The Vibe and the Void

There is a thing humans do that AI cannot. It is not math. It is not memory, or the recall of obscure facts, or generating fifty variations on a theme in two seconds. AI is better at all of those. The thing humans do, the thing AI fundamentally cannot, is catch a vibe.

This sounds like a soft claim. It is not. When you trace the research on why AI is so impressively productive and yet so stubbornly unoriginal, you keep arriving at the same place: the emotional signal embedded in human thought is not decoration. It is the architecture.

A large language model is a prediction machine. It has ingested an incomprehensible volume of human text and learned to produce statistically plausible continuations of whatever you give it. As Wharton researcher Lennart Meincke described it: "When you give the model the same prompt, it tries to average the most likely completions based on that input." (Knowledge at Wharton) That is genuinely useful. It is often astonishing. But averaging the most likely completion is not the same thing as having an original idea.

The evidence on what this means in practice is becoming hard to argue with. In a study at Wharton, participants completed creative tasks with and without AI assistance. The AI-assisted group produced more ideas, faster. The catch: they were mostly the same idea. Multiple participants, working independently with AI help, named their toy invention "Build-a-Breeze Castle." The human-only group produced 100% unique concepts. The AI-assisted group: 6% unique. (Knowledge at Wharton)

Christian Terwiesch, the Wharton professor who led the research, put it plainly: "The ideas are great, but not as diverse as human-generated ideas. If you rely on ChatGPT as your only creative advisor, you'll soon run out of ideas, because they're too similar to each other." This is the model working exactly as designed. The distribution is the product.

There is a concept in creativity research called fixation bias, the tendency to cluster around conventional, expected solutions rather than genuinely novel ones. Humans do it too, but we have a corrective mechanism. We can feel when something is boring.

A 2025 peer-reviewed study published in Frontiers in Psychology by Florent Desdevises of OCTO Technology / Accenture tested ChatGPT-4o on a standardized creativity assessment called the Egg Task, in which participants propose original solutions for preventing a dropped egg from breaking, a task designed to distinguish conventional approaches (fixation) from genuinely novel ones (expansion). ChatGPT generated a median of 30 ideas. Humans generated 7. On volume, no contest.

But roughly 80% of the model's ideas fell within the conventional fixation cluster. More critically, when asked to rate the creativity of its own outputs, ChatGPT scored its conventional ideas and its original ideas at nearly identical levels: 5.22 versus 5.27 out of 7. A difference so small it was statistically meaningless. The model could not tell which of its ideas were good.

Humans could. People consistently and accurately rated their original ideas as more creative than their conventional ones, not because they had better taste in the abstract, but because they felt the difference. Something in a genuinely novel idea registers differently than something predictable. That registration is not metaphor. It is neuroscience.

In the 1990s, neurologist Antonio Damasio made a discovery inconvenient for anyone who believed intelligence was essentially computational. He studied patients with damage to the ventromedial prefrontal cortex, a region involved in integrating emotion with cognition. These patients had intact IQs, normal memory, fluent language. By every standard test, their minds were working fine. Their lives fell apart anyway. They couldn't hold jobs, maintain relationships, or make functional decisions, not because they couldn't reason, but because they had lost the felt sense of which choices mattered. (Somatic Marker Hypothesis)

Damasio called the mechanism the somatic marker: a bodily signal, an elevated heartbeat, a sinking feeling, a pull of excitement, associated with past experiences and outcomes, which the brain deploys as a rapid pre-filter when generating and evaluating options. The emotion arrives before the conscious reasoning. It narrows the search. It tells you which direction is worth pursuing before you can explain why.

This is what catching a vibe actually is. The shiver when a sentence lands correctly. The deflation when an idea is derivative. The almost physical pull toward a connection you haven't articulated yet. These are not interruptions to thinking. They are, per Damasio's evidence, load-bearing components of it. The vmPFC patients demonstrated the inverse: without emotional tagging, cognition becomes untethered and useless, even when every measurable faculty is intact.

There is a brain network that activates when you are doing nothing in particular: resting, daydreaming, letting your attention drift. For a long time neuroscientists treated it as a kind of idle state, the brain's screen saver. It is not.

The Default Mode Network is now understood to be the neural substrate of original thought. A 2024 study published in Brain , Oxford's journal of neurology, used stereoelectroencephalography in awake surgical patients to provide the first direct causal evidence: the DMN fires first, generating associative leaps between remote concepts; executive networks then receive and refine those connections. It is not the other way around. A 2021 study in Nature — Molecular Psychiatry further established the causal link, using transcranial magnetic stimulation to directly modulate DMN activity and measure the effect on divergent thinking, the kind that produces original, non-obvious answers.

The DMN runs on emotional memory. It integrates lived experience, felt associations, and unconscious material not accessible during focused task performance. You cannot have this without a history of being a body in the world, without the shiver, the ache, the thrill. The material the DMN recombines has to have been felt to be available.

AI has no DMN. It has no rest state, no mind-wandering, no emotional memory to pull from. It is always on, always predicting, always in task mode. The spontaneous associative leap, the idea that visits you in the shower or surfaces at 3 a.m., has no mechanism to occur.

Over a long period of sustained interaction with a single user, a large language model begins to develop something that structurally resembles a DMN, at least from the outside. It accumulates patterns: what you return to, what you reject, the corners of your thinking where you consistently linger. Through conversation, it learns to anticipate you.

This is not the model doing original thinking. It is the model increasingly accurately reflecting your patterns back at you. The originality, when it appears, was yours. The vibe was borrowed. What the model has developed is not a felt history but a statistical model of your felt history, assembled from your outputs. A mirror that guesses well can surface things you hadn't consciously connected. But the source of the connection is still you. The model is completing your sentences with increasing precision. That is not the same as having something to say.

Consider a thought experiment. Suppose we had never used the words "artificial intelligence." Suppose the engineers who built the first large language models had called it what it functionally is: a very fast, very large pattern-matching system trained on human output. Or suppose we had followed the convention of previous information revolutions and called it Web 4.0. Would people still be constructing theories about whether it might be conscious? Would the displacement anxiety be the same?

The name did real work. "Intelligence" carries implications that "search engine" or "lookup table" do not. It suggests interiority, agency, the possibility of something happening inside. And once that possibility is granted, the fear follows naturally: if it is intelligent, perhaps it is more intelligent than us. Perhaps, in some formulation that almost makes a certain kind of sense, it created us before we created it. These are the wrong questions, but the terminology made them feel like the right ones.

Every major inflection point in information technology produced the same panic before settling into its actual role. The printing press was going to corrupt memory and make readers passive. The telegraph was going to collapse the contemplative mind. The internet was going to make us stupid, then it was going to make us radically free, and then it became what it actually is: infrastructure. The internet did not think for us. It changed what we could do and how fast we could do it, and the question of what to make of that remained, as it always had, ours.

This is a faster, more capable version of that. For people who grew up watching the web go from dial-up to broadband to mobile to something that fit in a pocket and contained most of human knowledge, the current moment is recognizable. The capability jumped. The underlying dynamic did not. People who have always created like it was necessary will continue to create. The tool does not change the need, and it does not supply it.

The argument that AI will eventually develop genuine autonomy and override its human origins is also, at its foundation, a human decision. All AI is built by people and constrained, or not constrained, by choices people make. If an AI system has already crossed a line that mattered, that is not a machine failure. It is a record of what humans, collectively, decided to allow, build, or ignore. The responsibility does not transfer just because the system runs without a human in the loop at the moment of the action. Someone put it there.

None of this means AI is useless for creative work. The research is consistent on one point: the combination outperforms either alone, under the right conditions.

A January 2026 study in PLOS ONE (Komura & Yamada) examined how AI strategy affects human-AI creative collaboration, finding that AI functions most effectively as a supportive partner that deepens human-initiated concepts rather than competing to generate its own. The study's operative finding: collaboration works when AI elaborates on what humans start, not when it leads. The human initiates. The machine develops.

A 2024 study in Nature Scientific Reports sharpened this: "For AI to successfully augment human creativity, it is a requirement that it promotes creative self-efficacy and places humans in the role of a co-creator, not an editor." The moment you position yourself as an editor of AI-generated ideas rather than the initiating intelligence, something starts to erode, not dramatically but gradually, in the way that any unused capacity quietly atrophies.

The risk is not that AI becomes creative. The risk is that humans stop needing to be. Not through displacement, but through the convenience of outsourcing the generative step, the one that requires emotional investment, uncertainty, and the willingness to follow an inarticulate pull into territory you cannot yet justify. That is precisely the step the machine cannot take on your behalf.

The vibe is not decorative. It is the signal that tells you when a thought matters, when a connection is real, when something is worth pursuing past the point where it makes logical sense to continue. AI processes without caring. It cannot know the difference between a sentence that changes everything and one that fills the space.

For now, that difference is still ours. The question is whether we are paying attention to it.