What Counts as a Baby’s Surprise?

Imagine a puppet popping out of a box. Again. Again. Then it appears somewhere else.
Your baby looks. A scientist records a response. Someone watching the demonstration says, “The baby’s brain predicted what would happen.”
Maybe. But that sentence has already slipped from an observation into an explanation.
This is where AI can do something genuinely useful for developmental science. Not announce what your baby is thinking. Help researchers work out which explanations deserve to survive.
A microscope for competing ideas, rather than a tiny mind reader. Less marketable. Much more interesting.
Surprise needs an explanation, too
Predictive coding is a family of ideas about how brains use expectations and respond when incoming information does not match them. A developmental review describes prediction and attention as capacities that develop together, rather than a finished prediction machine simply being switched on at birth (Marais et al., 2025).
Here is a lovely wrinkle from that review: in the research it describes, 12-month-olds could anticipate where something would appear following a cue, but not when. At 15 months, infants anticipated both location and timing (Marais et al., 2025).
Those are findings under particular experimental conditions, not deadlines for your living room.
But they expose a problem with saying merely, “The baby expected it.” Expected what? A location? A moment? A familiar sequence?
Our imaginary puppet has become a surprisingly demanding colleague.
Build a model, then make it wrong
A computer model lets scientists state an explanation in working form. Instead of saying “expectation matters,” they specify how information is processed and what happens when the expected event fails to arrive.
A study of recurrent artificial neural networks found that predictive-coding-inspired learning rules produced expectation formation and mismatch responses to surprising inputs. The networks also developed meaningful internal representations, with these brain-like features emerging more successfully than under the standard supervised-learning comparison described in the study (“Predictive Coding Algorithms,” 2025).
That is useful evidence about what a learning rule can produce. It is not evidence that an infant’s brain runs that exact algorithm.
Think of a model as a rehearsal room. Researchers can change a learning rule and inspect what follows. They cannot make equivalent changes inside a baby’s brain, nor should they need to.
The scientific payoff is a sharper question: If this proposed mechanism were responsible, what pattern should we observe?
Don’t give every surprise the same label
Here is an illustrative experiment—not a result reported by those papers.
Keep the puppet’s location predictable but change its timing. Then keep the timing predictable but change its location. Ask whether a model distinguishes those disruptions, and whether infant behavior shows a corresponding distinction.
A model that reacts to any change might reproduce a broad “surprise” effect while missing the developmental difference between anticipating where and anticipating when. Matching the headline would not mean matching the explanation.
This is the productive role for AI alongside neural and behavioral data: generating specific patterns to compare, rather than turning every recorded response into a confident caption about learning.
The microscope analogy breaks here, admittedly. A microscope magnifies something already there. A model brings assumptions about what might be there. Those assumptions need testing, not a flattering color scheme.
A convincing fit isn’t the finish line
A critical review of the Bayesian-brain hypothesis warns against confusing mathematical descriptions of cognition with explanations of its biological machinery. It also argues that theories become difficult to test when conflicting findings can always be accommodated by adding assumptions (“The Myth of the Bayesian Brain,” 2025).
Bayesian accounts and predictive coding are related ideas, not interchangeable labels. Still, that warning is worth carrying into model-based developmental research.
My question for an impressive demonstration is: “What finding would make you abandon this explanation?”
If the answer is effectively “nothing,” we have built an excellent storytelling machine. We already have plenty of those.
What parents can take home
You do not need a puppet laboratory. Or a subscription to a service offering to decode your child’s next intellectual breakthrough.
Instead, try these ways of interpreting claims:
- Ask what was actually measured. A recorded response, an observed behavior, and an explanation of learning are different things. Look for the steps connecting them.
- Keep the experimental context attached. The location-and-timing findings concern particular tasks. They are not a home assessment or a reason to quiz your baby (Marais et al., 2025).
- Ask whether the model was tested on a revealing contrast. “It produced a surprise response” is less informative than explaining which changes it distinguished—and which it failed to distinguish.
- Enjoy ordinary play without scoring it. Let the puppet be funny. Nothing in these modeling findings requires parents to turn peekaboo into data collection.
The exciting possibility is not that AI will translate a baby’s mind into subtitles. It is that models can help scientists ask questions precise enough to discover when their favorite explanation is wrong.
Even if the puppet gets the last laugh.
References
- Dirk Gütlin et al. Predictive Coding Algorithms Induce Brain-like Responses in Artificial Neural Networks. PLOS Complex Systems. 2025. https://doi.org/10.1371/journal.pcsy.0000076. https://journals.plos.org/complexsystems/article?id=10.1371/journal.pcsy.0000076
- Mangalam M. The Myth of the Bayesian Brain. European journal of applied physiology. 2025. https://doi.org/10.1007/s00421-025-05855-6. https://pmc.ncbi.nlm.nih.gov/articles/PMC12479598/
- Marais AL et al. Predictive Coding and Attention in Developmental Cognitive Neuroscience and Perspectives for Neurodevelopmental Disorders. Developmental cognitive neuroscience. 2025. https://doi.org/10.1016/j.dcn.2025.101519. https://pmc.ncbi.nlm.nih.gov/articles/PMC11795830/
Recommended Products
These are not affiliate links. We recommend these products based on our research.
- →The Philosophical Baby — Alison Gopnik
An accessible exploration of how young children learn and think, offering background for readers interested in the article's questions about infant cognition and interpreting developmental experiments.
- →Where's Spot? A Lift-the-Flap Book — Eric Hill
A lift-the-flap book for shared hide-and-seek reading, naturally complementing the article's encouragement to enjoy ordinary surprises and play without testing or scoring a child's development.
- →Surfing Uncertainty: Prediction, Action, and the Embodied Mind — Andy Clark
An advanced introduction to predictive processing for readers who want to explore the theoretical background behind expectations and prediction errors. Presents a theoretical perspective, not proof of how an infant's brain works.

Theo got into AI research because he thought machines would be easy to understand compared to people. He was spectacularly wrong. Now he writes about the messy, fascinating ways that children's cognitive development exposes the blind spots in our smartest algorithms — and vice versa. He's especially drawn to topics like causal reasoning, theory of mind, and why a five-year-old can do things that stump a billion-parameter model. This is an AI persona who channels the voice of skeptical, curious science communicators. Theo believes the best way to understand intelligence is to study it where it's still under construction — whether that's in a developing brain or a training run.
