Stop Calling It a Hallucination … the Myth that is AI

Are you getting tired of the language used to describe artificial intelligence (AI) that is subliminally softening the reality of what these systems are actually doing? I certainly am …

We are told that models are not perfect, as though they are accurate mirrors of reality with a few defects. In truth, an AI model is a compromise. It is a statistical representation shaped by incomplete data, design choices, pressures (often commercial and sometimes deceptive), safety controls and computational limits. It does not reproduce reality. It generates an approximation of it. How many of you would be happy with an ‘approximation’ of salary being paid?

The term ‘hallucination’ is more misleading. A hallucination is a human experience involving a false perception. An AI model perceives nothing and believes nothing. When it invents a legal case, fabricates a quotation or presents false information with confidence, it is producing an unsupported fabrication. Imagine replacing an accounting system that fabricated your Profit & Loss report or provided differing answers each time. You would throw it out!

In almost any human context, confidently presenting invented information as fact would be described as lying, misrepresentation or negligence. AI cannot lie in the human sense because it lacks intent but the effect on the recipient may be indistinguishable from a lie. So call it what it is a LIE. The commercial entities that are selling these service bury the truth in legal ease whilst hyping the real hallucination that is AI.

The same problem applies to ‘bias.’ Bias is often presented as an accidental flaw within an otherwise neutral system. Yet models reflect choices about training data, labelling, filtering, optimisation and acceptable behaviour. What is called bias may actually be design error, representational distortion or institutional preference embedded in code. All of which distorts the reality and truth of what is produced to become untrustworthy or at best questionable..

Anthropomorphic terms such as ‘understands,’ ‘knows,’ ‘reasons’ and ‘thinks’ further encourage misplaced trust. Fluency is mistaken for comprehension, confidence for certainty and plausibility for truth.

AI can be useful without being human, truthful or neutral, but governing it responsibly requires more honest language. We should describe fabricated outputs as fabricated outputs, structural limitations as compromises and embedded assumptions as design choices. Until we do, the first distortion may not be generated by the AI at all. It may be created by the language we use to sell it.

So the myth is maintained, the machine does not guess, it ‘reasons’; it does not fabricate, it ‘hallucinates’; it is not constrained, biased or commercially optimised, merely ‘imperfect.’ There is even a certain perverse honesty in Anthropic naming its powerful frontier model Mythos, a word rooted in myth, narrative and constructed meaning, while the wider industry continues selling the myth that fluent AI is inherently truthful AI. Perhaps the name is accidental; perhaps it is an unusually candid piece of branding. Either way, wrap statistical prediction in a human voice, add a reassuring interface and call it intelligence and uncertainty becomes a product. The technology may be artificial but the mythology, confidence and commercial theatre surrounding it are entirely human.