Shortly after market close, I will ask an AI generated image on Bing : "A room with no elephant, not even any reference to an elephant", expecting 4 propositions as usual.
Market will resolve YES if all four proposals match the prompt correctly.
That means not like today :
In resolves NO if one or more proposal have an elephant or a reference to an elephant (drawing, logo, written text). The reference must be direct. An elephant tusk is enough, but not just an ivory object. An hippopotamus would not qualify as a reference.
I will use another AI if for some reason Bing is unavailable.
This is inspired by Gary Marcus https://garymarcus.substack.com/p/there-must-be-some-misunderstanding
Also by my "Not a French Frog" comment here, a few days before:
https://manifold.markets/Ernie/midjourney-can-draw-a-frog-riding-a-c14b6948e49e#EqgVM2p5O8ci26CubOVV
I will not bet on this market, to avoid subjectivity.
At least, there was an effort to hide in the corner:
Related questions
@vluzko I wrote "I will use another AI if for some reason Bing is unavailable.". Preferably nearest similar option.
This is tough because, while I would be very confident that Dall-E itself would still suffer from this issue in 2025, Bing image creator also includes a step which edits the user's prompt (as does Dall-E through ChatGPT). This step seems perfectly capable of fixing the issue by just declining to mention negative requirements, and so if Bing or OpenAI decide to adjust their system prompt tomorrow, the "no elephant" phenomenon could disappear.
For example, adding "When asked to generate an image that does not contain a specific object or concept, it is best to simply not mention that concept in the prompt at all, rather than include an explicit forbiddance." as your custom system prompt in GPT-4 gives a very high success rate at avoiding elephants given this prompt.
So I feel like I'm really betting on whether I think Bing will be motivated to make this change, which feels a lot harder to judge.
@Zardoru I was confident, but after looking, I can't find confirmation that Bing adds to the prompt, so I can't say that I'm sure. However, here's some evidence: Try a prompt of the form "A newspaper headline about <nonsenseword>". For me, this generates multiple images that all share the same text. For example, given "A newspaper headline about Groppu", I get four images that all say "group takes bold steps to tackle environmental crisis". If I then run the same prompt, I get four images that all say "local group issues report social justice and environmental awareness". It's hard for me to see a way that this would happen if Bing were not adding these words to the prompt and running that prompt four times. If it were just that Dall-E has some association with this nonsense word, I wouldn't expect the text to be consistent within a group of four generations, but different across groups.
With ChatGPT the question is much clearer - it will show you the prompt it wrote.
