💾 Archived View for tilde.club › ~verdantmoss › ai-content.gmi captured on 2024-06-16 at 12:57:40. Gemini links have been rewritten to link to archived content

View Raw

More Information

⬅️ Previous capture (2023-01-29)

-=-=-=-=-=-=-

AI-Generated Content

When it comes to Machine Learning systems generating "content" - a term I dislike but is suitably broad - the cat is quite confidently out of the bag. There exist ML models that can generate prose, images, code, video, audio, and perhaps - by the time you read this - even more things. Out of those, prose, images, and code are approaching something that is now quite difficult to tell apart from creations of flesh and mind - although the images are often still a /little/ wonky.

This is, naturally, somewhat alarming. A computer can now, in mere moments, spit out a "creative work" indistinguishable from something created by flesh and blood. And I'd like to clarify that the word "created" is doing quite a bit of heavy lifting in that sentence - for a lot of the things that ML models can replicate, the equivalent act of human "creation" is a great investment of time and passion, not to mention the lifetime of skill and lived experience that finds some form in the finished work. What would a world look like where we are flooded with AI generated content? It might be easy to dismiss now, but an endless stream of tailored content would be just perfect for keeping people hooked on some garbage social media platform, and maximizing engagement is the name of the game these days.

But these models cannot, by design, be /truly/ creative. They can, on a surface level, appear creative - certainly, in most cases, the output from these models is novel, never before extant. But this is deceiving - these systems can only remix and repeat that which they were trained on - that which already exists. They will, without thought - because they do not truly think - regurgitate every bias on which they are trained. If AI generated content becomes mainstream, it will only serve to reinforce - and reinforce strongly - existing bigotry, existing roles, existing rigidity. Guess what comes out of an image generation model when you ask for a picture of a programmer, or a nurse? This is not a "bug", nor a feature - it is a fundamental limitation of these models, because they cannot be /truly/ creative. What would a world flooded with AI generated content look like? Frozen in time?

There are also more immediate, practical problems with AI content generation as well - the perfect example being Microsoft's Copilot:

Github Copilot

I'll spare you the need to open that link - it's a code generation model that can suggest entire chunks of code, for the low-low cost of $10/month. What code is it trained on to produce those suggestions? Everything on GitHub, and some other public code as well. What's that? That would include a bunch of GPL licensed code? Oh, it can spit out fragments of the code it's trained on verbatim? That doesn't sound right.

The point I'm trying to make is that these models are, right now, being trained on data without the consent of the people that made it. And I'd like to be clear - even though calling it "data" might make it sound a little abstract, this is countless hours of labor and passion that are being resold. I hope the Copilot lawsuit strikes a prodigious blow against not just Microsoft, but any other company that feels that it's okay to resell other peoples work in this way.

It sucks that my feelings towards this technology are so poor. This technology is - ostensibly - quite cool. It's loaded with great potential, and even were it not to have some practical application, it would be a great stride forward anyway. But it does have practical applications, plenty of them. Right now, text generation models can read a text (think: news article) and generate a coherent and accurate summary of that text. AI Dungeon sucks, but it's a great example of a fun and novel application of the same technology. Code generation models can even be used to find bugs, by finding code that looks strange and uncommon. It sucks that I think this technology will do much more harm than good, but we live in a late-capitalist cyberpunk hell-hole, so I suppose it feels foolish to think anything else.

last updated: 2022-11-28