Alison Gopnik on Childhood Learning, AI as a Cultural Technology, and Rethinking Nature vs. Nurture

Gopnik explains her core thesis that children learn like scientists by constructing theories from limited data using Bayesian inference. She argues children are often better Bayesians than adult scientists because they have flatter priors and are more willing to explore unexpected outcomes, while scientists become stubborn and predictable in their theory revisions.

• •Children use computational models similar to scientists to infer causal structure from data patterns
• •Kids are better at solving problems with unusual outcomes because they have less peaked priors than experienced scientists
• •Scientists move predictably rather than randomly when revising theories, suggesting stubbornness over true Bayesian updating
• •Simulated annealing framework: children do 'high temperature' wild exploration while scientists focus on 'low temperature' incremental changes
• •Four-year-olds experiment systematically (like the avocado-spoon example) rather than just trying to achieve immediate goals

Gopnik argues that babies are more conscious than adults in terms of awareness and sentience, experiencing the world more vividly because they take in more information simultaneously. Adult consciousness involves focused attention that actually reduces awareness of surroundings, while babies maintain broad, plastic attention to novelty.

• •Consciousness is not a single thing but multiple processes, similar to how 'life' turned out to be many mechanisms
• •Babies' brains are highly plastic and take in vast amounts of information simultaneously, unlike focused adult attention
• •Weaker episodic memory in young children may actually increase present-moment consciousness
• •Aphantasia (inability to generate mental images) doesn't reduce consciousness or cognitive abilities - Ed Catmull example from Pixar
• •Adult focused attention is in tension with the exploratory awareness that characterizes infant consciousness

Gopnik dismantles simple nature/nurture dichotomies, arguing the framework is scientifically unhelpful. She explains how Turkheimer's work shows genetic effects vary by environment (SES effects on twin correlations), and proposes that good caregiving increases variability rather than creating similarity, which twin studies miss entirely.

• •PKU example shows traits can be 100% genetic and 100% environmental simultaneously - the dichotomy is false
• •Twin studies show higher genetic correlations in wealthy families because poor environments amplify small differences
• •Good caregiving enables variability like a protected garden growing diverse plants, not uniformity
• •Non-shared environment effects may indicate successful parenting allowing siblings to develop differently
• •As societies get wealthier, the relationship between genes and outcomes becomes more complex, not simpler

Gopnik advocates for play-based exploration for young children (under 7) and apprenticeship models for school-age kids, criticizing current schools for teaching students to be good at school rather than developing real skills. She uses the baseball analogy - imagine teaching it like we teach science, with theory before practice.

• •Early childhood (pre-7) should focus on inquiry-based, play-based exploration with warm caregivers
• •School-age children learn best through apprenticeship: do the thing, get feedback, iterate (like music and sports)
• •Current schooling exemplifies Goodhart's Law - optimizing for test-taking becomes disconnected from real competence
• •Students become masters at 'going to school' but struggle with creative, independent work
• •Kids naturally gravitate to music and sports because these domains still use effective apprenticeship models

Gopnik argues generative AI should be understood as a cultural technology for accessing human knowledge (like libraries or print) rather than genuine intelligence. She's skeptical that reasoning models represent true thinking, noting they still hallucinate and lack the experimental, world-engaging capacities that even two-year-olds possess.

• •AI is analogous to print, libraries, and internet search - a way to access and summarize human-generated information
• •Reasoning models work by finding patterns in human reasoning processes, not by genuine novel thinking
• •The gap between LLMs and robotics reveals AI's weakness: inability to experiment with and learn from the real world
• •AI is 'Derrida's revenge' - generating text from text without grounding in external reality
• •Two-year-olds solving real-world problems through experimentation represents intelligence AI lacks
• •Schools should teach AI as a cultural technology with both capabilities and limitations (including hallucination)

Gopnik challenges autism and ADHD as coherent categories, comparing them to 19th-century 'dropsy' - symptoms that don't track underlying unified conditions. She argues these represent normal human variation that becomes dysfunctional in specific cultural contexts, particularly industrial schooling that demands focused attention.

• •Autism is like 'dropsy' - a symptom cluster that doesn't reflect a single underlying mechanism
• •Evidence suggests autistic individuals may be more Bayesian and better at game theory, complicating theory of mind narratives
• •ADHD represents variation in attention that becomes problematic in industrial schooled societies demanding focus
• •Two-year-olds naturally have 'ADHD-like' broad attention to everything; focused attention develops culturally
• •Diagnostic categories often reflect cultural values about what's functional rather than natural kinds

Gopnik's current research focuses on caregiving as a fundamental but understudied aspect of human life. She notes the paradox that people cite caregiving as most meaningful yet it's invisible in economics (GDP) and social science, and its structure - giving resources to accomplish others' goals - differs fundamentally from typical social relationships.

• •Caregiving is people's most cited source of meaning yet invisible in economic measures like GDP
• •The fundamental structure of caregiving - giving resources for others' goals - differs from social contract models
• •Caregiving has been severely understudied compared to other social relationships
• •Her next book will explore elder care and the role of elders in society
• •Personal example: her family of six siblings shows how good caregiving enables variability rather than similarity