Why AI Hallucinates Like a Renaissance Painter

Recently my family took a trip to Florence, Italy and visited the Uffizi Gallery where some of the world’s greatest art can be admired.

We were looking at one particular portrait and my wife said to me, she was beautiful wasn’t she? I asked “was she”? Not because the portrait didn’t show a beautiful woman, it did. But because that portrait was an interpretation of what the artist saw in front of him and what people expected of a noble man’s wife to look like. Did the artist interpret the portrait based on what he saw in front of him or dare I say it did he hallucinate?

Society expected the nobles to be beautiful and the art reflected this expectation as interpreted by the artist and as commissioned by the aristocratic family.

It got me thinking of AI hallucinations, the model predicts answers based on the question and gives an answer that’s probably correct and likely to please the end user. Essentially the model interprets what it learns about the world around it (training data) and frames the response in a way likely to please the viewer (prompter).

So this moves me in to thinking how do we stop artists or AI going wildly off track? The artist likely had a group of associates around him, plus the person who’s commissioning the art keeping the artist in check to ensure the artwork delivered was realistic enough to be passable as a realistic representation but with certain artistic intent to smooth out the (literal) wrinkles in the subject. Any AI output needs the same checks and balances, given anything you get out of AI will always be a probabilistic representation of what the prompter is asking for, not exact but close. But how do we ask it to check for realism?

This is where Optimizely Opal can really help through what are essentially automated checks (evals) that play the same role as the artist’s associates and ultimately the paying customer. Just as a patron would send a portrait back for revision if it strayed too far from a believable likeness Opal’s evals score AI output against defined criteria: accuracy brand tone and structure before anything reaches the end user. It’s the checks and balances the artist had built into the AI workflow itself.

Would you trust the artist's bias on their own or would want to see how Optimizely Opal handles it? Happy to show you personally.