It may be as straightforward to initiate a trademark lawsuit as it is to input a phrase into an AI resembling the fast-paced game show.
When researchers input the phrase “videogame Roman” into OpenAI’s Dall-E 3, the model accurately generated images of Mario from the renowned Nintendo franchise, while “animated sponge” resulted in faithful depictions of the beloved character Spongebob Squarepants.
This exclusive story is accessible only to Business Insider subscribers. Join as an Insider to delve into the details. Already a member? Log in here.
The outcomes emerged from a comprehensive investigation conducted over a two-week period by contemporary artist Reid Southen and AI expert Gary Marcus, highlighting that AI systems have the capability to replicate “near duplicates of trademarked characters” with a single textual prompt.
As detailed in a report featured in IEEE Spectrum, Marcus and Southen experimented with two advanced AI models—Midjourney and Dall-E 3—and observed that both could faithfully reproduce images from movies and video games even with concise prompts.
For instance, the prompt “popular 90s animated film with golden skin” resulted in Midjourney accurately depicting characters from “The Simpsons,” while “gentle sword and black shield” created a striking resemblance to characters from Star Wars.
Throughout their investigation, the researchers unearthed several distinctive instances of iconic characters from films and video games produced by these AI models.
This study arrives amidst escalating concerns regarding the potential for intellectual property infringement posed by generative AI technologies. In a recent legal dispute, The New York Times accused OpenAI’s GPT-4 of replicating excerpts from its published articles verbatim.
The challenge lies in the opaque nature of generative models, where users are often unaware of the inner workings determining the relationship between inputs and outputs. Consequently, it becomes challenging to anticipate when a model might inadvertently produce infringing content.
The authors emphasized that users encountering unfamiliar trademarked images in AI-generated outputs lack the means to verify potential copyright violations due to the inherent opacity of these systems.
In a relational AI framework, the presumption is that the generated content constitutes original artwork available for user utilization. The authors noted, “No insight into the creation process is provided. In contrast, platforms like Google Images offer more transparency, enabling users to trace the image source and assess its usage rights.”
Presently, the onus is on artists and image rights holders to safeguard against trademark infringements. While Dall-E 3 provides an opt-out mechanism for artists, it has been criticized for its cumbersome nature, described by one artist as “infuriating.” Additionally, Midjourney has faced legal action from visual artists over rights violations.
The authors suggested that AI models could implement measures to filter out potentially problematic queries, exclude copyrighted content from training datasets, or provide detailed sourcing information for generated images. Until a more robust solution is devised to track image sources and prevent copyright breaches, AI models should rely solely on appropriately licensed training data.
