Nightshade: Can Artists Who ‘Poison’ Their Works Prevent Infringement By AI?

Digital artists are facing an unprecedented dilemma: they must promote their art using social media, however, doing so potentially leads to their work being scraped and exploited by AI companies without consent.

In an attempt to deter this problem, earlier this year University of Chicago computer science professor Ben Zhao and his team released Nightshade—a tool that makes subtle changes to the pixels in an image that are imperceptible to the human eye but may trick AI models into misinterpreting the image. The ‘poisoned’ image then causes the AI model to create images that are different from their corresponding prompts.

For example, Prof. Zhao showed how an image of the Mona Lisa modified with Nightshade was perceived by an AI system to be a cat wearing a dress, despite being identical—to the human eye—to an unaltered image. Subsequently, when that same AI system was prompted to create an image similar to the Mona Lisa, it created an image of a cat wearing a dress.

Several images protected with Nightshade among millions of other images in a training set are unlikely to cause a disruption to an AI system due to dilution. However, as the number of poisoned images increases, it may be possible to, for example, confuse an AI system to the point where it believes a dog is a cat or vice versa. Therefore, it is critical to understand that the effectiveness of Nightshade depends on widespread adoption.

Supporters of Nightshade claim that digital artists will now be able to establish secure platforms where they can share and monetize their artwork without the fear of infringement by scraping programs. However, critics of the tool point out that AI algorithms continuously evolve to the point where it may anticipate Nightshade and adjust accordingly to minimize or negate its effects, or simply determine that pixel shading is irrelevant and ignore it altogether.

Ultimately, Nightshade and similar tools such as Glaze (also developed by Prof. Zhao and the University of Chicago) are reminders that AI (and financial) models are only as good as their inputs. If a model uses flawed training materials or incorrect assumptions, then the output will be flawed as well. To put it bluntly, garbage in, garbage out.

While the efficacy of Nightshade is debatable, what is clear is that we are currently at an intersection of technological progress and intellectual property protection. Without proper guidance and enforcement from lawmakers, copyright holders must rely on well-meaning computer scientists such as Prof. Zhao to fight back against infringement.

CONSOR offers a range of highly specialized services to help clients understand the value of their intellectual property portfolio and how that value is impacted as a result of market conditions. We work closely with clients and the courts on time-sensitive projects to unlock the value contained within the intellectual property. Our work is done according to the strategy that best suits a client’s specific needs. Whether that be licensing, partial sale, reorganization, merger, joint venture, or liquidation, simply let us know your needs. Email us at info@consor.com or call us at (858) 454-9091.