Meta-analytic techniques reveal that corvid causal reasoning in the Aesop's Fable paradigm is driven by trial-and-error learning

Why we asked

In the Aesop’s fable, a thirsty crow drops pebbles into a pitcher until the water rises high enough to drink. Researchers have built lab versions of this task. Rooks, jays, and crows drop stones into a water-filled tube to raise a floating worm within reach, and many birds succeed. Several research teams have taken this as a sign that corvids reason about cause and effect. A recent reanalysis suggested that an initial preference for the right objects, plus ordinary trial-and-error learning, could explain the results. We wanted a finer-grained look at how the birds actually learned.

What we did

Earlier analyses counted each trial as one chance to learn. However, a single trial contains a variable number of stone drops, and each drop moves the food closer or leaves it where it was. We pooled data from five published articles (33 tasks in all) and treated each drop as a learning opportunity. We fit multilevel models, with drops nested within birds and birds within articles, and asked four things. First, did birds start out preferring the useful option? Second, how quickly did they learn? Third, did they learn from what happened on the previous drop? Fourth, did learning carry over from one task to the next?

What we found

In the water versus sand task and the sinking versus floating task, birds did not begin with a preference for the useful option, and they became more likely to choose it with each drop (odds ratios of 1.04 and 1.05 per drop). In the hollow versus solid task, birds chose solid objects nearly every time from the start, so there was little left to learn.

Birds learned from successful drops but not from unsuccessful ones. After a bird dropped a stone into the water tube, its odds of choosing that tube again were 5.0 times the odds of choosing the sand tube. After a drop into the sand tube, we found no reliable change. In other words, a bird that watched the worm move closer tended to repeat itself, and a bird that saw nothing happen did not seem to learn much from it. Finally, the order in which birds met the tasks did not predict how quickly they learned, so we found no evidence of transfer between tasks.

What it means

This pattern is what we would expect from learning based on feedback. Our conclusion concerns the tasks. They cannot tell trial-and-error learning apart from reasoning about mass, volume, and displacement.

Future studies should say what would count as evidence for causal reasoning before they collect the data.