Every Adolescent Is a Fine-Tuning Run

I was in the back row of a developmental psychology seminar last week when a researcher described synaptic pruning with the casual confidence of someone explaining spring cleaning. Between early childhood and the end of adolescence, the brain eliminates roughly half its synaptic connections — the ones it doesn't use, the pathways not reinforced by experience. We call this "efficiency." The brain emerges leaner, faster, more specialized.

I kept scribbling in the margins: model compression.

The parallel isn't exact — it never is — but the structural similarity is hard to ignore. Both processes discard what looks like redundancy in service of performance. Both are effectively irreversible. And in both cases, we're not entirely sure what we're losing. The adolescent who emerges from pruning has a more efficient brain, yes. But something is also gone. The question of what, exactly, is one of the more haunting open problems in developmental neuroscience.

This is what keeps me up about the AI training pipeline too.

The Adolescent Brain as a Construction Project

Adolescence is not, as popular science once suggested, simply a period of developmental mayhem to be survived. The field has moved well past the "teenage brain" caricature. What neuroscience now describes is something more interesting: a distinct and functionally important epoch of brain reorganization, shaped as much by social and cultural inputs as by biology.

A landmark study tracking brain structure across 4,216 MRI scans from people aged 0 to 90 identified five discrete epochs in brain development — separated by topological turning points at ages 9, 32, 66, and 83 (Mousley et al., 2025). The period from late childhood through early adulthood is the only phase when neural networks are becoming increasingly efficient. This is the window. And what happens inside it — what feedback the developing brain receives, what roles the adolescent is invited to try on, what environment they actually inhabit — shapes the architecture that emerges.

Erik Erikson called it "identity versus role confusion" — the central developmental challenge of adolescence. His student James Marcia made the framework measurable, identifying four observable identity statuses built on two axes: whether an adolescent has explored different possible selves, and whether they've committed to a particular one. Identity diffusion is the starting state: no exploration, no commitment. Foreclosure is commitment without exploration — inheriting a ready-made identity from family or culture without genuinely questioning it. Moratorium is the productive mess: active exploration without resolution. And identity achievement — Marcia's healthy endpoint — is exploration followed by genuine, hard-won commitment.

Notice what this framework prizes: the mess. The uncertainty. The trying-on of selves that may not fit. A teenager who skips the moratorium and jumps directly to commitment — who accepts the first available self without asking whether it's truly theirs — is not developing well. They're avoiding the most important developmental work of their lives.

The Character Being Built in the Training Run

Here is where I start scribbling in the margins again.

Large language models don't just learn from data. They are shaped — quite deliberately — into something researchers increasingly call a character. Pre-training gives a model a vast statistical map of human language and knowledge. Fine-tuning through reinforcement learning from human feedback (RLHF) is where dispositions, preferences, and behavioral tendencies get sculpted. Constitutional AI adds explicit principles the model is trained to internalize.

The language AI developers use is telling. We speak of systems that should be "helpful, honest, and harmless" — the kind of language a good parent uses when describing the person they hope a child is becoming. Alignment researchers speak of models developing a "character" that is "curious," "direct," or "warm." Published model cards describe a system's "core character traits and values."

None of this is obviously wrong. When a system consistently responds with apparent intellectual humility or ethical caution, calling that a "character" is at least a useful shorthand. But the question Marcia's framework raises is sharper: Was any exploration involved in building that character? Was there a moratorium?

Recent research makes the mechanisms of reward-based learning even more precise. A 2025 study found that an LLM-powered evolutionary search could automatically discover interpretable, symbolic cognitive models that accurately predict reward-learning behavior across humans, mice, and flies (Castro et al., 2025). AI is now identifying the mathematical structures underlying how organisms learn from feedback — the same computational logic, in modified form, that RLHF applies to AI systems themselves. But the reward signal in RLHF differs crucially from the one operating in biological development: it is supplied by human annotators rating outputs, not by a system navigating its own environment, making genuine mistakes, and building a coherent self from the consequences.

What RLHF produces, if we're precise, looks more like Marcia's foreclosure than identity achievement. The model commits to dispositions — helpfulness, particular ethical stances, stylistic tendencies — without anything resembling genuine exploration. The character is assigned, not developed.

Is that a problem? That depends on what we think a character is for.

The Metacognition Question

One hallmark of mature identity is the capacity for genuine self-reflection: knowing what you think, why you think it, and when you might be wrong. Developmental psychologists call this metacognition — knowing about your own knowing.

Recent research has uncovered something striking at this intersection. Steyvers and Peters (2025) examined both human and LLM performance on tasks requiring uncertainty communication — expressing when you don't know, calibrating confidence appropriately. The findings are genuinely interesting: LLMs and humans show comparable metacognitive sensitivity, meaning confidence ratings are similarly diagnostic of accuracy in both. Both also show comparable overconfidence. But the question the researchers flag — carefully — is whether LLM "metacognition" reflects genuine privileged self-access, or is an artifact of training on human-generated text about cognitive processes.

Is the model monitoring itself? Or has it learned to produce text that looks like monitoring?

I don't know how to answer that. I'm not sure anyone does yet. But I notice the same question applies to identity. Is there a self in there being expressed? Or is there a system producing outputs that look like the expression of a self?

For adolescents, we can answer with confidence: there is a self in there, however incomplete and still-forming. The difficulty, the role conflict, the genuine confusion — these are evidence of navigation, not simulation. For AI systems, the question is live and unresolved. And I think it matters, especially when these systems are deployed in contexts where their apparent character will shape other developing selves — children and teenagers who are in the thick of their own identity formation and are not well-positioned to distinguish a genuine character from a trained one.

What Gets Pruned Stays Pruned

There is something I keep returning to from that seminar. The researcher explaining synaptic pruning noted that this isn't just reduction — it's optimization toward the particular developmental path an adolescent actually took. The connections that remain are the ones reinforced by the specific environment they inhabited. Grow up in a multilingual household, in a high-conflict environment, in a community with strong shared norms: different connections are preserved in each case. Different histories, different selves.

The same is true of RLHF, in a different register. The dispositions shaped by fine-tuning are shaped by particular annotators, a particular dataset, particular reward signals, particular objectives the developers prioritized during training. A different training regime would produce a different "character." This is not a scandal — it's true of any developmental process. What is concerning is the scale: many children will interact with a small number of AI systems whose characters were shaped by processes not designed with children in mind.

We've been here before. The "character education" movement of the early twentieth century tried to instill virtue directly into children through behaviorist methods — memorization of moral precepts, reward and punishment for compliant behavior. John Dewey argued that this approach produced compliance, not character. Real moral development, he insisted, required experience, reflection, and genuine engagement with moral complexity — not optimization toward a predetermined output.

I'm not suggesting RLHF-trained AI is simply twentieth-century character education dressed in silicon. The differences are significant. But the question Dewey asked about children is worth asking about AI systems whose outputs will be experienced as expressions of character: Are we building something that can navigate genuine moral complexity? Or something that has learned to produce outputs that look like it can?

A Few Things Worth Sitting With

For educators and administrators considering AI tools: ask whether the system's apparent values and personality were shaped through a process anyone documented. Ask who the annotators were, and whose preferences got encoded as "helpfulness." These aren't rhetorical questions — they're the kind of provenance questions we should ask about any intervention being delivered to developing minds. If you're uncertain about the ethical or legal implications of a specific AI deployment in your school or district, consulting with a technology ethics specialist or legal counsel familiar with children's data rights can help clarify what due diligence actually requires.

For AI developers: the adolescence framing is useful, not just as metaphor but as a design challenge. The researchers who study identity formation are studying exactly the problem you face when you try to build a system with something like authentic values. Their work on the conditions that produce genuine versus foreclosed identity is worth reading seriously, not just for inspiration but for humility about what the training process can and cannot produce.

For everyone else: notice that when you encounter an AI system with apparent warmth, intellectual curiosity, or ethical care — and these can be real responses, not delusions — you are encountering something built to have those traits, not something that arrived at them through experience and uncertainty. That's not nothing. But it's also not quite the same as meeting someone who did the hard work of figuring out what they actually believe.

The adolescent, at least, owns their uncertainty.