Feels like these days it would be easier for an AI to deal with a captcha than for a human being to complete it.
Today I had possibly my worst captcha-related experience.
This was not deemed acceptable.
I had to also click the middle one on the right. Even though it needs quite a lot of imagination to see it as a set of traffic lights.
Then I was allowed to proceed.
Feeling like a proper grumpy old man having to deal with this nonsense.
I can imagine an LLM could hallucinate that the middle right image is a traffic lights. But a human being would have to be pretty high to come to the same conclusion.
There’s been a lot of talk about how ChatGPT (and similar) hallucinate when they generate text. It’s an unfortunate choice of words to use because it sounds like some kind of side-effect. But in reality it’s central to the way that Generative AI works.
Do you know why they’re called hallucinations?
Before we had ChatGPT making text generation popular, creating images was the the big thing in Generative AI. When researchers looked into how these images were being generated they saw something that looked, well, like hallucinations.
And of course we had many, many cases where image generation tools would “hallucinate” extra fingers or lopsided ears etc.
When you’re talking about image generation, a word like “hallucinate” works quite well. If you see a generated image that looks like an acid trip or has wonky ears in or whatever then you tend to consider the whole image a hallucinated image. You treat the whole image as not real.
But when you’re seeing a ChatGPT output with 3 incorrect facts in, you don’t think of the whole text as a hallucination. You tend to consider those 3 incorrect facts as a hallucination, distinct from the rest of the text. Weird how people tend to treat pictures as “one thing” but text as “a bunch of different things stitched together that we can consider separately”.
There have been attempts to use a different word when talking about text generation: Confabulation. According to the National Institute of Health, confabulation is:
a neuropsychiatric disorder wherein a patient generates a false memory without the intention of deceit
But too late! “Hallucination” had already taken its grip on the public’s imagination.
When you look at a ChatGPT output you should expect that the whole thing is one confabulation. This confabulation is often useful but it’s still just made up text, however cleverly it is made up.
There are all kinds of guardrails that people have to build around the LLMs to help guard against these confabulations/hallucinations. But ultimately, when you’re generating images or text with GenAI, you are asking a machine to hallucinate / confabulate for you. Please remember that.
Be good, work hard, get lucky 