Scientists Put AI Through A Repeated Experiment And Discovered A Troubling Pattern In Its Creativity

Artificial intelligence can generate millions of images, write endless stories and produce seemingly limitless variations of almost anything. But researchers have discovered something unsettling when AI systems are left to create without human intervention. After repeatedly converting images into descriptions and turning those descriptions back into images, the systems began producing almost identical-looking results. The original meaning gradually disappeared, leaving behind generic scenes that researchers described as “visual elevator music.”

The findings, published in January 2026, raise serious questions about what happens when artificial intelligence increasingly influences the content people consume. The experiment involved researchers Arend Hintze, Frida Proschinger Åström and Jory Schossau, who examined how generative systems behave when repeatedly processing their own outputs. Their results suggest that AI can gradually lose the distinctive details that make creative material interesting, even without being retrained on its previous creations. The implications could extend far beyond image generation, particularly as AI becomes increasingly involved in producing and transforming digital content.

Researchers Discovered AI Has A Surprising Creative Weakness

The January 2026 study examined what happens when artificial intelligence systems are allowed to communicate through a continuous creative loop. The researchers connected two different technologies: a text-to-image system that generates pictures from written descriptions and an image-to-text system that examines pictures and converts them into captions. They then allowed the systems to exchange outputs repeatedly, with each newly generated description becoming the instructions for another image. There was no human intervention to correct the direction or restore details that had disappeared along the way.

What happened next was unexpected. Despite beginning with different prompts and allowing randomness into the process, the generated images gradually became increasingly similar. Distinctive details disappeared, while familiar visual patterns took over. The systems repeatedly gravitated toward conventional scenes, including impressive buildings, atmospheric cityscapes and peaceful countryside landscapes. Although the resulting images often looked polished, they increasingly lacked the specific meaning and unusual characteristics contained in the original instructions.

The researchers described these increasingly generic outputs as “visual elevator music.” The comparison captures the strange quality of the images: they were aesthetically acceptable but lacked anything particularly distinctive. The experiment also suggested that the systems were losing important information during repeated transformations. Even when the starting prompts contained complicated scenarios, the resulting images gradually became simpler and more conventional. The process demonstrated how easily a creative system can drift toward familiar patterns when repeatedly processing its own material.

An Experiment Revealed How Quickly AI Loses Meaning

One example from the research makes the problem particularly easy to understand. The researchers began with a detailed description of a politician dealing with an enormous amount of pressure. The prompt described a prime minister examining strategy documents while attempting to persuade the public to support a fragile peace agreement. The politician was also dealing with the pressure of impending military action. It was a scenario containing political conflict, an identifiable central character, emotional tension and several important details that gave the image a specific narrative.

The image-generation system translated those instructions into a visual representation. Another AI system then examined the image and produced a written description of what it saw. That description became the next prompt, and the process continued repeatedly. Each new image became the subject of another caption, which then provided the instructions for the following generation. As the cycle continued, the original political situation became increasingly difficult to recognize.

After multiple rounds, the original scenario had practically vanished. The final image showed a formal interior space with elaborate furnishings and red curtains. The prime minister was gone, along with the political documents, the fragile peace agreement and the looming military conflict. A detailed situation involving a leader facing an important decision had effectively become an ordinary-looking room. The researchers documented this progression in a collage showing how the original image changed over successive rounds.

The experiment revealed several important details that disappeared during the process:

• The central character: The prime minister eventually disappeared from the image.
• The political circumstances: The fragile peace agreement and public negotiations were lost.
• The emotional tension: The pressure surrounding the politician’s responsibilities disappeared.
• The immediate threat: The approaching military action was no longer represented.
• The original narrative: The final image no longer communicated the situation described in the initial prompt.

This example demonstrates why the researchers’ findings attracted attention. The systems continued producing images throughout the experiment, but the original meaning became increasingly difficult to preserve.

Why Researchers Called The Results Visual Elevator Music

The phrase “visual elevator music” describes something familiar to anyone who has browsed through generic stock photography. A peaceful landscape, an impressive office building or a dramatic city skyline might look attractive, but these images rarely communicate a complicated story on their own. They provide something recognizable without necessarily offering a distinctive perspective. That was the direction the AI systems appeared to favor as the experiment continued.

The underlying issue involves how generative AI systems recognize and reproduce patterns. These technologies learn relationships between words, images, styles and visual characteristics from enormous collections of existing material. Those patterns allow them to generate new content based on instructions. However, the researchers’ experiment suggests that repeated conversions can favor details that are easier to recognize and describe, gradually pushing unusual characteristics out of the resulting material.

The researchers also explored whether introducing randomness could prevent the systems from converging toward familiar imagery. Their findings suggested that simply allowing more variation did not prevent the overall drift. This raises an interesting distinction between generating different outputs and producing genuinely diverse creative ideas. A system might create hundreds of images with different lighting, colors and compositions while remaining within a relatively narrow range of familiar visual themes.

The Bigger Concern Extends Beyond AI Images

The experiment might initially sound like a technical curiosity with little connection to everyday life. Most people do not spend their time repeatedly converting images into descriptions and back again. However, the underlying process has similarities to how digital information increasingly moves between different formats. A photograph becomes a caption, a news report becomes a summary, a written description becomes an illustration, and videos are transcribed before being transformed into other forms of content.

Artificial intelligence is becoming involved in many of these processes. The concern is that important details could gradually disappear as information passes through repeated AI-mediated transformations. Consider a complicated news event that begins with a detailed report containing specific people, circumstances and events. An AI system summarizes it, another system uses that summary to generate an image, and a separate platform then categorizes or summarizes the resulting content. At every stage, some information could be lost or simplified.

The original material might contain unusual details that make the story interesting. Yet those details may not survive repeated transformations, leaving behind a simplified version that is easier for another system to process. This is where the experiment connects with broader concerns about the future of online creativity. If more digital content is created, summarized, ranked and regenerated through similar systems, the preferences built into those technologies could influence what people eventually see.

Scientists Have Warned About AI Creating A Cultural Feedback Loop

For several years, researchers and critics have raised concerns about artificial intelligence eventually training on material generated by other AI systems. The concern is relatively straightforward: AI models learn from existing material, and if AI-generated content increasingly fills the internet, future systems could end up learning from outputs produced by earlier generations of AI. This creates the possibility of a feedback loop in which synthetic material increasingly influences the systems responsible for producing the next generation of content.

The concern is that repetitive patterns, errors and biases could become amplified over time. If future models depend heavily on material that has already been generated and processed by artificial intelligence, they could receive a narrower representation of human creativity. Original perspectives and unusual forms of expression might become less visible within the material used to develop future systems. However, there is an important distinction between this broader concern and what the January experiment actually demonstrated.

The researchers did not retrain their models on AI-generated images. They did not introduce new training data during the experiment. The convergence happened through repeated use of the systems themselves, meaning the process produced increasingly generic images without the additional influence of retraining. This distinction changes how the findings should be interpreted. It suggests that homogenization can emerge before the potential problems associated with training future models on synthetic material even begin.

Retraining could potentially amplify this effect, but it was not responsible for the results observed in this particular experiment. The findings therefore provide a different perspective on the cultural stagnation debate. The concern extends beyond what happens when AI learns from its own creations. It also involves what happens when existing systems repeatedly transform information and gradually favor familiar details over distinctive ones.

AI Could Be Producing More Content Without More Original Ideas

One of the most interesting questions raised by the research concerns the difference between producing content and producing creativity. Generative AI can create enormous quantities of material in a relatively short period. A user can request hundreds of images, dozens of article outlines or numerous variations of a particular concept. However, the quantity of output does not automatically demonstrate genuine creative diversity. A system can produce thousands of variations while repeatedly relying on similar underlying patterns.

Imagine asking an AI system to generate 100 different pictures of a city. The results might feature different buildings, lighting conditions, weather and architectural styles. Some could look futuristic, while others might resemble historic European streets or modern commercial districts. Yet they could all follow familiar visual conventions. The images would look different at first glance while exploring relatively similar creative territory. This is one of the central concerns raised by the researchers’ findings.

The original researcher also argued that creative AI systems need incentives to move away from conventional outputs. That could involve encouraging less common forms of expression and designing systems that reward genuine deviation from familiar patterns. Without such considerations, generating more variations might simply mean producing additional versions of what already exists. A system can appear endlessly creative while repeatedly returning to the same limited collection of ideas.

The issue becomes particularly relevant when people use AI-generated material as the starting point for further creative work. If writers, designers and content creators increasingly choose from suggestions generated by similar systems, their decisions could become influenced by the same preferences. The result could be an online environment containing enormous amounts of material that feels strangely repetitive, even when individual pieces appear different.

Previous Research Has Raised Similar Questions About AI Creativity

The January 2026 experiment is not the only research raising questions about artificial intelligence and creative diversity. The source analysis also points to research involving AI-generated lesson plans, which reportedly identified a similar drift toward conventional and uninspiring material. Educational lesson plans require more than presenting information in an organized format. They can involve unusual examples, creative teaching approaches and different ways of explaining difficult concepts to students.

If AI-generated lesson plans repeatedly favor conventional structures, teachers could receive material that looks useful but offers relatively little originality. Similar concerns could extend to advertising, entertainment, journalism and digital design. In each case, artificial intelligence might make content production faster while encouraging people to rely on familiar structures and ideas. The resulting material could be perfectly functional without offering much that feels genuinely different.

There is also a question about how people respond to AI-generated suggestions. When a system presents several possible headlines, designs or story concepts, users naturally make selections based on their preferences. But if the available options are already shaped by similar underlying patterns, those choices could reinforce the same creative tendencies. Over time, people might become accustomed to seeing and selecting conventional material because genuinely unusual alternatives appear less frequently.

This does not establish that AI systems are making every creative industry less original. However, it provides another reason to examine how these technologies influence the material people encounter. The difference between producing more content and expanding the boundaries of creativity deserves considerably more attention as AI becomes part of everyday digital work.

Why This Does Not Mean Human Creativity Is Finished

There is an important reason to avoid treating the research as proof that artificial intelligence will inevitably destroy creativity. New technologies have repeatedly changed how people produce and consume art. Photography transformed visual culture, cinema created entirely new storytelling possibilities, and digital editing gave artists access to tools that previous generations could barely imagine. These developments changed creative industries without eliminating the human desire to experiment with unfamiliar ideas.

Artificial intelligence could also become a valuable creative tool. Artists can use it to test unusual concepts, experiment with unfamiliar styles and produce visual material that would otherwise require significant resources. Writers can use it to generate alternative approaches to a story, while designers can explore possibilities before committing to a final direction. The technology offers opportunities for experimentation, provided people remain willing to question its suggestions and push beyond predictable results.

Human direction can make a difference in how these systems are used. An artist who uses AI to explore an unconventional concept can still produce work that reflects a distinct creative vision. The problem becomes more complicated when people rely on automated suggestions without questioning why certain outputs appear repeatedly. If the easiest option is always the most familiar one, unusual ideas may receive less attention, even when they have greater creative potential.

The researcher argued that creative AI systems should be designed with incentives to move beyond statistically average outputs. This could involve encouraging systems to preserve distinctive details, supporting less mainstream forms of expression and giving greater consideration to unusual creative directions. Human judgment could also remain part of important stages in the process, helping prevent original ideas from disappearing through repeated automated transformations.

The Future Of Online Creativity Could Depend On What We Choose To Preserve

The January 2026 findings offer a striking look at a problem that could become increasingly relevant as artificial intelligence spreads across creative industries. The experiment did not establish that all AI-generated content is generic or that cultural stagnation is inevitable. It did demonstrate that repeated transformations between images and text can cause distinctive meaning to disappear, even without retraining the systems on their previous outputs.

The internet can produce more content than ever before, but the value of creative expression depends partly on the unusual details, perspectives and ideas that make individual works worth experiencing. If technology makes producing familiar material easier while allowing distinctive ideas to disappear, the sheer volume of content could conceal a growing lack of originality.

The real challenge for AI developers and the creative industries is to build systems that can preserve what makes an idea unusual. Otherwise, the future of digital creativity could become an endless stream of polished images and stories that look different on the surface but keep telling the same story.