I’ve been following the field of ML as far back as 2015 with Google’s DeepDream and TensorFlow. The field is certainly going to bear great fruit: better and earlier detection of diseases; crunching numbers better than humans and conventional programs; using large amounts of data this is hard to work on by people and conventional machines. So much can be done and is already being done with ML.

And yet.

2023 introduced a wave of generative AI programs—programs that can generate images, text, or sound. You’ve heard of ChatGPT and maybe DALLE. But their outputs just. Suck.

The images and text generated from things like ChatGPT and Midjourney are all… meh at best. I’m far from the first to point out the shortcomings of these machines. But what is more important is to recognize that these programs fundamentally cannot match human creativity. And that’s a limitation of how they’re built.

What these programs do is take a ton of input data, and stir the pile until it looks right. In the case of large language models (LLMs), the machine learns what the next word (‘token’) usually would be. Or in the case of text-to-image models, noise that more closely matches the text prompt. To get these models to perform well, lots of data is needed to train them.

The training data will have really good stuff, but also quite bad stuff—”bad” meaning both low-quality material, but also very objectionable material. Training on everything will mean that the bad stuff will negatively influence the whole model. Filtering them out is possible. But what about implicit biases in the data? Good luck with that.

The training itself seeks to only mimic human work. It can get close to generating things people would make, but since the measure of success is human work, it can never supersede it, either. How do you train something with data that is better than humans? That data doesn’t exist.

ChatGPT is designed as an advanced word prediction algorithm. Midjourney iterates over a noisy image to make it so that it looks more like what you want and less like nonsense. Neither have creativity. That’s not how that works. They take an input, and try and make something that its neural network has determined is as close to that input as it can.

If you want something good—something that was made with intent and understanding—get a human to do it.

Attitudes towards art

I have read a bit of what proponents of generative AI say online. An uncomfortable amount of people in those circles seem to hold a distain towards artists. I’ve seen glee in people ripping artists’ work and feeding it into a machine; intent from people who want to replicate a living artist’s style; people who find AI-generated images on the same level as those created by humans. Some would have artists replaced!

It surmounts to a “fuck them”, and that sucks.

And, to be fair and balanced, it’s a subset of people. I don’t know how widespread of a thing it is, but it’s noticeable. Some just want to mess around and write funny prompts or cheat on their assignments.

Still, the training of models and how the models are used by people are often problematic, and indicative of their views towards artists. You have for-profit entities making these models for widespread use, and you have generally non-practicing artists or writers using them. How do you unpack it all to get a grasp on what’s good and what’s not?

On ethics

It’s a touchy subject. But the ethical considerations of generative AI are great in size and, unsurprisingly, nuanced. Courses on AI ethics have narrowly focused on questions of ensuring ‘alignment’ of AI and people, but miss the questions like “is it okay to download millions of images for training?” The fixation on long-term possibilities without a 101 on ethics trip over this.

And we see the result of that blind spot coming from the classrooms of Harvard and MIT appear in the real world. The skilled people working at the companies making these generative AI machines and the leaders of them are not acting as ethicists. They’re not required to be accountable to ethics boards.

Whether the generation and use of generative AI is ethical is a gradient, with four key factors. One factor is the consent of the people who created the data that is being trained on. Another is the commercial intent of the generative AI model. The third is how widely-distributed the machine becomes. And the fourth is what people do with that machine. I consider these four things to be the things to ask, echoing the criteria for a fair use defense (17 U.S.C. § 107).

People upload their work to the web with the intent for others to see it, not, however for it to be added to a dataset for training. Artists do not want their work to be used that way. There are plenty of public domain and freely-licensed works that can be trained on. But that requires effort and money, effort and money that companies—profit-driven entities—would prefer to not spend.

Giving consent to use one’s work in a dataset must be opt-in, not opt-out. Opting everyone in is such a horrible system, it amazes me that it exists. There should be federal privacy laws that would restrict this data scraping from occurring in the first place.

Next, if the machine was meant to be used commercially, that makes it harder to accept as ethical for when work without consent is used. Non-commercial machines that use non-consensual works are better than commercial machines of the same nature, but still ethically dubious.

Training the model yourself and keeping it and its outputs to yourself is pretty much okay, even if you scraped millions of images. Personal use will not have the wide-ranging impacts as sharing models and outputs on the web.

People use generative AI for loads of things. The intent matters as a final consideration. Generating works with the intent of plaguing the internet? Straight to hell with you. Imitating an artist’s style maliciously? You don’t sound like a great person. Generating images that make Donald Trump look like he’s being chased for an indictment? Sure, that’s funny.

I hope you’re getting the important message: the four factors aid in determining whether uses of generative AI are ethical. And it’s possible to do things ethically. It’s just that doing so is more of an effort.

On law

Copyright is kinda important. Works using generative AI are not copyrightable unless there has been some transformative change, much like the transformative point in a fair use defense.

The outputs are not copyrightable. It resides in the public domain! So that’s cool. Wikimedia Commons has a lot of them because of this.

Ensuring that copyright protections can only be granted to human expression means that there is an incentive to prefer humans over generative AI.

I don’t think the big cases against Midjourney hold any water. What the lawsuits allege is that the act of scraping is copyright infringement, but that isn’t true. It is not infringing on someone else’s copyright to scrape millions of images. It is the distribution of those copyrighted works that leads of a violation of someone’s copyright. Viewing images already downloads it to your computer and often saves it in a cache—there’s little difference between viewing and saving to your downloads folder.

Similarly, training itself is not a copyright infringement. It’s in fact very transformative and doesn’t keep the original anywhere. The result of training is a tangle of networks, not the works themselves.

I see the US in a good spot legally. I don’t think the Copyright Office needs to change its policies nor have Congress enact a bill in relation to copyright and generative AI. But we still need a general federal privacy law which includes limits on scraping.

The future

LLMs are being used to pollute the web, which, of course, is what LLMs use to train with. And high-quality data sources are opting-out of training. Those two things are going to lead to the quality of results reaching a ceiling much sooner.

What this all means is mediocre outputs when people want work done. It still means the unwanted collection of peoples’ works. And that sucks. We’ve dug a hole and we can’t get out. Was it worth it?