Edited by: Luigi Baciadonna, Queen Mary University of London, United Kingdom
Reviewed by: Elisabetta Versace, Queen Mary University of London, United Kingdom; Michael Griesser, University of Zurich, Switzerland; Zhongqiu Li, Nanjing University, China
This article was submitted to Comparative Psychology, a section of the journal Frontiers in Psychology
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Like many predatory species, humans have pronounced individual differences in their interactions with potential prey: some humans pose a lethal threat while others may provide valuable resources. Recognizing individual humans would thus allow prey species to maximize potential rewards while ensuring survival. Previous studies on corvids showed they can recognize and remember individual humans. For instance, wild American crows produced alarm calls toward specifically masked humans up to 2.7 years after those humans had caught and ringed them while wearing that mask. However, individual behavior of the crows or the impact of social features on their responses, was hardly examined. Here, we studied predator learning and social effects on responses, using a similar method, in captive common ravens (
Learning about new predators allows individuals to adapt existing anti-predator behavior to new threats. Many animal species are able to recognize conspecifics on an individual level (
Most birds use mobbing as an anti-predator behavior. Mobbing is a coordinated action of multiple individuals of a weaker species against one or more individuals belonging to a more powerful species (
Taken together, these studies provide experimental evidence of predator learning in corvids, specifically when using masked humans as novel predators. Training events like catching or presentation of dead conspecifics (for American crows), handling of the nests or playback of alarm calls (for jackdaws) were restricted to single events or periods lasting no more than 3 days. Yet in all cited studies, obvious differences in behavioral response to the different masks were documented, indicating quick learning capabilities. Because several of these studies have been conducted on wild populations, the control over individual exposure intensity was intrinsically limited (e.g., for crows), or the tests were restricted to short time periods only (e.g., for jackdaws). Hence, individual variation in birds’ anti-predator responses have hardly been investigated for consistency over time and different social settings.
The current study focuses on another member of the corvid family, the common raven (
Similar to the work on crows and jackdaws (e.g.,
We predicted that the ravens would quickly learn to discriminate between masks, leading to higher scolding intensities (i.e., longer duration of alarm calling) for the dangerous mask than for the neutral mask. Based on previous reports and own pilot observations, we also predicted substantial individual variation in alarm calling intensity, potentially explained by individual-specific features like sex, raising type, and kinship, and/or by social features like group composition and dominance. Based on previous findings in corvids, we hypothesized that ravens would continue discriminating between the masks over a long time period, possibly years, without reinforcement (i.e., without the pairing with a dead raven). Furthermore, we expected that individual variation in scolding would be consistent across experimental presentations, as long as the group composition remained stable.
This experiment was approved by the animal ethics and experimentation board of the University of Vienna under the license number 2018-011. The entire data collection was non-invasive.
Study subjects were 16 captive ravens (
List of individuals. For ease of identification, single-individual sib-groups were named after individuals.
Name | Initial group | Sex | Year hatched | Raising | Sib-group |
---|---|---|---|---|---|
Anton | A | Male | 2010 | Parent-raised | 3 |
Ellen | A | Female | 2010 | Parent-raised | 4 |
Heidi | A | Female | 2010 | Parent-raised | 3 |
Jakob | A | Male | 2010 | Parent-raised | 4 |
Jonas | A | Male | 2010 | Parent-raised | 2 |
Klara | A | Female | 2010 | Parent-raised | 4 |
Lena | A | Female | 2010 | Parent-raised | 1 |
Sophie | A | Female | 2010 | Parent-raised | 1 |
Astrid | B | Female | 2010 | Hand-raised | 2 |
Joey | B | Female | 2010 | Hand-raised | Joey |
Lellan | B | Female | 2011 | Hand-raised | Lellan |
Matte | B | Male | 2011 | Hand-raised | Matte |
Orm | B | Male | 2011 | Hand-raised | Orm |
Ray | B | Male | 2011 | Hand-raised | Ray |
Skadi | B | Female | 2011 | Parent-raised | 5 |
Thor | B | Male | 2011 | Parent-raised | 5 |
All birds were marked with colored rings for individual identification. Each aviary had smaller chambers attached that provided opportunity for shelter and visually isolated retreating opportunities, but remained closed during experiments. Multiple branches provided enrichment and perching opportunities. The ground substrate consisted of gravel, wood chips, and sand. The birds were fed twice a day with a diet of meat, grain products, fruits, and vegetables and had access to water
The experiment lasted from October 2011 to October 2015 and consisted of three phases. In the initial control phase (October 2011), human presenters wore standardized clothing (gray poncho, rubber boots, and gloves) and one of two masks (
Mask presenter in standardized clothing holding a dead raven. Clothing consists of black rubber boots, white rubber gloves, and an olive plastic poncho. On the right are the black-haired dangerous mask and the red-haired neutral mask.
Plan of the aviaries. A barn on the far right provides visual cover for the start of the presentation, marked with “S.” Numbered circles show the presentation locations per aviary. The leftmost aviary was only used after the training phase, when groups got split into pairs. Presentations were only carried out in front of occupied compartments.
In the following training phase (October 2011–November 2011), the black-haired (hereafter “dangerous”) mask was presented together with a dead raven. The dead raven was collected at our field site in the Alps close to the Konrad Lorenz Research Station; it was an adult wild bird killed by captive wolves at the Cumberland Wildpark and thus unfamiliar to our captive ravens at Haidlhof Research Station. The dead raven was shaken; its wings spread and then dropped and picked up at each location. This was an opportunity for the ravens to associate a potential outcome of predation with the “dangerous” mask. There were four trials where a dead conspecific was presented with the dangerous mask. In contrast, the presentation of the neutral mask was performed empty-handed, i.e., neither a dead raven nor any other object was carried by the person when dressed up with this mask. Between every training trial, there was one additional trial where both masks were presented without the dead raven to test for learning speed. Two trials occurred per week.
In the final test phase (November 2011–October 2015), the precision and persistence of these associations were tested by further presentations of both masks without the dead raven. Trials occurred twice per month until May 2012, once a month until November 2013, three times in 2014, and once in 2015.
Across the entire data collection period, both presentations per trial were carried out on the same day and by the same person. We used 17 different presenters for a total of 39 trials. We documented individual scolding durations using video recordings (Canon Legria HF S10, Canon Legria HF S30). Video analysis was performed on PC with the use of Solomon Coder (
Analysis was conducted in R (version 3.6.1;
During our data collection, the size and number of our groups changed and some additional compartments were included while others were empty and skipped. This resulted in different durations where the mask was in view of the subjects (mean = 223.0 s,
Prior to analysis, we
As response, we used proportion of time spent alarm calling (as described above). As test predictors, we included mask type (dangerous or neutral), sex (male or female), raising type (hand- or parent-raised), and kinship of subject (families indicated by numbers, individuals without siblings by names), and size and sex ratio of the group as fixed effects. As control predictors, we included further fixed effects for order of presentation (first or second presentation of the day), age of the subjects, and days since the last training presentation. As random intercept effects, we included individual and presenter. To reduce type 1 errors, we included theoretically identifiable random slopes (
We used the function overdisp.test (provided by Roger Mundry) which returned a dispersion parameter of 0.72 and therefore smaller than 1, confirming that the model is not overdispersed. Slight underdispersion potentially leads to conservative test results and is not generally considered problematic. Collinearity of test predictors was determined for a standard linear model lacking the random effects and appeared to be no issue (maximum VIF: 3.1;
We conducted a full-null model comparison (
To investigate potential influences of dominance on alarm calling behavior, we used a second model including calculated Elo ratings based on won vs. lost conflicts (
The model formula is similar to model 1, with the addition of a fixed effect for Elo ratings as the only test predictor. As random intercept effects, we again included individual and presenter. We included random slopes of Elo ratings in both individual and presenter but could no longer identify them for age and group size in presenter, so we removed them (this is explained by the reduced sample size covering a smaller number of presentations). Both the dispersion parameter (0.68) and the maximal VIF (3.7) were within acceptable limits. We conducted a full-null model comparison following the same procedure as for model 1 with the null model lacking a fixed effect for Elo ratings, but being otherwise identical to the full model.
Both previous models investigate effects on the overall scolding participation per predictor. To test if the distinction between the masks changed over time, i.e., persistence, we ran a third model using as response the proportion of scolding the bad mask minus proportion of scolding the neutral mask. We again linearly scaled the response between 0 and 1 and fitted a third beta model using the same approach as described above. As test predictors, we included time since training, sex, raising and kinship of the subject, and size and sex ratio of the group as fixed effects. As random intercept effects, we included individual and presenter with random slopes of time since training in individual, raising and sex in presenter, and group size and sex ratio in both. Sibling-group was originally included in presenter but was dropped due to convergence issues. Sample size spanned 361 observations.
There were no issues with overdispersion (dispersion parameter 0.80) or collinearity (maximum VIF 3.1). The null model used for model comparison included only the random intercept effects with the random slopes, but no fixed effects.
All but one raven (male Ray) participated in active scolding of a human wearing a mask in the test phase, even though neither mask was paired with a dead raven at that time any longer. However, individuals varied strongly in their overall scolding participation (whether or not they engaged in scolding;
Scolding participation in test phase. Bars show participation in scolding as the proportion of presentations in which the individual produced at least one alarm call. Whiskers show SEs. Across the entire test phase, one individual never participated (male Ray).
Scolding intensity in the test phase per mask. Individuals are ordered by participation. Black diamonds show means. Only one individual (female Klara) scolded the neutral mask more than the dangerous mask.
When plotting group averages of scolding response per mask type across time, visual inspection of the graph indicates learning and memory effects (
Group averages of scolding response per mask across all phases. The control phase consisted of four trials over 2 weeks, the training phase of four trials with and three trials without dead raven in alternating order over 4 weeks (trials where the dangerous mask was presented while carrying a dead raven are marked with vertical lines) and the test phase consisted of 28 trials over 4 years.
Violin plots of scolding duration as proportion per mask type (dangerous vs. neutral), raising type (hand-raised vs. parent-raised), and group sex ratio (more females than males vs. equal or more males than females). Horizontal lines within the violin plots show quantiles set at 0.25, 0.5, 0.75, and 0.95. Black diamonds show means.
Overall, our test predictors (mask type, sex, raising and kinship of subject and size and sex ratio of the group) had a significant impact on scolding response (full-null model comparison:
Output from Model 1 on long-term memory.
Fixed effects | Estimate |
|
|
|
|
---|---|---|---|---|---|
(Intercept) | −4.67 | 0.98 | −4.76 | <0.001 | *** |
Dangerous mask | 0.43 | 0.18 | 2.40 | 0.017 | * |
Order 2nd | −0.26 | 0.17 | −1.53 | 0.126 | |
Sex male | −0.08 | 0.16 | −0.48 | 0.631 | |
Age | 1.82 | 1.73 | 1.05 | 0.293 | |
Sib-group 2 | 0.41 | 0.27 | 1.54 | 0.123 | |
Sib-group 3 | 1.72 | 0.21 | 8.17 | <0.001 | *** |
Sib-group 4 | −0.04 | 0.18 | −0.20 | 0.843 | |
Sib-group 5 | 2.86 | 1.93 | 1.48 | 0.139 | |
Sib-group Joey | 0.38 | 0.35 | 1.08 | 0.282 | |
Sib-group Lellan | 2.50 | 1.93 | 1.29 | 0.197 | |
Sib-group Matte | 2.54 | 1.95 | 1.30 | 0.193 | |
Sib-group Orm | 2.53 | 1.95 | 1.30 | 0.195 | |
Sib-group ray | 2.47 | 1.95 | 1.27 | 0.205 | |
Raising parent | 0.78 | 0.29 | 2.68 | 0.007 | ** |
Sex ratio | 3.24 | 0.73 | 4.44 | <0.001 | *** |
Group-size | −0.11 | 0.04 | −2.66 | 0.008 | ** |
Time since training | −1.66 | 1.48 | −1.13 | 0.260 |
General linear mixed model (GLMM) output showing fixed effects with response as proportion of scolding. Age and time since training were z-transformed, the rest dummy coded with the reference categories being neutral (for mask), first presentation (for order), 1 (for sib-group), and hand-raised (for raising). Higher sex ratios indicate more males.
Boxplots of scolding duration as proportion per mask and sib-group. Black diamonds show means. Sib groups indicate only kinship and individuals of the same family were not necessarily housed in the same aviary-compartment or raised together.
The full-null model comparison, with Elo ratings being the only test-predictor, was significant (
Output from Model 2 on dominance.
Fixed effects | Estimate |
|
|
|
|
---|---|---|---|---|---|
(Intercept) | −9.74 | 6.57 | −1.48 | 0.138 | |
Mask dangerous | 0.28 | 0.12 | 2.29 | 0.022 | * |
Order 2 | −0.15 | 0.10 | −1.42 | 0.156 | |
Sex male | −1.56 | 0.41 | −3.85 | <0.001 | *** |
Age | −3.04 | 1.80 | −1.68 | 0.092 | |
Sib-group 2 | 0.16 | 0.33 | 0.47 | 0.635 | |
Sib-group 3 | 1.56 | 0.29 | 5.37 | <0.001 | *** |
Sib-group 4 | −0.26 | 0.22 | −1.18 | 0.240 | |
Sib-group 5 | −3.86 | 3.41 | −1.13 | 0.257 | |
Sib-group Joey | 0.95 | 0.51 | 1.87 | 0.061 | |
Sib-group Lellan | −4.49 | 3.33 | −1.35 | 0.178 | |
Sib-group Matte | −3.57 | 3.37 | −1.06 | 0.289 | |
Sib-group Orm | −4.07 | 3.36 | −1.21 | 0.226 | |
Sib-group ray | −3.78 | 3.36 | −1.12 | 0.261 | |
Parent-raised | 1.42 | 0.99 | 1.44 | 0.150 | |
Sex ratio | −1.63 | 7.17 | −0.23 | 0.820 | |
Group-size | 0.92 | 0.50 | 1.86 | 0.063 | |
Time since training | 0.98 | 0.91 | 1.08 | 0.279 | |
Dominance | 2.41 | 0.51 | 4.75 | <0.001 | *** |
GLMM output showing fixed effects with response as proportion of scolding. Elo ratings for dominance were scaled from 0 to 1, age and time since training were z-transformed, the rest dummy coded with the reference categories being neutral (for mask), first presentation (for order), 1 (for sib-group), and hand-raised (for raising). Higher sex ratios indicate more males.
Violin plots showing scolding duration as proportion per mask type (dangerous vs. neutral) and dominance (top 50% of dominant individuals vs. bottom 50%). Horizontal lines within the violin plots show quantiles set at 0.25, 0.5, 0.75, and 0.95. Black diamonds show means.
Other than models 1 and 2, we now used as response the difference in scolding duration between the masks (dangerous minus neutral), rather than scolding duration in general. The combination of test predictors (time since training, sex, raising and kinship of the subject and size and sex ratio of the group) had a significant effect on mask-distinction (full-null model comparison:
Output from Model 3 on memory.
Fixed effects | Estimate | SE |
|
|
|
---|---|---|---|---|---|
(Intercept) | −0.85 | 0.37 | −2.29 | 0.022 | * |
Sex male | 0.06 | 0.12 | 0.51 | 0.609 | |
Sib-group 2 | 0.40 | 0.15 | 2.59 | 0.010 | ** |
Sib-group 3 | 0.63 | 0.12 | 5.43 | <0.001 | *** |
Sib-group 4 | −0.07 | 0.11 | −0.64 | 0.522 | |
Sib-group 5 | 0.08 | 0.12 | 0.65 | 0.515 | |
Sib-group Joey | 0.50 | 0.19 | 2.61 | 0.009 | ** |
Sib-group Lellan | 0.46 | 0.19 | 2.39 | 0.017 | * |
Sib-group Matte | 0.36 | 0.25 | 1.42 | 0.156 | |
Sib-group Orm | 0.40 | 0.26 | 1.57 | 0.115 | |
Sib-group ray | 0.39 | 0.26 | 1.53 | 0.125 | |
Raising parent | 0.57 | 0.16 | 3.52 | <0.001 | *** |
Sex ratio | 1.35 | 0.44 | 3.04 | 0.002 | ** |
Group-size | −0.03 | 0.03 | −0.87 | 0.387 | |
Time since training | 0.09 | 0.07 | 1.29 | 0.197 |
GLMM output showing fixed effects with response as difference in proportion of scolding per mask (dangerous – neutral). Age and time since training were z-transformed, the rest dummy coded with the reference categories being neutral (for mask), first presentation (for order), 1 (for sib-group), and hand-raised (for raising). Higher sex ratios indicate more males.
Captive ravens quickly learned to distinguish human experimenters wearing one of two masks, whereby the “dangerous” mask was initially paired with the presentation of a dead conspecific and the neutral mask was not. In subsequent tests without a dead raven, ravens scolded more toward humans wearing the dangerous mask than the neutral mask; furthermore, they continued to do so over a 4-year period without further experimental reinforcement. Despite having received the same amount and quality of exposure, individual birds differed strongly in how often and/or how long they participated in scolding the masked humans. This inter-individual variation was largely explained by social factors and fairly consistent across experimental presentations in socially stable situations. Later changes in the individuals’ scolding participation and/or intensity coincided with changes in group composition and pair formation.
Ravens quickly learned to distinguish between humans based on their facial features, which is in line with the results of previous studies on other corvids (
A noteworthy result of our study was the high individual variation in scolding participation, despite the equal and highly controlled exposure experienced by all birds. This variation could be explained by a mix of factors: Model 1 revealed effects of kinship, i.e., sibling groups participating either strongly or weakly in scolding (
Unlike
We also observed high-ranking individuals to aggressively challenge low-ranking individuals for producing intense scolding bouts (personal observation). However, because individuals tended to be close to the presenter while scolding, an alternative explanation would simply be redirected aggression toward the nearest subordinate group member (instances of re-direction have been observed in captive and free-ranging ravens, but not systematically studied). These dilution or suppressor effects could be responsible for the low scolding responses and failure to distinguish between the masks in some individuals, rather than a failure in learning to identify the masked human as potential threat. Disentangling these effects is not possible in our paradigm, but would be an interesting line of investigation for a follow-up study testing participating individuals in separation. If individuals distinguish between the masks in isolation, it would rule out a failure to learn, and support the presence of dilution or suppressor effects while in the group. By testing focus individuals in dyads with higher vs. lower ranking individuals, one could investigate dominance effects in more detail.
Finally, Model 1 also revealed an effect of rearing style, with parent-raised birds scolding the human presenters more readily and intensively. This is in accordance with the substantial literature on early life experiences, often showing long-term effects (
Scolding intensity (to either of the masks) was rather low during training, and at the beginning of testing, but increased throughout the testing phase (
The discrimination between masks was hardly affected by the time elapsed since training in the experiment, suggesting that (at least some) ravens remembered the putative predation events for 4 years. While the dangerous mask elicited longer scolding durations throughout the study, we did notice some generalization, and thus increased calling, toward the neutral mask toward the end of the study period. This has also been observed in related studies on other corvids (
While in all social constellations the dominant males of the groups took the lead in scolding, the majority of group members participated at low levels. The dominant males were accompanied in scolding by their siblings before they reached maturity (first 1–2 years of the study) and, after pair formation, by their female partners. Pair formation seemed to boost participation in scolding of (previously) subordinate females and males alike, which fits the finding that pair formation accompanies a rise in dominance status (
Literature on heterospecific individual recognition is relatively rare, with the exception of recognition of human faces, which has been shown in variety of species, ranging from mammals, birds, and reptiles to invertebrates like octopuses and honeybees (
Although ravens regularly exploit human resources (
All datasets presented in this study are included in the article/
The animal study was reviewed and approved by Animal Ethics and Experimentation Board Faculty of Life Sciences University of Vienna.
TB and CB designed the study. CB collected and analyzed the data and drafted the manuscript under the supervision of the other authors. TB and WF provided critical revisions to the manuscript. All authors approved the final version of the manuscript for submission.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We thank Barbara Haidn, Miriam Sima, Martina Schiestl, Sarah Deventer, Martina Stocker, Lisa-Claire Vanhooland, Katharina Wenig, Pauline Schaffner, Tanja Hampel, Kerstin Pölzl, Nadine Kauntz, Alexandra Feigl, Stephan Reber, Alexandru Munteanu, Christine Schwaab, Siegrid Balvin, Mark O’Hara, Lisann Heyse, Kathrin Weigerstorfer, Sebastian Dörrenberg, Stephan Reber, Theresa Matzinger, Andrius Pasukonis, and Arianne Veidt for their help with data collection and video coding. We thank Roger Mundry and József Arató for statistical advice and Delphine Soulet, Raffaela Lesch, Jim McGetrik, Magdalena van Buuren, and three reviewers for feedback on the manuscript.
The Supplementary Material for this article can be found online at: