AI Doesn’t Hallucinate

I have also started to push back on the phrase “AI hallucinations”.

(There’s a theme developing here.)

Not because these errors are not real, but because the use of the language is misleading.

Hallucination is an anthropological and clinical term. It describes a human experience. One rooted in perception, consciousness, and biology. When we apply it to machines, we quietly smuggle in the idea that something mind-like is happening.

It is not.

Large language models are not seeing things. They are not confused. They are not imagining. They are producing statistically plausible outputs based on patterns in data. When those outputs are wrong, fabricated, or misleading, the issue is not perception. It is error, uncertainty, and probabilistic completion doing exactly what it was designed to do.

The danger of using the word hallucination is that it sounds accidental, almost endearing. As if the system briefly drifted from reality rather than confidently producing something untrue.

Subscribe

I would much rather we used language that reflects what is actually happening. Fabrication. Confabulation. Unverified output. Model error under uncertainty. None are catchy. All are much more honest.

The words we use shape how seriously we take the risks, how we design controls, and how willing people are to challenge outputs rather than defer to them.

As with mindset and superpowers, this is not pedantry (well possibly a wee bit). It is about the clarity offered by future literate leaders.

And clarity really matters when systems sound convincing even when they are wrong.

(Couldn’t find source for this image)

More to come on this error type and its consequences.

What language do you use when explaining this to non-technical audiences?

#scottspeaks #futureliteracy