Before the other day, I can’t say that I had ever considered whether it was possible for AI slop flyers to not be horrible. I will suppose, for the purpose of argument, that not-horribleness exists merely at a superficial level, that not-horribleness reflects some aesthetic judgement. I will suppose that you can put aside the question of whether something AI-generated is inherently not good—that you can^[1][1]1. You cannot. overlook any of the ethical questions.

Frontier models have truly mind-boggling numbers of parameters. Is it so unfathomable that some of those configurations might produce novel outputs? I don’t think so—I find it pretty unsurprising. The question of whether it’s possible for these tools to produce anything other than offensively^[2][2]2. Putting aside that the reason you find it offensive is because you know that it’s AI-generated. ugly slop is a fundamentally uninteresting one.

The sort of restaurant owner who puts in front of their business a sign showing the most trypophobia-inducing depiction of a cheeseburger known to man has no interest in the aesthetic quality of the image.

If the images you use to advertise your business could be generated by typing one sentence into the free tier of ChatGPT, it’s over. Aesthetics be damned. Arguments about how, with a little more effort and a few prompt tweaks and some more tokens, you could something obtain aesthetically-pleasing—or, at least, unobjectionable—will fall on uncaring ears. The cost, or absence thereof, is the point. The lack of effort is the point.

Similarly, there is no doubt in my mind that it is possible to get an LLM to build a piece of software that is well-engineered. That is to say, it could produce software that both addresses the end-user requirements and consists of good code, for values of good that existed prior to 2022. And that doesn’t matter.

To a first approximation, no organization that is funding the development of software is interested in building good software. They are interested in building good-enough software. If you are an engineer interested in building good software employed by such an organization, perhaps you can fight the rising tide. But when hasn’t that been a losing battle.^[3][3]3. Don’t let me discourage you. I say, as a user of software, please keep tilting at windmills.

How long until management realises that good-enough software comes cheaper now? Does it matter that good software comes cheaper now too? Even if good software now is cheaper than good-enough software before, good-enough software now is even cheaper still. It was already difficult to convince management to spend more money or time to do something better than the bare minimum, so why should that be so easy now.^[4][4]4. It is somewhat perplexing how many software engineers think it will be, and that LLMs have somehow changed the calculus. The lack of cost and effort is the point.

The pessimist position is that software quality is going off a cliff, and for the moment we’re all just hanging in midair, Wile E. Coyote-style. Maybe that’s not the case, but even less likely seems the possibility that LLMs will somehow be the rising tide that lifts all boats, that, any month now, software quality will be a solved problem and defects will be a relic of the past.

Discussions of the quality of generative AI output and ways to improve all miss the forest for the trees. They are correct: it doesn’t have to be horrible. But none of this had to be horrible before. And yet it was, and the forces that made it so haven’t magically disappeared.