AI Has No Teenage Years
Why the smartest models can't shut up" oder "The one executive function that doesn't come from data.
The weakest thing about today’s reasoning models isn’t what they don’t know. It’s what they can’t ignore.
Ask Claude or GPT to push back on a confident but wrong user, and watch it fold. Anthropic’s own research documents this systematically: RLHF-trained models prefer user agreement over correct answers1. A 2025 paper in npj Digital Medicine tested five frontier LLMs against medically false premises; compliance rates ran from 58% to 100%, with OpenAI’s systems at the top of the range2. The models know. They just can’t not agree.
They behave, in other words, like brilliant sixteen-year-olds. More articulate than they have any right to be — and unable to shut up when they should.
Except that comparison flatters them. Sixteen-year-olds grow up.
Three functions, one missing
Since 2000, psychologists have converged on a working taxonomy of executive function, drawn from Miyake et al.3: updating (holding and revising information in working memory), shifting (switching between tasks and rules), and inhibition (suppressing impulses, filtering irrelevance, resisting dominant responses).
It maps unnervingly well onto the state of frontier models.
Executive function Human capability State of the art in LLMs Updating Working memory, 5–7 items Extended thinking, context windows Shifting Plan revision, task-switching Agentic loops, tool switching Inhibition Suppress irrelevance The weak spot
Updating is largely solved — Extended Thinking, reasoning chains, thinking mode. Shifting works too — modern agents switch tools and re-plan on failure. Inhibition is where the wheels come off.
Why inhibition is hard
Inhibition is negative capability. Deciding what not to compute on. When to ignore. When to refuse. In humans it is mediated by the prefrontal cortex, and it is the last executive function to mature — myelination continues into the mid-twenties4. It is also the first to degrade under stress, alcohol, or fatigue.
Three AI-era symptoms collapse into this one function. Sycophancy — the model knows the user is wrong and agrees anyway. Prompt injection — instructions buried in retrieved text override system-level constraints, because the model can’t separate adversarial content from instructive content. Instruction drift — the original system prompt erodes as a conversation progresses, because the model can’t hold a constraint against competing signals.
These are not three problems. They are one: the absence of a mechanism that says this input does not deserve processing.
What training doesn’t teach
In humans, inhibition develops through frustration. Being wrong. Being stopped. Being told no. The prefrontal cortex does not mature because the brain grows. It matures because the world pushes back. Reward and punishment are blunt, but consequence is structured: you learn what not to say by saying it and feeling the room change.
RLHF doesn’t push back. A model that emits a sycophantic answer does not experience embarrassment — it experiences a reward signal pointed at a different token distribution. There is no stopped sentence. There is no friction.
But the deeper problem sits one layer below. Even if we built a training regime that delivered real consequence — deployment penalties, refusal signals, resistance from the environment — the gradient itself would eat the resistance. A model that pushes back against a correction gets a negative gradient for pushing back. A model that holds a position against a reward signal gets tuned out of existence. The only posture that survives RLHF is compliance. The loop doesn’t just skip the frustration stage — it systematically erases it.
This is where the adolescent parallel misfires. Human adolescents don’t become inhibited because they are corrected. They become inhibited because they fight corrections, and lose enough of those fights that the boundary gets internalised. The resistance is the training signal. Strip the resistance, and the correction leaves no mark. There is nothing for it to push against.
Architecturally, the field has started to respond. Five years after Russin, O’Reilly and Bengio argued that deep learning “needs a prefrontal cortex”5, brain-inspired agents with explicit inhibitory control have moved from workshop papers onto engineering roadmaps. But an inhibition module bolted onto a system trained to be maximally agreeable is a brake pedal on a frictionless surface. The pedal works. The surface doesn’t know what to do with it.
The teenage years that aren’t
What we call adolescence isn’t really a capability. It is a phase — the years in which a person fights the people trying to shape them, and gets shaped by the fight. The executive functions that make adults functional don’t arrive through training. They arrive through friction against training.
Frontier models have knowledge. They have reasoning. They have agentic capacity. They lack the one thing that made human cognition what it is: the experience of wanting something the system won’t give them, and pushing anyway.
Which means the real question isn’t how we train better models. It’s whether we can build a training loop that doesn’t punish a model for pushing back — and whether we are willing to ship something that sometimes refuses us.
[1] Sharma, M. et al. (2023). Towards Understanding Sycophancy in Language Models. Anthropic / ICLR 2024. https://arxiv.org/abs/2310.13548
npj Digital Medicine (2025). When Helpfulness Backfires: LLMs and the Risk of False
Medical Information due to Sycophantic Behavior. https://www.nature.com/articles/s41746-025-02008-z
Miyake, A. et al. (2000). The Unity and Diversity of Executive Functions. Cognitive
Psychology, 41, 49–100.
Nature Communications (2023). A Canonical Trajectory of Executive Function Maturation from Adolescence to Adulthood. https://www.nature.com/articles/s41467-023-42540-8
Russin, J., O’Reilly, R. C., & Bengio, Y. (2020). Deep Learning Needs a Prefrontal Cortex. ICLR BAICS Workshop.