AI Is (Not) a Service Problem

Generative AI has remained a sore topic in online art communities, leading to the development of adversarial solutions like Glaze which aim to make images untrainable or have undesirable effects on models. Unfortunately, these solutions are not without their downsides—downsides which are visible to human viewers as well.

I recently observed a conversation on an artist Discord server. One artist asked for advice on whether he should “glaze” the commissioned work he had done for a client before sending it. The unanimous response from the other artists was to just send the glazed work. One even (falsely) said it only affects AI.

Unfortunately, the Glaze algorithm, at least at this time, produces visible distortions on images it processes. These strange artifacts appear as an odd mixture of JPEG compression and the shine of soap bubbles.

A couple of days later, the artist who asked the question returned, stating that he had to send the unglazed version of the work. The client had pointed out that the distortion was obvious, especially when zoomed in, and did not want the work glazed. A cursory glance at the artist’s website shows that the price of the commission would likely have been around $600, so the outcome of the artist’s experiment is rather unsurprising.

Imagine if you had hired a developer to make a website for you, and you paid over half a thousand dollars. This developer, however, was so adamant about “protecting” “their” code that they added heavy DRM to your site. Because of the DRM, your site was so broken that it was unusable half the time. Even worse, it’s as if the developer you hired thought of you more as a threat than a valued client. Would you consider this acceptable? Is it reasonable for the developer to cling so tightly to work that you paid for? Would you even pay him at all, or if you had already, would you have requested a refund or chargeback?

The saddest part of this story is that artists will be causing more harm to themselves by trying tricks like this. History has shown that if a customer is provided with a worse experience by paying, they’ll simply go somewhere else. Gabe Newell of Valve Software famously stated that “piracy is almost always a service problem and not a pricing problem.” Of course, “piracy” can be replaced by any form of undesired competition, legal or not. I believe he has drilled straight to the core issue: if I have to deal with distortions in my images regardless, why would I commission a work rather than generate it with an AI model? By applying these “protection” schemes, artists remove some of their primary advantages over AI: quality and consistency.

The animosity toward AI also seems to be quite inconsistent. Opponents of generative AI often complain that the datasets gathered contain the works of authors who did not decide to be included. In spite of this reasoning, AI/ML-based machine translation remains popular, even among those who claim to oppose generative AI. These models are typically trained on the same sorts of mass-gathered datasets. Opponents will also point to possible market effects, or the fact that artists typically spend years honing their craft. Professional translators spend just as much if not more time on their craft than artists, yet is their time different? Their job market has shrunk significantly, but there is no uproar for them.

Let me ask you: why would there be? We can all agree that it is usually difficult to learn another language, and it requires time and effort many of us don’t have. Because of this, machine translation is a reasonable alternative for increasing our ability to communicate with others. Is generative AI not the same? Are pictures not their own language?

At the end of the day, this conflict isn’t about AI, or copyright, or ethics, or jobs, or even good business. It’s about vanity, and that’s OK. People will always be entitled to their own opinions, but please, whatever decision you make, don’t accidentally hurt yourself in the process.