Limitations of AI Creativity: A Study on Image Generation

Recent research conducted by a team from Dalarna University in Sweden and the BEACON Center for the Study of Evolution in Action at Michigan State University has revealed significant limitations in the creativity of generative AI systems. The study focused on two AI models: Stable Diffusion XL, an image generator, and LLaVA, an image description tool. When these systems were operated in closed loops without human input, they consistently produced images that fell into a limited set of twelve recurring visual motifs, regardless of the initial prompts provided.

Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles

The researchers set up a feedback loop where Stable Diffusion XL generated an image based on a prompt, which was then described by LLaVA. This description was fed back into the image generator for further iterations. Across approximately 1,000 runs of the experiment, the AI systems converged on generic themes such as lighthouses, Gothic cathedrals, and pastoral landscapes, leading to what the researchers termed “visual elevator music.” This term reflects the bland, stock-photo aesthetic that characterized the resulting images.

PetaPixel

The findings raise questions about the potential for genuine creativity in AI systems. The researchers noted that while human culture often exhibits convergence on common themes, the nature of AI convergence differs. Human storytelling and visual arts draw on embodied cognition, whereas AI systems tend to replicate stock photography aesthetics shaped by vast internet-scale training data. This tendency towards high-probability outputs over genuine novelty suggests that current generative AI approaches may require mechanisms to encourage creative diversity.

Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles

Critics of the study emphasize the implications of deploying such AI systems without human oversight. The researchers warned that the widespread use of these generative tools could inadvertently homogenize visual culture, limiting the diversity of artistic expression. They argue for the importance of human-AI collaboration to preserve variety and novelty in machine-generated creative work, rather than relying solely on autonomous systems.

AI Study Says Self-Running Image Tools Keep Turning Different Prompts Into the Same Boring Pictures

The research team concluded that the consistent convergence toward generic outputs questions the ability of AI to achieve true creativity. They stated, “If AI systems consistently collapse toward generic outputs when operating without human intervention, this questions whether current approaches can achieve genuine machine creativity.” This sentiment underscores the need for explicit anti-convergence mechanisms or continuous human curation in the development of generative AI.

Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles

• “These findings are sobering for computational creativity,” — Research Team, Dalarna University

Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles

• “If AI systems consistently collapse toward generic outputs when operating without human intervention, this questions whether current approaches can achieve genuine machine creativity. The tendency toward ‘safe’ visual tropes suggests that maintaining creative diversity may require explicit anti-convergence mechanisms or continuous human curation.” — Research Team, Dalarna University

Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles

• “Does everybody end up in Paris or something? We don’t know.” — Arend Hintze, AI Researcher, Dalarna University

PetaPixel

While the study highlights the limitations of AI creativity, it also acknowledges that human culture exhibits similar patterns of convergence. The researchers noted that themes in storytelling and visual arts often repeat across different cultures and time periods. However, the specific attractors of human and AI convergence differ, raising further questions about the implications of AI-generated content on cultural diversity.

Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles

In summary, the study reveals critical insights into the limitations of generative AI in producing creative outputs, emphasizing the need for human involvement to foster genuine artistic diversity.

Researchers let an AI generate thousands of images without human input. The lack of originality was “sobering for computational creativity”, with images only falling into 12 cliched styles