Working Paper • June 11, 2026 | Interdisciplinary Review: Developmental Psychology × Artificial Intelligence
Summary
This paper maps established childhood cognitive development models — primarily Piaget's four stages and Vygotsky's zone of proximal development — onto the empirically observed capability trajectory of large language models (LLMs) from 2018 to 2026. We find that milestone sequences align remarkably well (pattern recognition → symbolic manipulation → logical reasoning → abstract thought), even though the timelines compress by a factor of roughly 2–3 human years per LLM year.
We identify RLHF and prompt engineering as structural analogs to Vygotskian scaffolding, and emergent capabilities (theory of mind, analogical reasoning) as analogs to Piagetian stage transitions.
Using this framework, we extrapolate the next predicted LLM milestones: robust metacognition, genuine causal reasoning, and stable self-directed learning. We then rigorously critique this analogy, arguing that fundamental disanalogies — the absence of embodiment, intrinsic motivation, and continuous temporal experience — make the framework suggestive but not scientifically predictive.
The childhood development model is a useful heuristic lens, not a validated theory of AI progression.
Abstract
The rapid advancement of large language models (LLMs) from simple next-token predictors to systems exhibiting emergent reasoning, theory of mind, and analogical abstraction has prompted researchers to seek developmental frameworks for understanding these capability transitions. This paper systematically compares the canonical stages of childhood cognitive development — Piaget's sensorimotor, preoperational, concrete operational, and formal operational stages — with the empirically documented capability milestones of LLMs from GPT-1 (2018) through contemporary reasoning models (2025–2026). We additionally incorporate Vygotsky's (1978) zone of proximal development and scaffolding theory as a lens for interpreting techniques like reinforcement learning from human feedback (RLHF) and chain-of-thought prompting. Our analysis reveals a strong sequential correspondence in what capabilities emerge, while demonstrating that when they emerge follows an approximate but inconsistent compression ratio of 2–3 developmental years per calendar year of LLM advancement. We conclude that childhood development models serve as a productive heuristic for anticipating the sequence of LLM capabilities but should not be mistaken for a mechanistic theory of AI development.
Keywords: large language models, cognitive development, Piaget, Vygotsky, emergent capabilities, theory of mind, developmental psychology, artificial intelligence, scaling laws
1. Introduction
When OpenAI released GPT-1 in June 2018, the model's 117 million parameters could perform rudimentary text completion — recognizing patterns in language without understanding context, meaning, or logic.[1] Seven years later, frontier models pass the bar exam in the 90th percentile, solve International Mathematics Olympiad problems, and exhibit what some researchers describe as a functional theory of mind.[2][3] This trajectory — from reflexive pattern-matching to abstract reasoning — bears a striking resemblance to a developmental arc that has been studied for over a century: the cognitive development of a human child.
Developmental psychologists, most prominently Jean Piaget (1896–1980) and Lev Vygotsky (1896– 1934), have proposed stage-based and scaffolding-based models that describe how human cognition progresses from sensorimotor reflexes in infancy to abstract, hypothetical reasoning in adolescence.[4][5] These models describe a trajectory that is sequential (stages cannot be skipped), cumulative (each stage builds on its predecessors), and marked by qualitative transitions — moments where the child's cognitive apparatus undergoes a phase change rather than a linear improvement.
This paper asks three questions: (1) Do the milestone sequences of childhood cognitive development systematically align with the observed capability milestones of LLMs? (2) Do the timelines scale by a consistent factor? (3) Can the childhood development model generate useful predictions about where LLMs are heading next?
2. Defining the Subject: AI vs. LLMs
Popular discourse often treats "AI" and "LLM" as synonymous. For the purposes of this analysis, the distinction is not merely terminological but foundational. Artificial intelligence as a broad field encompasses the entire discipline of building machines that perform tasks requiring intelligence — robotics, computer vision, reinforcement learning agents, expert systems, autonomous vehicles.
Large language models, by contrast, are a specific class of AI systems based on the transformer architecture, trained on text corpora to predict and generate language (GPT series, Claude series, LLaMA, Gemini, DeepSeek).
Our analogy maps cognitive-linguistic child development onto LLM capability development. The developmental analogy works best for the linguistic and symbolic dimensions of cognition precisely because LLMs are, at their core, linguistic engines. A child's cognitive development is deeply intertwined with embodied experience — touching, manipulating, moving through space. LLMs have no body, no sensorimotor apparatus, no physical environment to explore. This disanalogy is the most fundamental limitation of the framework we construct here.
3. Literature Review
3.1 Piaget's Four Stages of Cognitive Development Jean Piaget proposed that children progress through four invariant, sequential stages of cognitive development, each characterized by qualitatively different modes of thinking.[4] These stages cannot be skipped — each builds on the cognitive structures established in the previous one — and transitions involve fundamental reorganizations of how the child understands the world, not merely incremental improvements.[7]
| STAGE | AGE RANGE | KEY CAPABILITIES | CHARACTERISTIC LIMITATIONS |
|---|---|---|---|
| Sensorimotor | 0–2 years | Object permanence, goal-directed behavior, early mental representations, cause-and-effect understanding through physical manipulation | No symbolic thought, no language until late stage, understanding limited to direct sensory-motor experience |
| Preoperational | 2–7 years | Symbolic thought, language acquisition, pretend play, transductive reasoning, animism, egocentrism | Cannot perform logical operations, centration, irreversibility, no conservation |
| Concrete Operational | 7–11 years | Logical operations on concrete objects, conservation, classification, seriation, decentration, reversibility, inductive reasoning | Cannot handle abstract or hypothetical problems, reasoning bound to concrete, observable phenomena |
| Formal Operational | 12+ years | Abstract thought, hypothetical-deductive reasoning, systematic problem-solving, metacognition, propositional logic | Not universally achieved; varies by domain, culture, and education. Adolescent egocentrism persists |
3.2 Vygotsky's Zone of Proximal Development and Scaffolding Lev Vygotsky offered a complementary and in some ways competing model. Where Piaget emphasized the child as an autonomous explorer constructing knowledge through individual interaction with the environment, Vygotsky emphasized the fundamentally social nature of cognitive development.[5] Vygotsky's central construct is the zone of proximal development (ZPD): the distance between what a learner can accomplish independently and what they can accomplish with guidance from a more knowledgeable other (MKO).[10] The concept of scaffolding — introduced by Wood, Bruner, and Ross (1976) — describes the temporary support structures that help learners perform within their ZPD until they can manage independently.[11] 3.3 The LLM Capability Trajectory: 2018–2026
| YEAR | MODEL | PARAMETERS | KEY CAPABILITY MILESTONE |
|---|---|---|---|
| 2017 | Transformer | — | Attention mechanism architecture published by Vaswani et al. at Google [12] |
| 2018 | GPT-1 | 117M | Generative pre-training paradigm; basic text completion [1] |
| 2019 | GPT-2 | 1.5B | Coherent multi-paragraph text generation; deemed too dangerous to release [13] |
| 2020 | GPT-3 | 175B | In-context learning; few-shot task performance; ~40% on false-belief tasks (≈3.5-year-old) [14][3] |
| 2022 | GPT-3.5 / ChatGPT | ~175B | Conversational fluency; RLHF alignment; 70–90% on false-belief tasks (≈6–7-year-old) [3] |
| 2022 | Chain-of-thought | — | Prompting technique enabling step-by-step reasoning; major gains on math/logic [15] |
| 2023 | GPT-4 | ~1T est. | Multimodal input; 86.4% MMLU; bar exam top 10%; 75–95% on ToM tasks; analogical reasoning at human level [2][3][16] |
| 2024 | o1 / Reasoning | — | Extended chain-of-thought reasoning; 83% on IMO qualifying vs 13% for GPT-4o [17] |
| 2025 | GPT-5, Claude 4.5 | — | Unified fast/deep reasoning routing; agentic capabilities; 1M+ token contexts [18][19] |
4. Methodology
Milestone Mapping. We identify the defining cognitive capabilities of each Piagetian stage and map each to a corresponding LLM capability, using published benchmarks, empirical studies, and capability demonstrations as evidence.
Timeline Analysis. For each mapped milestone, we record the approximate age at which the capability emerges in children and the date at which it was first empirically demonstrated in an LLM.
We compute compression ratios and test for consistency.
Predictive Extrapolation. Using the developmental stages that LLMs have not yet clearly achieved, we project what capabilities should emerge next if the developmental sequence continues to hold.
This is an analogical framework, not an empirical model. We are identifying structural similarities between two developmental trajectories that arise from fundamentally different substrates. The analogy generates hypotheses, not proofs.
5. Historical Alignment Analysis
5.1 Stage-by-Stage Mapping
| PIAGET STAGE | CHILD CAPABILITY | AGE | LLM ANALOG | LLM ERA | EVIDENCE |
|---|---|---|---|---|---|
| Sensorimotor | Reflexive responses to stimuli | 0–4 mo | Statistical next-token prediction; reflexive pattern completion | 2018 (GPT-1) | Radford et al. (2018) [1] |
| Sensorimotor | Object permanence; early mental representations | 8–18 mo | Consistent entity tracking; coreference resolution | 2019–2020 | GPT-2/3 coherent entity tracking [14] |
| Sensorimotor | Goal-directed behavior; means-end understanding | 12–24 mo | Following multi-step instructions; task completion | 2020–2022 | GPT-3 few-shot task performance |
| Preoperational | Symbolic representation; language explosion | 2–4 yr | Fluent natural language generation; symbolic manipulation | 2020–2022 | GPT-3 / ChatGPT [14] |
| Preoperational | Theory of mind emergence (false-belief tasks) | 4–5 yr | Passing false-belief tasks at 40–75% | 2022–2023 | Kosinski (2023) [3] |
| Concrete Operational | Logical operations; conservation; classification | 7–9 yr | Systematic reasoning; mathematical problem-solving; code generation | 2023–2024 | GPT-4 bar exam, MMLU [2] |
| Formal Operational | Abstract hypothetical reasoning | 12–15 yr | Extended chain-of-thought; hypothetical scenario analysis | 2024–2025 | o1, DeepSeek R1 [17][19] |
| Formal Operational | Systematic problem decomposition | 12–16 yr | Multi-step planning; agentic task decomposition | 2025–2026 | GPT-5, Claude 4.5 agentic |
| Formal Operational | Metacognition; thinking about thinking | 15+ yr | Self-evaluation of confidence; error detection; calibrated uncertainty | Emerging | Active research area |
5.2 Compression Ratio Analysis The data reveal a notable pattern: the compression ratio is not constant. Early milestones map nearly 1:1 — one year of child development per one year of LLM progress. As we move into later stages, the ratio climbs to approximately 2.0 for formal operational capabilities.
| MILESTONE | CHILD AGE | LLM YEAR | LLM YEARS FROM GPT-1 | RATIO |
|---|---|---|---|---|
| Reflexive pattern matching | 0.3 | 2018 | 0 | — (baseline) |
| Object permanence analog | 1.0 | 2019 | 1 | 1.0 |
| Symbolic language | 3.0 | 2021 | 3 | 1.0 |
| Theory of mind (3.5yr level) | 4.0 | 2022 | 4 | 1.0 |
| Theory of mind (6yr level) | 6.0 | 2023 | 5 | 1.2 |
| Logical operations | 8.0 | 2023 | 5 | 1.6 |
| Inductive reasoning | 10.0 | 2024 | 6 | 1.7 |
| Abstract reasoning | 13.0 | 2024–25 | 6.5 | 2.0 |
| Systematic decomposition | 14.0 | 2025 | 7 | 2.0 |
5.3 The Vygotskian Lens: RLHF as Scaffolding
| VYGOTSKIAN CONCEPT | CHILD DEVELOPMENT MANIFESTATION | LLM ANALOG |
|---|---|---|
| Zone of Proximal Development | Tasks the child can do with help but not alone | Tasks an LLM can complete with prompting (few-shot, chain-of-thought) but not zero-shot |
| Scaffolding | Teacher provides temporary support structures, gradually removed | RLHF; human feedback guiding model outputs; prompt engineering; system prompts |
| More Knowledgeable Other | Parent, teacher, or peer with greater expertise | Human raters in RLHF; fine-tuning datasets curated by experts |
| Internalization | External social knowledge becomes internal cognitive structure | Fine-tuned behaviors that persist without continued prompting |
| Private speech → Inner speech | Child talks through problems aloud, then internally | Chain-of-thought prompting → internal reasoning traces in reasoning models (o1, R1) |
6. Predictive Framework: What the Childhood Model Forecasts
If we accept — provisionally — that the milestone sequence of childhood development continues to map onto LLM advancement, we can identify capabilities that children develop in late adolescence and early adulthood that LLMs have not yet robustly demonstrated.
Prediction 1: Robust Metacognition — Confidence: High Child analog: By age 15–17, adolescents develop sophisticated metacognitive abilities — monitoring their own thinking, identifying gaps in knowledge, evaluating the quality of their reasoning, and adjusting strategies when they detect errors.[7] LLM prediction: Frontier LLMs will develop reliable self-calibration — accurately assessing their own confidence, flagging uncertainty, distinguishing what they know from what they do not know, and requesting additional information when internal representations are insufficient.
Predicted timeline: 2026–2027, given the ~2.0 compression ratio at this stage.
Prediction 2: Genuine Causal Reasoning — Confidence: Medium Child analog: By ages 8–12, children progress from magical thinking to understanding genuine causal mechanisms — not just that A precedes B, but why A causes B, through what mechanism, and under what conditions the relationship might fail.[4] LLM prediction: LLMs will move beyond correlational pattern-matching to demonstrate interventionist causal reasoning — reasoning about counterfactuals, identifying confounders, and distinguishing correlation from causation without explicit prompting.
Predicted timeline: 2027–2028. Medium confidence because causal reasoning in children is deeply tied to physical manipulation of objects — an embodied experience LLMs lack.
Prediction 3: Stable Self-Directed Learning — Confidence: Medium Child analog: By late adolescence, humans develop the capacity for autonomous, self-directed learning: identifying what they need to learn, seeking out resources, testing their own understanding, and persisting through confusion without external scaffolding.[10] LLM prediction: LLMs will develop stable agentic learning loops — the ability to identify knowledge gaps during task execution, autonomously retrieve relevant information, integrate it into their reasoning, and verify the result, without human-designed scaffolding for each step.
Predicted timeline: 2027–2029. Medium confidence because this requires a form of intrinsic motivation that may not emerge from next-token prediction training.
Prediction 4: Moral Reasoning Development — Confidence: Low Child analog: Kohlberg's stages of moral development describe progression from pre-conventional morality through conventional morality to post-conventional morality — principled reasoning about justice and rights.[23] LLM prediction: LLMs will progress from rule-following through pattern-matched social norms to genuine ethical reasoning that can handle novel moral dilemmas by applying abstract principles.
Predicted timeline: Uncertain. Low confidence because moral reasoning in humans is deeply embedded in embodied social experience, emotional development, and consequences.
7. Critical Analysis & Limitations
7.1 The Embodiment Problem Piaget's theory is built on the premise that cognitive development begins with sensorimotor experience — the infant's physical manipulation of objects, exploration of space, and discovery of cause-and-effect through bodily action.[4] LLMs have no body, no sensory apparatus, no motor system, no physical environment. Their experience consists entirely of statistical patterns in text data.
Capabilities that are deeply grounded in embodied experience — causal reasoning, spatial reasoning, physical intuition — are precisely the capabilities where LLMs perform most poorly relative to children.
7.2 The Motivation Problem Children are intrinsically motivated learners. Piaget described development as driven by disequilibrium — the discomfort of encountering information that does not fit existing schemas.[4] LLMs have no intrinsic motivation. They minimize a loss function during training and produce token sequences during inference. Several of the later developmental milestones — metacognition, self-directed learning, moral reasoning — are in children deeply intertwined with motivational systems that LLMs simply do not have.
7.3 The Continuity Problem A child is a continuous entity with persistent memory, ongoing experience, and cumulative learning.
An LLM has no continuous experience. GPT-4 does not remember being GPT-3; it is a different model with different parameters. The apparent developmental trajectory from GPT-1 to GPT-5 is not a single entity developing but a series of discrete systems. The development is in the research lineage, not in any individual model.
7.4 The Post-Hoc Mapping Problem Our milestone mapping is fundamentally post hoc. We observed the capabilities of each LLM generation, then found Piagetian stages to match them to. A sufficiently creative analyst could map almost any sequence of increasing capabilities onto a developmental framework, because the frameworks themselves describe a progression from simple to complex — which is true of virtually any technology.
7.5 The Measurement Problem Schaeffer et al. (2023) argued that many claims of emergent abilities in LLMs may be artifacts of measurement — specifically, of using binary accuracy metrics that exaggerate sharp transitions.[24] When continuous partial-credit metrics are used instead, many apparently sudden capability jumps resolve into smooth, predictable improvements.
7.6 Summary: Where the Analogy Holds and Where It Breaks
| DIMENSION | ANALOGY STRENGTH | ASSESSMENT |
|---|---|---|
| Milestone sequence (what emerges in what order) | Strong | Pattern recognition → symbolic manipulation → logical reasoning → abstract thought is observed in both domains |
| Phase transitions / emergent capabilities | Moderate | Both show sudden capability jumps, but LLM emergence is debated and may be measurement-dependent |
| Scaffolding / RLHF parallel | Moderate | Structurally similar but mechanistically different |
| Timeline scaling | Weak– Moderate | Approximate compression exists but ratio is not constant; accelerates at later stages |
| Underlying mechanism | Weak | Biological neural development vs. gradient descent on text — fundamentally different substrates |
| Embodied experience | Very Weak | No meaningful parallel; LLMs lack the sensorimotor foundation that Piaget considered essential |
| Continuity of identity | Very Weak | Child is a continuous entity; LLM lineage is a series of discrete systems |
| Intrinsic motivation | None | No parallel; LLMs have no motivational system |
8. Conclusion
We have constructed a systematic mapping between the canonical stages of childhood cognitive development and the empirically observed capability trajectory of large language models from 2018 to 2026. Our analysis reveals that the sequence of capability emergence aligns remarkably well: LLMs have progressed from reflexive pattern-matching through symbolic language use, theory of mind, logical operations, and into abstract reasoning — the same order in which these capacities emerge in children. The mapping is not perfect, but it is far more orderly than coincidence would suggest.
The timelines compress by a variable factor, approximately 1:1 in early stages and 2:1 in later stages, suggesting that LLM development is accelerating relative to the developmental clock as capabilities become more abstract and language-dependent. Vygotsky's scaffolding framework provides an additional lens, with RLHF, fine-tuning, and chain-of-thought prompting serving as structural analogs to the guidance provided by more knowledgeable others in child development.
The childhood development analogy illuminates; it does not prove. The analogy has served its purpose not by being right, but by being specific enough to be wrong in informative ways.
References
[1] Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). Improving language understanding by generative pre-training. OpenAI.
[2] OpenAI. (2023). GPT-4 technical report. arXiv preprint arXiv:2303.08774.
[3] Kosinski, M. (2023/2024). Evaluating large language models in theory of mind tasks. PNAS, 121(45), e2405460121.
[4] Piaget, J. (1952). The origins of intelligence in children. International Universities Press.
[5] Vygotsky, L. S. (1978). Mind in society. Harvard University Press.
[6] Wei, J., et al. (2022). Emergent abilities of large language models. arXiv:2206.07682.
[7] Piaget, J. (1972). Intellectual evolution from adolescence to adulthood. Human Development, 15(1), 1–12.
[9] Fischer, K. W. (1980). A theory of cognitive development. Psychological Review, 87(6), 477–531.
[10] Vygotsky, L. S. (1987). Thinking and speech. Plenum Press.
[11] Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. J. Child Psychology and Psychiatry, 17(2), 89–100.
[12] Vaswani, A., et al. (2017). Attention is all you need. NeurIPS, 30.
[13] Radford, A., et al. (2019). Language models are unsupervised multitask learners. OpenAI Blog.
[14] Brown, T. B., et al. (2020). Language models are few-shot learners. NeurIPS, 33.
[15] Wei, J., et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35.
[16] Webb, T., Holyoak, K. J., & Lu, H. (2023). Emergent analogical reasoning in large language models. Nature Human Behaviour, 7(9), 1526–1541.
[17] OpenAI. (2024). Learning to reason with LLMs. OpenAI Blog.
[18] OpenAI. (2025). GPT-5 system card. OpenAI.
[19] DeepSeek. (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv:2501.12948.
[20] Kosoy, E., et al. (2023). Comparing machines and children. arXiv:2305.11243.
[23] Kohlberg, L. (1981). Essays on moral development. Harper & Row.
[24] Schaeffer, R., Miranda, B., & Koyejo, S. (2023). Are emergent abilities of large language models a mirage?
NeurIPS, 36.
[25] Strachan, J. W. A., et al. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8(7), 1285–1295.
This document is published by Gyre Holdings LLC d/b/a Gyre Research for informational purposes only and does not constitute investment advice or a solicitation to buy or sell any security. Readers should consult a qualified financial professional before making any investment decision. All content is the intellectual property of Gyre Holdings LLC d/b/a Gyre Research and may not be reproduced or distributed without prior written consent.