False alarms and information transmission in grouping animals

(a) Predator cue ambiguity

When predators are harder to identify, it is likely that false alarms will be more common (Hilton et al., 1999). The difficulty of identifying predator cues accurately is ultimately what causes false alarms. Identifying predators accurately will depend on three key factors: (i) the similarity between non-target background information and the target predator cues; (ii) background complexity; and (iii) factors which obscure predator cues.

(b) Cost of responding

False alarms are likely to have both an energetic cost of producing the alarm behaviours and a subsequent opportunity cost, where time is lost that could have been spent on other activities such as foraging or roosting/sleeping (Beauchamp & Ruxton, 2007). This creates a trade-off between the anti-predator benefits of responding to a perceived predator cue and the loss of energy and opportunities if the cue has been misidentified. When faced with ambiguous predator cues, an animal must identify the optimal trade-off between false positives and false negatives. If an animal responds too frequently it wastes valuable resources, but if it responds too infrequently it becomes more vulnerable to predation (see Section II.2.c).

Thus, where the costs of responding are high, the false alarm rate is predicted to be lower (Beauchamp, 2010), and where alarm behaviours are more costly to produce, an animal is likely to be less sensitive to ambiguous cues (Fig. 2), both leading to decreased false alarm rates. Beauchamp (2010) presented empirical evidence for this effect in flocks of staging semipalmated sandpipers where individuals can double their body mass over 10–14 days. False alarms were less common later in staging, when body mass was higher and thus the energetic costs associated with alarm flights had increased. As referred to above, tadpoles of red-eyed treefrogs sometimes hatch prematurely in response to vibrational predator cues. This behaviour is more costly earlier in development when hatchling mortality is higher. False alarms in response to non-predator vibrations were found to be more common later in development when the cost of premature hatching was lower (Warkentin et al., 2019).

The cost of responding varies both within and among species depending on the type of alarm behaviour and other factors such as body condition. For example, in birds, producing an alarm call is likely a less energetically costly behaviour than an escape flight. This would lead to the prediction that false alarm calls will be more common than false alarm flights, however we are unaware of any empirical tests of this prediction. Among different species, the energetic costs of movement can vary substantially, even between animals that are a similar body mass. For example, Winter & Von Helversen (1998) found that bats expend 20–25% less energy in flight than bird species of a similar body mass. Differences in energy expenditure are also present across locomotory modes: the energetic costs of swimming are substantially lower than flying per unit distance travelled. In flight, energy is required to generate lift, whereas swimming animals can use buoyancy to maintain their position in the water column with minimal energy expenditure (Schmidt-Nielsen, 1972). Thus, the costs of responding to alarms are likely to vary substantially among different behaviours and different species, with the consequences of false alarms also likely to differ.

(c) Vulnerability to attack

False alarm rate is also likely impacted by the likelihood of being captured if an animal ignores a predator-like cue that represents a real threat (i.e. a false negative response; Table 1). Higher predation rates will therefore lead to higher levels of false alarms. This is exemplified by the willow tit (see Section II.1) which have very high false alarm call rates (81%) and demonstrate a very low criteria for alarm (Fig. 2B). In another example, flocks of house sparrows (Passer domesticus) respond more quickly to social alarm cues (rapid departures) when in larger groups (Boujja-Miljour, Leighton & Beauchamp, 2017). Boujja-Miljour et al. (2017) suggested that their predators may preferentially target larger flocks leading to higher vulnerability in larger groups, higher sensitivity to social cues, and hence to more false alarms. High vulnerability to predation means that the costs of an alarm response (see Section II.2.b) are offset by the risks associated with not responding to a true predator attack. This idea is exemplified in error management theory, a model based on signal detection theory (Johnson et al., 2013). Error management theory models the trade-offs in different decision-making scenarios where the costs of false negatives and false positives are asymmetric. In the context of predator detection, false alarms may be costly in terms of wasted energy or lost opportunities, but false negatives could result in immediate death. Therefore, in scenarios where predator attacks are invariably successful, the trade-off will be heavily weighted towards the avoidance of false negatives. In this situation, error management theory shows that biasing towards the least costly error (here, false alarms) may be the best option (Johnson et al., 2013).

Several factors could affect an animal's vulnerability to attack, for example, joining a group may reduce an individual animal's vulnerability to attack through the dilution effect. If a group of five animals is attacked, the probability of each individual being targeted is 1/5, however, in a group of 20 the probability is just 1/20. There are exceptions to this rule: some animals in a group may be more likely to be predated than others (Ward & Webster, 2016), for example, non-vigilant individuals (Roberts, 1996; Turner & Pitcher, 1986), ‘odd’ individuals (Ohguchi, 1981; Rutz, 2012), or those at the edges of the group [Krause, 1994; Ioannou et al., 2019; but see Lambert, Herbert-Read & Ioannou (2021) for moving groups]. While, generally, in a larger group individual vulnerability to attack is likely to decrease (groups can be subject to higher attack rates than lone individuals due to increased conspicuousness; Turner & Pitcher, 1986). Any effect that decreases the probability of predation is thus predicted to reduce the false alarm rate, as the costs associated with a false positive response would increase relative to those of a false negative response.

(a) Access to social information

A key benefit to group living is access to social information. Social information provides individuals in a group with access to a broader range of information than any one individual could gain from direct experience. This can include information about predators or the distribution, quality, and quantity of resources (Pitcher et al., 1982; Webster & Laland, 2015; Ward & Webster, 2016). When reacting to a predator or engaging with a resource, animals produce cues that can alert other group members (mostly passively, but sometimes intentionally; Evans, 1982; Haftorn, 2000; Freeberg & Lucas, 2002), for example, rapid escape flights and ‘wing clapping’ in pigeons (Columba livia) alert conspecifics to danger (Davis, 1975) and striking at food resources in predatory fish alerts conspecifics to food (Webster & Laland, 2015; Webster et al., 2019). By responding to cues from conspecifics that have successfully located resources, a group can search over a greater area than would be possible for a single individual (Pitcher et al., 1982), and the greater the chance that resources will be found. Larger groups can locate resources more rapidly, as was shown empirically in fish shoals (Pitcher et al., 1982) and bats (Cvikel et al., 2015).

In addition to more rapid resource location, access to social information also allows larger groups to forage more efficiently due to an increase in feeding rate enabled by decreased individual vigilance: ‘the-many-eyes-effect’ (Lima, 1995). When two different activities are mutually exclusive, an animal must budget the time it dedicates to each behaviour. If remaining vigilant for predators is incompatible with feeding, the animal must alternate between scanning the environment for predators and foraging (Barbosa, 2002). However, when in a group, the burden of vigilance is shared across multiple members. As other group members are also vigilant for predators, each individual can allocate more time to feeding without decreasing the group's collective vigilance. In larger groups, at any particular moment there are more animals scanning for danger, therefore, it is more likely that one will identify an approaching predator, making larger groups more effective at detecting predators sooner and at a greater distance (Siegfried & Underhill, 1975; Lazarus, 1979; Treherne & Foster, 1981; Boland, 2003). Decreased individual vigilance in larger groups (the ‘group size effect’) has been reported widely in mammals and birds (Elgar & Catterall, 1981; Elgar, 1989; Lima & Dill, 1990; Beauchamp, 2008; Beauchamp et al., 2021).

(b) Information transmission and alarm signals

Animals can benefit from the ‘many eyes effect’ because they are able to recognise when a conspecific has detected a predator and this information is transferred through the group. Many animal species have intentional alarm signals (e.g. alarm calls) but in the absence of these, indirect cues such as cessation of feeding, adoption of scanning postures, orientation changes, or fleeing can be monitored to infer detection of a predator (Ward & Webster, 2016). For example, pigeons actively fleeing a predator as opposed to departing for other reasons use fewer preparatory motions before initiating flight and this is exploited by other group members to differentiate an alarm from a regular departure (Davis, 1975). By responding to cues from other group members, not all individuals need to observe a predator directly for a whole-group response to occur, as tested empirically in ocean skaters (Halobates robustus; Treherne & Foster, 1981). This information travels through groups from neighbour to neighbour in a wave of local communication.

(c) Local communication and self-organisation

Most animal groups do not require a top-down decision-making process but instead are self-organised using local interactions (Sumpter, 2005; Ward & Webster, 2016). In self-organised systems, each group member bases its behaviour on that of its close neighbours (e.g. speed, orientation, or discrete behaviours such as a startle response) and information is transferred locally from one neighbour to the next in waves across the group (Rosenthal et al., 2015). Thus, interactions of local individuals within the group lead to the global properties of the system in emergent patterns including the group's shape, speed, and direction (Reynolds, 1987; Giardina, 2008; Rosenthal et al., 2015). Simulation models of these systems have been created. Reynolds (1987) modelled a three-dimensional (3D) landscape with dots (‘Boids’) acting as model birds. Reynolds gave each Boid three simple rules to follow: (i) maintain a close distance to your neighbour; (ii) do not move too close to your neighbour; (iii) follow your neighbour's travel direction. The model resulted in a cohesive group structure that responded to objects in the environment and behaved remarkably similarly to bird murmurations. Ultimately, the model showed that these cohesive structures could be produced by simple rules concerning only an individual's local neighbourhood (Reynolds, 1987). The assumptions of this model were tested empirically by Giardina (2008) in large European starling (Sturnus vulgaris) murmurations. The study recorded and mapped the locations and movement patterns of individual starlings in murmurations, and the flock was then digitised using points that followed the exact movement of the individual birds. Modelling using these data provided evidence for topological response zones where the birds maintained their position by following rules for repulsion and attraction zones and orientating with respect to their neighbours (Fig. 3). Giardina's (2008) analysis suggested that individual starlings responded to their nearest six or seven neighbours.

Similar studies of local interactions have shown how startle behaviour in fish is spread rapidly and effectively through local communication. Radakov (1973) first quantified these waves of information transfer in shoals of silversides (Atherinomorus sp.) by filming startled groups and monitoring the rate of behavioural cascades. The fright response began with a small number of individuals turning, this behaviour was copied by their neighbours, and so on until there was a whole-group response. The speed of this informational wave was 15 ms⁻¹ which is faster than their own maximum swimming speed (1 ms⁻¹) and, crucially, faster than the swimming speed of their predator (Radakov, 1973). Herbert-Read et al. (2011) analysed local interactions and escape waves in Pacific blue-eyes (Pseudomugil signifer) startled with an artificial predator. They found that changes in the direction and speed of a small percentage of individual fish that detect the predator initiate the beginning of an escape wave. The ‘front’ of this behavioural wave is a tightly packed band of individuals which causes the other fish to turn and move in the same direction. This front passes through the group in a wave, forming a group escape response (Herbert-Read et al., 2011).

(d) Reliability of social information

Although group information can be advantageous for both locating resources and responding to predators, with each animal basing its behaviour on that of other group members, vulnerability to the propagation of poor information is introduced. Behavioural cascades based on false information can spread through a group, with a whole-group response being generated by just a few misinformed individuals (Giraldeau, Valone & Templeton, 2002). For example, some redshank whole-group alarm flights were generated from a few individuals in the group mistaking a non-raptorial corvid for a bird of prey (Cresswell et al., 2000). In some circumstances, even when social information conflicts with individual experience, individuals follow the behaviour of the majority of the group members. For example, sticklebacks follow their group to a foraging site even if they have previously experienced that area to be low in resources (Webster & Hart, 2006). For this reason, blind copying of group members with no individual validation of socially acquired information could be maladaptive. In some species, the risks of misinformation lead to avoidance of the use of social information entirely unless personal information gathering is not possible. For example, in European starlings social information about resource distribution was only used if personal information gathering was difficult or costly (Templeton & Giraldeau, 1996).

When considering the use of socially acquired information on predators (as opposed to individual detection), it is important to note that alarms may not be universally ‘true’ or ‘false’ since what constitutes a threat is not universal across all individuals. For example, juvenile willow tits are vulnerable to corvids, while adults are not (Haftorn, 2000), therefore, a juvenile willow tit alarm calling in response to a corvid is a true alarm, but from the perspective of an adult it is false. The threats facing one individual may be genuine, making the alarm true, but the same stimulus may be irrelevant to another individual, creating a false alarm from the perspective of the receiver. In degus (Octodon degus), adults are less likely to respond to alarm calls from juveniles, likely because predator identification by juveniles is less likely to be reliable, and they likely face different threats (Nakano et al., 2013). This difference in threat perception may apply particularly to animals that form mixed-species groups. Ruddy turnstones (Arenaria interpres) use alarm information from both conspecifics and other wader species such as oystercatchers (Haematopus ostralegus) (Metcalfe, 1984). When turnstones are associating with conspecifics, they rely on social predator information more often than when they associate with oystercatchers (Metcalfe, 1984). It is possible that this is, in part, due to these two different species having different perceptions of what constitutes a threat. Oystercatchers are three times larger in body mass and are attacked less often by smaller birds of prey compared to the smaller wading turnstones (Whitfield, 1985). Oystercatchers, therefore, may not produce alarms in response to the same predators as turnstones, resulting in less shared vigilance. Even between members of the same species, state effects likely influence how an animal perceives risk. For example, an animal that is injured or hungry may approach threat differently, with trials in 15-spined sticklebacks (Spinachia spinachia) showing that starving individuals take more risks when presented with predator cues than do satiated individuals (Croy & Hughes, 1991). Whether an animal is the producer or receiver of an alarm thus may alter the validity of that alarm, and the trade-offs associated with the use of socially derived predator information, and false alarms may be altered by this variation in the perception of threat.

(a) Altering alarm reactivity

In some animal species, it is known that individuals can alter their reactivity/sensitivity to conspecific alarm signals. In these species, reactivity to conspecific alarm cues is a malleable trait, with individuals reducing their responsiveness to signallers whose alarm calls have been previously unreliable. This has been tested empirically in Richardson's ground squirrels (Spermophilus richardsonii; Hare & Atkins, 2001), yellow-bellied marmots (Marmota flaviventris; Blumstein, Verneyre & Daniel, 2004), Western Australian magpies (Cracticus tibicen dorsalis; Silvestri, Morgan & Ridley, 2019), and vervet monkeys (Chlorocebus pygerythrus; Cheney & Seyfarth, 1988). In these studies, animals were more responsive to conspecific alarm calls from individual conspecifics whose calls were previously correlated with the presentation of a model predator (compared to individuals whose alarm calls were not). In house sparrows, where social alarm information is often unreliable (Table 2), individuals self-validate alarm signals from individual conspecifics by increasing reaction times, likely to scan for danger before copying alarm behaviours (Boujja-Miljour et al., 2017). Similarly, yellow-bellied marmots were more likely to respond to alarm calls when multiple calls were played concurrently, likely because the alarm is more likely to be accurate if it has been verified by multiple animals (Blumstein et al., 2004). In an alternative mechanism to adjusting sensitivity based on reliability, in mixed-species reef fish shoals, modelling suggested that fish dynamically alter their sensitivity to socially transmitted cues depending on group density to supress the transmission of false alarms (false alarms can be more common at higher densities; Fahimipour et al., 2022).

Optimal reactivity to a conspecific alarm should theoretically vary between different species depending on their specific social and ecological niche. An animal's optimal reactivity level to alarm will change according to the risks and benefits associated with responding. This could range from using a ‘better-safe-than-sorry’ principle where animals respond to most alarms (e.g. in willow tits; Haftorn, 2000), to contexts of low predator threat where the response to any alarms could be muted. Theoretically, the decision to respond to an alarm should be based on its likely reliability, the cost of responding, and the cost of a false negative (the risk of predation if a true alarm is ignored) (Fig. 4). If alarms are typically reliable and the cost of responding is low, an individual should respond to most alarms. However, if an alarm is likely to be false and there is a high cost of responding, it could be more beneficial to ignore the alarm. The gradient and position of the optimal response line should vary depending on the costs of ignoring a true alarm. If this cost is high, then an individual may need to respond to most alarms – the ‘better-safe-than-sorry-principle’. However, if the costs of ignoring an alarm are low, it may be beneficial to ignore most alarms, even if there is a reasonable likelihood of the alarm being true. Whether these dynamics play out in real animal systems could be investigated in trials that experimentally alter these factors (Fig. 4B).

(b) The quorum response

A second mechanism that may act to limit the spread of false information is the quorum response (Sumpter & Pratt, 2009). In a quorum response, the likelihood that a group member copies a behaviour shows a non-linear relationship with the number of other group members that are performing that behaviour (Sumpter & Pratt, 2009; Fig. 5). When a behaviour is performed by one group member (e.g. an escape flight), the behavioural response of the group is first suppressed to lower the likelihood of a whole-group response to a small number of misinformed individuals. As more group members perform the behaviour, it becomes more likely that they are responding to true information, therefore, the propensity of other members to copy the behaviour rises sharply after a certain threshold. Once this threshold is met, a quorum has been reached, and a whole-group response is initiated. The initial suppression of the group response thus impedes a cascade of poor information as several individuals must validate the behaviour before other group members begin to copy it (Sumpter & Pratt, 2009). The quorum response has been applied to a variety of contexts including foraging and group predator detection, with empirical data available for insects (Pratt, 2005; Seeley & Visscher, 2004), fish (Ward et al., 2008), birds (Cresswell et al., 2000; Collins & Sumpter, 2007), and mammals (Sueur, Deneubourg & Petit, 2011). Cresswell et al. (2000) found that foraging redshanks will only initiate an immediate escape response if a certain threshold of members departs simultaneously. Even in a real attack, when just a single redshank departed there was a longer delay before other group members responded, suggesting that they were first assessing the reason for departure before copying (Cresswell et al., 2000). By allowing time to assess the reason for single departures, redshanks can avoid missed feeding opportunities caused by false alarm flights. While quorum responses are effective in suppressing the impacts of poor information, they still cannot fully eliminate the risk of cascading false information (Laland & Williams, 1998; Dall et al., 2005).

(c) Wave attenuation

The propagation of behavioural cascades through animal groups can vary in terms of the magnitude and distance travelled by the cascade. Radakov (1973) described and measured this effect in fish, showing that some behavioural cascades propagate throughout groups of fish while others attenuate rapidly. The distance travelled by a behavioural cascade was correlated with the level of risk, for example, in higher risk scenarios such as a barracuda attack, fright responses propagated further through the group (Radakov, 1973). Further work by Rosenthal et al. (2015) showed that most individual startle behaviours do not cascade through the whole group but instead dampen rapidly, only affecting a portion of the group. Following these studies, many analyses have attempted to decipher the mechanisms behind why some waves propagate fully through a group while others attenuate. Sosna et al. (2019) investigated the underlying mechanisms by exposing golden shiners (Notemigonus crysoleucas) to the chemical alarm cue Schreckstoff. After exposure, the average nearest-neighbour distance decreased substantially, and the group clustered together. Golden shiners school using visual cues, and this change in structure significantly altered the number of neighbours in their visual field and subsequently how they responded to the behaviour of their nearest neighbours. Sosna et al. (2019) suggested that these changes in the physical group structure indicated a structural encoding of risk; in riskier environments the fish arrange themselves in a structure that allows behavioural fright cascades to propagate more easily through the group. This mechanism allows groups to alter their alarm reactivity depending on the perceived risk of their environment, lowering reactivity when alarm information is more likely to be false, and increasing reactivity when predators are more likely to be present (Sosna et al., 2019).

(a) Modelling the costs of false alarms

Despite the surprisingly high proportion of false alarms recorded in some animal species (Table 2), the consequences of false alarms are not well understood. It is presumed that false alarms are costly in terms of the energetic cost of producing alarm behaviours and subsequent loss of opportunities (Beauchamp & Ruxton, 2007), with this cost of false alarms dependent on the current activity of the animal. For example, foraging animals may lose valuable feeding time (Kahlert, 2006), roosting animals may lose sleep (Beauchamp, 2010), and courting animals may lose mating opportunities (Martín et al., 2003). However, very few empirical studies have quantified these costs. In greylag geese, it was estimated that each false alarm prevented feeding for 19 min, creating a relatively substantial reduction in total foraging time given that false alarms occurred multiple times per day (Kahlert, 2006). Beauchamp & Ruxton (2007) modelled the consequences of false alarms on behaviour using a genetic algorithm model to investigate how animals should adapt their behaviour to optimise vigilance. Their model showed that the increased likelihood of misclassifications that arises with increasing group size would lead to an increase in false alarm rates, and that this added a non-negligible cost to foraging in larger groups. To compensate for this additional cost, individual vigilance of group members should decrease as group size increases. They also found that it would be beneficial to rely less on collective detection in larger groups (Beauchamp & Ruxton, 2007).

Most traditional models of group vigilance assume that alarms are perfectly accurate (Pulliam, 1973; Pulliam, Pyke & Caraco, 1982; McNamara & Houston, 1992), however, it is now clear that this is not true. While many factors contribute to the costs of living in large groups, such as competition for resources, false alarms are likely an additional cost of group living that may significantly impact fitness trade-offs.

(b) The fire-drill hypothesis

The surprisingly high rate of false alarms observed across animals poses a question: why do false alarms occur so often, when they are likely to be costly? As explained in Section II.2.b, it has been argued that false alarms inevitably involve a trade-off between the benefits of socially acquired predator information and energetic/lost opportunity costs. However, a recent paper (Root-Bernstein, 2021) argued that false alarm flights in birds can be explained by the so-called ‘fire-drill hypothesis’. In this hypothesis, false alarm flights are not maladaptive because they allow birds to practice the mechanical actions of emergency evasive flight that may increase their chances of survival in a real predator attack. Since the body mass of a bird can fluctuate dramatically (e.g. Beauchamp, 2010), Root-Bernstein (2021) argues that false alarm flights allow birds to practice and adjust their take-off manoeuvres to compensate for differences in body mass over time. This could explain the seemingly contradictory high rates of false alarm flights observed in birds by providing a significant associated benefit. This new hypothesis provides a clear alternative to the classical explanation that false alarms are unavoidable when identifying predators using ambiguous cues (as described by signal detection theory; Fig. 2) and thus occur as an inevitable cost associated with the use of using social information. We agree with Root-Bernstein (2021) that there is a lack of theoretical modelling to support this hypothetical trade-off. However, the costs associated with false alarms are evidenced by the widespread use of mechanisms to impede the cascade of false information (see Section III.3). If false alarms are beneficial, it would be difficult to explain the evolution of these mechanisms. Root-Bernstein (2021, p. 31) also argued that false alarm flights cannot be maladaptive as evolution cannot select against conditional outcomes (‘the bird is alive given the real presence of a predator at time x’) but rather acts only on absolute outcomes (‘the bird is alive at time y’). However, this line of reasoning considers only a single flight, neglecting the costs of repeated false alarm flights over a longer term. With false alarms occurring multiple times over a given period, a bird that responds to all alarms (true or false) will survive predation attempts for longer than a bird that does not respond. However, a bird that filters out and responds only to true alarms will both avoid predation and expend less energy on unnecessary departures. The latter strategy would likely lead to a higher probability of survival. There will therefore be an inevitable trade-off between the responses to false alarms and the cost of these responses. These two conflicting arguments await detailed modelling before we can resolve their differences.