Writer: Cameron Serkland
Contributors: Ryan Baldwin, Luis Bencomo
Artists and Artificial Intelligence
To aspiring and professional artists alike, the emergence of image generation through artificial intelligence and its continued growth provides a challenging new landscape to navigate through. As images generated through large language models rapidly outpace human output by covering searches and feeds, how can traditional artists adapt? As companies scrape the internet for images to train their models, how can individual artists protect their intellectual property from being used without consent? In this series of articles, we will explore how best to adapt to these challenges.
Spotting AI Image Generation
In the obfuscation of the modern internet, it can become important for artists to themselves recognize when a work has been generated by AI. There are telltale signs that can be used to spot AI generated images. Bad anatomy, especially within the hands, and a yellow tint are infamous signs of generated images. Alongside these, muddy coloration, unnatural lighting, repetitive patterns, inconsistent sizes on repeating objects and indecipherable text are all possible signs of AI image generation.
If we were in the early days of image generation, I would send you off with this list, content in the fact that you’d be able to spot whatever generated image crosses your path. The reality, however, is more complicated than that. As the capabilities of image generation grow so too do the avenues at which we can spot it narrow. This presents us with a conundrum; how can we even be sure if what we are looking at has been generated?
One method could be asking the source directly. Platforms such as Google’s Gemini and their SynthID have an invisible watermark on their generated content. By simply sending the platform the image and asking if it was generated by their model, you can potentially get the information directly from the source.
Another potential method would be to use websites with AI image detection tools to determine whether an image has been generated. Problems arise however when false positives and negatives cloud these tools, with some being only slightly better than a coin flip in their detection. Tools such as AI or Not fare much better in comparisons such as those done by originality.ai, boasting a 97.14% accuracy within their testing criteria.
The concept of false positives, of thinking an image is AI generated when it is not, provides a conundrum to artists. Budding artists might not have the greatest grasp on anatomy or lighting, leading those online to potentially decry their work as generated when it is not. The quick gut reaction of internet denizens to call out any artwork with flaws as being generated paints a bleak landscape for legitimate artists trying to hone their craft.
Unfortunately, artists in this climate may need to keep evidence of their process. Art timelapses, streaming your process, layered work files and early versions of art can all help in providing legitimacy to your art career. Luckily, these records can also help you as an artist by giving you a way to investigate your own process and see how you can improve. That said, not everyone has the tools, space or workflow to produce these receipts, and this burden of proof creates a hostile landscape for any but the most entrenched artists.
This leaves us with a message to anyone reading this, be cautious not to spread baseless accusations of the legitimacy of other’s work. A climate of witch-hunts does little to protect artists, and people must stick together if we are to stave off a tide of generated content. Any method used to find whether an image is generated should be used to inform your approach to that work, not as a sign that it is ok to harass its creator.
Sources and Further Reading:
https://roblaughter.medium.com/is-that-image-ai-here-are-14-telltale-signs-to-look-for-d40e5cff2d0a
https://deepmind.google/models/synthid
https://originality.ai/blog/do-ai-image-detectors-work-accuracy-study
Protecting Your Art
As AI image generators need an absurd amount of data to train their models, their creators have used data scraping to scour the internet for viable training data. In terms of replicating art, this means that the intellectual property of countless artists has been used to train these models. Though ongoing battles wage on about the legality and ethics of this process, this article exclusively deals with how you, the artist, can have a say in how your own work is involved or not in this process.
On the technical end, websites throughout the internet often have a robots.txt, a file that can be accessed through their domain that determines what permissions bots have in scraping the data on the site (for example, google.com/robots.txt). Checking the robots.txt of the host of your art can show if a web crawler following the standard can access your images for the purpose of training data. This is only one small measure against data scraping, as while many companies comply with the standard, malicious bots are not impeded at all by the standard.
Another method is using services such as HaveIBeenTrained, which investigates whether an image you have uploaded has been used to train AI image generation models and gives you the opportunity to opt out. Though it is another step towards protecting your art, organizations must individually use the API provided by such services to respect requests to opt out, leaving us with the same conundrum where there is no protection against bad faith actors who ignore whether you consent to your work being used or not.
So how can an artist adapt to rampant data scraping gobbling up their artwork? Much like in nature, it seems that to avoid being eaten, one must become poisonous. More dedicated tools like invisible watermarks or other alterations can be added to images, changing their metadata in a way that makes them useless or detrimental to models that use them as training data. These methods too use advances in machine learning to make alterations to the image invisible to the human eye that aren’t removed by transformations like cropping, filters or compression.
Of these tools, the University of Chicago’s Glaze deters style mimicry, while their Nightshade seeks to disrupt models to their core. These utilities act as protection from data scraping without consent, the negative effects of training on the image outweighing the potential gains of training with them. With tools like these, artificial intelligence can give artists the tools to once again take back control over their work.
But what if you want to find ways to benefit yourself from ai training on your data, to live in symbiosis with image generators? Though Adobe has been in hot water with artists over their services many times before, Adobe Firefly does seem to be a step towards more ethical sourcing of training data. Using images from compensated artists through Adobe Stock, Firefly is the first large scale model that can lay some claim to an ethical model of image generation. Only time will tell whether others will follow suit in protecting the intellectual property of artists, but, in the meantime, artists can use the many tools at their disposal to combat any system that treats them unfairly.
Sources and Further Reading:
https://www.robotstxt.org/robotstxt.html
https://glaze.cs.uchicago.edu/what-is-glaze.html
https://nightshade.cs.uchicago.edu/whatis.html
https://www.adobe.com/ai/overview/firefly/gen-ai-approach.html