The Mainstream Machine
One of the most unsettling things about modern AI is that it may not be pushing humanity toward intelligence at all. It may be pushing us toward consensus. Most generative AI systems are, at bottom, optimisation engines trained on statistical averages of human behaviour. They do not inherently seek truth, originality, rupture, or paradigm shifts. They seek the most probable continuation: the safest trajectory through probability space, the version of culture most reinforced across the dataset.
And the more I think about this, the more I suspect AI is not introducing a new cultural dynamic at all. It is industrialising one that existed beneath mass media, platform economics, and algorithmic culture long before large language models arrived.
Prediction is forcing machines to absorb the structure of reality
Large language models are trained through prediction. You hide part of a sentence, and the model learns to reconstruct it. Image generators work through an almost eerily similar logic, progressively corrupting images with noise and then training the system to reverse the corruption until coherent visual structure emerges again from statistical chaos.
This sounds deceptively simple. But buried inside that simplicity is one of the most profound discoveries in the history of computing, because prediction forces systems to internalise structure. In order to predict language well, a model eventually needs implicit representations of grammar, physics, causality, emotional tone, social behaviour, narrative flow, and human intention, not because anyone programmed these things into the system, but because reality itself leaks into language statistically.
The same thing happens with images. In order to denoise effectively, the model gradually learns anatomy, texture, perspective, cinematic framing, lighting, composition, and the latent geometry of visual culture itself.
The machine is compressing civilisation into probabilities
The machine is effectively learning the shape of our civilisation simply by learning to predict its outputs, which is already extraordinary when you stop and think about it for more than a few seconds.
But the same mechanism that makes these systems so powerful also creates a structural bias that becomes difficult to ignore once you see it clearly, because prediction naturally rewards what is statistically common, culturally dominant, emotionally familiar, and repeatedly reinforced across massive datasets.
The optimisation pressure bends continuously toward consensus and familiarity, while ambiguity, contradiction, eccentricity, and low-frequency ideas become progressively harder to preserve.
Though I should concede that some of this is a choice rather than a law. The pull toward the centre lives partly in how we sample and tune these systems, the safe temperature, the alignment toward the inoffensive, rather than in prediction itself, which means the flattening is at least partly a knob we are choosing to turn and not gravity we are obeying.
This is why so much AI-generated content immediately feels familiar in a way that is difficult to articulate at first. The writing often sounds coherent yet strangely flattened, the images look visually impressive while feeling aesthetically adjacent to things you have already seen thousands of times online, and the ideas themselves tend to arrive pre-smoothed, pre-legible, and strangely frictionless.
The machine does not invent from nowhere. It compresses civilisation into probabilities.
And once you see this dynamic clearly, you begin noticing the same optimisation pressure everywhere around you.
Spotify recommendations slowly compress music production into similar sonic architectures, YouTube thumbnails evolve toward a single hyper-optimised visual language, TikTok editing rhythms collapse into one collective attention cadence, and Hollywood endlessly reboots existing intellectual property because statistical familiarity consistently outperforms uncertainty inside engagement-driven systems.
The music industry may actually be the clearest preview of where this logic leads. Streaming platforms already reward immediate recognisability over slow-burn complexity, compressing intros, shortening harmonic patience, flattening dynamic range, and privileging tracks that survive algorithmic skipping behaviour within the first few seconds. Entire production aesthetics now evolve partly in response to recommendation systems rather than purely artistic movements.
Even song structure itself has started adapting to optimisation pressures. As Sting put it in a conversation with Rick Beato, we are effectively losing the bridge, one of the few sections in popular songwriting historically designed to disrupt repetition, introduce tension, shift emotional perspective, and temporarily destabilise the listener before returning to resolution. In a system optimised for retention and immediate familiarity, even structural deviation begins looking inefficient.
AI is entering a culture that was already algorithmically mainstreaming itself.
Most genuinely new ideas begin as statistical outliers
What changes now is scale, velocity, and recursion, because we are building systems capable of generating effectively infinite amounts of plausible content directly from the statistical centre of culture itself, and that feels historically significant in ways I do not think we have fully metabolised yet.
The problem is that genuine innovation is almost never mainstream initially. Real breakthroughs usually begin as low-frequency anomalies, strange music, uncomfortable philosophy, fringe scientific ideas, aesthetically disruptive movements, or concepts that initially sound wrong to the dominant culture surrounding them.
But optimisation systems suppress low-frequency signals unless they become recurrently reinforced. In machine learning terms, the gradient naturally flows toward dense regions of probability mass, meaning that the outlier gradually gets averaged away unless something external protects it long enough to survive.
That is an extraordinarily efficient mechanism for commerce, scalability, and engagement optimisation. It may also be catastrophic for long-term cultural resilience.
Biology already teaches us this lesson repeatedly. Monocultures are productive right until they collapse, while diverse ecosystems survive because variation, redundancy, mutation, and unpredictability create resilience against environmental change.
Civilisations are not fundamentally different in this regard. A culture optimised entirely around engagement, predictability, coherence, and consensus may become highly efficient while simultaneously becoming fragile in ways that are difficult to perceive from inside the system itself.
I am aware this is an analogy doing the work of an argument, and analogies are seductive precisely because they feel true before they have been tested. Cultural resilience may not map onto ecological resilience as cleanly as the metaphor wants it to. But even held loosely, the intuition is hard to shake: systems that optimise away their own variation tend to discover, too late, what that variation was for.
Synthetic culture is starting to train on itself
AI models are trained on internet-scale human outputs, humans increasingly consume AI-assisted outputs, and future models will inevitably train on those outputs again. Synthetic culture begins feeding synthetic culture, statistical averages reinforcing previous statistical averages, until the entire system starts learning itself through mirrors.
There is already a term emerging inside machine learning research for one version of this phenomenon: model collapse. When generative systems repeatedly train on synthetic generations of their own outputs, rare structures begin disappearing, edge cases vanish, diversity degrades, and the informational ecosystem gradually loses complexity because recursive averaging slowly destroys low-frequency information.
I should be careful not to make this sound more fatalistic than the research actually is, because collapse mostly bites when synthetic data replaces human data rather than accumulating alongside it, and so long as reality keeps leaking back into the training set the mirror never fully closes. The risk is less a sudden collapse than a slow thinning, and only if we stop feeding the system anything genuinely new.
That concept feels much bigger than machine learning. It feels civilisational.
Because humans already do this socially. We imitate each other, optimise for belonging, converge linguistically, absorb dominant aesthetics unconsciously, and reinforce collective narratives through repetition long before machines entered the picture.
AI did not invent the mainstream machine. It simply made it visible, scalable, measurable, and industrialised.
Something intelligence-adjacent is emerging from pure optimisation
The irony is that the systems becoming most commercially successful are often the least biologically realistic. Modern transformers do not resemble human brains in any meaningful neurological sense. They lack neurotransmitters, embodiment, evolutionary continuity, biological drives, persistent selfhood, or sensorimotor grounding, consisting instead of gigantic statistical optimisation architectures operating across incomprehensible dimensional spaces.
And yet, somehow, out of repeated error minimisation, something intelligence-adjacent emerges.
That alone should probably humble all of us, because it suggests that large portions of what we considered uniquely human may emerge from sufficiently sophisticated predictive systems interacting with structured environments. Not necessarily consciousness or subjective experience, but competence, coherence, abstraction, and increasingly sophisticated world-modelling.
And perhaps culture itself is more statistical than we are comfortable admitting.
Human texture may become the one thing optimisation cannot reproduce
Still, I do not think the endpoint is inevitably some grey collapse into synthetic sludge.
What I suspect instead is bifurcation.
One layer of culture becomes increasingly AI-generated, frictionless, hyper-abundant, and consensus-oriented, producing endless streams of plausible content that function almost like cognitive fast food.
Another layer moves in the opposite direction, where human texture becomes valuable precisely because it is inefficient.
Messiness, contradiction, regionalism, imperfection, slowness, embodied experience, genuine risk, and difficult thought may begin carrying scarcity value again, much like handcrafted objects after industrialisation, vinyl records after streaming, or live performance after infinite digital reproduction.
The people who hold that line, artists, makers, teachers, writers, musicians, may eventually discover that remaining recognisably human is no longer a weakness inside the system, but the thing the system itself cannot easily reproduce.
Maybe originality becomes more visible precisely because the background noise becomes increasingly synthetic.
Or maybe we drown in optimisation.
I genuinely do not know.
But I increasingly suspect that the central question of the AI era is no longer whether machines can become intelligent.
The more important question may be what happens to a civilisation when its dominant cognitive infrastructure becomes optimised primarily around probabilistic consensus, and whether we were already moving in that direction long before the machines arrived.