Chatbots Do Not Hallucinate, They Confabulate

If you have any familiarity with ChatBots and Large Language Models (LLMs), like ChatGPT, you know that these technologies have a major problem, which is that they “hallucinate.” That is, they frequently generate responses that are nonsensical or factually inaccurate. Research estimates suggest that this problem emerges between 3% and 27% of the time, depending on context and model.

For example, when my family was playing around with ChatGPT, we wanted to see if it “knew” who my father was. My dad, Dr. Peter R. Henriques, is a retired professor of history who has written several books on George Washington. ChatGPT respond correctly that my dad was a biographer of Washington; however, it also claimed, wrongly, that he wrote a biography on Henry Clay. This is an example of a hallucination.

Where do hallucinations like these come from? LLMs like ChatGPT are a type of artificial intelligence that run algorithms that decode content on massive data sets to make predictions about text to generate content. Although the results are often remarkable, it also is the case that LLMs do not really understand the material, at least not like a normal person understand things. This should not surprise us. After all, it is not a person, but a computer that is running a complicated statistical program.

Hallucinations can emerge from the way LLMs combine source data from different domains, misinterpret the query, or lack the necessary specific knowledge. The more basic issue is that it is running a predictive program, rather than engaged in human understanding.

And it is here that we come to the point of this blog, which is to explain why describing these errors as “hallucinations” is a mistake. It probably reflects the failure of the computer science community to understand the human mind. I don’t fault them, because my field, psychology, has not done a great job clarifying its subject matter. Nonetheless, the fact that these LLM errors came to be called hallucinations is diagnostic of a profound misunderstanding of the difference between a computer and a human mind.

Let’s start with an example of a hallucination in a real person who was dealing with schizophrenia that I shared in my recent book, A New Synthesis for Understanding the Problem of Psychology: Addressing the Enlightenment Gap:

“There they are!” she said earnestly, pointing to bugs that I could not see. “They just crawled out of my skin and now are flying around the room!”

I was sitting across from a woman I was working with who was diagnosed with schizophrenia. She had both haptic (touch) and visual hallucinations in the form of bugs that would crawl out from under her skin and fly around and, she feared, would infect other people.

At the moment, we were sharing a meal from McDonalds. She was convinced that she was infectious and that she would infect people with these insects if she ate a meal with others. This was a form of cognitive behavior therapy to help her see that she was, in fact, safe to be around. We were sharing the meal to test out the validity of her belief. Thankfully, I was not infected by insects.

Her belief that she would infect others was technically called a delusion. A delusion is a belief about the state of the world that is idiosyncratic and highly unlikely to be accurate. These delusions were based on false perceptions she experienced in her subjective conscious experience. That is, they were based on hallucinations.

As detailed in my book, I have developed a new, unified approach to psychological science which divides mental behaviors into three domains or layers (see here). In my system, hallucinations take place in the domain of Mind2, which is the domain of subjective conscious experience in animals and humans. Mind2 emerges out of the domain of Mind1, which is the domain of neurocognitive activity that operates subconsciously. Finally, the talking portion of the human mind is the domain of Mind3.

The insects my patient experienced were in her Mind2, and she reported on them via the domain of Mind3. She saw and felt them as being present and real, even though those perceptions did not correspond with reality from an exterior point of view. That is the nature of hallucinations. They are events in Mind2 that are experienced as real, but are not.

Artificial Intelligence Essential Reads

The Rise of the Techno-Energy Complex

How Leaders Can Use AI to Bring Business Objectives to Life

With this definition, we can see both why computer scientists attached the word “hallucinations” to describe the LLM errors, and why it is wrong to do so. The reason it makes some sense is because both the errors and the actual hallucinations humans experience involve misconstruing something as being there or true or real when it is not.

However, the word is a misnomer because hallucinations involve the subjective world of Mind2, and emerge out of our embodied way of being in the world. LLM’s are not primates, and they don’t have anything like a Mind2. They are what philosophers call “zombies,” which means that they have no inner experience whatsoever. If they have no inner experience, then they cannot hallucinate.

If hallucinations are not the right word for these errors, what is a better word? If we are pulling from the world of psychiatry and clinical psychology, a much better word is confabulation. Confabulation is when individuals generate false content without meaning to deceive, often to fill in some expected social role.

Some classic examples of confabulation come from split brain research. This is when folks had an operation that separated the line of communication between the left and right hemispheres. Researchers found that they could flash a command to the right hemisphere, which the person would start to follow. However, the left hemisphere, which houses the language system, would not know what was happening or why. If the person was asked what they were doing, they would often “confabulate,” which means they would generate a reason that sounded ok, but was basically made up.

For example, the researchers might have flashed a command to the right hemisphere to go get a drink. The person would then start to follow that command. However, if the researchers asked why the person was getting up, the left hemisphere would not know the answer, but would just confabulate and say, “I want to get a drink as I am feeling a little thirsty.” Verbal confabulations thus involve generating false or confusing content without intent to deceive based on faulty memories or problems with neurocognitive functioning.

In sum, calling LLM errors hallucinations is a misnomer. They do not have a Mind2, and thus are not reporting on perceptual experiences that they have but do not correspond to reality. What they are doing is generating language modeling that is nonsensical or false, without any intent to deceive, based on problems with retrieval, source, and comprehension.

Given this analysis, it is clear that when we see errors like the claim my dad wrote a biography on Henry Clay, we should say that LLMs confabulate rather than hallucinate.