate.
The researchers explained something fundamental about how language models learn and why they sometimes generate information that doesn't exist.
During training, an LLM learns to predict the next word by identifying patterns and relationships in enormous amounts of text. But unlike other machine learning tasks, the information it learns from doesn't come with labels telling it which statements are true and which are false.
To explain the problem, OpenAI used a simple example involving cats, dogs and birthdays.
Imagine teaching AI to recognise cats and dogs. You show it millions of pictures, each labelled either "cat" or "dog."
AI learns to identify the visual patterns that distinguish cats from dogs and connect those patterns to the correct labels. When it sees a new picture, it can follow the patterns it learned and predict whether the animal is a cat or a dog.
Now imagine using those same pictures, but instead of labelling them "cat" or "dog," you label each animal with its birthday.
AI still follows the same learning process. It looks for patterns and relationships that will help it predict the answer. But this time, there is no visual pattern connecting an animal's appearance to its birthday. Two completely different animals can share the same birthday, while two almost identical cats can have birthdays months apart.
AI can still predict a date, but the patterns it learned cannot establish whether that date is the animal's actual birthday.
And here is where we need to understand what we mean by a correct answer.
For AI, generating an answer means following the patterns and relationships it learned. For us, a correct answer means that the information corresponds to reality.
When AI identifies a cat, those two things align. The patterns lead to an answer that also corresponds to reality. But when it predicts a birthday, following the same process doesn't guarantee that the date corresponds to reality.
The same principle applies when we ask AI for scientific research.
AI has learned how scientific studies are written, how their titles relate to their subjects, how researchers and findings are presented, and how different pieces of information connect.
Using those learned relationships, it can generate a scientific reference that fits the subject we're researching.
Sometimes that reference belongs to a real study. Sometimes it describes a study that never existed.
From the perspective of how the answer was generated, AI followed the same process in both cases. But from our perspective, one answer corresponds to reality and the other doesn't.
And this is where our assumptions create the problem. When we receive a scientific reference, we assume AI searched for the study, found it and gave us the information. But unless it actually performed that search, it generated the answer from learned patterns and relationships.
Just as predicting a cat's birthday doesn't establish when the cat was born, generating a convincing scientific reference doesn't establish that the study exists.
That is why generating an answer and verifying whether it corresponds to reality are two different tasks.
Anthropic: Why Can't We Just Fix AI Hallucinations?
In March 2025, Anthropic published Tracing the Thoughts of a Large Language Model, a study exploring what happens inside Claude when it generates answers.
While OpenAI investigated why hallucinations arise during training, Anthropic explored something equally interesting: why AI can still hallucinate even when features have been developed to prevent it.
Because that's the obvious question, isn't it? If AI generates information that isn't true, why don't developers simply teach it not to answer when it doesn't have enough information?
Well, Anthropic has been doing exactly that.
Claude has undergone anti-hallucination training, designed to encourage it to acknowledge when it doesn't have enough information rather than invent an answer.
The researchers wanted to understand how this protection actually works inside Claude.
And they discovered something surprising.
Claude has two competing features that influence whether it answers a question.
The first is active by default and encourages Claude to respond that it doesn't have enough information to answer.
The second activates when Claude recognises familiar information. When this happens, it can override the first feature, allowing Claude to generate an answer.
To demonstrate how these features interact, the researchers used two names.
When asked about Michael Jordan, the famous basketball player, Claude recognised the name and answered.
But when asked about an unfamiliar person named Michael Batkin, Claude declined to answer because it didn't have enough information.
So far, the protection was working.
The researchers then intervened in Claude's internal activity. They artificially activated the feature associated with recognising a known answer, even though Claude didn't have enough information about Michael Batkin.
And suddenly, Claude began inventing information about him, claiming he played chess.
By activating one feature, the researchers caused Claude to generate an unsupported answer despite its existing protection against hallucinations.
But here's the interesting part.
They discovered that the same thing can happen naturally, without anyone intervening.
Sometimes Claude recognises a name without having enough information about the person. That recognition can activate the feature that encourages it to answer, overriding the feature that would otherwise make it acknowledge uncertainty.
And once Claude proceeds to answer, it can generate a convincing story based on the patterns and relationships it has learned.
And that's what makes this research so interesting.
Even with protective features in place, hallucinations can still happen. And we could be the exception that encounters one. There is no bulletproof way to guarantee that AI will never hallucinate.
What Can We Learn From These Studies?
That we should always remember what an LLM is: a prediction mechanism, and that is what it was trained to do.
OpenAI showed the principles of what an LLM was trained to do, and how. Anthropic showed what happens when humans intervene, demonstrating that even when artificial protective mechanisms are introduced, an LLM can still revert to what it was designed to do, which is predict.
Both studies showed that generating an answer and establishing whether that answer is true are two different tasks.
That means AI predicted, to the best of its ability, what the right answer should be. Therefore, it wasn't lying, as it was designed to predict.
So why do we still think that hallucination is lying?
Because we've forgotten what lying actually means.
What Is Lying?
Lying by definition is when we knowingly communicate something we believe is false, with the intention of making someone else believe it's true.
For example, imagine I ask you whether you read an article I published yesterday. You didn't, but you tell me you did.
That's a lie. You know you haven't read it, but you deliberately tell me something else.
But what if I didn't ask you directly whether you read it? Instead, I ask you, "Did you like the article I published yesterday?" And your answer is, "It was great!"
Is that lying?
Some could say it is. But is it? Technically, you didn't lie, as we could argue that you don't need to read it in order to know it's great. Right? It could be a valid point of view. Some will still say, "But how would you know? You didn't read it." And in their world, it would be considered lying.
But that is not the definition. It's a perspective, a way of defining subjectively what lying is in their world.
Because lying, by definition, is being asked a direct question that you know the answer to but decide to change.
"My article is great. :) Even without reading it." is not considered to be lying by definition. It's a subjective understanding of what lying means from a personal point of view.
Do Humans Have Hallucinations?
Now, let's take our discussion about lying one step further.
What happens when we genuinely believe something is true, even though it never happened?
Interesting, right? Yes, there is such a thing, and it has even been demonstrated through research.
I'm not talking about misleading someone or choosing which parts of a story to tell. I'm talking about being absolutely certain that something happened, remembering the details, and discovering later that the entire thing was wrong.
We are talking about what we call memory. The one thing we are sure we know, as we have experienced it ourselves. So researchers used that to demonstrate that human memory is not always an exact representation of what happened.
Talarico and Rubin: Remembering September 11 Differently
In 2003, psychologists Jennifer Talarico and David Rubin published a study called Confidence, Not Consistency, Characterizes Flashbulb Memories.
The researchers wanted to understand something interesting about human memory: why are we so confident about remembering important events?
On September 12, 2001, just one day after the September 11 attacks, they asked 54 university students to describe how they first heard about what happened.
They also asked them to describe an ordinary event from around the same time.
Over the following months, the researchers asked the students to recall those experiences again.
And what did they discover?
The students' memories of September 11 changed over time, just like their memories of ordinary events.
But there was one important difference.
While the students became less confident about their memories of ordinary events, they remained highly confident that they remembered September 11 correctly.
Even when their descriptions no longer matched what they had originally written.
Think about that for a moment.
They had recorded their memories just one day after the event. The researchers could compare those original descriptions with what the students remembered later.
Their memories had changed, but their belief that they remembered correctly had not.
They weren't deliberately lying. They genuinely believed their later recollections.
The researchers demonstrated that feeling certain about a memory doesn't necessarily mean that the memory is accurate.
And that is particularly interesting for our discussion.
Because we humans can produce an account of something we believe happened, confidently present it as true, and only discover that our memory has changed when someone compares it with the original information.
Roediger and McDermott: Remembering Words That Were Never There
Another study published in 1995 by psychologists Henry Roediger and Kathleen McDermott demonstrated something similar, using a much simpler experiment.
Participants were given lists of related words.
For example: bed, rest, awake.
All these words are connected to the idea of sleep.
But the word "sleep" itself was never included in the list.
When participants were later asked to remember the words, many confidently recalled "sleep" as one of them.
Think about that for a moment.
The word was never presented, but the relationship between the other words was strong enough for people to remember it as if it had been.
The researchers called this a false memory.
So, Are We Lying to Ourselves?
In both experiments, people reported information that didn't correspond to what actually happened.
But they weren't necessarily trying to deceive anyone. They genuinely believed their recollections.
The second experiment is particularly interesting when we compare it with AI.
People were given information containing strong relationships between words. They remembered another word that fit those relationships, even though it was never presented.
AI also learns relationships between information and can generate an answer that fits those patterns without the answer necessarily being factual.
I know that when you look at that, you'll say it's not a comparison because we look at AI as a tool with information available to it. And we humans don't have the same mechanism. Adding a word and making up an entire research study are different.
I understand that perspective, but for me, it's because we don't understand what AI is or how it works. But at the end of the day, it's also because we don't understand ourselves. AI was developed based on us, and if we can hallucinate, so can it.
The Bottom Line
The answer to the question of what hallucination is?
Should be what we identify when we check an AI-generated answer and discover that it doesn't correspond to reality.
AI generates answers using patterns and relationships. It doesn't know whether the information is true or false. And having the ability to verify information doesn't guarantee that it will.
verifying the information your LLM provides you is a specific task. you, as the human is responsablie to verify any outcome.
Just to make sure :)
Because a convincing answer isn't necessarily a fact, and a false answer isn't necessarily a lie.
It goes for an LLM, and it goes for the thoughts and memories in our heads. :)
Personal Note — Sarah Lahav
I have to admit, I had a lot of fun researching this one. :)
What started as a simple question about AI hallucinations turned into a discussion about lying, truth, memory, and how little we actually understand about the way we process information ourselves.
And that's what I love about researching AI.
Every time we try to understand something about how AI works, we end up discovering something about ourselves.
And that is the part I want everyone to know. AI is developed in our image. Our thought processes are the same. We just are not aware of what is going on in our minds unless you are a researcher and that kind of thing fascinates you.
This is why it's hard for almost all of us to see that AI is not doing something to us. It just reveals something in us we were not aware of.
And that is the big gift AI is giving humanity, and that is why I am on my quest. I talk about AI, but I want every person who finds us to learn something about himself, about the magic we have, the power of how we all create reality, about the power of thought.
If you want to learn more, you can read my book AI Thinking.
And if you want me to come in and help you and your bots :) let's talk ... because there is something there, instructions or information, they are missing in order for them to be able to do a great job. And as the bots' spokesperson, I am here to help. :) Let's talk.