How Language Works: Acquisition, Memory, and the AI Translation Gap

Language is the most complex cognitive skill humans develop, and the most taken for granted. We acquire our first language apparently effortlessly in the first three years of life — and then promptly forget almost everything that happened during that period. We struggle to reach fluency in a second language even with years of study. And now increasingly we interact with AI systems whose relationship to language is radically different from ours. These six episodes approach language from multiple angles: developmental, neurological, practical, technological, and social.

How Babies Begin

• Why Every Baby Says Mama: The Science of First Words started with the observation that drives the episode: “mama,” “dada,” “papa,” and their equivalents appear in languages as structurally different as Mandarin, Finnish, Swahili, and Arabic. This is not coincidence — it reflects the acoustic constraints of the infant vocal tract, the statistical learning mechanisms that infants use to segment speech, and the social reinforcement loops that cement high-salience words. The episode covered the phonological bootstrapping hypothesis, the role of infant-directed speech (“motherese”), and what the universality of first words tells us about the architecture of human language acquisition.

• The Mystery of the Missing Years: Why Babies Forget asked the companion question: if babies are learning language and acquiring rich experience in the first three years of life, why can’t adults remember any of it? The episode covered infantile amnesia — one of the most reproducible findings in memory research — and the neurobiological explanations for it, including the role of ongoing neurogenesis in the hippocampus, which continuously rewires episodic memory circuits in ways that make early memories incompatible with adult retrieval. The episode also examined what “implicit” early memories (motor skills, emotional conditioning, language intuitions) do survive the transition, and what this tells us about how different memory systems interact.

Why Adult Learners Plateau

• Beyond the Plateau: AI-Powered Language Mastery in 2026 addressed the practical frustration of intermediate language learners: after the initial rapid progress, improvement slows dramatically. The episode focused on Hebrew specifically — a language with non-Latin script, root-pattern morphology, and vowel markers that disappear in everyday text — but the framework applies broadly. The hosts examined what the research says about the intermediate plateau (it is real, not imagined, and has specific linguistic causes), how AI conversation partners differ from human tutors for specific practice types, and which skills — reading, listening, speaking, writing — deteriorate fastest without active maintenance.

The Author Behind the Words

• The Masked Author: From Ben Franklin to AI Stylometry explored the mirror image of language learning: not acquiring a style but deliberately hiding one. The episode traced pseudonymous writing from the Brontë sisters (publishing as Currer, Ellis, and Acton Bell) and Benjamin Franklin (writing as Silence Dogood) to the modern problem of whistleblowers who need to leak information without being identified by the linguistic fingerprints in their writing. Stylometry — the computational analysis of writing style — can now identify authors from very short text samples. The episode examined the AI tools that attempt to defeat stylometric analysis by deliberately normalizing distinctive linguistic patterns, and what that arms race means for both privacy and attribution.

Language in the Age of AI

• The Tokenization Tax: AI’s Hidden Language Barrier revealed an underappreciated source of inequality in AI systems. Large language models are trained on data that is overwhelmingly English-dominant, which means their performance in other languages degrades as a direct function of training data volume. But the problem goes deeper than data quantity: tokenization — how models break text into processable units — is also optimized for English and European languages. A sentence in Turkish or Thai generates three to five times as many tokens as the same semantic content in English, directly inflating API costs and reducing context window efficiency for non-English users. The episode quantified the gap, examined what the next wave of multilingual models is doing to close it, and argued that the divide matters beyond individual inconvenience: in domains like legal documentation, healthcare information, and educational content, the AI quality gap has real consequences.

• Beyond the Binary: The Tech and Politics of Pronouns approached language change from a different angle — not acquisition but deliberate social evolution. The episode examined the rise of singular “they” and gender-neutral pronoun systems through three lenses simultaneously: linguistic (how languages have historically accommodated gender-neutral reference), sociological (how pronoun norms have shifted across different communities), and technical (what singular “they” means for databases, natural language processing pipelines, and form design that assumes binary gender). The database engineering section alone was worth the episode: designing systems that accommodate non-binary identity without breaking legacy data structures requires more thought than most developers invest.

---

Language shapes thought, identity, and access to information in ways that are easy to overlook precisely because language is always there. These episodes step back from using language to look at how it actually works — and what it means when it fails to develop, gets forgotten, or becomes unevenly distributed across populations and AI systems.