Designing book covers in half an hour, redux

Long-time readers may remember that back in 2013, Matt and I played a game where we each designed a cover, in half an hour, for a book whose name was randomly generated. Here’s what I came up with for The Name of the Names:

I really enjoyed that process and even toyed with the idea of offering it as a service for hire, for people creating self-published books.

But new we live in the future, and generative “AI” can do this stuff for us. Right?

Off I went to DALL-E 2, which OpenAI offers as a free demo. I entered this prompt:

Cover for a high fantasy book titled “The name of the names”. The title should appear prominently on the cover, along with the subtitle text “Book one of the False Names trilogy” and the author name “Michael P. Taylor”.

Here are its four offerings:

Screenshot ” data-medium-file=”https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=212″ data-large-file=”https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=480″ class=”size-full wp-image-21826″ src=”https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=480&h=679″ alt=”" width=”480″ height=”679″ srcset=”https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=480&h=679 480w, https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=960&h=1358 960w, https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=106&h=150 106w, https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=212&h=300 212w, https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=768&h=1087 768w, https://svpow.files.wordpress.com/2024/03/overview.jpeg?w=723&h=1024 723w” sizes=”(max-width: 480px) 100vw, 480px” />

Screenshot

And each one in full detail:

Leave aside minor matters like the use of square aspect ratio for book covers, the cropping that shows partial words, and the absence of anything resembling artwork. What’s happened to the text here is the really startling thing. I’ve written before about generative art’s problems with text, but I find it striking that across four text-heavy covers, the only words that are comprehensible are several instances of “the” and one or two “of”s.

I don’t doubt that this performance will improve over time, and DALL-E 3 (which OpenAI wants you to pay to upgrade to) is probably better already.

But I think this is a really nice illustration of the fundamental flaw in what we’re all suddenly calling AI for some reason. There is literally no comprehension in there — and so, no intelligence in any meaningful sense of the word. An image-based “AI” isn’t good at producing text because it literally doesn’t know what text is — only what it looks like. And in the same way, a text-based “AI” literally doesn’t know what meaning is — only what it looks likes. What sequence of words resembles meaning.

We have got to stop fooling ourselves about these things. In particular, the idea that LLMs could be used for peer-reviews is nonsense. What they can be used for it to produce sequences of words that resemble peer-reviews — which is literally worse than nothing.