The idea that artificial intelligence might one day rival human creativity has become a familiar theme in public conversation. Generative models can rapidly produce images, stories and designs, which makes it tempting to assume that they possess something like imagination. A new study published in Advanced Science challenges that assumption in a direct and illuminating way. By examining how humans and AI generate images from abstract prompts, the researchers show that what looks like creativity on the surface is often a sophisticated form of imitation. Their findings, echoed in a plain‑language article in TechXplore, offer a clearer picture of what AI can and cannot do when it comes to creative thought.
The project began in 2024 during a workshop hosted by Fundació Èpica La Fura dels Baus, an organization known for bringing together artists, scientists and technologists. This interdisciplinary setting helped shape a study that approached creativity not as a single output but as a process that unfolds from the first spark of an idea to the final drawing on a page. That shift in perspective is important. Much of the public discussion about AI creativity focuses on the result, such as a striking image or a clever paragraph. The researchers wanted to look deeper, asking how those results come to be and whether the underlying process resembles anything like human imagination.
To do this, they designed a visual creative imagination task based on abstract shapes. These shapes served as starting points, much like the way a cloud might inspire a child to imagine a dragon or a ship. Human participants were divided into two groups, trained visual artists and members of the general population. Each person viewed the abstract stimuli, imagined possible images and then selected one to draw. This approach was intended to capture the full arc of human creative ideation, from internal visualization to physical execution.
The researchers then asked a Stable Diffusion model to perform the same task. Stable Diffusion is a text to image system that begins with random noise and gradually shapes it into a picture based on a written prompt. To make the comparison fair, the model was fine-tuned using human drawings. It was then given two types of prompts. In the human guided condition, the prompt included one of the ideas generated by a human participant. In unguided condition, the model received only a basic instruction. This allowed the researchers to test how much the model relied on human direction to produce something that might be judged as creative.
Once all the images were generated, a large group of human raters evaluated them using five criteria, including liking, vividness, originality, aesthetics and curiosity. Two AI systems also rated the images, one operating independently and one using examples of human ratings as guidance. This produced a rich dataset that captured both the creative process and the perception of creativity.
The results were consistent across every group of evaluators. Visual artists produced the most creative images. The general population followed. The human guided AI model came next. The unguided AI model performed worst by a wide margin. Even though the model had been trained in human drawings, it struggled to produce creative images on its own. When deprived of human guidance, its performance dropped even further. As one of the study’s authors explained, the model did worse when left to its own devices.
These findings help clarify what AI creativity means. Many people assume that if an AI system can produce something new, it must be creative. In practice, generative models work more like extremely advanced collage machines. They recombine patterns learned from vast datasets. This can produce impressive results, but it is not the same as forming an original idea. Human creativity involves weaving together perception, memory, emotion and intention. It is shaped by lived experience and personal meaning. AI systems do not have these ingredients. They operate through statistical prediction, not imagination.
The visual creative imagination task makes this distinction easier to see. When a person looks at an abstract shape and imagines a scene, they draw on a lifetime of sensory and emotional associations. An artist might see movement, symbolism or narrative potential. A non-artist might see something simpler but still personal. The AI model, by contrast, does not imagine anything. It responds to prompts by assembling visual features that resemble patterns it has seen before. Without human guidance, it lacks direction and purpose.
The study also sheds light on how creativity can be measured. By using multiple criteria and multiple types of raters, the researchers avoided the common pitfall of relying on a single metric. Creativity is not just novelty. It is also coherence, appeal and the ability to spark curiosity. The five part rating system captured these nuances, and the consistency of the results across human and guided AI rates strengthened the conclusions.
One of the most important insights from the study is the role of human guidance. When the AI model received a human generated idea, its output improved significantly. It still did not match human creativity, but it performed at a level like the general population. This suggests that AI can amplify human creativity when used as a tool, but it cannot replace the human spark that initiates the process. The researchers emphasize that human intervention is required at every stage, from training to ideation. Without it, the model falters.
These findings have several implications for how society understands and uses AI in creative fields. First, they highlight the limits of current technology. Generative models excel at producing variations on existing patterns, but they do not originate ideas. Second, they suggest that evaluating AI creativity through verbal tasks alone gives a misleading impression. Language based tests play to the strengths of large language models, which are trained on enormous text corpora. Visual imagination tasks reveal a different picture, one that exposes the gap between imitation and genuine creativity.
The study also points toward practical improvements. If AI systems are to be used responsibly in creative industries, their dependence on human guidance should be acknowledged and built into workflows. Artists, designers and educators can use these tools as partners rather than competitors, directing them with clear intent. Understanding the limits of AI creativity can help prevent overreliance on automated systems and encourage more thoughtful integration.
In the broader landscape of AI research, the study underscores the importance of examining creative processes rather than just creative outputs. Creativity is multifaceted. It involves perception, imagination, decision making and execution. By studying each component, researchers can better understand where AI aligns with human abilities and where it diverges. This approach can guide future work on models that support, rather than mimic, human creativity.
As generative AI becomes more widespread, the need for clarity about its capabilities grows. This study provides that clarity. It shows that while AI can produce visually appealing images, it does not imagine them. It depends on human ideas, human training and human interpretation. Creativity remains a distinctly human strength, shaped by experience and intention. AI can assist, but it cannot replace the imaginative process that defines human art and innovation.
Red Sky Alliance is a Cyber Threat Analysis and Intelligence Service organization. For questions, comments or assistance, please contact the office directly at 1-844-492-7225, or feedback@redskyalliance.com
Weekly Cyber Intelligence Briefings:
- Reporting: https://www. redskyalliance. org/
- Website: https://www. redskyalliance. com/
- LinkedIn: https://www. linkedin. com/company/64265941
Weekly Cyber Intelligence Briefings:
REDSHORTS - Weekly Cyber Intelligence Briefings
https://attendee.gotowebinar.com/register/5504229295967742989
https://six3ro.substack.com/p/when-machines-imitate-imagination
Comments