Hallucinations in brains and machines
We all have a blind spot where the optic nerve passes through the retina. But we’re generally unaware of it: the rest of the visual pathway “fills in” the gaps by inferring what’s probably out there, based on what we can see.
One common eye disease amongst older people is age-related macular degeneration, where large parts of the retina stop working, causing multiple large blind spots. But as with the blind spots we’re all born with, most people with macular degeneration aren’t aware of these blind spots because the brain continues to fill in the missing information.
But at some point, there’s so little information coming in that the brain’s inference goes haywire: it has nothing to work from, so people start seeing random patterns like grids or geometric shapes, or more complex scenes like faces, little people in costumes or whole landscapes. As an aside, this can be quite terrifying: there you are, minding your own business, when you suddenly see a family of tiny Edwardians looking up at you from the table. Or you look at the empty armchair, where your spouse routinely sat for many years until their death – and now you can see them again.
But these visual hallucinations are not a sign of mental health problems. They are just the brain filling-in and guessing what’s probably there, even if your retina is silent on the matter. (Many other forms of hallucination are symptoms of mental health problems, drug use etc.)
You can read more about visual hallucinations caused by sight loss, also known as Charles Bonnet syndrome, from the RNIB, the Macular Society or the NHS.
Turning to generative AI and GPT-3, as seems inevitable these days, there’s been some debate about whether they can be said to “hallucinate” when they generate output with claims that have no basis in reality, and in particular, no basis in their training data.
For example, Carl Bergstrom argues that ‘Your chatbot is not “hallucinating”’ because this behaviour isn’t a pathology – it’s baked into the very design of large language models. My somewhat pedantic (hello!) response is that not all hallucinations are pathological: perceptual filling-in is the brain working correctly, even if the perception is false.
Bergstrom also argues that using words like “hallucinate” leads us to anthropomorphize not-actually-intelligent AI models, and treat them as if they have sense impressions of the world, albeit imperfect ones. To me, it’s a useful analogy rather than a literal statement: describing the prompt given to generative AI models as a stimulus, and the generated text as a response seems like a sensible way to discuss their behaviour. Prompts and training data are the nearest equivalent to human sensory experiences for generative AI models.
For what it’s worth, I strongly agree with most of Bergstrom’s position: GPT-3 is best thought of as a large-scale bullshit-producer, and the Big Tech companies behind it and it’s competitors are hiding their own responsibilities by blaming hallucinating (and so implicitly autonomous) models or malevolent users for any bad output.
But here’s a final twist: people have more confidence in filled-in inferred perception than they do in directly perceived stimuli (at least in laboratory studies). So perhaps we shouldn’t be surprised at the confidence that GPT-3 has in its own bullshit!