Habr AI: Why LLM Hallucinations Look Less Like a Math Bug and More Like a Human Lapse

A Habr AI column argues that model hallucinations should be viewed not only as an engineering defect but also as a reflection of familiar human thinking failures. The author approaches the problem like a clinician: instead of outrage, he proposes working out where exactly the system loses the frame of the task and why it confidently produces a wrong answer.

Why this irritates everyone

Reactions to LLM hallucinations are usually sharp: the user expects the model to stick to the context, but gets a confident yet wrong statement. In response, developers long hid behind a classic set of explanations such as GIGO and RTFM, then began reinforcing products with RAG and additional constraints. But the complaint itself never went away: people expect a machine to behave like a "healthy" interlocutor who does not lose the thread of a conversation and does not invent facts on the fly.

The column flips this logic. The author writes that the first failures of neural networks did not surprise him, because working with disordered thinking is an ordinary part of clinical practice. Hence the main thesis: the irritation comes not only from the error itself, but from the inflated expectation that AI should think more cleanly and with more discipline than a human.

• The user expects literal compliance with the instruction
• The engineer points to the quality of input data
• Product teams add RAG, filters and checks
• But the model itself still inherits familiar failure patterns

A bug in the protein neural network

The strongest part of the text is an everyday example in which it is no longer the model that errs, but a human. The author gives a simple premise: "According to my passport I'm Olga. At home they call me Alena. Pick one of the two." Formally there are only two options, yet the "protein neural network" often answers: "You're Lena." This, in the author's view, is what a breakdown of human prompt engineering looks like on an elementary task.

The point of the example is that the brain does not like holding contradictory constraints for long. Instead of a strict choice from the given set, it quickly drifts into associations (Alena, Elena, Lena) and substitutes the statistically familiar answer. The error comes not from malice and not from a complete failure to understand language, but from an urge to cut a path to a "plausible" conclusion. So the complaint "the neural network ignores context" in a sense comes back to the human as well.

What breaks in the head

The author describes this failure as a constraint violation, a breach of the task's parameters. Viewed through psychology, the brain drops the original prompt, pulls out the most convenient associative option and delivers it with complete confidence. In terms of cognitive biases, this resembles jumping to conclusions: leaping to a verdict without adequately checking the conditions. To a reader from the AI industry, this sounds almost like a familiar LLM bug: the constraint existed, but the system did not hold it to the end of generation.

"AI is not broken. Maybe we have recreated our own bug."

The author's practical conclusion is unexpectedly down-to-earth. Some of these failures are fixed not only by new architectural tricks but also by proper discipline in interaction. If the task is unclear, it is better to ask again than to guess. This logic works for people and for models alike: the better the boundaries are set, the lower the chance that the answer veers into confident improvisation. That is exactly why the debate about hallucinations cannot be reduced to the question of whether there is enough data in the system.

What this means

The column is useful because it removes unnecessary drama from the topic. LLM hallucinations remain a serious product problem, but the clinical view shows that part of their nature may be closer to human cognitive shortcuts than to a mysterious "madness of the machine." For developers, this is an argument not only to improve models and retrieval, but also to design interfaces in which it is easier for the system to clarify a request than to err gracefully. For users, it is a reminder that a confident tone does not yet equal understanding.