PSYCHOLOGY
Don’t Think About Orange Juice!
What we can learn from ChatGPT about the dangers of association bias
--
ChatGPT has an interesting quirk.
If I tell it to “Create an image of a completely empty glass with no orange juice in it”, ChatGPT will create an image of a glass of orange juice.
That’s exactly the prompt I used to create the header image above.
Similarly, the prompt “Create an image of an empty plate with no banana on it” creates the following:
If you’ve got time and your hands and want to waste some of it screwing with ChatGPT and watching it fail, try it for yourself. Once the image has been created — in other words, once the association has been made — it can be difficult to get ChatGPT to remove it from subsequent images.
The examples above were inspired by a post I saw on BlueSky featuring the following screenshot:
As hilarious and entertaining as these examples are (I’m easily pleased — I should probably get out more) it does raise an interesting and more serious issue.
Chat GPT and other AI models use keyword weighting and association to create images. They find it hard to conceptualise absence or nothingness.
When ChatGPT sees “glass” and “orange juice” in your request, it immediately references the gazillion pictures depicting glasses of orange juice that were used in its training. Searching that database of glasses of orange juice to find an empty glass is a task that leaves the AI scratching its virtual head.
Association bias in humans
In a way, we humans are not that different.
If I say to you, “Don’t think of a beachball”, I’ve immediately implanted that image in your head, no matter how fleetingly. A lot of people will pride themselves on being able to keep the image out of their head by concentrating on some other image, but ask them what they’re not thinking of and they’ll quickly reply, “A beachball”. It’s in there!
If I say, “Imagine a girl on a beach with no palm trees in sight,” your mind will almost certainly flash an image of a palm tree, no matter how briefly, so you know what to exclude from your mental picture. As an aside, I wouldn’t be surprised if you also pictured her with a beachball!
This phenomenon has far-reaching social and psychological consequences.
Ironic Process Theory
Studies conducted by social psychologist Daniel Wegner and others have shown that attempting to avoid certain thoughts (or emotions) in this way is not only doomed to failure but actually tends to increase the frequency and intensity of such thoughts. This is known as “ironic process theory” or “ironic rebound”.
This can make any attempts to suppress depressive or anxiety-producing thoughts counterproductive, a phenomenon well-known to clinical psychologists and therapists. It’s far easier to focus on creating new positive habits than to focus on avoiding old negative habits.
Advertisers, too, are well aware of the ironic process theory. That’s why you’re more likely to see a product labeled as containing “Pure, natural ingredients” rather than “Contains no harmful ingredients.”
It’s also why politicians are more likely to spend time talking about all the good things they’ve done rather than denying accusations of scandal. Declaring “I am NOT corrupt” is a sure-fire way for a politician to reinforce any scandal in the public consciousness.
Dangers of association bias
Memories are malleable. They are not stored fully-formed like a video recording. Instead, they are reconstructed each time we remember a situation or event. As such, they are susceptible to association bias in much the same way.
Imagine a scenario where a witness is being interviewed at a police station. “Did you see a red car in the car park during that time?”
The witness might not have seen any red car, but to consider the question, they first have to picture a red car there. The more times they do this (“Think carefully. Are you sure you didn’t see a red car?”), the less sure they feel and the more likely it becomes that their brain will create a false memory around that imagined red car.
In researching this article recently, I was struck by how association bias and ironic rebound might be affecting something very close to my own heart — animal advocacy and vegan outreach. I hesitate to include it here for fear of introducing and reinforcing the very association bias I’m about to mention, but I think it’s an important issue to raise and something that all vegans and animal rights advocates should consider.
Friends, not food
“Friends, not food” is an extremely popular slogan within the vegan/animal rights movement. You see it on tee-shirts, bags, badges, placards and in countless social media posts. You’ll also see similar memes such as pictures of cute piglets, for example, with the caption, “I am NOT bacon”.
It’s a nice sentiment. While we shouldn’t have to see someone as a friend to refrain from exploiting them or having them killed, I believe we do need to shift the popular perception away from seeing animals as food or commodities. I’d much rather someone see a pig as a friend than as just a piece of meat.
However, applying the principles discussed above raises the question of how effective such memes are or whether they are even counterproductive. If we keep posting pictures of pigs along with the word ‘bacon’, are we involuntarily reinforcing the association of pigs with bacon in people’s minds? If so, are we making it even harder for people to see farmed animals as individuals rather than commodities?
The words we use and the images they create are perhaps more powerful than we think, but not necessarily in the way we might expect.
As ChatGPT has clearly illustrated, it’s far easier for people to “Imagine a pig who is your friend” than to “Imagine a pig who is NOT bacon”.
