ManyBabies 3 Found No Sign of Infant Rule Learning. That Is the Result.

Thirty laboratories, 839 infants, a pre-registered analysis plan. The result: no effect. Not a weak effect, not an ambiguous one. A Bayesian analysis of the data indicated strong evidence for the null. [12]

The study tested 839 infants between 5 and 12 months of age, 650 of whom survived exclusions, across 30 laboratories contributing 31 samples. The primary analysis compared looking times on test trials consistent with a rule against trials inconsistent with it. There was no significant difference. [12] A 1999 study had reported that 7-month-old infants could pull an abstract rule out of a string of syllables and apply it to syllables they had never heard. That finding became a fixture, cited and extended across modalities, stimuli, and species. ManyBabies 3 set out to check it properly.

The logic being tested

The paradigm rests on a simple inference. An infant hears a series of syllable sequences built on an abstract pattern, the same structure recurring with different sounds. After enough exposure, the infant is assumed to have extracted the pattern rather than memorised the specific syllables. Then come test trials: some new sequences follow the same abstract pattern, some violate it. If the infant has the rule, the violating sequences should be more surprising, and surprise in infants is measured by looking time. Longer looking at the inconsistent trials is the signature of rule learning. [12]

This is an indirect measure, and that matters for how the null should be read. Looking time is a proxy, the best the field has for pre-verbal infants, but a proxy. A failure to find a looking-time difference is a failure to find the behavioural signature of rule learning under these conditions. It is not a direct readout of what the infant's brain did or did not encode.

What the study found, in detail

The pre-registered design included three moderators, and the abstract reports that none of them showed significant effects. [12]

• Mono- versus multilingualism. Infants raised with more than one language did not perform differently from monolingual infants.
• Age. The sample spanned 5 to 12 months, a wider range than the original 7-month-old sample. Age did not moderate the result.
• Experimental paradigm. The study varied the paradigm, and the variation did not rescue the effect.

The age moderator is worth pausing on. Developmental researchers routinely treat age as a linear predictor of cognitive change, and a meta-meta-analysis of 25 phenomena in early linguistic and cognitive development found that assumption largely unsupported: for most phenomena, effect sizes did not change meaningfully across age. [3] So the absence of an age effect here is not automatically a sign that something went wrong. It may be the ordinary case.

The authors also ran robustness analyses, checking whether different choices in preprocessing the data changed anything. They did not. And a follow-up analysis lab by lab found no evidence of rule learning in 30 of the 31 contributed samples. [12]

That last number should settle the question of whether this is a power problem. A single lab with 20 infants failing to replicate could be noise. Thirty labs failing, with one exception, is a different kind of result.

Why a null result here is informative

Multi-lab studies exist because single-lab results are hard to interpret. A small study that finds an effect might be detecting something real, or might be detecting a quirk of its sample, its stimuli, its experimenter, or its analysis choices. A small study that finds nothing might simply be underpowered. Neither outcome resolves the question on its own.

Coordinated replication changes the arithmetic. The ManyLabs initiative demonstrated that replication efforts could be coordinated across many laboratories, and subsequent work has built on that model to accelerate cumulative evidence in fields that rely on small, individualised designs. [9] The same logic has been proposed in sport science, where researchers have argued for big-team approaches, shared data repositories, and privacy-preserving pooled analyses, while acknowledging the real difficulties around data ownership and context. [7]

There is a broader methodological point here too. A cross-taxon review of reversal learning tasks found that 99% of studies examined only one taxon, and that learning criteria and outcome measures diverge so much across groups that it becomes hard to tell genuine cognitive differences from methodological artefacts. [13] The same worry applies within developmental psychology: when every lab runs its own version of a task, a literature of mixed results is hard to read. Frameworks such as LEVANTE are being built to capture variability in children's learning across individuals, groups, tasks, and contexts rather than relying on group averages and single snapshots. [1] ManyBabies 3 is an attempt to remove that ambiguity for one specific claim, and the answer it returns is clean.

What this does not show

It does not show that infants cannot learn abstract rules. It shows that this paradigm, in this age range, with this measure, did not detect rule learning in this sample. Those are different statements, and the gap between them is where the interesting science lives.

Three specific limits are worth naming.

First, the measure. Looking time is a coarse instrument. An infant could encode the rule and still not look longer at the violating sequences, if the violation is not salient enough, or if attention is captured by something else in the display. The study cannot distinguish no learning from learning that this measure did not pick up.

Second, the population. The sample spans 5 to 12 months, wider than the original 7-month-old sample but still a narrow window of development. The abstract reports that age was not a significant moderator within that window. It says nothing about what happens before 5 months or after 12 months.

Third, the framing. This is described as a conceptual replication, not an exact one. Conceptual replications test the idea rather than the precise procedure, which is a strength for generalisability and a limitation for pinpointing what went wrong. If the original effect was real but fragile, dependent on a specific stimulus set, a specific timing, or a specific lab environment, a conceptual replication across 30 labs might miss it for reasons that have nothing to do with whether infants can learn rules.

The abstract notes that the finding raises fundamental issues about rule-learning that are presented in the Discussion section. We cannot see those from the abstract. Whatever they are, they are the authors' interpretation, not the data.

Where the evidence runs out

Our sources do not include the original Marcus et al. (1999) paper. Everything we can say about it here comes from how ManyBabies 3 describes it: that it reported 7-month-olds learning abstract rules over syllable sequences and generalising to novel sequences. We cannot independently assess the original methods, sample, or analysis.

We also cannot say whether the null result generalises beyond looking-time paradigms. The study varied the experimental paradigm as a moderator and found no significant effect of that variation, but the abstract does not tell us which paradigms were tried or how different they were from the original.

And we cannot say what the correct interpretation of the broader literature is. The abstract states that dozens of studies replicated and extended the original finding. ManyBabies 3 found no evidence for it in a large, pre-registered, multi-lab test. Both of those things are in the record. Reconciling them, deciding whether the earlier literature was measuring something real that this design missed or whether it was a collection of underpowered studies converging on a false positive, is not something this study resolves on its own, and it is not something we can resolve from the abstract.

What the study does establish is narrower and firmer: under a pre-registered protocol, across 30 laboratories, in 839 infants, the looking-time signature of abstract rule learning was absent, and the data favoured the null. That is a result. It should be reported as one.