How Ravens Plan: The Strange Cognitive Time Travel of Corvids

Planning for the future is one of the more philosophically loaded cognitive abilities in the inventory of human capacities. It requires holding a representation of a state of the world that does not currently exist, comparing it to other possible future states, selecting one as a goal, and taking actions in the present that further the goal at some cost to current well-being. For a long time, this was thought to be a uniquely human ability, requiring symbolic language and a sense of self that was thought to be absent in other animals.

The first cracks in the consensus appeared in the 1990s with work on chimpanzees, but the strongest evidence has come from corvids, particularly ravens (Corvus corax) and Eurasian jays (Garrulus glandarius). A series of experiments conducted in the early 2000s and continuing into the 2020s has demonstrated future-oriented behavior in corvids that is difficult to explain without attributing some form of mental time travel to them. The implication is that the cognitive infrastructure for planning is older and more widespread than the standard story suggests, and that the bird-mammal cognitive distance is much smaller than 300 million years of evolutionary divergence would suggest.

The original problem

The classical objection to attributing future planning to non-human animals is what William Roberts called the "Bischof-Köhler hypothesis": animals can act on present needs but cannot anticipate future needs that differ from current ones. A jay caching food in autumn is acting on the autumn need to cache, not on the winter need to eat; a squirrel storing nuts is responding to immediate availability of nuts, not to a representation of future winter starvation.

The challenge for any experimenter wanting to claim future planning in animals is to design a task where the animal has to take an action that is costly in the present and only beneficial in the future, where the future state is not predictable from current cues, and where the action requires choosing among multiple possible futures rather than simply executing a fixed program. This is harder than it sounds; many candidate experiments turn out to be explainable by associative learning or current-state responding.

Raby et al. 2007: the first decisive experiment

The 2007 paper by Caroline Raby, Dean Alexis, Anthony Dickinson, and Nicola Clayton in Nature is the experiment that made the case strongly enough to shift the field's consensus. The study used Western scrub jays (Aphelocoma californica), close relatives of ravens with similar caching behavior.

The jays were given experience in two compartments: in one (the "breakfast room"), food was available in the morning; in the other (the "no-breakfast room"), food was not. After several days of this routine, the jays were given access to food in the evening with the opportunity to cache it. They cached preferentially in the no-breakfast room, anticipating tomorrow's hunger in a location and time that was not currently relevant.

A control experiment varied the type of food available in each room (kibble in one, peanuts in the other) and showed the jays cached more of the missing food type, demonstrating that the planning was specific to anticipated future needs rather than a generalized "cache more in deprived locations" rule. This was the kind of result that could not be explained by simple associative learning: the jays were behaving as if they had a representation of tomorrow's breakfast options and were taking action in the present to address it.

Kabadayi and Osvath 2017: ravens and tools

The Raby experiment used food caching, which corvids do as part of their natural behavioral repertoire. A natural objection is that corvid caching evolved precisely as a future-oriented behavior, so the experiment is detecting an evolved specific competence rather than general planning.

The 2017 paper by Can Kabadayi and Mathias Osvath, published in Science, addressed this concern with ravens and a task that was novel and not part of natural raven behavior. Five ravens were trained that a particular tool could be used to extract food from a puzzle box. Then they were given the choice of selecting and keeping the tool when the puzzle box was unavailable, with the box appearing 17 hours later. The ravens chose the tool over a high-value immediate food reward at rates well above chance, and used the tool successfully when the box reappeared.

The same study tested bartering: the ravens learned that a particular token could be exchanged for food with a human partner. Given a choice between an immediate small reward and the token (which could be exchanged for a larger reward 15 minutes later), the ravens preferred the token and exchanged it successfully. The tasks were both novel, both required selecting an item with no current utility, and both required maintaining the relevant memory across hours.

The result is hard to explain without attributing some form of forward-looking representation to ravens. The behavioral repertoire of caching is irrelevant; the ravens were planning for the use of tools and tokens that have no analog in their natural ecology.

Boeckle and Clayton 2019: episodic-like memory

Planning for the future is closely tied to remembering the past in a specific way. The neuroscientist Endel Tulving distinguished between semantic memory (knowing facts) and episodic memory (remembering specific past events as having happened to you). Episodic memory is thought to be the substrate that planning runs on: the same neural machinery that lets you mentally re-experience the past lets you mentally pre-experience the future.

If corvids have planning, the prediction is that they should also have episodic-like memory. The 2019 paper by Markus Boeckle and Nicola Clayton in Animal Cognition demonstrated this in Eurasian jays. The jays cached perishable and non-perishable food items at different times, and were tested for retrieval after intervals where the perishable items would have spoiled. The jays preferentially retrieved non-perishable items after long intervals and perishable items after short intervals, indicating they remembered both what they cached, where they cached it, and when.

This is a "what-where-when" memory that has the structural properties of episodic memory, even though the question of whether it has the same subjective character as human episodic memory cannot be answered. The point is that the cognitive machinery to integrate type, location, and time of past events exists in birds, which is the same machinery that planning would require.

The neuroanatomy: convergent rather than homologous

The puzzle that corvid planning poses is anatomical. Mammalian planning is associated with the prefrontal cortex, an area of the cerebral cortex that birds do not have in the same form. Birds have a forebrain region called the nidopallium caudolaterale (NCL) that occupies a similar functional role: it has dopaminergic inputs, it is involved in working memory and decision-making, and lesions to it produce planning and executive function deficits.

The NCL and the mammalian prefrontal cortex are not homologous (they did not evolve from a common ancestral structure) but are functionally analogous. They are an example of convergent evolution at the neural level: two lineages, separated by 300 million years, independently arrived at similar neural solutions to similar cognitive problems.

This is a striking finding because it suggests that the architectural requirements for planning are constrained: there are not many ways to build a planning machine, and the specific way the brain does it has been re-derived multiple times. The anatomical convergence is matched by behavioral convergence: corvids and apes show many of the same cognitive abilities (planning, tool use, mirror self-recognition in some species, transitive inference, theory of mind in some tasks) despite the deep evolutionary distance.

What the experiments do not show

The strong claims about corvid planning need to be tempered with what the experiments leave open. The ravens in the Kabadayi-Osvath experiment plan over hours, not days; whether they can plan over weeks or months is not established. The jays in the Raby experiment plan one meal ahead; whether they can plan a sequence of multiple future events is not established.

The subjective character of corvid planning, if it has one, is not accessible. The behavioral evidence is consistent with several internal architectures, including ones that might not feel like anything from the inside. The philosophical question of whether ravens have something like the human felt experience of imagining the future is not answered by behavior alone.

The generalization to other birds is also not automatic. Corvids are unusually large-brained relative to body size and are unusual among birds in their cognitive achievements. The same experiments would probably not produce the same results in pigeons or chickens. The "birds plan" generalization is more accurately "some birds plan, in some respects, under some conditions" and the species-by-species inventory is still being filled in.

The deeper observation

The corvid planning literature is one of the cleaner examples of how the comparative cognition field has shifted in the last twenty-five years. The default assumption was that human cognitive abilities are unique until proven otherwise; the new default is that human cognitive abilities are extreme expressions of capacities widely distributed across vertebrates. Planning, episodic memory, theory of mind, tool use, and several other abilities once thought to be human-specific have all been found in some form in corvids, apes, dolphins, elephants, or octopuses.

The implication for the inventory of intelligent life on Earth is that complex cognition has evolved multiple times via different anatomical routes, and that the universe of possible minds is much larger than the small region currently occupied by the most cognitively similar primates. The raven's solution to planning runs on different neural substrate than the human solution but produces recognizable planning behavior, which suggests that the abstract problem of planning has a small number of good solutions and biology has found several of them.

The further implication is that the assumption "this animal cannot do X" should be approached with care. The Raby experiment was met with skepticism when first published because the prevailing theory was that animals could not plan; the skepticism was wrong, and the wrong assumption had probably been distorting the interpretation of corvid behavior for decades before the experiment forced a reconsideration. The honest position about cognition in non-human animals is that it is much harder to design clean negative experiments than positive ones, and that absence of evidence has been confused for evidence of absence many times in this literature.