The modern conversation around artificial intelligence no longer unfolds quietly within academic labs or policy briefings. It spills across headlines, social media, corporate pitches, and dinner table debates. And more often than not, it swings between two exaggerated poles: AI is either the savior of civilization or the destroyer of human thought. This seesaw of enthusiasm and dread has become a kind of theater, where the audience is kept in a state of suspense, rarely invited to reflect deeply on what is actually being performed.
It is tempting to believe that AI is accelerating so rapidly that we must constantly update our beliefs in dramatic ways. One month, we are told that machines can pass bar exams and write poetry better than most humans. The next, we are warned that those same machines are eroding our cognitive capacities, turning us into passive consumers of algorithmic wisdom. This pendulum, swinging from miracle to menace, reveals less about AI itself and more about how our cultural and media systems metabolize novelty.
What gets lost in this spectacle is any steady ground. Public understanding is shaped not by careful reasoning but by headlines that flatten nuance into fear or fantasy. And within this distorted landscape, even sincere research can become ammunition for stories that confuse more than they clarify. Two recent examples illustrate this pattern well: the “Potemkin Understanding” paper on large language models, and the MIT Media Lab’s study on AI’s effects on brain activity.
The Potemkin Paper: Performing Insight
Published in mid-2025, the paper titled “Potemkin Understanding in Large Language Models” proposes a category of failure where a language model appears to grasp a concept by answering key definitional questions correctly but fails when asked to use that concept in more natural, generative ways. The authors describe this as a Potemkin phenomenon; a reference to the historical tale of fake villages constructed to impress Catherine the Great. In their view, the model’s correct answers are a façade, hiding the absence of genuine understanding.
On the surface, the study is methodologically sound. It attempts to test models not just on rote factual knowledge, but on their ability to apply that knowledge in context. The idea is not new; researchers have long known that language models simulate understanding rather than possess it. But the paper’s value lies in making that failure visible in structured ways. It pushes the reader to think not just about what a model can answer, but how coherent its behavior is across tasks.
Still, something about the framing feels off. The very choice of the word “Potemkin” carries a charge of deception. It suggests that the model is not just limited, but pretending. And while the authors do not accuse the model of lying, such a thing would be nonsensical, they do set a tone that invites that conclusion. The metaphor’s historical weight is not neutral. It implies political theater, hollow performance, and a kind of intellectual fraudulence.
Once picked up by the media, the paper was presented not as a refinement of evaluation metrics, but as a revelation: AI doesn’t really understand. The irony, of course, is that this was never in doubt for those who understood how LLMs work. They generate plausible continuations of language based on training data. They do not form beliefs, and they do not comprehend in any human sense. By giving a catchy name to a well-known limitation, the paper became fuel for another round of performative shock.
The deeper concern is not that the paper was wrong, it wasn’t, but that it contributed, however unintentionally, to a kind of theater it should have resisted. The spectacle of “Potemkin intelligence” now joins a growing gallery of metaphors, stochastic parrots, mindless oracles, algorithmic charlatans, that draw more heat than light. They entertain. They alarm. But they rarely clarify.
The MIT Brain Study: Manufacturing Moral Panic
Around the same time, another study made waves across news platforms. Conducted by researchers at the MIT Media Lab, the project used brainwave monitoring to measure how students performed on essay-writing tasks under three conditions: using ChatGPT, using Google, or using their own memory. The results, like those from many psychological studies, were complex and cautious. When participants relied on ChatGPT early in the writing process, their neural activity decreased. They also felt less ownership of the content and retained less information afterward.
The researchers framed their findings with care. They did not claim that AI harms the brain or causes long-term damage. They emphasized that what they observed was a temporary shift in cognitive effort, possibly a form of mental offloading. The data did not show that people became less intelligent, only that their engagement changed depending on how and when they used the tool.
None of this stopped the headlines. Articles declared that AI was “making us dumb,” that ChatGPT “rots your brain,” and that the act of using language models could stunt cognitive development. Once again, a nuanced finding was repackaged into a moral panic. The tone suggested that using AI is not just risky, but corrupting; that we should fear not just what AI does, but what it might turn us into.
This response mirrors earlier fears about calculators, television, search engines, and smartphones. Every new medium that changes how we think is first met with anxiety about decline. But what makes this cycle more dangerous now is the pace and scale of the amplification. A preliminary preprint with a small sample size and modest claims can become, within hours, a global headline. The study was rigorous. The media treatment was not.
The Incentive to Sensationalize
Neither of these cases is isolated. They point to a structural issue that affects how AI research is conducted, reported, and understood. There is a powerful incentive to make ideas go viral. Researchers are encouraged to give their findings memorable names. Journalists are rewarded for framing studies in ways that provoke reaction. Corporate actors often quietly benefit from both ends of the spectrum, using fear to justify control, and hype to drive adoption.
What this produces is not an honest map of technological change, but a series of distorted landmarks. The public sees not the complexity of how tools work or how people engage with them, but cartoon versions: the brilliant AI genius or the brain-rotting shortcut. These images are vivid and memorable. But they are also misleading.
Even within the academic community, there is a tendency to reward spectacle. Papers that frame their contributions as breakthroughs or red flags are more likely to be cited. Those that offer modest refinements or challenge existing metaphors are often overlooked. In this environment, clarity becomes a liability. Complexity is a hard sell.
There is also the issue of narrative fatigue. When everything is either a revolution or a warning sign, we become desensitized to both. We stop asking what is actually useful, and we lose our ability to reason carefully. This is not just bad for public understanding. It is bad for science.
Toward a Culture of Honest Engagement
So what might a healthier discourse look like?
It would begin with restraint. Not every paper needs a metaphor. Not every finding needs a headline. Researchers can contribute by describing their work in plain language, acknowledging its limits, and resisting the urge to frame every result as a turning point.
Journalists, too, bear responsibility. They can ask better questions. What does this study actually show? What does it not show? How does it fit with what we already know? If science is to inform the public, it must do so through trust, not spectacle.
And readers, whether policymakers, professionals, or curious observers, must slow down. The most interesting questions about AI are not whether it is smarter than us or making us stupid, but how it is changing our habits of thought, our modes of collaboration, and our expectations of knowledge. These are not easy things to measure. They unfold over time, through experience and reflection.
What we need is a culture of patient observation. We need space for tools to be tools, for systems to be evaluated on their function, not their resemblance to minds. The more we treat language models like people, the less we understand them. And the more we reduce ourselves to numbers on a brain scan, the less we understand ourselves.
Beyond the Spectacle
Artificial intelligence is not magic, and it is not a monster. It is a mirror with strange properties, reflecting us back in ways we do not always expect. When we exaggerate its brilliance or fear its influence, we project our own insecurities onto the surface. And when we allow research to become theater, we trade accuracy for impact.
The two papers examined here are not flawed in their methods, but in how they were received. They deserve a better conversation; one that is sober, attentive, and committed to intellectual honesty. The real challenge is not what AI will do to us, but what we allow our stories about AI to do to our thinking.
It is not intelligence that is being automated. It is the narrative itself. And if we are not careful, we will find ourselves repeating headlines instead of forming thoughts. The true risk is not that machines are replacing our minds. It is that the spectacle is replacing our seriousness.
Image: Pixabay
Tom’s Blog 