Hallucination is an old term for a new concept.
I first heard the term “hallucination” used in connection with AI during a presentation we organised earlier this year with Greg Shove, CEO of Section, hosted by Journey. Greg explained it (and reiterated in Professor Scott Galloway’s No Mercy / No Malice newsletter):
“The leading AI models still don’t work reliably — and they’re prone to ketamine-like hallucinations, Silicon Valley-speak for ‘they make shit up.’”
While various studies offer percentages of how much AI platforms hallucinate (more on that below), I would bet that I—and most humans—“hallucinate” more often. Not because we intend to deceive but because the way our brains work sets us up for creativity and innovation, but also making memory accuracy difficult.
In this article, we’ll explore how AI hallucinates, how humans do it, and how both can have surprising upsides. The downsides? Well, we’ll cover those in another piece.
How much does AI hallucinate?
There’s no universal figure for AI hallucination rates; they vary depending on the complexity of tasks and the quality of training data. For example, research on large language models like GPT-3 and GPT-4 shows hallucination rates ranging from 3% (GPT-4) to 27% (Google PaLM-Chat) when answering factual questions or summarising content.
Interestingly, while researching for this article, I asked ChatGPT-4 how often AI hallucinates. It claimed a range of 10% to 20%, but I struggled to find external sources to back that up. This inconsistency is a perfect example of why AI hallucinations are a hot topic.
Another example: I asked ChatGPT-4 why Venus has a retrograde rotation, and it confidently stated that Venus is the only planet with this feature. Every astronomy nerd would jump in to say, “That’s not true!” Uranus also has a weird retrograde rotation, and even Pluto rotates backwards. (Yes, Pluto is a planet!)
Why humans are unreliable narrators
We humans often “hallucinate”, too, but it can be a result of cognitive processes rather than deliberately making things up.
My pig fell into the pool.
As a teenager, I tried to avoid mowing the lawns by fencing them off and letting our sheep graze instead. While leading them past the pool, our pet Kunekune pig, Rosie, tagged along but fell in, sinking to the bottom. Fortunately, Rosie knew how to swim and made it to the surface. We both got a huge fright.
Over time, I began telling the story as if I was alone when it happened, but my sister later reminded me she was there. Somehow, I’d conflated two memories—the first time I fenced off the lawn and the time Rosie fell in. This mix-up is a good example of confabulation, where our brains unintentionally fill in memory gaps and a flashbulb memory, which is vivid but not consistently accurate.
What is confabulation?
Confabulation sounds much more exciting than it is. It might evoke thoughts of elaborate deceptions, but in reality, it refers to the brain’s unconscious attempts to “fill in the blanks” when memory fails. These false memories feel accurate to the person recounting them and are not deliberate lies.
Flashbulb memory: “I know exactly where I was when Princess Diana died”
My flashbulb memory of Princess Diana’s death is another example of our flawed memory systems. I distinctly remember my mum waking me up from a midday nap to break the news and us watching it unfold on the kitchen TV. But flashbulb memories, while vivid and detailed, are often inaccurate. They stick with us because of the event’s emotional weight, but details shift over time.
Police officers must contend with unreliable human memories, especially in high-stress situations. Cognitive interviewing techniques using open-ended questions and sensory prompts help witnesses recall events more accurately. Witnesses are encouraged to take their time and recount details without pressure, helping to mitigate the errors introduced by flashbulb memories.
Creative mistakes
While in the early part of writing an article draft with ChatGPT, I stumbled over my words and said, “Let’s start again.” ChatGPT took it literally and produced an article titled “On starting again.” The first draft was not great; I rewrote the whole thing from scratch more or less, but the process felt like a creative collaboration. I frequently use AI (ChatGPT, Claude, etc.) to tighten and improve my work iteratively. While this instance was more of a miscommunication than a hallucination, it got me thinking: can AI hallucinations be good?
Humans and AI hallucinating together might be good
Could AI hallucinations—like human ones—spur creativity? These errors can sometimes lead to unexpected insights or innovative connections. AI’s associative capabilities, while not perfect, are still powerful and exciting. After all, mistakes and misinterpretations often drive creativity forward.
Collective hallucinations
Ultimately, humans and AI alike are prone to “hallucinations,” but perhaps that’s not always bad. These errors—whether through confabulation, flashbulb memories, or misunderstood prompts—can spark creativity, push boundaries, and lead to unexpected ideas. Creative “collective hallucinations” may be part of how we innovate, and I’m excited to see what emerges.