Asep Darmawan is a reporter for Tech Ledgers covering Artificial Intelligence & Big Data. She/He is based in Indonesia.
24 August 2026 • 9 min read
For the vast expanse of human history—spanning at least 100,000 years—human children stood alone as the universe’s singular entities capable of acquiring fluent, generative language. Today, that exclusivity has vanished. Large language models (LLMs) such as OpenAI’s GPT series, Anthropic’s Claude, and DeepSeek have achieved a fluency and syntactic flexibility that allows them to convincingly masquerade as human conversationalists.
Executive Overview
Yet, looking beneath the computational hood reveals an astronomical imbalance. Teaching an AI model to speak, read, and write requires a staggering amount of data. State-of-the-art models gorge themselves on trillions of "tokens"—word-like computational chunks—before they ever utter a coherent sentence. By contrast, a human toddler masters their native tongue after experiencing a mere fraction of that input, often starting to grasp grammar and vocabulary by their first birthday after hearing only tens of millions of words.
This yawning chasm between artificial and biological language acquisition is known to cognitive scientists and artificial intelligence researchers as the data efficiency gap. While current industry strategies rely on raw computational scale—pushing model architectures to ingest nearly all available text on the internet—this brute-force approach is running toward a hard wall. With the pool of high-quality human text projected to run dry by the early 2030s, the AI community is forced to reevaluate how language is learned.
By reverse-engineering the developmental milestones of children, researchers hope to design more data-efficient models. Simultaneously, testing theories of human learning within these silicon networks provides cognitive scientists with an unprecedented sandbox to answer enduring questions: Are humans born with innate grammatical instincts, or is language an emergent statistical property of experiential learning? Ultimately, closing the data gap promises to democratize AI development, preserve endangered minority languages, and offer a mirror into the inner workings of the developing human mind.
Detailed Chronology: From Chomsky’s Rulebooks to the Transformer Revolution
The philosophical and technical battle over how language is acquired has raged for over half a century, shifting pendulum-like between purely environmental and strictly biological paradigms.
The 1950s–1960s: The Chomskyan Revolution and Symbolic AI
In the mid-20th century, MIT linguist Noam Chomsky revolutionized the study of language by challenging behaviorist psychologists like B.F. Skinner. Skinner argued that language was learned purely through environmental conditioning and reinforcement—much like a dog learning tricks for treats. Chomsky countered with the "poverty of the stimulus" argument. He asserted that human syntax is infinitely recursive and far too complex for children to master purely from the "impoverished" fragments of speech they hear in daily life. Instead, Chomsky argued that humans possess a hardwired, innate "universal grammar"—a biological template that allows children to deduce complex rules from minimal exposure.
As early AI research gained funding from the Pentagon during the Cold War, computer scientists adopted Chomsky’s symbolic, rule-based approach. Rather than immersing computers in natural language, researchers tried to explicitly code grammatical rules and dictionaries into software programs. This era of symbolic AI ultimately stagnated. Handicapped by weak hardware and rigid rules, these systems buckled under the ambiguity of human speech, plunging natural-language processing into the decades-long "AI winter" of the 1970s and 80s.
The 2010s: The Resurgence of Neural Networks
As computing power cheapened and the internet exploded with user-generated content, neural networks made a roaring comeback. Rather than programming explicit rules, researchers built statistical learners designed to recognize patterns across massive datasets.
The breakthrough arrived in 2018 and 2019 with the advent of the transformer architecture and foundational models like BERT and GPT-2. By ingesting billions of tokens, these models proved that raw statistical pattern matching could yield staggering linguistic competence. By 2022, OpenAI’s release of ChatGPT shattered any remaining skepticism, demonstrating that statistical engines could write poetry, pass exams, and converse with convincing depth.
The 2020s: The Birth of BabyLM and Multimodal Milestones
Realizing that commercial LLMs were consuming inhuman quantities of data, researchers began pushing back against the scaling orthodoxy. In August 2022, linguist Alex Warstadt—then a PhD student at NYU—sparked a conversation on social media that led to the creation of BabyLM, an annual competition challenging researchers to train language models on "developmentally plausible" datasets of just 10 million to 100 million words.
Concurrently, developmental psychologists began outfitting infants with head-mounted cameras (headcams) to capture their daily visual and auditory environments. Projects like Stanford’s SAYCam and Princeton’s expansive 1,000-day infant recording initiatives provided researchers with raw, egocentric video datasets. These datasets enabled artificial intelligence researchers to train multimodal models on the actual sights and sounds experienced by young children, marking the first steps toward bridging the sensory gap between minds and machines.
Supporting Context & Metrics: The Staggering Scale of the Divide
To truly comprehend the data efficiency gap, numerical metrics fail; scale can only be understood through stark physical analogies.
The Llama 3.1 Benchmark: Meta’s open-weight model, Llama 3.1, chewed through an astonishing 15 trillion tokens during its pretraining phase. Frontier models currently under development are projected to consume ten times that amount.
The Paper Stack Analogy: If you were to print out all the text used to train a modern frontier LLM on standard paper, the resulting stack would reach past the International Space Station.
The Human Counterpart: A pre-teen raised in a linguistically rich household will experience roughly 100 million words by age 12, scaling to approximately 300 million words by age 20 (including literacy). Printed out, a human pre-teen’s lifetime word exposure would stack a mere 20 meters high.
The Curriculum Paradox: Early BabyLM competitions heavily favored "curriculum learning"—starting models on simple data (baby talk) and scaling to complex syntax over time, mimicking human education. Surprisingly, experiments revealed that transformer models do not require ordered data sequencing to learn effectively; models like the 2024 champion, GPT-BERT, proved that mixing masked-language modeling with next-token prediction on a 100-million-word diet could outperform models trained on data 15,000 times larger.
Official Statements and Expert Perspectives
The friction between cognitive science and artificial intelligence has fostered a dynamic dialogue among leading researchers across both fields:
"The progress recently has been amazing. But we still have to burn down a forest and scrape the entire sum of all human knowledge to re-create this milestone that happens in our living rooms over the course of a year."
— Michael C. Frank, Cognitive Scientist, Stanford University
"Claude has seen the amount of language that an entire city will experience in one generation… It’s just totally miraculous. If you train GPT-2 on 30 million words, you get a nonsense generator; you don’t get a kid."
— Ethan Gotlieb Wilcox, Linguist and Cognitive Scientist, Georgetown University
"No matter how skeptical you are about AI, the thing that everyone has been really impressed with is: These things learn syntax. I didn’t think that was going to turn out to be true. And I think most people didn’t think that you could just look at the statistics of a large sample of language and figure out grammar."
— Alison Gopnik, Developmental Psychologist, University of California, Berkeley
"Children are not only reasoning about the evidence they’re being told. They’re reasoning about the teacher, about the teacher’s knowledge, and about why the teacher is telling [them] this particular information."
— Elizabeth Bonawitz, Developmental Cognitive Scientist, Harvard University
"For the last 100,000 years or however long human language has existed, humans have been the only entities in the universe that use language. Now there’s this other linguistic entity… Finally we have a model; not in the sense of a language model, but in the sense of a model organism."
— Alex Warstadt, Linguist and Data Scientist, University of California, San Diego
Future Outlook: The Next Frontier of AI and Cognitive Science
As the race to scale LLMs encounters physical limitations—including power grid constraints, data scarcity, and exorbitant training costs—the insights harvested from developmental psychology offer a roadmap for the next generation of artificial intelligence.
1. Active Exploration and "Empowerment"
Current LLMs learn passively and in isolation from static text corpora. In contrast, children are active experimenters. Research from Alison Gopnik’s lab demonstrates that children explore environments (such as interactive game spaces) to maximize their "empowerment"—the ability to make predictable, causal impacts on their surroundings. Future AI architectures may incorporate intrinsic motivation loops, allowing models to actively seek out information, probe their own knowledge gaps, and test hypotheses against physical or simulated environments.
2. Democratization and Minority Languages
Closing the data efficiency gap is not merely an academic exercise; it carries profound socioeconomic implications. David Samuel, a machine-learning researcher at the University of Oslo, notes that minority and low-resource languages—such as Sami, Czech, or regional dialects—often possess only tens of millions of available tokens, equivalent to a toddler’s linguistic exposure. Developing data-efficient architectures means smaller universities, independent labs, and indigenous communities can train capable models without requiring hyperscale data centers.
3. AI as a Model Organism for the Mind
While biological brains are constructed of living neurons, dynamic blood supplies, and continuous physiological growth—whereas LLMs pretrain once and remain static—cognitive scientists are increasingly viewing AI models as "computational model organisms." Much like comparative psychologists study animal behavior to understand human biology, researchers can deploy LLMs as linguistic lab rats. By artificially simulating bilingualism, depriving models of specific grammatical exposures, or introducing sensory constraints, scientists can test theories of human cognition that would be ethically impossible to study with real children.
Ultimately, artificial intelligence has ceased to be merely a tool for automation. By holding up a mirror to the way humans acquire language, AI has transformed into an investigative lens, illuminating the deepest mysteries of the developing human mind.
