Reid Southen, a movie concept artist based in Michigan, entered the realm of A.I. image generators with a sense of curiosity. Initially captivated by the technology's capacity to transform text prompts into visual creations, Southen soon found himself entangled in ethical dilemmas upon discovering that these systems were trained on existing artwork, potentially exploiting artists and breaching copyright boundaries.
Intrigued by online experiments, Southen decided to scrutinize Midjourney, an A.I. image generator. When prompted to create an image of Joaquin Phoenix as "The Joker," he was astonished when the system generated an image strikingly similar to a frame from the 2019 film.
Expanding his exploration with various prompts, Southen observed that terms like "Videogame hedgehog" led to the creation of Sonic, "Animated toys" depicted characters from Pixar's "Toy Story," and "popular movie screencap" resulted in an image of Iron Man striking a familiar pose.
Reflecting on the situation, Southen stressed, "What they're doing is clear evidence of exploitation and using intellectual property they don't have licenses for," highlighting the A.I. companies' use of copyrighted material.
These experiments, replicated by other artists, A.I. watchdogs, and journalists at The New York Times, raise crucial questions about the training data employed by A.I. systems and the potential infringement of copyright laws by these companies.
Legal disputes, including lawsuits filed by notable figures like Sarah Silverman and John Grisham, have brought the matter into the courtroom. A.I. companies argue that using copyrighted material falls within the realm of "fair use," a provision within copyright law permitting specific use of material. They also acknowledge a flaw known as "memorization," where A.I. systems reproduce copyrighted material too closely, a challenge they claim to be actively addressing. A.I. experts propose that memorization occurs when the training data is saturated with numerous similar or identical images, even extending to rare instances such as emails.
For example, when Southen requested Midjourney to generate a "Dune movie screencap" from the "Dune movie trailer," the model may have had limited options in its training data. The result was a frame almost identical to one from the movie's trailer, underscoring the difficulties faced by A.I. systems in avoiding such close reproductions.