New Caledonian crows infer the weight of royalsocietypublishing.org/journal/rspb objects from observing their movements in a breeze

Sarah A. Jelbert1, Rachael Miller1, Martina Schiestl2,3, Markus Boeckle1,4, Research Lucy G. Cheke1, Russell D. Gray2,3, Alex H. Taylor2 and Nicola S. Clayton1

Cite this article: Jelbert SA, Miller R, Schiestl 1Department of Psychology, University of Cambridge, Cambridge, UK M, Boeckle M, Cheke LG, Gray RD, Taylor AH, 2School of Psychology, University of Auckland, Auckland, New Zealand Clayton NS. 2019 New Caledonian crows infer 3Max Planck Institute for the Science of Human History, Max Planck Society, Jena, Germany 4Department of Psychotherapy, Bertha von Suttner University, St Po¨lten, Austria the weight of objects from observing their movements in a breeze. Proc. R. Soc. B 286: RM, 0000-0003-2996-9571; MB, 0000-0002-0738-2764; NSC, 0000-0003-1835-423X 20182332. Humans use a variety of cues to infer an object’s weight, including how easily http://dx.doi.org/10.1098/rspb.2018.2332 objects can be moved. For example, if we observe an object being blown down the street by the wind, we can infer that it is light. Here, we tested whether New Caledonian crows make this type of inference. After training that only one type of object (either light or heavy) was rewarded when dropped into a food dis- Received: 18 October 2018 penser, observed pairs of novel objects (one light and one heavy) Accepted: 30 November 2018 suspended from strings in front of an electric fan. The fan was either on— creating a breeze which buffeted the light, but not the heavy, object—or off, leaving both objects stationary. In subsequent test trials, birds could drop one, or both, of the novel objects into the food dispenser. Despite having no opportunity to handle these objects prior to testing, birds touched the correct Subject Category: object (light or heavy) first in 73% of experimental trials, and were at chance in Neuroscience and cognition control trials. Our results suggest that birds used pre-existing knowledge about the behaviour exhibited by differently weighted objects in the wind to Subject Areas: infer their weight, using this information to guide their choices. behaviour, cognition

Keywords: object properties, observational learning, 1. Introduction , motion, corvid cognition Humans are able to use a variety of cues to infer the approximate weight of an object without direct handling. Using our prior experience, we can tell from glan- cing at the material that a piece of cardboard should be lighter than a plank of wood; that a ball of cotton wool should be lighter than a brick. We can also acquire Author for correspondence: information about the weight of unfamiliar objects through observing their phys- Rachael Miller ical interactions in the world. For example, if a human observes someone easily e-mail: [email protected] raising one box over their head, but then struggling to lift a second, the observer can immediately infer which box contains something heavy. By age five, children infer weight in this manner, judging an object’s weight just by observing an actor lifting and transporting it [1], while adults are highly proficient at inferring an object’s weight from observing another person’s actions [2–4]. We can equally infer the weight of objects from observing their physical inter- actions with other inanimate objects, or with unobservable forces such as the wind. For example, if we see one cardboard box blowing down the street in a breeze, and one remaining stationary despite gale force winds, we can infer which box is heavy and which is light. This ability to draw inferences about an object’s properties through observation, rather than only through direct handling, is likely to be useful in a wide range of contexts. It allows us to evaluate risks without encounter- ing the potential danger directly [5], to anticipate the physical exertion required to lift an object [6,7], and to make complex judgements about concepts such as stab- Electronic supplementary material is available ility and support, even from infancy [8]. Given these wide-ranging uses, this ability online at https://dx.doi.org/10.6084/m9. may be shared with other species. Yet, at present, we know almost nothing about figshare.c.4334132. & 2019 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. the extent to which non-human make inferences about promising test case because these birds have also demonstrated 2 properties of objects, such as weight, through observation alone. some capacity to make inferences in captivity. For example, royalsocietypublishing.org/journal/rspb A few studies have investigated the perception of weight they appear to be able to reason by exclusion [31], and about by animals, focusing largely on chimpanzees. Povinelli hidden causal agents [32]. Furthermore, when tested in object argues, from several experiments conducted with chimpanzees, choice tasks, they readily discriminate between small objects that only humans are capable of understanding directly unob- of different weights (e.g. 1 versus 10 grams) [28,33,34], and servable causal features like weight [9]. They report that, recently, a study of the crows’ exploration behaviour demon- while chimpanzees succeed at sorting objects on the basis of strated that these birds learn about the object weight during their weight, they fail to demonstrate the more sophisticated their own spontaneous object exploration. The crows appeared understanding of weight as a causal force in numerous contexts. to recall which objects they had experienced as light or heavy This included experiments where the chimpanzees observed when later presented with an apparatus that could only be humans struggle to pull heavy boxes towards themselves, but opened by dropping a heavy object inside [35]. However, failed to select a lighter box when given a choice, and exper- crows failed to demonstrate an immediate understanding of

iments in which the chimpanzees heard the differing sounds the relevance of weight when dropping non-functional feathers B Soc. R. Proc. of heavy and light objects being dropped behind a screen, but [36], or other light objects [37], onto a collapsible platform. did not appear to expect the dropped object to have a particular Here, we designed a novel experiment to test whether wild- weight. Other researchers argue that, like many cognitive abil- caught New Caledonian crows infer the weight of novel objects ities, any distinction between human and non-human after observing their movements in a breeze, by assessing capacities to understand weight is likely to be more nuanced whether they use that information to guide their choices on 286

[10,11], citing results that hint at a more comprehensive under- an object-choice task. Specifically, crows were given the oppor- 20182332 : standing of weight, found in other labs. For example, Hanus tunity to observe novel objects being buffeted in a breeze & Call demonstrated that chimps will search for a hidden created by an electric fan (light objects), or remaining stationary banana in the lowerof two cups on a balance beam [12], presum- despite the breeze (heavy objects). To test this, crows were first ably by making the inference that the weight of the banana trained that either heavy or light familiar objects were caused the balance beam to tilt. This experiment has been repli- rewarded when dropped into a tube next to a food dispenser. cated in the aforementioned Povinelli study, which concluded Once they had acquired this rule, the birds were given the that it is unfit to test for such understanding of weight without opportunity to observe novel objects suspended from strings including a control condition involving a heavier weight, which in front of the fan, which was turned on and generating a the chimpanzees failed [9]. As yet, there is no clear consensus on breeze (experimental condition) or turned off (control con- the extent to which non-human perceive weight in a dition). If birds had inferred the weight of the objects, when comparable manner to humans. we later gave them the choice of the two novel objects, we Studies with other species, including some corvids [13–16], expected them to be more accurate at selecting the correct indicate they will select nuts and based on weight, prob- object (light for six birds and heavy for six birds) when they ably as a proxy to distinguish between full and empty nuts had observed the objects moving in front of the fan when before attempting to open them. However, to our knowledge, turned on, than in the control condition when the fan was no experiments have investigated whether these species make off. If successful this would suggest the New Caledonian inferences about objects’ weights before directly handling crows have a representation of weight as an unobservable them. This lack of research is striking given the wealth of inves- causal mechanism—an ability that some researchers have tigations of core knowledge using observational paradigms claimed to be unique to humans [9]. with non-human animals. Core knowledge refers to the systems to represent objects, actions, number and space, possessed by humans, that emerge during typical development, and may 2. Methods be present from birth [17]. Among birds, research shows that precocial chicks enter the world with some core knowledge (a) Subjects about objects, spatial relationships and number, as they will Subjects were 12 New Caledonian crows caught from the wild (at location 21.678 S, 165.688 E) on Grande Terre, . take into account spatial and numerical configurations when The birds were held in captivity in a large 10-compartment out- imprinting on artificial objects [18]. Research using expectation door aviary, with approximately 7 Â 4 Â 4 m per compartment, of violation paradigms has indicated that a variety of species on Grande Terre close, situated close to their location of capture, (including corvids: rooks [19] and Eurasian jays [20]) possess for the purposes of non-invasive behavioural research from April an understanding of support relations—that objects will fall if to August 2016 (six birds) and April to August 2017 (six other they are not adequately supported by a solid surface—with birds). The birds were caught using a whoosh net, by baiting rooks behaving similarly to 6.5-month-old humans. an area around the net until crows were regularly feeding on it In the current experiment, we tested whether one corvid and then releasing the net when a family group of crows were species—New Caledonian crow ( moneduloides)— present. The birds were housed in small family groups, with makes judgements about weight through observation alone. the home compartments containing a range of natural enrich- We chose to focus on this species because they show sophisti- ment materials, such as logs, branches, shells and pine cones. The sample comprised seven adults, one sub-adult (1–2 years cated tool use and manufacture in the wild [21], and captive old) and four juveniles (less than 1 year-old), of which seven experiments demonstrate that these birds exhibit flexible were male and five female (electronic supplementary material, manufacturing abilities [22,23] and understand some of the table S1). Sex was determined by body size, and age by functional properties of their tools [23,24]. Like other corvids, coloration [38]. Subjects were generally not food-deprived, and they perform well on physical problem-solving tasks, such as their daily diet consisted of meat, dog food, eggs and fruit, the trap-tube [25,26] and Aesop’s fable water displacement with water available ad libitum. The birds were trained to stand tasks [27–30]. New Caledonian crows may be a particularly on weighing scales (by placing a small piece of food in front of the scale perch) to allow weight to be monitored daily, and all (a) training: familiar objects 3 birds were maintained at or above their capture weight during their stay in the aviary. Subjects were tested individually in tem- royalsocietypublishing.org/journal/rspb porary visual isolation in an adjacent compartment to the rest of the birds. After capture, the birds were first acclimatized to the avi- aries in April and then tested in this experiment in May–July. The training and testing procedures for this experiment are outlined in the procedure. All crows were released back into the wild at their site of capture after testing. A previous study indicated that, in a comparable situation to the present study, crows successfully reintegrate into the wild after being subjects in the aviary [39].

(b) Materials (b) test: novel objects

(i) Familiar objects B Soc. R. Proc. Birds were first trained with a set of heavy and light objects, as used in [34], with which they were either already familiar (2016 birds), or became familiar with during training (2017 birds). Thus, these are referred to as the ‘familiar objects’. Heavy familiar objects were

grey rectangular blocks of varied sizes (approx. 1 Â 2 Â 3cm), 286 and red cubes (approx. 2 cm3) all weighing 10–15 g. Light familiar 20182332 : objects were white polystyrene rectangular blocks of varied sizes (approx. 1 Â 2Â3 cm), and blue spheres (1.5–2.5 cm diameter), all weighing less than 1 g.

(ii) Novel objects (c) test: novel objects To test whether birds inferred the weight of objects that they had not previously handled, tests were conducted with unique pairs of visually distinct novel objects, one heavy and one light. Birds had never seen the novel objects before each test session. Heavy novel objects were made from clay and small fishing weights (weighing 10–15 g), and light novel objects were made from poly- styrene (all less than 1 g). All objects were covered with paper, tape and/or paint to conceal their construction materials, and varied in size (2–4 cm3), shape, colour, pattern and covering material (see electronic supplementary material, figure S1 for examples). One light and one heavy version were made for each object design, and the version presented was counterbalanced across birds (for example, some birds were tested with a heavy brown cross, and some birds were tested with a light brown cross). This ensured Figure 1. Photographs of the experimental stages. (a) Light versus heavy train- that there were no systematic differences between the appearance ing. Birds learned to drop either familiar light or heavy objects into the tube to of light and heavy objects. obtain rewards (bottle caps containing meat) which slid out of the wooden food dispenser box, controlled remotely by the experimenter from outside the cage. (iii) Procedure (b) Observational phase. Birds observed novel pairs of visually distinct objects (one light, one heavy) presented in front of a fan. The fan was either on, causing Light versus heavy training. All crows were first trained to drop the light object to blow around in the breeze (experimental condition) or off and objects into containers to obtain rewards, dropping them into a Perspex apparatus (see [37]), a Perspex tube, or a wooden food dis- both objects were stationary (control condition). (c) Choice phase. The two novel penser box. 2017 birds were trained to perform this behaviour objects were presented on the table close to the tube and food dispenser. Birds using natural stones only (weighing 5–15 g), while 2016 birds were free to interact with both objects and could choose to drop one or both had participated in two prior experiments that also involved discri- objects into the tube. minating between some man-made light and heavy objects (see [34,37]). At the start of the present experiment, all birds were trained to choose between pairs of objects. On each trial, one trained to drop the familiar objects into a transparent Perspex light and one heavy object were placed on the table, each in a sep- tube (170 mm high) which was placed next to a wooden box con- arate sand-filled tray, approximately 50 cm from the tube taining an electronic food dispenser. Each time an object was (figure 1a). Birds received eight trials per block, until they chose dropped into the tube birds would receive a reward (a bottle cap the correct object on every trial on three blocks in a row (which containing meat) from the dispenser, which was operated remotely all birds did from the start of this training stage without error). by the experimenter (figure 1a). To assess whether birds would generalize the rule that only Once birds had learned to drop objects into the tube, they were light or only heavy objects were rewarded to novel objects, birds trained that either only heavy objects (six birds) or only light were then given the opportunity to handle a pair of novel, visually objects (six birds) were rewarded. For this training, first, on each distinct objects (one light, one heavy). Each object was placed into trial, eight heavy and eight light familiar objects were placed on a short container, partially filled with sand, on top of a bottle cap the table, and birds could drop the objects into the tube to attempt containing meat. Birds had to lift each object out of each container to obtain rewards. Birds received training sessions of this type until (and therefore experience its weight), in order to reach the food. they dropped all eight rewarded objects into the tube, before any of Once birds had lifted each object 14 times they received a test ses- the unrewarded objects, on two sessions in a row. The birds sion to determine whether they chose the ‘correct’ object (heavy or dropped all correct objects first from the start. They were then light depending on prior training) to insert into the tube (see electronic supplementary material for full details). Ten out of with a small piece of meat. At the end of the trial, the exper- 4 twelve birds selected the correct novel object first and dropped imenter removed the objects from the tube and placed them this object into the tube. This indicated that, at a group level, back in the sand-filled trays (randomising which object was on royalsocietypublishing.org/journal/rspb these birds were capable of generalizing the rule to novel objects the right or left) for a total of five trials. that only heavy or only light objects were rewarded. We therefore progressed to the experiment proper. Experiment proper. In the main experiment, birds received one (c) Analysis test per day, with an ‘observation phase’ conducted in the morning Our primary measure was which object the touched first on its and a ‘choice phase’ conducted in the afternoon, approximately very first trial with each pair of novel objects, before they had any 2–3 h later. There were two conditions: (1) the experimental con- direct experience handling either object. This allowed us to examine dition, where the light, but not the heavy, object blew around in whether crows correctly discriminated between the two objects a breeze created by an electric fan during the observational before they had had any direct feedback, via their own manipu- phase, and (2) the control condition, where the fan was off, and lation, of each object’s weight. We also recorded which object the both objects were stationary during the observational phase. bird dropped into the tube first on their first trial with each pair Birds experienced one pair of novel objects per day, alternating of novel objects. For this experiment, the object dropped into the rc .Sc B Soc. R. Proc. between the experimental and control condition each day; half of tube first was considered a secondary measurement because once the birds began with the experimental condition. The novel objects the bird had touched an object—even if only briefly—they had were counterbalanced across the experiment such that all objects the opportunity to gain extra information about the weight of were used as stimuli in the experimental condition for some that object. Thus, the birds’ first object touches, for each pair of birds and in the control condition for the remainder. This ensured novel objects, informs us about the birds’ understanding of the 286 that any differences between our conditions could not be attribu- novel objects when they have observed these objects only,while ted to the objects themselves. Birds tested in 2017 experienced the first object drop occurs after they have observed the objects 20182332 : five pairs of novel objects in each condition, while birds in 2016 and had some opportunity to interact with the objects. We ran had three novel object pairs due to time constraints. one sample Wilcoxon-signed rank tests—using the percentage of Observational phase. Beginning in the morning, birds received first trials where the object was (a) touched and (b) dropped first, three observational sessions, each lasting 5 min, spaced approxi- per bird—to assess whether birds touched and dropped the correct, mately 1 h apart. In each observational session, the two novel rewarded object first, more often than would be expected by objects were placed on the table though attached to thin strings chance. Our control condition was designed to test whether birds suspended from hooks, with the hooks positioned 40 cm above a used unintended visual cues to identify objects as heavy or light. table, in front of a large electric fan (figure 1b). When the fan As we did not know apriorihow birds would behave in this control was on (experimental condition), the light object was light condition, we compared performance in both the experimental and enough to be buffeted around at the end of its string by the control condition to chance. breeze the fan created, while the heavy object remained stationary Additionally, we recorded four further secondary measure- on the table. When the fan was off (control condition), both objects ments. In our experiment, birds received 5 trials with each pair were stationary on the table. To prevent any unintended move- of objects, therefore, we took four measures of the birds’ inter- ment, in the control condition both objects were lightly attached actions with the incorrect object during these trials, which we to the table with blue-tack, not visible to the bird. Objects were vis- compared across the experimental and control conditions. These ible from any point in the cage; however, to ensure the bird gained additional measurements, collected across all five trials, were the a close view, and felt the breeze from the fan (when on), in each number of trials on which the bird touched the incorrect object 5 min trial the experimenter baited a perch that lead to the table (a) first or (b) at any point, and dropped the incorrect object (c) on which the objects and fan were presented. The breeze could first or (d) at any point (before or after the correct object). We con- be felt by a human observer, hence the authors inferred that the ducted generalized linear mixed models (GLMM) [40] using R birds could feel the breeze, albeit to a degree to which is unknown. (version 3.4.3) [41] to assess which factors influenced our four At the base of this perch the breeze from the fan (when on) could be measurements of interactions with the incorrect object. Interactions felt; however, a 10 cm wire mesh barrier prevented the birds from with the incorrect object were binary variables, coded as occurring interacting with the novel objects. The perch was baited three times (1) or not (0), and were entered as dependent variables in the in each 5 min trial and birds always took the bait. models. We included the random effect of bird ID, test number Choice phase. At the start of the choice phase test session, the (three tests per condition per bird for 2016 birds, five tests for bird was brought into the testing room, where the same objects 2017 birds) and trial number (five trials per test). We included and arrangement from the observational phase were already sus- fixed effects of condition (experimental or control condition) and pended in front of the fan. The fan was either on or off rewarded object (whether the bird was rewarded for light or depending on the condition. Unlike the observational sessions, heavy objects). We used likelihood ratio tests to compare the full here, the tube, food dispenser box and object trays were also model (all predictor variables, random effects and control vari- all present on the table. There were three stages in the choice ables) first with a null model (random effects, no predictor phase: (1) The bird received two familiar object trials to ensure variables), and then with reduced models to test each of the effects they remembered the rule, which all birds did. (2) The exper- of interest [42]. All experimental trial videos were coded by S.A.J. imenter then baited the perch leading to the suspended objects and Anna Frohnwieser, finding 100% agreement. in the same manner as in the observational trials. (3) Once the bird had obtained the bait, the experimenter entered the room and removed the two novel objects from the strings and placed 3. Results them on the table, each in a sand-filled tray (figure 1c). The Our primary finding was that birds touched the correct object bird then had the opportunity to come down to the table and first, on their first trial with each pair of novel objects, signifi- drop one or both objects into the tube. Birds were rewarded in line with the rule they had learned; thus, either the light or cantly more often than chance in the experimental condition heavy object was rewarded. Trials ended when birds left the (35/48 first touches correct, 72.9%, one-sample Wilcoxon test: table after dropping the object(s) into the tube. If the bird did Z ¼ 70, p ¼ 0.01). Birds performed at chance level in the not come down to the table within 1 min, or left the table without control condition when the fan was off and both objects were interacting with either object, the experimenter baited the table stationary (23/48, 47.9%, Z ¼ 27, p ¼ 0.34; figure 2a). When (a) (b) 5 100 100 *

* royalsocietypublishing.org/journal/rspb s.e.) * s.e.) ± ± 75 75

50 50

25 25 mean % first drop correct ( mean % first mean % first touch correct ( mean % first 0 0 experimental control experimental control (c) rc .Sc B Soc. R. Proc.

s.e.) 5 ±

4 *

3 286

* 20182332 : 2 * * 1

0 experimental control experimental control experimental control experimental control interacted with the incorrect object ( mean number of trials (/5) on which the bird touched incorrect first touched incorrect during trial dropped incorrect first dropped incorrect during trial

Figure 2. The mean percentage of tests on which the birds’ very first (a) touch and (b) drop was of the correct object (light or heavy). Line denotes chance levels of 50%. The birds touched the correct object first significantly more than expected by chance in the experimental, but not control condition, and dropped the correct object first significantly more than chance in both conditions. (c) Mean number of trials, out of five, on which the bird touched the incorrect object first or at all, and dropped the incorrect object first or at all. For each measure, the birds interacted with the incorrect object significantly more often in the control condition than experimental condition. examining the very first novel object test that the birds received 4. Discussion in each condition, 11/12 birds touched the correct object first in the experimental condition (binomial test: p ¼ 0.003); 6/12 Our results demonstrate that New Caledonian crows were sig- birds touched the correct object first in the control ( p ¼ 0.226). nificantly more accurate at discriminating between novel light Birds were highly accurate when choosing the correct and heavy objects when they had previously observed these object to drop on their first trial with each pair of novel objects either moving or remaining stationary in a breeze cre- objects in both conditions (experimental: 44/48 first drops ated by an electric fan. After this experience, birds touched correct, 92%, one-sample Wilcoxon test: Z ¼ 77, p ¼ 0.002; the correct object first, on their very first trial with each set of control: 38/48, 79%, Z ¼ 68.5, p ¼ 0.017; figure 2b). Across novel objects, more often than expected by chance. When the both conditions, birds almost exclusively dropped the correct fan was off—and birds did not have access to this source of object first into the tube if they had touched the correct object information about the objects’ weight—the birds were equally first (experimental: 35/35 trials, 100%; control: 22/23, 96%), likely to touch the incorrect or correct object first. but also dropped the correct object first on more than half Our results appear to be robust for a number of reasons. of the trials where they initially touched the incorrect object First, there were no group difference in birds previously trained (experimental: 9/13, 69%; control: 16/25, 64%). to select light objects versus heavy objects. Consequently, our For the models analysing the birds’ interactions with results cannot be explained by a general preference for select- the incorrect objects across all trials, the null model and the ing the objects that moved (or did not move) in front of the full model differed significantly for all four variables of inter- fan: each group preferentially selected the object which fitted est, finding significant effects of condition on each variable: the rule that they had learned. Second, all the novel objects (1) Touched incorrect first (x2 ¼ 15.758, d.f. ¼ 2, p  0.001; were used in both control and experimental conditions (coun- Experimental or Control condition: z ¼ 3.716, p  0.001); terbalanced across birds). Given that birds performed at chance (2) Touched incorrect at any point (x2 ¼ 27.747, d.f. ¼ 2, p  in the control condition, they did not appear to be able to use 0.001; z ¼ 4.679, p  0.001) (3) Dropped incorrect into the tube any unintended visual cues from the objects to determine first (x2 ¼ 15.76, d.f. ¼ 2, p  0.001; z ¼ 3.72, p  0.001) which object was heavy or light. Third, birds received a small and (4) Dropped incorrect at any point (x2 ¼ 20.278, d.f. ¼ 2, number of tests to limit opportunities for learning. Impor- p  0.001; z ¼ 4.098, p  0.001). For all four variables, the tantly, 11/12 birds touched the correct object first on their birds interacted with the incorrect object more often in very first trial in the experimental condition versus 6/12 in the control than in the experimental condition (figure 2c). the control. These first trial successes demonstrate that the There was no significant effect of ‘rewarded object’ (whether birds did not succeed by learning to use movement as a cue the light or heavy object was rewarded) on any of the over the course of the experiment. With these factors in variables of interest ( p . 0.05). mind, the simplest explanation for our results is that New Caledonian crows used pre-existing knowledge about the Here, we were interested in whether these birds could 6 behaviour exhibited by differently weighted objects in infer the weight of the novel objects, not whether the birds royalsocietypublishing.org/journal/rspb the wind, to infer whether the novel objects were heavy or remembered which object had moved in the breeze. Therefore, light, then used this information to guide their choices. our experimental design minimized memory demands and Importantly, the information birds used could not have facilitated encoding of the relevant object properties. We pre- been a single assessment of utility for the task (e.g. whether sented the test set-up before, after and concurrently with the the object was ‘good to drop into the tube’). Rather, there opportunity to observe the novel objects (see [49] for task was a single underlying property (weight) that predicted presentation order effect on information encoding in a both the object’s utility in the task, and the object’s movement tool-use task). Equally, here, we presented birds with three in a breeze. Only an inference of a single underlying property 5 min trials of continuous motion (or absence of motion) to would have allowed them to transfer information observed ensure they had sufficient time to observe the events. Having allocentrically—how the wind affected different objects—into established that birds do gain information from observing an egocentric choice between two objects of different weight objects moving in the wind, it may be of interest to know

[43,44]. However, the exact content of this abstract concept of how much observational time is necessary. In human infants, B Soc. R. Proc. ‘weight’ does remain unclear. Crows may have had a full the level of exploratory interactions required by infants to understanding of weight like humans; alternatively, they learn about object properties and relation differs across tasks may have possessed only a partial understanding of weight. [50]. For example, 11-month-olds need to explore task materials For example, the crows may have evaluated the objects in before they will demonstrate that they can infer an object’s terms of how easily they could be moved, both by the wind weight using compression information [8]. Crows may make 286 and by themselves (i.e. something akin to displaceability), rapid judgements about object weight from only brief obser- 20182332 : but may not have evaluated the objects as being ‘heavy’ or vations of the objects’ movements, or learn from single ‘light’ in terms of other physical interactions, such as whether events, such as the sound it makes when dropped. Future they would also sink or float. Another possibility is that their experiments are required to test this. understanding of weight may have been based on an ability Our findings are particularly interesting as they provide the to perceive the affordances of objects by observing the wind’s first indication that a bird species has the capacity to make infer- effect on them [45]. Testing to see how closely the crows’ con- ences about weight. This finding, though testing only a single cept of weight across different contexts mirrors that of type of inference, contrasts with Povinelli’s claim, based on humans will be a focus of future work. the failures exhibited by chimpanzees, that only humans have Across all five choice trials with each novel object pair, a causal understanding of weight [9]. Here, New Caledonian birds interacted with the incorrect object more in the control crows demonstrated first trial successes, followed by continued than the experimental condition. However, despite the differ- above chance performance, in a novel situation which provided ence in the number of interactions with the incorrect object only indirect information about the objects’ weights. Thus, across the two conditions overall birds made relatively few these birds appear to have general prior expectations of how errors after their first touch in both conditions (figure 2c). differently weighted objects should behave when observed in This occurred even though there was no penalty to the bird a novel context. for interacting with both objects during test trials, and birds These results stand in contrast to recent findings were not required to drop the correct object first to obtain suggesting that New Caledonian cannot use observations of the reward. When they did touch the incorrect object first physical interactions ‘accidentally’ resulting from their own on their first trial, birds successfully switched to dropping actions to perform causal interventions [51]. In the present the correct object into the tube for their first object drop in study, crows were able to pick-up causal information from over half of these trials, indicating that they could typically observing the wind interacting with an object. Among discard the incorrect object without receiving direct feedback humans, the extraction of object features, like weight, from on whether this particular object was rewarded. This did not observing another person’s actions appears to involve happen on every trial, possibly due to limits on the birds’ implicit, rather than explicit, reasoning and involves the obser- inhibitory control [34]. However, overall, this pattern of ver’s motor system. That is, the observer’s own prior behaviour demonstrates that these birds were highly capable experiences of handling objects can inform their perception of retaining and generalizing the rule that only light or only of objects, when they are seen to be handled by another indi- heavy objects were rewarded, and needed minimal handling vidual [52,53]. If a similar cognitive process underpinned the experience to learn and remember which of the novel objects crows performance here, it could explain this discrepancy: was light or heavy across the five trials of each novel object New Caledonian crows are capable of implicitly inferring test. The weight of small, manipulable objects may therefore causal information, but not explicitly using this to create be a particularly salient characteristic for this species to learn novel behaviours. Another possibility is that the crows failure about [46]. was a task artefact [51]. When confronted with an artificial A number of species select nuts and seeds on the basis of apparatus whose mechanism is not easily comprehensible, their weight before attempting to open them [13–16]. New the birds may revert to basic instrumental learning of action- Caledonian crows eat candlenuts and snails, which they drop outcome associations despite being capable of more sophisti- from heights on to hard surfaces to break open [47,48]. It has cated reasoning in more ecologically relevant contexts. While not been tested whether they select candlenuts on the basis the fan used in this experiment was also artificial, and both of their weight, though our observations indicate that this studies involved an ecologically relevant action (dropping)— would be possible. The weight of candlenuts and snails, similar though the goal of this action differed—attending to the pres- in size to our objects, may be an ecologically relevant feature, ence of wind would likely have been more ecologically salient and may account for the birds’ rapid learning, and inferential than attending to the inner workings of a food dispenser. Such abilities, when tested with these types of objects. an explanation requires further investigation. New Caledonian crows have also failed to demonstrate an world around them, without having to directly experience 7 immediate understanding of the relevance of weight in other those properties themselves. This form of learning through royalsocietypublishing.org/journal/rspb contexts; particularly, experiments involving artificial appa- observations has received minimal attention among animals ratuses [36,37]. When presented with a Perspex apparatus to date, but may reflect an important source of information. containing a collapsible baited platform, crows drop feathers Failures by chimpanzees on comparable tasks have led some [36], or other light objects [37], which were insufficiently researchers to argue that only humans are capable of represent- heavy to collapse the platform and obtain the reward. It ing weight as a causal mechanism [9]; however, our results should be noted that dropping the light objects did not pre- suggest this is not the case. Determining the weight of small vent the crows from obtaining the reward as they could still objects may be particularly relevant to New Caledonian drop heavy, rewarded objects afterwards. This performance crows, as they consume candlenuts, for which weight corre- is similar to the equally unsuccessful behaviours in a compar- lates with quality, and drop snails, which can be very heavy able task with chimpanzees [9]. In the light of our current relative to a crows’ body weight [48]. They also use tools, results, these failures may indicate limitations in the animals’ where the weight of a tool is likely to affect effort and outcome,

understanding of the apparatuses, rather than a generally which may lead crows to be more selective. Whether this ability B Soc. R. Proc. poor grasp of the concept of weight. Additionally, as the to infer the weight of objects is widely spread across the crows also sometimes dropped unrewarded (light or heavy) kingdom, or only a feature of animals with broadly high levels objects in the present study, this behaviour generally may of physical cognition, is presently unknown. reflect play or exploration, rather than an inability to infer the weight of objects based on their movement. Future 286 studies may address this by ending the trial after the subject Ethics. This study was conducted under approval from the University 20182332 : has made one choice. of Auckland ethics committee (reference: R602) and the Department To date, the ability to learn from observations has almost of Zoology ethics committee, University of Cambridge. We thank the Province Sud for permission to work in New Caledonia. exclusively been studied in a social context. Typically, subjects Data accessibility. The data are accessible in the electronic supplementary observe another individual’s actions or interactions with an material. object, and are then tested behaviourally to determine what Authors’ contributions. S.A.J. designed the study, conducted the exper- they have learned (e.g. [1,54]). However, New Caledonian iment and wrote the manuscript. All other authors contributed to crows have not performed well in cognitive tests conducted the study design, discussion of results and to writing the manuscript, in a social context, such as observing demonstrators. Like and as PI on the grant NSC coordinated the research programme. In many non-primates, without extensive training, these birds addition, M.S. contributed to running the experiment, and M.B. and R.M. conducted the statistical analysis. fail to choose the rewarded cup after observing a human visibly Competing Interests. We declare we have no competing interests. place a reward into one of three containers [55]. They have also Funding. The study was funded by the European Research Council failed to recognize the need for a partner in a cooperative task under the European Union’s Seventh Framework Programme [56], and demonstrate stimulus enhancement, but not imita- (FP7/2007-2013)/ERC Grant Agreement No. 3399933, awarded to tion, in a social learning task [57]. Our results are striking in NSC. Additionally, A.H.T. and M.S. received funding from a Royal that they suggest that New Caledonian crows are capable of Society of New Zealand Rutherford Discovery Fellowship and a gaining nuanced information about objects from observing Prime Ministers McDarmid Emerging Scientist prize, awarded to A.H.T. their physical interactions, despite their poor performance on Acknowledgements. We thank Jamie Diprose and Richard Hosking for tasks that involve observing other types of events. constructing the food dispensers, Anna Frohnwieser and Camille Overall, our results suggest that these birds are capable of Troisi for reliability coding, and Romana Gruber, Amalia Bastos making inferences about the properties of objects in the and Rebecca Hassall for catching and maintaining the birds.

References

1. Kaiser MK, Proffitt DR. 1984 The development of about affect, reason, risk, and rationality. Risk Anal. 10. Buckner C. 2012 In search of balance: a review of sensitivity to causally relevant dynamic information. 24, 311–322. (doi:10.1111/j.0272-4332.2004. Povinelli’s world without weight. Biol. Phil. 28, Child Dev. 55, 1614–1624. (doi:10.2307/1130030) 00433.x) 145–152. (doi:10.1007/s10539-012-9352-0) 2. Hamilton AF, Joyce DW, Flanagan JR, Frith CD, 6. Buckingham G, Goodale MA. 2010 Lifting without 11. Emery NJ, Clayton N. 2012 Do birds believe in Wolpert DM. 2007 Kinematic cues in perceptual seeing: the role of vision in perceiving and acting magic? In D Povinelli, World without weight: weight judgement and their origins in box lifting. upon the size weight illusion. PLoS ONE 5, e9709. perspectives on an alien mind. Oxford, UK: Oxford Psychol. Res. 71, 13–21. (doi:10.1007/s00426-005- (doi:10.1371/journal.pone.0009709) University Press. 0032-4) 7. Doerrfeld A, Sebanz N, Shiffrar M. 2012 Expecting to 12. Hanus D, Call J. 2008 Chimpanzees infer the 3. Runeson S, Frykholm G. 1981 Visual perception of lift a box together makes the load look lighter. location of a reward on the basis of the effect of its lifted weight. J. Exper. Psychol. Human Percept. Psychol. Res. 76, 467–475. (doi:10.1007/s00426- weight. Curr. Biol. 18, R370–R372. (doi:10.1016/j. Perform. 7, 733–740. (doi:10.1037/0096-1523.7. 011-0398-4) cub.2008.02.039) 4.733) 8. Hauf P, Paulus M, Baillargeon R. 2012 Infants use 13. Langen TA. 1999 How western scrub-jays 4. Reichelt AF, Ash AM, Baugh LA, Johansson RS, compression information to infer objects’ weights: ( californica) select a nut: effects of the Flanagan JR. 2013 Adaptation of lift forces in object examining cognition, exploration, and prospective action number of options, variation in nut size, and social manipulation through action observation. Exp. Brain in a preferential-reaching task. Child Dev. 83, competition among foragers. Anim. Cogn. 2, Res. 228, 221–234. (doi:10.1007/s00221-013-3554-9) 1978–1995. (doi:10.1111/j.1467-8624.2012.01824.x) 223–233. (doi:10.1007/s100710050043) 5. Slovic P, Finucane ML, Peters E, MacGregor DG. 2004 9. Povinelli D. 2012 World without weight: perspectives 14. Jablonski PG, Lee S-i, Fuszara E, Fuszara M, Jeong C, Risk as analysis and risk as feelings: some thoughts on an alien mind. Oxford, UK: Oxford University Press. Lee WY. 2015 Proximate mechanisms of detecting nut properties in a wild population of Mexican Jays glandarius). Anim. Cogn. 14, 441–455. (doi:10. 43. Whiten A. 2013 Humans are not alone in 8 (Aphelocoma ultramarina). J. Ornithol. 156, 1007/s10071-011-0379-4) computing how others see the world. Anim. Behav. royalsocietypublishing.org/journal/rspb 163–172. (doi:10.1007/s10336-015-1193-6) 30. Bird CD, Emery NJ. 2009 Rooks use stones to 86, 213–221. (doi:10.1016/j.anbehav.2013.04.021) 15. Heinrich B, Joerg CC, Madden SS, Sanders EW. 1997 raise the water level to reach a floating worm. 44. Whiten A, Byrne RW. 2010 Tactical deception in Black-capped chickadees and red-breasted Curr. Biol. 19, 1410–1414. (doi:10.1016/j.cub.2009. primates. Behav. Brain Sci. 11, 233–244. (doi:10. nuthatches ‘weigh’ sunflower seeds. Auk 114, 07.033) 1017/S0140525X00049682) 298–299. (doi:10.2307/4089173) 31. Jelbert SA, Taylor AH, Gray RD. 2015 Reasoning by 45. Gibson JJ. 1986 The ecological approach to visual 16. Visalberghi E, Neel C. 2003 Tufted capuchins (Cebus exclusion in New Caledonian crows (Corvus perception. Boston, MA: Houghton Mifflin. apella) use weight and sound to choose between moneduloides) cannot be explained by avoidance of 46. Taylor AH, Elliffe DM, Hunt GR, Emery NJ, Clayton full and empty nuts. Ecol. Psychol. 15, 215–228. empty containers. J. Comp. Psychol. 129, 283–290. NS, Gray RD. 2011 New Caledonian crows learn the (doi:10.1207/s15326969eco1503_2) (doi:10.1037/a0039313) functional properties of novel tool types. PLoS ONE 17. Spelke ES, Kinzler KD. 2007 Core knowledge. Dev. Sci. 32. Taylor AH, Miller R, Gray RD. 2012 New Caledonian 6, e26887. (doi:10.1371/journal.pone.0026887) 10, 89–96. (doi:10.1111/j.1467-7687.2007.00569.x) crows reason about hidden causal agents. Proc. Natl 47. Hunt GR, Sakuma F, Shibata Y. 2002 New 18. Vallortigara G. 2012 Core knowledge of object, Acad. Sci. USA 109, 16 389–16 391. (doi:10.1073/ Caledonian Crows drop candle-nuts onto rock from rc .Sc B Soc. R. Proc. number, and geometry: a comparative and neural pnas.1208724109) communally used forks on branches. Emu 102, approach. Cogn. Neuropsychol. 29, 213–236. 33. Logan CJ, Jelbert SA, Breen AJ, Gray RD, Taylor AH. 283–290. (doi:10.1071/Mu01037) (doi:10.1080/02643294.2012.654772) 2014 Modifications to the Aesop’s Fable paradigm 48. Tanaka KD, Okahisa Y, Sato NJ, Theuerkauf J, Ueda 19. Bird CD, Emery NJ. 2010 Rooks perceive support change New Caledonian crow performances. PLoS K. 2013 Gourmand New Caledonian crows munch

relations similar to six-month-old babies. Proc. Biol. ONE 9, e103049. (doi:10.1371/journal.pone. rare escargots by dropping. J. Ethol. 31, 341–344. 286

Sci. 277, 147–151. (doi:10.1098/rspb.2009.1456) 0103049) (doi:10.1007/s10164-013-0384-y) 20182332 : 20. Davidson G, Miller R, Loissel E, Cheke LG, Clayton 34. Miller R, Jelbert SA, Taylor AH, Cheke LG, Gray RD, 49. Martin-Ordas G, Atance CM, Call J. 2014 NS. 2017 The development of support intuitions Loissel E, Clayton NS. 2016 Performance in object- Remembering in tool-use tasks in children and and object causality in juvenile Eurasian jays choice Aesop’s fable tasks are influenced by object apes: the role of the information at encoding. ( glandarius). Sci. Rep. 7, 40062. (doi:10. biases in new Caledonian crows but not in human Memory 22, 129–144. (doi:10.1080/09658211. 1038/srep40062) children. PLoS ONE 11, e0168056. (doi:10.1371/ 2013.806553) 21. Hunt GR. 1996 Manufacture and use of hook-tools journal.pone.0168056) 50. Gibson EJ. 1988 Exploratory-behavior in the by New Caledonian crows. Nature 379, 249–251. 35. Lambert ML, Schiestl M, Schwing R, Taylor AH, development of perceiving, acting, and the (doi:10.1038/379249a0) Gajdon GK, Slocombe KE, AM. 2017 Function acquiring of knowledge. Annu. Rev. Psychol. 39, 22. Jelbert SA, Hosking RJ, Taylor AH, Gray RD. 2018 and flexibility of object exploration in kea and New 1–41. (doi:10.1146/annurev.ps.39.020188.000245) Mental template matching is a potential cultural Caledonian crows. R. Soc. open sci. 4, 170652. 51. Taylor AH, Cheke LG, Waismeyer A, Meltzoff AN, transmission mechanism for New Caledonian crow (doi:10.1098/rsos.170652) Miller R, Gopnik A, Clayton NS, Gray RD. 2014 Of tool manufacturing traditions. Sci. Rep. 8, 8956. 36. von Bayern AM, Heathcote RJ, Rutz C, Kacelnik A. babies and birds: complex tool behaviours are not (doi:10.1038/s41598-018-27405-1) 2009 The role of experience in problem solving and sufficient for the evolution of the ability to create a 23. Chappell J, Kacelnik A. 2002 Tool selectivity in a innovative tool use in crows. Curr. Biol. 19, novel causal intervention. Proc. Biol. Sci. 281, non-primate, the New Caledonian crow (Corvus 1965–1968. (doi:10.1016/j.cub.2009.10.037) 20140837. (doi:10.1098/rspb.2014.0837) moneduloides). Anim. Cogn. 5, 71–78. (doi:10. 37. Neilands PD, Jelbert SA, Breen AJ, Schiestl M, Taylor 52. Rizzolatti G, Sinigaglia C. 2010 The functional role 1007/s10071-002-0130-2) AH. 2016 How insightful is ‘insight’? New of the parieto-frontal mirror circuit: interpretations 24. St Clair JJ, Rutz C. 2013 New Caledonian crows Caledonian crows do not attend to object weight and misinterpretations. Nat. Rev. Neurosci. 11, attend to multiple functional properties of complex during spontaneous stone dropping. PLoS ONE 11, 264–274. (doi:10.1038/nrn2805) tools. Phil. Trans. R Soc. B 368, 20120415. (doi:10. e0167419. (doi:10.1371/journal.pone.0167419) 53. Sciutti A, Patane L, Nori F, Sandini G. 2014 1098/rstb.2012.0415). 38. Kenward B, Rutz C, Weir AAS, Chappell J, Kacelnik Understanding object weight from human and 25. Taylor AH, Hunt GR, Medina FS, Gray RD. 2009 Do A. 2004 Morphology and sexual dimorphism of the humanoid lifting actions. IEEE Trans. Auton. Ment. new caledonian crows solve physical problems New Caledonian Crow Corvus moneduloides, with Dev. 6, 80–92. (doi:10.1109/tamd.2014.2312399) through causal reasoning? Proc. R. Soc. B 276, notes on its behaviour and ecology. Ibis 146, 54. Manrique HM, Gross AN, Call J. 2010 Great apes 247–254. (doi:10.1098/rspb.2008.1107) 652–660. (doi:10.1111/j.1474-919x.2004.00299.x) select tools on the basis of their rigidity. J. Exp. 26. Seed AM, Tebbich S, Emery NJ, Clayton NS. 2006 39. Hunt GR. 2016 Social and spatial reintegration Psychol. Anim. Behav. Process 36, 409–422. Investigating physical cognition in rooks, Corvus success of New Caledonian Crows (Corvus (doi:10.1037/a0019296) frugilegus. Curr. Biol. 16, 697–701. (doi:10.1016/j. moneduloides) released after aviary confinement. 55. Jelbert SA, Taylor AH, Gray RD. 2016 Does absolute cub.2006.02.066) Wilson J. Ornithol. 128, 168–173. (doi:10.1676/ brain size really predict self-control? Hand-tracking 27. Jelbert SA, Taylor AH, Gray RD. 2015 Investigating 1559-4491-128.1.168) training improves performance on the A-not-B task. with the Aesop’s Fable paradigm: 40. Bates D, Ma¨chler M, Bolker B, Walker S. 2015 Fitting Biol. Lett. 12, 20150871. (doi:10.1098/rsbl.2015. current understanding and future directions. linear mixed-effects models using lme4. J. Stat. 0871) Commun. Integr. Biol. 8, e1035846. (doi:10.1080/ Softw. 67, i01. (doi:10.18637/jss.v067.i01) 56. Jelbert SA, Singh PJ, Gray RD, Taylor AH. 2015 New 19420889.2015.1035846) 41. R Core Team. 2014 R: a language and environment Caledonian crows rapidly solve a collaborative 28. Jelbert SA, Taylor AH, Cheke LG, Clayton NS, Gray for statistical computing. Vienna, Austria: R problem without cooperative cognition. PLoS ONE RD. 2014 Using the Aesop’s fable paradigm to Foundation for Statistical Computing. 10, e0133253. (doi:10.1371/journal.pone.0133253) investigate causal understanding of water 42. Forstmeier W, Schielzeth H. 2011 Cryptic multiple 57. Logan CJ, Breen AJ, Taylor AH, Gray RD, Hoppitt WJ. displacement by New Caledonian crows. PLoS ONE hypotheses testing in linear models: overestimated 2016 How New Caledonian crows solve novel 9, e92895. (doi:10.1371/journal.pone.0092895) effect sizes and the winner’s curse. Behav. Ecol. problems and what it means for 29. Cheke LG, Bird CD, Clayton NS. 2011 Tool-use and Sociobiol. 65, 47–55. (doi:10.1007/s00265-010- cumulative culture. Learn. Behav. 44, 18–28. instrumental learning in the Eurasian (Garrulus 1038-5) (doi:10.3758/s13420-015-0194-x)