Step 1: feed ChatGPT an album cover and ask it to describe it

Step 2: take its response and remove mentions of the album name

Step 3: start a new chat and ask it to generate its own image description

Step 4: I post the results to Manifold, asking, "Which album is ChatGPT recreating with this image?" I will make it a multiple choice market which allows people to add answers. The answer I add will always be Trout Mask Replica, and it will never be correct. Each market will stay open for two weeks.

Resolution: I will be doing this once a day for the next five days. I will consider each image to "be recognizable" if the highest probability answer for a given market is correct. So, if in the Abbey Road market, the Abbey Road choice had the highest probability, I would consider Manifold to have recognized it.

For this market to resolve as yes, I will need three or more of the five markets to converge on the correct answer. Otherwise, it resolves as no.

Considering I've already used the most famous album art of all time, the examples I use will be more difficult. I have no perfect way of quantifying what makes an album "iconic", but I will do my best to choose albums that music nerds are largely aware of. I will at least promise to only include albums with at least one song that has more than ten million streams on Spotify.

I will be taking the first response from the AI for both the image description and generation in order to avoid cherry-picking.

I am allowing the AI to include descriptions of text given that it would be difficult at times to remove. For example, here is the description it gave for Automatic for the People:

Here is what I edited the prompt to, along with the results:

As always, feel free to drop any suggestions or questions in the comments.