Self Defense for Artists in the Age of AI

Self Defense for Artists in the Age of AI

Artists^[1] have good reason to be wary of generative AI. AI is a billion dollar industry^[2] and some generative AI models have been found to scrape the portfolios of far poorer artists^[3] in order to create their images. This is threatening for a number of reasons – not only are the AI-generated images a plausible dilution to an artist’s brand and reputation, there’s a possibility of loss of commissions, other work opportunities, and basic income.^[4]

While many artists have found limited success defending their work from AI in the courts, there is new technology that may provide artists the tools to fight back– by poisoning the data sets that Generative AI uses to create its output.

Generative AI uses the input of user’s prompts to generate text, images, and other media, including video.^[5] Generative AI “models ‘learn’ the patterns and structure of their training data (input) through an iterative process, and then generate new works (output) that have similar characteristics.”^[6]

The work that these models put out can be extraordinarily impressive – mimicking the look of the old masters, or comic book characters, or popular tv shows.^[7] The issue, of course, is where that training data comes from.

Many generative image models “rely on human-made images for training data, which are collected by the millions from various online sources.”^[8] This is done by “scraping,” by which a software collects data from the web, in this case images, from the aforementioned various online sources.^[9] Some of those online sources have on occasion included human artist’s entire portfolios, with “people just reproducing [an artist’s work] for free, with no context as to how it was made.”^[10] Google was hit by a class-action lawsuit over its data scraping to train its AIs, although the case was dismissed in July 2024.^[11]

This dismissal, like the dismissal of a similar case involving OpenAI^[12], isn’t uncommon. Although large numbers of lawsuits have been filed against many generative AI companies^[13], there seems to be a problem with how, exactly, to make those claims. The OpenAI case was dismissed due to “contain[ing] swaths of unnecessary and distracting allegations making it nearly impossible to determine the adequacy of the plaintiffs’ legal claims”^[14], and the Google case was dismissed in light of that ruling.^[15]

One might take the stance that data scraping seems like a possible case of copyright infringement^[16], but as seen above, it isn’t so simple. There have been claims filed alleging that scraping copyrighted works is infringement, but the “primary defense to copying for training purposes is fair use.”^[17] More specifically to the issue of copying artist’s work, some courts have “rejected the claim that all outputs produced by a generative AI program are or must be a derivative work of one or more of the works in the training data.”^[18]

So what’s an artist to do?

It makes sense that artists would want some way to defend against this. Even if there was an easy access claim to legal recourse, many working artists do not have the time, means, or knowledge to take on a protracted legal battle,^[19] What artists do have, however, is the full power to decide how they share the work they make – and what they put on it. Once upon a time, this could have simply meant bigger and bolder watermarks, or not sharing the work at all.^[20]

Now, there are tools like Glaze.^[21] Glaze is an app that seeks to protect artist’s work against AI scraping by “computing a set of minimal changes to artworks, such that it appears unchanged to human eyes, but appears to AI models like a dramatically different art style.”^[22] Glaze acts as a sort of self-defense for the artist worried about their portfolio being subject to data scrapers. Glaze isn’t perfect. It’s certainly not future proof – there are already two tools out there that specifically attack “glazed” images in order to allow them to be entered into data sets free of the filter.^[23] There are also tools like Photoguard^[24] and Mist^[25], which offer similar effects as Glaze. But, as everyone knows, the best defense is a good offense.

Enter Nightshade.^[26] Nightshade was produced by the same team that created Glaze, and while Glaze is a purely defensive tool, Nightshade is more aggressive.^[27] Nightshade “transforms images into ‘poison’ samples, so that models training on them without consent will see their models learn unpredictable behaviors that deviate from expected norms.”^[28]

An artist’s work with Nightshade applied, if scraped without their permission, would then slowly warp the dataset being used to feed or train the generative model. Again, it’s not a perfect solution. Nightshade has a more conspicuous effect or filter, visible to the human eye, than Glaze does.^[29] Some artists may find that filter more detrimental to their work than a splashy watermark, or simply want to risk their image being scraped in order to share their vision in full.

There’s also a push, as more people become aware of how these models are trained, for the engineers who build these models to self-police.^[30]  Popular art websites like DeviantArt and ArtStation offer artists the choice to opt out, although how effective that checkmark is remains to be seen.^[31]. OpenAI, creators of Dall-e 3, now offers an “Artists and Creators Opt Out Form,” which allows “owners [who] may not want their publicly available works used to help teach our models” to remove their images from “future” datasets.^[32] The operative word there is “future” – presumably, once an image is in a dataset, it’s not coming out.

Although it’s heartening to see that some of these opt out options are becoming available, one must be aware that opting out is an option in the first place. Until there are more stringent regulations in place on what goes into a training data set, tools like Glaze and Nightshade may provide some peace of mind.