25

reliably predicts feeding opportunities. It is possible that B would entertain doubts about A's gastronomic judgment, but it is unlikely that they would consider their companion to be behavingdeceptively . The implication of this logic is that a compelling case for deception will require an analysis not only of the immediate circumstances of signal production, but also of historical context (Smith 1977, 1981, 1991), that is, of the animal's prior experience of pairings between the environmental events that normally elicit a call and the other stimuli reliably present at the time of production.

A final point that arises from the Gyger and Marler (1988) study concerns the frequency of 'deceptive' signalling. It was claimed that fully 45% of food calls were deceptive. This estimate is based on a flawed statistical analysis (see above), and the actual value may well be somewhat lower. Nevertheless, there is a striking contrast with classical theoretical analyses of animal signalling which suggest that 'deceptive' usage should be rare (Dawkins and Krebs, 1978). It is often assumed that this is a general rule in animal communication, so that claims for deceptive signalling at an appreciable rate are inherently not credible. Many of the original analyses of deceptive signalling were, however, based upon game theory models of aggressive interactions in which the costs of deception (severe physical injury) are potentially very high (Maynard Smith and Price, 1973; Maynard Smith, 1982). The payoff matrices describing signalling behaviour such as food calling are almost certainly quite different. It is intriguing to note that in instrumental conditioning paradigms, animals will continue to work on very 'lean' schedules of reinforcement. Under these conditions, less than one response in a hundred might produce a delivery of food. This analogy suggests that, under more natural conditions, animals might be prepared to expend energy (e.g., by approaching a sender) even if the link between the sender's behaviour and feeding opportunities has been quite tenuous. There is a real need for systematic studies of signal reliability with the goal of establishing the range of values that animals naturally experience for the correlation between signal production and presence of the putative referent.

XII

ULTIMATE QUESTIONS

This will be a short section. I have so far been concerned almost exclusively with proximate questions, not because I regard them as more interesting or more important than evolutionary ones, but because this has been the principal focus of most studies conducted to date. Work on referential signalling has been comparative only in the weak sense (Wasserman, 1993) of evaluating the characteristics of animal signals to determine whether they have properties in common with language. This strategy has generated a number of important findings, some of which call into question traditional assumptions about the degree to which humans are unique. Dogmatic statements about the special attributes of language have perhaps been particularly tempting targets because they can successfully be attacked with a very modest data set; logically a single contradictory result will suffice. But there have been no studies of referential signalling that are comparative in the strong sense, providing a series of systematic analyses of closely-related species that would permit us to reconstruct the evolution of a trait, as has been done successfully in studies of sexually-selected signals (e.g., Ryan and Rand, 1993; Basolo, 1995a,b).

XIIa. Evolution Of Referential Signals
Some clues about the way in which comparative studies might proceed are provided by comparisons between the alarm call systems of ring-tailed and black-and-white ruffed lemurs. These two species have quite different behaviour and ecology. Ring-tailed lemurs are highly terrestrial and occupy open habitat (e.g., Jolly, 1966; Tattersall, 1982), while ruffed lemurs are arboreal and live in dense rainforest (Pereira et al., 1988). Ring-tailed lemurs produce highly- specific alarm calls, both during natural encounters with predators and during experimental simulations of such events. Playback experiments demonstrate that call type encodes sufficient