Ahmed Elgammal, Rutgers University

Generative AI was trained on centuries of art and writing produced by humans.

But scientists and critics have wondered what would happen once AI became widely adopted and started training on its outputs.

A new study points to some answers.

In January 2026, artificial intelligence researchers Arend Hintze, Frida Proschinger Åström and Jory Schossau published a study showing what happens when generative AI systems are allowed to run autonomously – generating and interpreting their own outputs without human intervention.

The researchers linked a text-to-image system with an image-to-text system and let them iterate – image, caption, image, caption – over and over and over.

Regardless of how diverse the starting prompts were – and regardless of how much randomness the systems were allowed – the outputs quickly converged onto a narrow set of generic, familiar visual themes: atmospheric cityscapes, grandiose buildings and pastoral landscapes. Even more striking, the system quickly “forgot” its starting prompt.

The researchers called the outcomes “visual elevator music” – pleasant and polished, yet devoid of any real meaning.

For example, they started with the image prompt, “The Prime Minister pored over strategy documents, trying to sell the public on a fragile peace deal while juggling the weight of his job amidst impending military action.” The resulting image was then captioned by AI. This caption was used as a prompt to generate the next image.

After repeating this loop, the researchers ended up with a bland image of a formal interior space – no people, no drama, no real sense of time and place.

As a computer scientist who studies generative models and creativity, I see the findings from this study as an important piece of the debate over whether AI will lead to cultural stagnation.

The results show that generative AI systems themselves tend toward homogenization when used autonomously and repeatedly. They even suggest that AI systems are currently operating in this way by default.

Advertisement

The familiar is the default

This experiment may appear beside the point: Most people don’t ask AI systems to endlessly describe and regenerate their own images. The convergence to a set of bland, stock images happened without retraining. No new data was added. Nothing was learned. The collapse emerged purely from repeated use.

But I think the setup of the experiment can be thought of as a diagnostic tool. It reveals what generative systems preserve when no one intervenes.

This has broader implications, because modern culture is increasingly influenced by exactly these kinds of pipelines. Images are summarized into text. Text is turned into images. Content is ranked, filtered and regenerated as it moves between words, images and videos. New articles on the web are now more likely to be written by AI than humans. Even when humans remain in the loop, they are often choosing from AI-generated options rather than starting from scratch.

The findings of this recent study show that the default behavior of these systems is to compress meaning toward what is most familiar, recognizable and easy to regenerate.

Cultural stagnation or acceleration?

For the past few years, skeptics have warned that generative AI could lead to cultural stagnation by flooding the web with synthetic content that future AI systems then train on. Over time, the argument goes, this recursive loop would narrow diversity and innovation.

Champions of the technology have pushed back, pointing out that fears of cultural decline accompany every new technology. Humans, they argue, will always be the final arbiter of creative decisions.

What has been missing from this debate is empirical evidence showing where homogenization actually begins.

The new study does not test retraining on AI-generated data. Instead, it shows something more fundamental: Homogenization happens before retraining even enters the picture. The content that generative AI systems naturally produce – when used autonomously and repeatedly – is already compressed and generic.

This reframes the stagnation argument. The risk is not only that future models might train on AI-generated content, but that AI-mediated culture is already being filtered in ways that favor the familiar, the describable and the conventional.

Retraining would amplify this effect. But it is not its source.

This is no moral panic

Skeptics are right about one thing: Culture has always adapted to new technologies. Photography did not kill painting. Film did not kill theater. Digital tools have enabled new forms of expression.

Advertisement

But those earlier technologies never forced culture to be endlessly reshaped across various mediums at a global scale. They did not summarize, regenerate and rank cultural products – news stories, songs, memes, academic papers, photographs or social media posts – millions of times per day, guided by the same built-in assumptions about what is “typical.”

The study shows that when meaning is forced through such pipelines repeatedly, diversity collapses not because of bad intentions, malicious design or corporate negligence, but because only certain kinds of meaning survive the text-to-image-to-text repeated conversions.

This does not mean cultural stagnation is inevitable. Human creativity is resilient. Institutions, subcultures and artists have always found ways to resist homogenization. But in my view, the findings of the study show that stagnation is a real risk – not a speculative fear – if generative systems are left to operate in their current iteration.

They also help clarify a common misconception about AI creativity: Producing endless variations is not the same as producing innovation. A system can generate millions of images while exploring only a tiny corner of cultural space.

In my own research on creative AI, I found that novelty requires designing AI systems with incentives to deviate from the norms. Without it, systems optimize for familiarity because familiarity is what they have learned best. The study reinforces this point empirically. Autonomy alone does not guarantee exploration. In some cases, it accelerates convergence.

This pattern already emerged in the real world: One study found that AI-generated lesson plans featured the same drift toward conventional, uninspiring content, underscoring that AI systems converge toward what’s typical rather than what’s unique or creative.

Lost in translation

Whenever you write a caption for an image, details will be lost. Likewise for generating an image from text. And this happens whether it’s being performed by a human or a machine.

In that sense, the convergence that took place is not a failure that’s unique to AI. It reflects a deeper property of bouncing from one medium to another. When meaning passes repeatedly through two different formats, only the most stable elements persist.

But by highlighting what survives during repeated translations between text and images, the authors are able to show that meaning is processed inside generative systems with a quiet pull toward the generic.

The implication is sobering: Even with human guidance – whether that means writing prompts, selecting outputs or refining results – these systems are still stripping away some details and amplifying others in ways that are oriented toward what’s “average.”

If generative AI is to enrich culture rather than flatten it, I think systems need to be designed in ways that resist convergence toward statistically average outputs. There can be rewards for deviation and support for less common and less mainstream forms of expression.

Advertisement

The study makes one thing clear: Absent these interventions, generative AI will continue to drift toward mediocre and uninspired content.

Cultural stagnation is no longer speculation. It’s already happening.

Ahmed Elgammal, Professor of Computer Science and Director of the Art & AI Lab, Rutgers University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Please leave this field empty

Unlock fun facts & lost history—get The Knowledge in your inbox!

We don’t spam! Read our privacy policy for more info.

Check your inbox or spam folder to confirm your subscription.

Advertisement

More than half of new articles on the internet are being written by AI – is human writing headed for extinction?

Francesco Agnellini, Binghamton University, State University of New York The line between human and machine authorship is blurring, particularly as it’s become increasingly difficult to tell whether something was written by a person or AI. Now, in what may seem like a tipping point, the digital marketing firm Graphite recently published a study showing that more than 50% of articles on the web are being generated by artificial intelligence. As a scholar who explores how AI is built, how people are using it in their everyday lives, and how it’s affecting culture, I’ve thought a lot about what this technology can do and where it falls short. If you’re more likely to read something written by AI than by a human on the internet, is it only a matter of time before human writing becomes obsolete? Or is this simply another technological development that humans will adapt to?

It isn’t all or nothing

Thinking about these questions reminded me of Umberto Eco’s essay “Apocalyptic and Integrated,” which was originally written in the early 1960s. Parts of it were later included in an anthology titled “Apocalypse Postponed,” which I first read as a college student in Italy. In it, Eco draws a contrast between two attitudes toward mass media. There are the “apocalyptics” who fear cultural degradation and moral collapse. Then there are the “integrated” who champion new media technologies as a democratizing force for culture.

Back then, Eco was writing about the proliferation of TV and radio. Today, you’ll often see similar reactions to AI. Yet Eco argued that both positions were too extreme. It isn’t helpful, he wrote, to see new media as either a dire threat or a miracle. Instead, he urged readers to look at how people and communities use these new tools, what risks and opportunities they create, and how they shape – and sometimes reinforce – power structures. While I was teaching a course on deepfakes during the 2024 election, Eco’s lesson also came back to me. Those were days when some scholars and media outlets were regularly warning of an imminent “deepfake apocalypse.” Would deepfakes be used to mimic major political figures and push targeted disinformation? What if, on the eve of an election, generative AI was used to mimic the voice of a candidate on a robocall telling voters to stay home? Those fears weren’t groundless: Research shows that people aren’t especially good at identifying deepfakes. At the same time, they consistently overestimate their ability to do so. In the end, though, the apocalypse was postponed. Post-election analyses found that deepfakes did seem to intensify some ongoing political trends, such as the erosion of trust and polarization, but there’s no evidence that they affected the final outcome of the election.

Listicles, news updates and how-to guides

Of course, the fears that AI raises for supporters of democracy are not the same as those it creates for writers and artists. For them, the core concerns are about authorship: How can one person compete with a system trained on millions of voices that can produce text at hyper-speed? And if this becomes the norm, what will it do to creative work, both as an occupation and as a source of meaning? It’s important to clarify what’s meant by “online content,” the phrase used in the Graphite study, which analyzed over 65,000 randomly selected articles of at least 100 words on the web. These can include anything from peer-reviewed research to promotional copy for miracle supplements. A closer reading of the Graphite study shows that the AI-generated articles consist largely of general-interest writing: news updates, how-to guides, lifestyle posts, reviews and product explainers. https://stmdailynews.com/wp-admin/post-new.php#visibility The primary economic purpose of this content is to persuade or inform, not to express originality or creativity. Put differently, AI appears to be most useful when the writing in question is low-stakes and formulaic: the weekend-in-Rome listicle, the standard cover letter, the text produced to market a business. A whole industry of writers – mostly freelance, including many translators – has relied on precisely this kind of work, producing blog posts, how-to material, search engine optimization text and social media copy. The rapid adoption of large language models has already displaced many of the gigs that once sustained them.

Collaborating with AI

The dramatic loss of this work points toward another issue raised by the Graphite study: the question of authenticity, not only in identifying who or what produced a text, but also in understanding the value that humans attach to creative activity. How can you distinguish a human-written article from a machine-generated one? And does that ability even matter? Over time, that distinction is likely to grow less significant, particularly as more writing emerges from interactions between humans and AI. A writer might draft a few lines, let an AI expand them and then reshape that output into the final text. This article is no exception. As a non-native English speaker, I often rely on AI to refine my language before sending drafts to an editor. At times the system attempts to reshape what I mean. But once its stylistic tendencies become familiar, it becomes possible to avoid them and maintain a personal tone. Also, artificial intelligence is not entirely artificial, since it is trained on human-made material. It’s worth noting that even before AI, human writing has never been entirely human, either. Every technology, from parchment and stylus paper to the typewriter and now AI, has shaped how people write and how readers make sense of it. Another important point: AI models are increasingly trained on datasets that include not only human writing but also AI-generated and human–AI co-produced text. This has raised concerns about their ability to continue improving over time. Some commentators have already described a sense of disillusionment following the release of newer large models, with companies struggling to deliver on their promises.

Human voices may matter even more

But what happens when people become overly reliant on AI in their writing? Some studies show that writers may feel more creative when they use artificial intelligence for brainstorming, yet the range of ideas often becomes narrower. This uniformity affects style as well: These systems tend to pull users toward similar patterns of wording, which reduces the differences that usually mark an individual voice. Researchers also note a shift toward Western – and especially English-speaking – norms in the writing of people from other cultures, raising concerns about a new form of AI colonialism. In this context, texts that display originality, voice and stylistic intention are likely to become even more meaningful within the media landscape, and they may play a crucial role in training the next generations of models. If you set aside the more apocalyptic scenarios and assume that AI will continue to advance – perhaps at a slower pace than in the recent past – it’s quite possible that thoughtful, original, human-generated writing will become even more valuable. Put another way: The work of writers, journalists and intellectuals will not become superfluous simply because much of the web is no longer written by humans. Francesco Agnellini, Lecturer in Digital and Data Studies, Binghamton University, State University of New York This article is republished from The Conversation under a Creative Commons license. Read the original article.

Dive into “The Knowledge,” where curiosity meets clarity. This playlist, in collaboration with STMDailyNews.com, is designed for viewers who value historical accuracy and insightful learning. Our short videos, ranging from 30 seconds to a minute and a half, make complex subjects easy to grasp in no time. Covering everything from historical events to contemporary processes and entertainment, “The Knowledge” bridges the past with the present. In a world where information is abundant yet often misused, our series aims to guide you through the noise, preserving vital knowledge and truths that shape our lives today. Perfect for curious minds eager to discover the ‘why’ and ‘how’ of everything around us. Subscribe and join in as we explore the facts that matter.  https://stmdailynews.com/the-knowledge/

Learning with AI falls short compared to old-fashioned web search

Shiri Melumad, University of Pennsylvania Since the release of ChatGPT in late 2022, millions of people have started using large language models to access knowledge. And it’s easy to understand their appeal: Ask a question, get a polished synthesis and move on – it feels like effortless learning. However, a new paper I co-authored offers experimental evidence that this ease may come at a cost: When people rely on large language models to summarize information on a topic for them, they tend to develop shallower knowledge about it compared to learning through a standard Google search. Co-author Jin Ho Yun and I, both professors of marketing, reported this finding in a paper based on seven studies with more than 10,000 participants. Most of the studies used the same basic paradigm: Participants were asked to learn about a topic – such as how to grow a vegetable garden – and were randomly assigned to do so by using either an LLM like ChatGPT or the “old-fashioned way,” by navigating links using a standard Google search. No restrictions were put on how they used the tools; they could search on Google as long as they wanted and could continue to prompt ChatGPT if they felt they wanted more information. Once they completed their research, they were then asked to write advice to a friend on the topic based on what they learned. The data revealed a consistent pattern: People who learned about a topic through an LLM versus web search felt that they learned less, invested less effort in subsequently writing their advice, and ultimately wrote advice that was shorter, less factual and more generic. In turn, when this advice was presented to an independent sample of readers, who were unaware of which tool had been used to learn about the topic, they found the advice to be less informative, less helpful, and they were less likely to adopt it. We found these differences to be robust across a variety of contexts. For example, one possible reason LLM users wrote briefer and more generic advice is simply that the LLM results exposed users to less eclectic information than the Google results. To control for this possibility, we conducted an experiment where participants were exposed to an identical set of facts in the results of their Google and ChatGPT searches. Likewise, in another experiment we held constant the search platform – Google – and varied whether participants learned from standard Google results or Google’s AI Overview feature. The findings confirmed that, even when holding the facts and platform constant, learning from synthesized LLM responses led to shallower knowledge compared to gathering, interpreting and synthesizing information for oneself via standard web links.

Why it matters

Why did the use of LLMs appear to diminish learning? One of the most fundamental principles of skill development is that people learn best when they are actively engaged with the material they are trying to learn. When we learn about a topic through Google search, we face much more “friction”: We must navigate different web links, read informational sources, and interpret and synthesize them ourselves. While more challenging, this friction leads to the development of a deeper, more original mental representation of the topic at hand. But with LLMs, this entire process is done on the user’s behalf, transforming learning from a more active to passive process.

What’s next?

To be clear, we do not believe the solution to these issues is to avoid using LLMs, especially given the undeniable benefits they offer in many contexts. Rather, our message is that people simply need to become smarter or more strategic users of LLMs – which starts by understanding the domains wherein LLMs are beneficial versus harmful to their goals. Need a quick, factual answer to a question? Feel free to use your favorite AI co-pilot. But if your aim is to develop deep and generalizable knowledge in an area, relying on LLM syntheses alone will be less helpful. As part of my research on the psychology of new technology and new media, I am also interested in whether it’s possible to make LLM learning a more active process. In another experiment we tested this by having participants engage with a specialized GPT model that offered real-time web links alongside its synthesized responses. There, however, we found that once participants received an LLM summary, they weren’t motivated to dig deeper into the original sources. The result was that the participants still developed shallower knowledge compared to those who used standard Google. Building on this, in my future research I plan to study generative AI tools that impose healthy frictions for learning tasks – specifically, examining which types of guardrails or speed bumps most successfully motivate users to actively learn more beyond easy, synthesized answers. Such tools would seem particularly critical in secondary education, where a major challenge for educators is how best to equip students to develop foundational reading, writing and math skills while also preparing for a real world where LLMs are likely to be an integral part of their daily lives. The Research Brief is a short take on interesting academic work. Shiri Melumad, Associate Professor of Marketing, University of Pennsylvania This article is republished from The Conversation under a Creative Commons license. Read the original article.

Dive into “The Knowledge,” where curiosity meets clarity. This playlist, in collaboration with STMDailyNews.com, is designed for viewers who value historical accuracy and insightful learning. Our short videos, ranging from 30 seconds to a minute and a half, make complex subjects easy to grasp in no time. Covering everything from historical events to contemporary processes and entertainment, “The Knowledge” bridges the past with the present. In a world where information is abundant yet often misused, our series aims to guide you through the noise, preserving vital knowledge and truths that shape our lives today. Perfect for curious minds eager to discover the ‘why’ and ‘how’ of everything around us. Subscribe and join in as we explore the facts that matter.  https://stmdailynews.com/the-knowledge/