The increasing prevalence of generative AI tools has led to a surge in accusations against game developers, who are frequently challenged over the authenticity of their visual assets, even when they express reservations about the technology themselves. This phenomenon extends beyond the gaming industry, impacting public discourse and the perception of digital content.
The nature of AI-generated imagery is fluid, lacking a fixed style or quality as it derives from vast training datasets. This ambiguity, compounded by the rapid evolution of AI tools, makes it increasingly difficult to distinguish between human-created and AI-generated art. Developers report that even when they actively condemn generative AI, they are still compelled to defend their work against claims of using such technology.
Failbetter Games, creators of titles like Fallen London, have experienced this firsthand. Hannah Flynn, the studio's communications director, noted that a common occurrence following significant announcements or widely shared videos is the query about generative AI usage. While some community members express genuine concern rooted in the studio's stated policy against AI, others are less invested "social media fly-bys." Failbetter maintains a firm stance, emphasizing their commitment to craft and disinterest in becoming "AI wranglers."
Similarly, David Oshry, CEO of New Blood Interactive, has faced persistent accusations regarding their upcoming horror game, Tenebris Somnia. The studio released a behind-the-scenes video showcasing the creation of physical props to counter claims that chatbot-generated assets were used. Oshry describes the accusations as "pretty much non stop," particularly on platforms like Instagram, which he believes may be contributing to a societal misunderstanding of digital art creation. Despite the exhausting nature of refuting these claims, the attention has positively impacted game wishlists. However, Oshry expresses a preference for creating content about the game itself rather than dedicating resources to proving it is not "AI slop."
Abandoned Sheep, the developers behind Schrodinger's Cat Burglar, opted to replace a piece of key art that was repeatedly labeled as AI-generated. Martin Binfield, the studio director, explained that the artwork, hand-painted by Tom Williams in mid-2022, predated generative AI's capability to produce work of that caliber. Although Binfield understands the vigilance of commenters concerned about AI, the studio found it more practical to update the art, which was already due for a refresh. He acknowledges that some accusations stem from a perception of art being "too good to be true," particularly when aspects like airbrushing or detailed reflections appear unnaturally polished.
The difficulty in identifying AI-generated content is exacerbated by the fact that the technology itself is constantly advancing, making established visual hallmarks quickly outdated. Advocates for generative AI often present its capabilities in broad, investor-focused terms, which further muddies the waters. This confusion places a burden on users who have become accustomed to casually labeling content as "AI slop."
Experts suggest that honing visual and technical literacy is crucial for navigating this landscape. Janna Sophia Koppenwallner, an illustrator researching human AI detection accuracy, noted that while initial studies suggested human detection capabilities were no better than chance, her ongoing research indicates an average accuracy of around 74%. However, she also observed a tendency for individuals to become overly suspicious, leading to misidentification of real art as AI-generated. Koppenwallner points out that AI models often exhibit distinct failure patterns, such as blending elements unnaturally, which differ from the types of errors human beginners make, like struggling with proportions or lighting.
Beyond generative AI, older machine learning tools used in graphics production can also be a source of misidentification. Abdou Bouam, a freelance artist, recounted an instance where a client questioned the authenticity of images due to perceived blurriness, which were actually the result of denoising methods used in rendering 3D scenes. Denoising, essential for cleaning up noisy renders produced by techniques like path-tracing, can sometimes introduce artifacts that resemble AI generation, such as lost details or smearing.
Bouam distinguishes these denoising tools, like Nvidia Optix, from generative AI. While they utilize machine learning, they predate current generative AI and are trained on specific datasets rather than indiscriminately scraped internet data. He argues that these tools do not alter artistic intent or character appearance, maintaining the artist's control over the final output.
As generative AI continues to integrate into various fields, developers and critics alike emphasize caution. Koppenwallner advises that visual detection is becoming increasingly difficult, with context playing a more significant role. She stresses the importance of carefully considering the circumstances surrounding an image before making accusations, as the technology continues its rapid and often inscrutable evolution.