Generative AI remains a polarizing topic of conversation. It feels like everyone has a strong opinion, and it’s either “love it, use it” or “hate it, deride it.” I often fall on the side of “NO, no way,” but I keep vacillating between that and “why bother?” and also “eh, maybe?”
As background: I initially played around with genAI for image creation but stopped once I realized that the datasets were produced by scraping the internet without regard for the creators involved. The idea that huge companies are profiting off of this sticks in my craw.
I can’t get over the way this helps steal labor and consolidate more wealth in the ruling class. I can’t understand how AI proponents don’t see the way this is the worst manifestation of our society in hyperdrive.
That said, I also find I cannot stand uniformly opposed to generative AI, either. I’m usually an early adopter to tech. I love the ways that technology have changed and grown throughout my lifetime.
What we now call “AI” has become an extremely large basin holding an extremely diverse array of technology. A lot of these uses are harmful. Others are useful. I don’t want to throw the robot baby out with the AI slop water.
Adobe Photoshop has had “content aware fill” for a long time now. The idea is that you select a part of an image, and Photoshop will fill in what it expects should occupy that space using the pixels around it. I’ve been using content-aware fill since at least Photoshop 2019. For personal photos, this has been a quick way to remove things like telephone poles from the sky. What it produces is not much more sophisticated than a clone stamp. I class this as a useful tool–though it’s also not generative AI.
Later, Adobe added “generative fill,” which is like content aware fill on steroids. It uses their family of generative AI models, called Firefly, to create a more complicated image that the average person will recognize as GPT-like.
According to Adobe’s FAQ about Firefly, the dataset is trained on public domain images and its large stock library. If you’ve contributed to the stock library, your work is in the dataset. This is part of the terms of service, although I don’t know how and when this usage was added.
How well did they inform contributors? Did everyone know in advance what this would mean? How many contributing artists belong to agencies whose managers made those decisions above their heads?
It’s hard to say, but it’s still better than Meta deciding it’s fair use to steal from authors because their books have no value.
Where does Microsoft’s experimental AI engine fall on this spectrum?
John Carmack, a game dev elder and co-originator of the most classic boomer shooters, describes this AI use as a useful tool. I’m not inclined toward authoritarianism — here meaning that an authority’s opinion is only an opinion, and not above scrutiny — but I think he’s right that AI algorithms will be a growing part of the workflow.
I tried playing Quake II (one of my all-time favorite shooters) in this format. It’s obviously not yet a playable commercial game. It’s dreamy, foggy, and forgetful. At best, we can say it’s recognizable as the original game, and you can move inside of it.
But this is the first time I’ve glimpsed something that genuinely feels like a future successor to current game engines. Can it become markedly better? Is the output always going to be worse? Will it deprive game devs of jobs? I don’t know yet. I do find it interesting, though. I don’t want to discount it out of hand.
Ultimately, I evaluate individual tools falling into the AI bucket like this:
- Does it reduce desirable work for creatives? I don’t think it’s a big deal when people use AI models for silly little personal projects. I don’t care if people want to see their face on a Bridgerton character. Also, using generative tools to make photo editing slightly less onerous will make a creative’s life more pleasant. I can’t tell you how grateful I am that it’s easier to isolate models and remove that one patch on their jacket sleeve these days. I don’t miss clone-stamping and painting that stuff away.
- Does it disrupt necessary career progression? Replacing junior developers with AI blasts apart the entire industry in counterproductive ways. How does a senior developer become a senior developer if she can’t make a living as a junior developer first? How will artists ever graduate from Art As A Side Job to Art As A Full Time Job if all the jobs freelancing are replaced by people whipping up some muddy AI garbage?
- Does its usage on an individual level harm other creatives? Using datasets with material that artists did not consent to including is harmful. When you are generating art in many models — especially when you use specific artists’ names in the prompts — you are personally inflicting harm on creatives. Whole-image generation is likelier to resemble preexisting work by a single artist, which should be regarded as plagiarism. Adobe Firefly is maybe okay to use where you would otherwise normally use Adobe Stock, although it’s a horrible gray area.
- How significant is the environmental impact? As with most environmental issues, individual use is never as impactful as institutional use. Microsoft’s Copilot has been forced onto vast numbers of desktop computers and devours computing power whether or not you like it. On the other hand, Apple’s on-device AI performs processing on your phone, and it might demand charging your phone slightly more. It’s probably not a big deal. And an individual doing a couple of prompts is a drop in the ocean compared to Google adding AI processing for every single search engine query. Over time, as processing power becomes overall cheaper, environmental impact will decline. It’s alarming right now. It will improve.
- Does it do a good job? Whole-image generation can look as glossy as you want these days, but the algorithm doesn’t have intention, purpose, and an individual’s life experience to create a distinctive lens. Even when you’re editing out errors, you’re still leaning on the most bland, generic, commercial imagery that is possible. Things tend to look plastic. Women are homogenized into their most offensively attractive forms. And I still haven’t read any AI-generated text that isn’t a circuitous, unfocused, tension-free disaster of word soup. Trying to make AI output usable takes just as much work as making the thing yourself.
I’m sure I’m missing a few points, but these are the ones off the top of my head. And with this litmus test applied, there isn’t a ton of common AI usage that I would consider appropriate.
But there is some.
Many artists agree that using AI to generate mock-ups is no big deal. Anything where an individual isn’t putting AI into a final, sellable product is probably okay. AI that makes parts of unappealing labor go faster (like finding the exact code you want in a library) is helpful rather than thieving. These are natural progressions of existing technology, and they will become less damaging as environmental concerns are addressed and (hopefully) more datasets are made with material provided consentingly by accredited, compensated creators.
The very fact I believe “appropriate AI usage” is up for personal evaluation makes me feel more generous to all the individuals involved. We’re all trying to figure out how to navigate a complicated world. The problem isn’t really the technology, but the fact a sickly society can only use new tools in sickly ways.