Adobe Firefly Generative AI & ChatGPT

This is a sentence I never, ever thought I would type, but it’s possible Adobe (Adobe!) might have done something to address my concerns about AI generated art.

This is because before I started using Photoshop during COVID in 2020, my previous history with Adobe products was long, troubled and unhappy, including (but not limited to), my computer crashing in 1994 the first time I tried to install Adobe Reader, many, many, many tech support difficulties related to Flash Player in the 2000s, many, many tech support difficulties related to maintaining Adobe Creative Suite programs on a computer lab, and all the many times that Adobe Updater crashed while trying to install updates to update Adobe Updater.

If you’ve read my previous posts on the topic of AI art, my chief concern is the ethics of it – I remain unconvinced that it’s doing anything other than doing “salami-slicing” plagiarism. The counter-argument is that the AI isn’t copying, it’s “learning” patterns. My counter-counter-argument to that is that a machine is incapable of learning, just applying a more refined formula that generates better results, and it does that by copying a million different images and creating the average of them. That’s why if you go to most of the AI image generations and type in a prompt like “Magic The Gathering Plains Card”, it will generate an image that is an average of all the MTG Plains cards in its data set, including wield symbols where the card would have text because a MTG Plains card is statistically likely to have text in that portion of the card. It’s not “creating” a new image and it’s not “learning” anything, it’s just averaging suitable images from its training data.

However, Adobe recently announced plans for a beta of an AI image generation tool they call Firefly. According to Adobe, their AI has only been trained on public domain images and stuff within the Adobe Stock collection that they have rights to use. If you’re thinking that public domain means the photos will be low quality, remember that every photograph taken by a US federal employee in the course of their duties is in the public domain, so there is a vast amount of high-quality public domain images to use. Adobe also claims that it is working on a payment scheme for people whose images were used in the training.

Obviously this is not a perfect solution – pirated stuff does sometimes turn up on Adobe Stock (and, well, everywhere else on the Internet), and it’s entirely possible Adobe is lying through their teeth in a “technically legally true” way about some of this. That said, this is a clever approach. Most of Adobe’s customer base is image/video-editing professionals, and they were the angriest about the potential abuses of AI art. Adobe doesn’t really have enough goodwill among its customer base to afford yet another viral uproar, but since all the big brains claim that AI is the Next Big Thing, the company needs to compete in the space to survive. So it was a smart move to let others rush into the space first, and then announce their own solution that addresses some of those concerns.

Anyway, I signed up for the beta, and I just got in today. I’ll test it out and share my thoughts.

If it does work and the promises are true, Firefly is the sort of generative AI product I feel I could use in good conscience. I would use it to generate assets I would include in my book covers and Facebook ad images. I already use various stock photos and DAZ 3D for that, so it would be another tool in the toolkit. Doubtless fiddling with the generator to find the exact correct prompt would be no different than scrolling endlessly through stock photo results or trying yet another render in DAZ to make the output look right.

On a related note, a reader asked if I was considering using generative text AI like ChatGPT to help with my writing output.

Short answer: absolutely not.

Longer answer:

1.) At this point, it would feel like committing fraud and cheating readers. Like, if you buy a book that says “Jonathan Moeller” on the cover, the implicit promise is that Jonathan Moeller wrote that book. And if you’re buying that book, you want it to be written by Jonathan Moeller. That means you have decided you like my writing, warts and all (I appreciate that!), and have decided to pay actual money for it. So buying a Jonathan Moeller book that was actually written by an AI is a bit like a restaurant offering french fries that are allegedly 100% potato but actually turn out to be tofu, soybeans, and sawdust.

Granted, if I did write a book with AI and the cover said “by Jonathan Moeller & CrapGPT,” that would be different. But I don’t want to do that any more than I want to eat a french fry made out of sawdust.

2.) Generative AI kind of sucks.

I have to admit that when I listen to people who are very impressed with and enthusiastic about AI, I kind of wonder about their credulity. It’s a like a wizard cast a spell on horse manure to make people think it was the finest pepperoni pizza, and people are wolfing down this horse manure and praising the flavor while the horse looks on in bemusement. A lot of excitement has been generated from the fact that you can tell ChatGPT to write text in the style of Sir Arthur Conan Doyle or whoever, but what comes out when you use a prompt like that tends not to be very good.

Granted, in certain applications, horse manure is actually highly valuable. It’s just not pepperoni pizza.

In past blog posts, I’ve joked that generative AI is the Infinite Crap Generator. Maybe it’s more accurate to say that generative AI is the Most Likely Random Crap Generator But Occasionally Not, but you have to iterate a whole lot with your prompt and results to find something that isn’t crap. You can use it to write things, but you would have to wade through a lot of crap to get there first, and then edit it, and then put it all together…

…at which point, you might as well just write it yourself for less hassle.

3.) Copyright.

As of this writing, the copyright situation around generative AI is highly unsettled, but trending towards negative. The official position of the US Copyright Office is that AI-generated images can’t be copyrighted, though legislation might end up changing that at some point. Someone tried to copyright an AI-generated comic book, which was revoked, and then revisited the decision to say that while the images themselves couldn’t be copyrighted, the arrangement and human-written text could.

So, overall, the opinion of the current authorities is that machine-generated art cannot be copyrighted, and presumably that will be the same with copyrighted text.

With those three reasons, I won’t be attempting an AI-written book anytime soon. But perhaps in another ten years, “writing a novel” will involve typing like a thousand different scene prompts into an AI generator and editing the output together.

I doubt it, though.

-JM