When Machines Fill In the Blanks

Why “Hallucination” Is the Wrong Word for the Right Phenomenon

I have never quite understood why the term “hallucination” causes so much confusion when applied to large language models. If you set aside the word itself and look at what the phenomenon actually involves, it turns out to be remarkably human in nature. Educators, in particular, should find it familiar.

Consider the following hypothetical scenario to illustrate the point. A teacher poses a question to a student. The student, not knowing the answer, does not respond with “I don’t know.” Instead, they reach for the most plausible answer available; the one that seems most likely to satisfy the teacher. The student fills in the gap with what feels correct, shaped by the context of the lesson and the tone of the question. This instinct to produce a probable response rather than admit uncertainty is so deeply embedded in classroom culture that we rarely stop to think about it.

This is, in functional terms, exactly what a large language model does when it “hallucinates.” It generates the most statistically likely continuation of a prompt, shaped by the vast patterns in its training data and the implicit expectation that it should produce an answer.

For a long time, this parallel seemed so intuitive to me that I assumed everyone interpreted the term the same way. That changed recently during a comment exchange on a LinkedIn post, where I found myself in a discussion with a fellow educator about AI reliability. As we went back and forth, I realized this colleague understood “hallucination” as something fundamentally different from what I meant by it. For them, a hallucination signaled an abnormality; a malfunction. If a brain hallucinates, the reasoning went, it is broken. Something has gone wrong at the level of the system itself. And if an AI hallucinates, by extension, the technology must be flawed in some deep, perhaps irreparable, way.

That conversation stayed with me. It made me realize that the problem with “hallucination” is not primarily technical. It is rhetorical. The word carries connotations that actively mislead educators about what these systems are doing and why.

The Term Has a Longer History Than Most People Think

The popular assumption is that “hallucination” was coined recently, perhaps by marketing departments looking to humanize chatbots and soften the perception of their errors. But the historical record tells a different story. The term has been in use within computer science for roughly three decades, and its origins have nothing to do with large language models.

The earliest consistent use appears in Computer Vision research, specifically in the field of super-resolution. In 2000, researchers Simon Baker and Takeo Kanade published a paper titled “Hallucinating Faces,” which addressed the problem of reconstructing high-resolution images from low-resolution, pixelated inputs. Standard methods at the time simply averaged surrounding pixels, producing images that were larger but inevitably blurry. Baker and Kanade proposed something more ambitious: if a system had been trained on enough high-resolution face images, it could infer the missing details of a new, blurry face by drawing on learned statistical patterns. The additional pixels generated through this process were, in their terminology, “hallucinated.”

In this original context, hallucination was a positive achievement. A system that could hallucinate effectively was considered superior to one that could not. The word described a form of constructive synthesis: the generation of plausible details from internal knowledge in the absence of external ground truth. In this way, the mechanism was inferential, Bayesian even. The system filled in the gaps intentionally, and that was the whole point.

From Useful Feature to Dangerous Flaw

The connotation shifted in 2015 when Google released DeepDream. This project explored the internal representations of deep neural networks by running them in reverse: rather than asking a network to classify an image, researchers optimized input images to maximally activate specific neurons. The results were the now-familiar swirling, psychedelic images of multi-eyed creatures emerging from clouds and static. The tech community, the press, and even the researchers themselves immediately described these outputs as “hallucinations.”

DeepDream matters for the history of this term because it made the mechanism visible. The network was literally “seeing” patterns that were not present in the input because its learned priors overwhelmed the actual sensory data. It was also when researchers began drawing explicit parallels between artificial neural networks and the biological visual cortex. Studies noted that DeepDream’s outputs bore a striking phenomenological resemblance to biological hallucinations induced by psychedelic compounds. The “Hallucination Machine” project by Suzuki and colleagues in 2017 used similar technology to simulate altered states of consciousness in virtual reality, further entangling the technical and biological meanings.

The term then migrated into Natural Language Processing through two intermediate steps. First, in Neural Machine Translation, researchers in the mid-2010s observed that translation models would occasionally produce fluent, grammatically perfect sentences that were entirely unrelated to the source text. A German sentence about a financial report might be translated into an English sentence about a different topic altogether, with impeccable syntax. Researchers described these as “hallucinations” because the model had become detached from its input and was generating text based on its internal language patterns alone.

Second, in 2018, Rohrbach and colleagues published a paper on “Object Hallucination in Image Captioning,” demonstrating that models describing images would frequently mention objects that were not actually present. A model might see a beach and describe a surfboard simply because those words co-occur frequently in training text. This work introduced the CHAIR metric (Caption Hallucination Assessment with Image Relevance) and established a formal framework for measuring the gap between generated output and source reality.

By the time GPT-3 arrived in 2020, the term was already deeply embedded in the AI research lexicon. It was imported into large language model discourse from Computer Vision, through Machine Translation and Image Captioning. This was a natural inheritance of technical vocabulary, not a marketing invention.

Why “Hallucination” Stuck, and Why It Misleads

The term persisted for three technical reasons. First, unlike a “bug” or “glitch,” which implies a breakdown of function, hallucinations are characterized by their fluency and confidence. The model is operating exactly as designed in terms of probability maximization; it is predicting the next token effectively. It simply fails in terms of truth. “Hallucination” captures this quality of “successful failure” more precisely than “error” does. Second, the generation mechanism parallels the original Computer Vision usage: the model creates details from internal patterns rather than retrieving verified facts. Third, the adjacent fields of Computer Vision and Image Captioning had already standardized the vocabulary. Researchers adopted existing terminology from their peers.

These are defensible reasons. The problem is that these are reasons appreciated primarily by researchers, not by the broader public. In public discourse, “hallucination” carries very specific medical and psychological associations that distort understanding.

In clinical psychiatry, hallucinations are perceptual experiences without external stimuli. They are symptoms of conditions such as schizophrenia and are associated with dysfunction, pathology, and a break from reality. When educators encounter the term within the context of AI, many understandably map it onto this clinical meaning. The result is the interpretation I encountered in that LinkedIn exchange: if the system hallucinates, it must be malfunctioning. Something is broken. The technology itself is unreliable in some fundamental sense.

But this reading misidentifies the nature of the phenomenon. A large language model that generates plausible but inaccurate text is not malfunctioning. It is doing precisely what its architecture was designed to do: predicting the most probable next token given all preceding tokens. The architecture is one of generation, not retrieval. These systems do not look up facts in a database and occasionally get the lookup wrong. They construct text, token by token, from statistical patterns learned during training. Generating plausible continuations is the feature. The lack of a built-in truth-verification mechanism is the limitation.

Confabulation: A More Precise Alternative

A growing consensus in cognitive science and medical AI research suggests that “confabulation” would be the technically superior term. In neuropsychology, confabulation describes the production of fabricated, distorted, or misinterpreted memories without the conscious intention to deceive. It occurs in conditions like Korsakoff syndrome, where patients fill in gaps in their memory with plausible but false narratives. They are not lying. They are not perceiving things that do not exist. Instead, they are reconstructing memories from incomplete information, and the reconstruction goes wrong.

This maps onto what large language models do with far greater precision. These systems do not store text; they store statistical weights. When they generate a response, they reconstruct a plausible output from those weights, filling in gaps with what is statistically likely rather than what is factually verified. Geoffrey Hinton, widely regarded as one of the founders of deep learning, has argued along these lines. Hinton contends that human memory itself is not a file storage system but a reconstructive process. We do not retrieve memories intact; we recreate them from distributed patterns, often filling in details that were never there. He suggests that the mechanism is functionally identical in both biological and artificial neural networks: reconstruction under uncertainty.

“Confabulation” captures this shared mechanism without the clinical baggage of perceptual breakdown. It removes the implication that the system is “seeing things” and replaces it with the more accurate description that the system is “filling in the blanks.” For educators, this reframing matters. A system that confabulates is one whose outputs need to be checked against evidence, the same way a student’s probable answer needs to be checked against the textbook. A system that hallucinates, by contrast, is one that cannot be trusted at all.

Living With an Imperfect Term

Despite its limitations, “hallucination” is unlikely to disappear. Thirty years of technical literature have cemented it. The research community uses it; the media uses it; users understand it. And there is even an argument that its emotional charge serves a purpose: it alerts people to the seriousness of the problem. A “confabulation” sounds clinical and perhaps dismissible. A “hallucination” commands attention.

What matters more than replacing the term is understanding it correctly. For educators, the essential insight is that hallucination is not a sign of a broken system. It is a structural property of how generative language models work. These systems were designed to predict the probable rather than retrieve the true. When Baker and Kanade taught a computer to “hallucinate” eyelashes onto a blurry face in 2000, that was a triumph of inference. When a large language model generates a plausible but fabricated legal citation in 2026, the underlying mechanism is the same: generative inference from learned patterns. The difference lies in our expectations, not in the engineering.

Understanding this changes how we respond. Rather than treating hallucinations as evidence that the technology is unreliable, we can treat them as a reminder that these tools require the same critical engagement we would apply to any source. The student who offers a probable answer in the classroom is not broken. Neither is the machine that does the same. Both require a teacher, or a user, who knows to ask: “Is that actually true?”

That, in the end, is the question that matters far more than what we call the phenomenon.

---

The images in this article were generated with Nano Banana Pro.

P.S. I believe transparency builds the trust that AI detection systems fail to enforce. That’s why I’ve published an ethics and AI disclosure statement, which outlines how I integrate AI tools into my intellectual work.