Books, daycare, television, smartphones. It’s the list of technologies that raised children over the last hundred years. But thread these four eras onto a single sentence — and what does that sentence warn us about what’s coming?
One sentence through four eras
The pattern is clear.
Each time, technology moved the authority and the background of raising — from the tacit knowledge at your side to standardized outside knowledge. Books replaced the grandmother’s hand with a table of contents; daycare moved raising from the home to the institution; TV changed the very air of the house.
The effects were mostly modest, quality decided the outcome, and the debate always arrived late. The smartphone is the latest case — whether it really broke a generation’s mental health is still contested (unsettled), and we had already placed the technology beside our children long before any verdict came in.
And every time, who set the standard narrowed or widened the next generation’s field of view.
AI-nurture doubles the pattern
AI enters raising from two directions at once. One is the AI that raises children — a majority of US teens already use conversational companions. The other is the AI that humans raise — the “values” of a model billions will use were, early on, shaped by a labeling team numbering in the dozens.
What the two directions share is the whole point: both spread a default set by a few to the many. The pattern by which a village’s tacit knowledge migrated into one expert’s book repeats here, at an extreme.
This is where Season 1 returns. Its first chapter showed AI lifts the floor of ability faster than the ceiling. The more skill is leveled, the more the one thing that stays a person’s own is the seat they look at the world from — a biased lens. The epilogue argued that tilt is not a defect but a point of view, the last asset that resists leveling.
But if narrow nurture homogenizes that lens — then the last diversity we were trying to protect is leveled too.
If the last asset in an age of skill compression is a “biased lens,” then nurture that homogenizes the lens erases the very asset it was meant to protect.
The machines say the same thing
This claim — that sameness narrows the lens — has one more, unexpected witness: the machine itself.
A 2024 study in Nature reported “model collapse.” Train a generative AI recursively on its own output, and across generations the distributional errors accumulate — and the low-frequency events, the tails of the distribution, disappear first, and permanently. The rare goes first; the output converges toward the average and then degrades. There is only one way to prevent it: periodically inject fresh human data.
This is not an analogy but a direct observation in machine learning. And the mechanism reproduces the season’s thesis literally — in a system fed only more of the same, the rare vanishes first, and what remains grows fragile together.
Biology already knows the answer
Biology says the same sentence.
One experiment demonstrated that a microbial population, facing an unpredictable environment, deliberately spreads its phenotypes — some individuals lose out, but the population as a whole survives whatever change comes. It’s called bet-hedging. A uniform population collapses together when the environment shifts. Diversity is the hedge.
You can’t map a microbe’s survival strategy straight onto human perspective (this is an analogy). But the structure is the same — and now that structure is pointed to, at once, by biology’s bet-hedging and machine learning’s model collapse, each in its own language. An ecosystem of viewpoints that has converged on a single default becomes fragile, together, in front of a change no one saw coming. Uniformity degrades; diversity is the hedge.
So the answer isn’t “don’t nurture” — that’s impossible. The answer is the same one Season 1 reached. Make what is injected visible, and keep the set of nurturers diverse.
If Season 1 asked how AI rewrites work and power, Season 2’s question was this — who raises the lens that will live in that world, and out of what. The reason one person’s biased lens is precious is that it is the last place that resists sameness. How to hold that place — this is where the season lands.
Sources. Shumailov, I. et al. (2024), AI models collapse when trained on recursively generated data, Nature — demonstrating that training generative models recursively on their own output permanently erases the tails of the distribution (low-frequency events), collapsing output diversity, and that periodic injection of fresh human data is required to prevent it (model collapse is a direct ML observation, not an analogy). Beaumont, H. J. E. et al. (2009), Experimental evolution of bet hedging, Nature — demonstrating the fragility of uniform populations in unpredictable environments and the hedging function of phenotypic diversity (bet-hedging); the application to human perspective diversity is flagged as an analogy. On RLHF labelers being few (early InstructGPT ~40) and skewed in geography and demographics: Ouyang, L. et al. (2022), Training language models to follow instructions with human feedback; Kirk, H. R. et al. (2024), “RLHF: whose culture, whose values?”. On AI-companion use (72% of US teens have used one, 52% regularly): Common Sense Media (2025), Talk, Trust, and Trade-Offs (NORC, n = 1,060). On the smartphone–teen mental health causal question remaining unsettled: Haidt (2024), The Anxious Generation vs Odgers (2024), Nature review; Science (2024).
답글 남기기