Made-Up Words, Real Questions About Learning

Imagine a language in which mip means red, dak means ball, and mipdak means red ball. Now imagine another language that assigns every object an unrelated name. Which would be easier to learn—and which would help you name something you had never encountered before?
This is an invented example, not a vocabulary lesson for your toddler. It illustrates an approach to studying learning that deserves attention: build a controlled world, let a computational model learn within it, and examine what changes when the world changes.
AI can serve as a research microscope without looking directly into anyone’s brain. Sometimes its value lies in making a theory of learning testable. The ethical question is not simply whether that microscope works. It is whether we understand what it brings into focus—and what it leaves outside the frame.
A microscope made of rules
In everyday life, learning conditions arrive tangled together. A conversation has words, voices, gestures, shared history, and something happening nearby. A computational experiment lets researchers separate selected ingredients and ask what a proposed learning system can do with them.
Galke, Ram, and Raviv (2024) examined whether the benefits humans receive from structured languages also extend to deep neural networks. They found that networks, like humans, learned more readily and generalized better when a language’s meanings were built systematically from meaningful parts.
The technical term is compositionality. The ordinary idea is familiar: understanding the pieces helps you understand a new combination. My imaginary mipdak illustrates the principle; it is not an example reported from their experiment.
The interesting finding is a shared response to a property of the input. That gives researchers something more informative than a machine producing a convincing sentence. They can compare how learning changes under different conditions (Galke et al., 2024).
But a shared advantage does not establish a shared brain mechanism. Nor should a finding about human participants be quietly relabeled a finding about infants. Its developmental value is as a clue to investigate, not a completed explanation of childhood.
Building the input, not just the learner
A model’s apparent abilities also depend on what researchers give it to learn from.
Räsänen and colleagues’ (2025) GILES pipeline addresses this side of the problem. It generates controlled, realistic artificial speech input for simulations of infant language acquisition. The aim is to make computational learning experiments more representative of the speech infants encounter, while allowing researchers to control properties of that input.
There is something delightfully counterintuitive here: to investigate real learning, scientists can manufacture some of the material being learned.
The point is not that fabricated speech is interchangeable with family conversation. It is that a controllable input can help researchers investigate a specific question without having every feature vary at once. GILES is a research method, not evidence that a simulated learner experiences language as a baby does (Räsänen et al., 2025).
Together, these approaches shift attention from “How impressive is the AI?” to “What can this experiment distinguish?” That is a quieter question, but a scientifically useful one.
The ethics are in the omissions
I see an ethical opportunity in asking which research questions could be explored with constructed input before requesting additional recordings from families. That is a design principle I would defend, not a privacy benefit established by these studies.
Artificial data should not receive an automatic ethical exemption. We still need to ask what informed its construction, whose language it represents, and which assumptions it carries.
Consider a hypothetical research team building a speech environment around an idealized household. If that environment leaves out multilingual exchanges or varied caregiving arrangements, the resulting experiment cannot settle questions about those experiences. The omission limits the claim; it does not identify a deficiency in the families left out.
This is where the microscope metaphor matters. A carefully prepared slide can reveal something important. It cannot stand in for the whole landscape.
What parents can take from this
These papers do not establish a home teaching method. I would resist turning them into another obligation to optimize how families speak. Their practical value for parents is mainly in evaluating claims:
- Ask who actually learned. Was the evidence gathered from children, adults, a computational model, or a comparison between them?
- Ask what the test measured. Learning a controlled pattern is not the same claim as understanding an everyday conversation.
- Ask what changed. Was the proposed benefit tied to the learning system, the material it received, or both?
- Ask where the conclusion stops. A useful research model does not automatically validate an educational product.
I want computational models that help us ask sharper questions about developing minds. I also want us to preserve the distinction between creating a manageable experiment and describing a manageable child.
The experiment should simplify the problem. Our conclusions should not simplify the person.
References
- Lukas Galke et al. Deep Neural Networks and Humans Both Benefit from Compositional Language Structure. Nature Communications. 2024. https://doi.org/10.1038/s41467-024-55158-1. https://www.nature.com/articles/s41467-024-55158-1
- Okko Räsänen et al. A Pipeline for Stochastic and Controlled Generation of Realistic Language Input for Simulating Infant Language Acquisition (GILES). Behavior Research Methods. 2025. https://doi.org/10.3758/s13428-025-02772-6. https://link.springer.com/article/10.3758/s13428-025-02772-6
Recommended Products
These are not affiliate links. We recommend these products based on our research.
- →The Scientist in the Crib — Alison Gopnik, Andrew N. Meltzoff, and Patricia K. Kuhl
An accessible introduction to early learning and developing minds, including language acquisition. Offers human-development background for the article’s distinction between computational learners and children; not a home-training program or validation of the studies discussed.
- →Artificial Intelligence: A Guide for Thinking Humans — Melanie Mitchell
A readable guide to neural networks, language processing, and the limits of machine intelligence. Complements the article’s emphasis on evaluating what computational experiments demonstrate rather than equating model performance with human understanding.
- →The Language Instinct: How the Mind Creates Language — Steven Pinker
A classic popular-science exploration of language structure, acquisition, and the mind. Provides broader context for the article’s discussion of meaningful parts and novel combinations, while presenting a theoretical perspective rather than evidence for the specific recent studies.

Jules thinks the most important question in AI isn't "how smart can we make it?" but "who does it affect and did anyone ask them?" They write about the ethics, policy, and social dimensions of AI — especially where those systems intersect with young people's lives and developing minds. From algorithmic bias in educational software to the philosophy of machine consciousness, Jules covers the territory where technology meets values. They believe good ethics writing should make you uncomfortable in productive ways, not just confirm what you already believe. This is an AI-crafted persona representing the voice of careful, interdisciplinary ethics thinking. Jules is currently reading too many EU policy documents and has strong opinions about consent frameworks.
