Why argue? Towards a cost–benefit analysis of argumentation

1. Introduction

If you are in a romantic relationship, you should be aware that arguing with your partner is not always the brightest idea – not to mention arguing with your in-laws, your teacher, your boss, or the policeman who just gave you a speeding ticket. Our frequent reluctance to argue is not a form of lamentable timidity, but rather the expression of strategic concerns: we do not engage in argument when doing so is likely to have an overall negative outcome. Arguing is a demanding activity for a variety of reasons, one that we do not undertake without purpose and due consideration – aside from specific and comparatively rare circumstances, e.g. argumentative outbursts. In particular, we estimate the costs, benefits, and dangers of argumentation in order to decide whether it is worth arguing with someone or for something. The same should apply to computational systems: building and using an argument-based technology requires careful assessment of the expected outcome and estimated costs of doing so. And argument-based technologies for open systems are likely to be more effective if they are not ‘doomed to argue’, but rather allowed to opt for different interactive modalities in different contexts, arguing only when it is expedient doing so.

In spite of the key relevance of a cost–benefit analysis of argumentation, the otherwise prolific field of argument theories did not produce much in this vein, neither with respect to human argumentation nor for argument-based technologies. We try to redress this lack of consideration here. It is important to appreciate that our aim is not to ‘humanise’ argumentation technologies, but rather to make them smarter according to the objective standards of efficiency, given the purposes and context of application of such technologies. Looking at how people strategically decide when and how to argue is instructive in this respect, i.e. to improve the performance of the computational system and not just for the sake of building Artificial Intelligence (AI) arguers which better approximates the performance of natural arguers. On the other hand, a formal and computational model incorporating strategic considerations on costs and benefits will also provide insights into human argumentation, since it will be a model (incomplete, simplified, and normative) of human argumentative planning. Having said this, in the rest of the paper we will focus mostly on the costs, benefits, and dangers of argumentation for natural agents, and only towards the end we will show how these concerns are relevant for and should be incorporated in the design of argumentation technologies.

Even though the cost–benefit character of argumentation has by and large been understudied in the literature,1

Strategic considerations on how to argue are explicitly addressed also by Amgoud and Maudet (2002), focusing at the level of single moves within an ongoing debate: since at each step of the dialogue the agent may be faced with several alternative moves, selecting the best move represents a key strategic concern. Some form of cost–benefit analysis enters the picture here, but the considerations bearing on the agent's decision are quite different from those we will be exploring in this paper. To start with, Amgoud and Maudet assumes that the agent is already engaged in argumentation and has to decide how to best proceed from there, whereas we mainly address the issue of whether arguing makes sense in the first place. Second, these authors focus mostly on benefits, understood as the strategic potential of a given move to foster the agent's dialogical goals (given a certain context), whereas costs are largely ignored or at best reduced to the (lost) benefits associated with alternative moves that the agent did not pursue. This implies confining their analysis to ideal agents that do not consume resources to argue, whereas we are interested to explore what happens to arguers who have to operate under constraints of limited resources and bounded rationality, as it happens to be the case for both human and artificial agents (on the importance of studying argumentation under assumptions of bounded rationality, see also Gabbay and Woods 2003; Paglieri and Woods in press).2

Finally, considerations on the computational costs of argumentation-based technologies are abundant in the literature on algorithmic and complexity issues related to such technologies (just to mention few key contributions, see Dung 1995; Dimopoulos, Nebel, and Toni 2002; Dunne and Bench-Capon 2002; Cayrol, Doutre, and Mengin 2003; Dung, Mancarella, and Toni 2007; Dunne 2007). Indeed, this research sub-topic has recently emerged as a major trend in studies on computational argumentation (Bench-Capon and Dunne 2007). However, in this case, the preoccupation is with the tractability of a given argumentation framework, whereas our concerns here are on a completely different scale: even assuming that a given argumentation framework is tractable, the individual agent still has to make a strategic decision on whether arguing is rational or not, given current goals, available resources and relevant context. Obviously, the tractability of the underlying argumentation framework impacts on the strategic considerations of the agent. If the framework makes argumentation intractable or typically very costly for the agent, the likelihood of considering such practice worthwhile is either non-existent or comparatively low. But the crucial point is that, even when argumentation is tractable and its costs are in principle affordable, it does not immediately follow that arguing is the best option for the agent. It is this crucial decision problem that so far has been largely overlooked in argumentation theories, both within and outside AI: our aim now is to move some preliminary steps to explore the issue.

We begin in Section 2 by stigmatising two biases that have plagued argumentation theories so far, somehow preventing a cost–benefit analysis of arguing. Then we provide working definitions of the main concepts used in this article, that is, benefits, costs, and dangers and contextualise such notions within the general framework of expected utility theory (Section 3). The following three sections (Sections 4–6) are dedicated to explore each one of these factors, while their implications for argumentation technologies are assessed in Section 7. We conclude by summarising our main results and outlining future directions for research on argumentative decision-making (Section 8).

Argumentation theories often make two crucial assumptions: that a failure of argumentation leaves the arguers in the same situation from where they started, and that dialogical goals are the paramount concerns for the arguers in deciding whether and how to engage in argument. We believe both assumptions to be substantially mistaken and highly misleading, for the reasons highlighted in what follows. This is relevant for our current purposes, because subscribing to these assumptions is what has impeded so far appreciating the importance of a cost–benefit analysis of argumentation.

2.2. In favour of argumentative instrumentalism

Dialogical goals are often instrumental to extra-dialogical goals: the car dealer wants to persuade you of the superiority of a given model in order to sell you that model, not just for the sake of discussion. In the terminology of the theory of goals developed by Castelfranchi and colleagues (Parisi and Castelfranchi 1981; Conte and Castelfranchi 1995; Castelfranchi 1998; Castelfranchi and Paglieri 2007), dialogical goals are rarely terminal, i.e. they tend not to be ends in themselves. There are many ways of skinning a cat and arguing is just one of them. This implies a form of argumentative instrumentalism: both the overall decision to argue and more fine-grained choices of specific moves are ultimately affected by the extra-dialogical goals of the arguer. Argumentative instrumentalism has three facets: first, dialogical sub-goals are instrumental to some dialogical end (e.g. lauding the design of a given car is instrumental to persuade you of its superiority as a purchasing option), which in turn is instrumental to some extra-dialogical goal (e.g. selling you the car); second, the extra-dialogical goal has priority over the dialogical one, so if a given move is likely to foster the former but not the latter, the arguer should choose that move (e.g. observing that only few chosen people can afford to buy that model may be instrumental to selling you the car by appeasing your vanity, even if this point has no bearing on the alleged superiority of that particular model over others); third, a given move or set of moves may be instrumental to some other extra-dialogical goal of the arguer, different from the one which motivated to argue in the first place (e.g. the seller may try to impress an attractive female client with his elocution not as an effective mean to sell her the car, but rather hoping to get a date with the client). The prototypical goal structure of an argumentation plan is illustrated in Figure 1, distinguishing between dialogical goals (D-goals) and extra-dialogical ones (ED-goals).

OPEN IN VIEWER

Figure 1.

Dialogical and extra-dialogical goals of argumentation.

This is in fact a general structure in a cognitive view of planning. The same figure might well represent my plan of cooking a delicious dinner to seduce Mary: some of the internal choices I will make might not be for reaching my culinary goal (i.e. for the dinner to be splendid), but rather aiming directly at the overall extra-culinary goal that motivated me to cook in the first place, that is, making progress in my relationship with Mary. For instance, I may choose to serve a very strong red wine, even if it is not well suited for the food I cooked, in order to get Mary a bit drunk and thus more amenable to concede herself to me tonight. Moreover, some other choices may be motivated by additional but completely independent goals, like not spending too much money for my attempt at seducing Mary or respecting my own dietary restrictions while cooking for her, and so on.

So what is somehow surprising is not that such a scheme applies to argumentation, but rather that so far the import of this (very simple) fact has been largely downplayed in argumentation theories (a notable exception is Gilbert 1997). Possibly due to their logical roots, argumentation studies have tended to treat argumentation as if it was a self-referential activity, with no tie to concerns that extended beyond the boundaries of the dialogue – or, at least, no tie that was relevant to understand the argumentative process under analysis. Goals are frequently discussed in argumentation theories, but this interest is typically limited to dialogical goals, as it happens in the well-known taxonomy of dialogue types suggested by Walton and Krabbe (1995; see also Walton 1998) or in pragma-dialectics (van Eemeren and Grootendorst 2004). Even those who advocated a broader and greater attention to goals in argumentation theories, like Gilbert (1997), tended to consider only the question of ‘What is the goal of the argument?’, whereas a better interrogative would be: ‘What are the goals of the arguer, such that prompted him/her to argue and to use certain arguments rather than others?’ (see also Hample 2005). So what is needed is a goal-oriented (motivated) framework for argumentation, including not only a self-referential representation, but pointing also to other external or higher-level non-argumentative goals that contribute to the origin, the structure, and the outcome of the argumentative process.

We argue for something; and we decide to argue or not in view of something, motivated by some expected outcome. Only reference to these goals can explain not only why we argue (for what, with what motive), but in part also how we do so. In other words, accepting an instrumental view of argumentation is necessary to properly develop its cost–benefit analysis, since benefits and dangers of arguing can only be assessed with respect to the extra-dialogical goals that prompted us to engage in such activity.

The most straightforward (and straightforwardly computable) way of defining benefits, costs, and dangers is in terms of utility. Benefits refer to whatever goals (dialogical and/or extra-dialogical) of the agent the act of arguing accomplishes, including both the motivating goal of the argument and any additional positive side effects the agent might obtain. In other words, all the positive utility achieved by arguing can be lumped under the heading ‘benefits’. Negative utility, on the other hand, is to be spelled out in two different categories: costs and dangers.4

Some additional remarks are in order. First, all these considerations enter strategic decision-making through the filter of the agent's expectations, insofar as they need to be considered ex ante, before the act of arguing takes place. Thus, here we are talking about the expected utility of argumentation, given the context and the agent's goals and beliefs. Second, there is no reason to assume that arguers calculate this expected utility by simultaneously and comprehensively taking into consideration all factors that might affect it – indeed, decades of studies on decision-making suggest that people are likely to focus only on a limited sub-set of such factors, depending on their cognitive make-up and contextual features (Simon 1955, 1956; Kahneman, Slovic, and Tversky 1982; Rubinstein 1998; Gigerenzer and Selten 2001). Nevertheless, spelling out all the main factors that might (and, in some circumstances, should) affect the decision to argue is a worthy analytic effort, especially if the final aim is to provide suggestions on how artificial agents may intelligently manage the same decision. Third, and more generally, broadly construing the arguer's strategic considerations in terms of expected utility theory does not commit us to the standard version of the theory. On the contrary, further sophistications (e.g. risk aversion, framing effects, mental accounting, etc.) can and must be integrated in the analysis, but they do not alter the basic rationale behind it. With all the limitations and distortions that arguers might suffer, they still try to ponder benefits, costs, and dangers of arguing, in order to decide whether to engage in it.

So the crucial challenge is to identify what factors are indeed relevant for making such a decision: what goals are likely to matter and/or should matter for the agent at that juncture and what beliefs determine expectations on satisfaction/frustration of such goals through argumentation. In what follows, we try to offer a principled classification of such factors in everyday argumentation between human agents. The emerging taxonomy will later be confronted with the priorities faced by several types of argumentation technologies.

Insofar as the dialogical goal of argumentation is instrumental to some extra-dialogical goal, there are at least two probabilities that modulate the agent's expected utility on arguing: the likelihood that argumentation will achieve its dialogical goal, e.g. persuading a potential buyer that an item is worth buying (probability of dialogical success, p^D henceforth), and the likelihood that this achievement will produce the satisfaction of the final extra-dialogical goal, e.g. having the person buying the item because he thinks it is worth buying, regardless how this conviction was achieved (probability of extra-dialogical efficacy, p^E from now on). These probabilities may well differ, and they often do. Imagine a determined but careful buyer, that is, a person who is committed to purchasing a certain item but also willing to buy it only when satisfied by the offer of the seller, even if this means taking a long time finalising the purchase. An inexperienced seller faced with such a client may well be sure that, if he manages to present a convincing case, he will sell the item, and yet be very skeptical about his chances of ever persuading the client (that is, p^E>p^D in this case). Now consider a gullible but hesitant buyer, who is easily swayed by the opinion of other people, and yet rarely turns his fancies into concrete actions. An experienced salesman will see immediately that the client is very open to persuasion, but may still despair of ever making a sale with such a chronically undecided type (that is, here p^D>p^E).

These two values combine to determine a third probability, which is the one that really matters in the decision to argue: the likelihood that, by arguing, one will satisfy the final extra-dialogical goal (probability of extra-dialogical success, p^S henceforth). In anticipatory considerations of expected utility, p^S can be represented as the expected probability that the arguer's extra-dialogical goal will be achieved, assuming that argumentation is employed to that end. This clearly depends on how likely the arguer is to win the argument (p^D), and how likely is this victory to achieve the desired extra-dialogical end (p^E). How exactly to combine these two values depend on the assumptions made on their relationship: e.g. considering p^S as the conditional probability of p^E given p^D or treating p^D and p^E as referring to independent events and thus determining p^S as their product. The details on how p^S is (or should be) assessed are immaterial to our argument, since all we need to show is that both dialogical and extra-dialogical concerns concur to determine the likelihood of getting what we want through argument.

But again, what is it that we want? In other words, if p^S provides an approximation of likelihood of success, what is the value we assign to such an outcome and that we need in order to compute the expected utility function of arguing? This clearly can be answered only on a case-by-case basis, since the extra-dialogical goals that motivate our arguments and that engender the value of the dialogical goal of winning such arguments vary wildly: from frivolous disputes on trivial matters that we can gladly afford to lose or even dismiss out of hand to matters of life or death, as it often happens for arguments occurring in the legal and medical domains. The stakes of arguing change dramatically across these scenarios and so does its expected utility. Still, there are general constraints that need to be taken into account in all cases, and they do not concern probability alone.

A key issue is the presence and quality of alternatives, that is, means other than argument through which the agent's extra-dialogical goal may be achieved. When alternatives are present, they will have their own expected utility, and a rational subject should weigh against each other's argumentative strategies and non-argumentative options that are viable and pick whatever it is that maximises his expected utility function or any other general mechanism used to ensure the agent's satisfaction. As we mentioned before, here we are not committed to any standard version of expected utility theory and have no problem in accommodating all the limitations and correctives that may seem convenient to adopt: e.g. boundedly rational arguers may be capable of considering only a limited subset of alternatives at any given time. The presence or absence of alternatives as well as their quality has a direct effect on the decision to argue, in particular on the assessment of expected benefits associated with argumentation. When the subject has no valid alternative to achieve his extra-dialogical goal other than arguing, renouncing to argue implies forsaking the extra-dialogical goal. In contrast, renouncing to argue in the presence of alternatives leave other options to the subject, such that he may still achieve his extra-dialogical goal after all. The obvious consequence is that in the former case, (no alternatives) the subject will think twice before forsaking argumentation as a viable means, lacking any other option. In other words, when the benefits of arguing are specific of such an activity (argument-specific extra-dialogical goals), they are more likely to motivate the agent's to argue.

Finally, the argumentative activity as a whole or specific argumentative moves made by the subject may well satisfy (or frustrate) other extra-dialogical goals of the agent, different from the one who would motivate arguing in the first place. As far as these additional side benefits (or collateral damages) are anticipated by the arguer, they might affect his decision to argue.5

To sum up: the probability of achieving the benefits that motivated the subject to argue (p^S) is assessed by combining the likelihoods of two distinct events, i.e. successfully carrying out one's argumentative efforts (p^D) and having the desired extra-dialogical effects as a consequence of that success (p^E). Given this probability, the subject's expected utility further depends, and crucially so, on the value of the extra-dialogical goal that would justify arguing, considering also alternative means to achieve that end, as well as alternative goals that arguing might foster or jeopardise. In Section 7, we will discuss the import of these considerations for computational models of argumentation. Now, it is time to move from the ‘pros’ column to the ‘cons’ list, in the balance sheet of the arguers: that is, we shall now explore the costs and dangers of arguing, as potential reasons to discourage such a course of action.

Arguing requires the subject to suffer some direct costs, in terms of time, breath, cognitive resources, social exposure, plus what economists call opportunity costs – all other things I could have done, if I was not stuck here arguing with you.6

Let us first focus on instances where costs are likely to increase while benefits remain stable over time, e.g. in persuasion and negotiation. Here prolonged argumentation is usually not in the agent's best interest, and this effect is cumulative: the more efforts I devote to convince you, the more I stand to lose if, at the end of the day, you are not convinced. This may be true even when also the benefits are likely to increase over time (e.g. certain instances of inquiry and of eristic confrontation), provided the rate of growth for the benefits is lower than that for costs: when this happens, the two curves are bound to intersect, and when this happens it indicates a drastic shift in the arguer's cost–benefit balance. In fact, at this point, argumentation has turned into a bad deal, even if the arguer is fully successful in it, because its costs have exceeded its benefits. What is worst, the longer the argumentation is further prolonged after this point, the more severe will be the loss suffered by the arguer – again, regardless whether he wins or loses. More generally, what we have in these cases is that the arguer's cost–benefit balance is deteriorating over time, even in the event of a final victory (which is the optimistic assumption we are making here). Prolonging the argument is never a good strategy under these circumstances, as it is graphically illustrated in Figure 2.

OPEN IN VIEWER

Figure 2.

Cost–benefit balance as a function of time.

So far, these considerations indicate that, in general and with very few exceptions, a golden rule of argumentation is ‘Keep it short!’ – not only for reasons of rhetorical efficacy, but also from a cost–benefit point of view. How does this affect the decision to start arguing in the first place? Insofar as the agent has expectations on how hard it will be to conclude successfully a given argumentative engagement, he is also likely to have a rough estimate of the time necessary to do so: together with an intuition of the cost and benefit functions, this is all the information needed to judge whether arguing is going to be either a bad or a good deal, from the standpoint of its costliness.8

Finally, time is relevant also in determining the perceived value of, respectively, costs and benefits, since the former are suffered immediately and with certainty by the arguer, whereas the latter, aside from being more or less uncertain, usually do not occur until the argument is over and often may be delayed even further (especially if we consider extra-dialogical benefits; see Section 4). It is a known fact that the utility of an outcome is discounted over delay, and we also know that temporal discounting in human agents is myopic – that is, we tend to overvalue events that are temporally proximal and undervalue those that are temporally remote (Laibson 1997; Ainslie 2001; Berns, Laibson, and Loewenstein 2007). This has two substantial implications on the effect of costs in the decision to argue: (i) the costs of arguing are likely to loom larger than benefits and dangers, since the latter are typically both delayed and uncertain, whereas the former are immediate and certain; (ii) in argumentative decisions, the agent's preferences may exhibit the same kind of dynamic inconsistency that plague intertemporal choices in general (Strotz 1956; Kirby and Herrnstein 1995), so that the arguer considers arguing the optimal strategy while the moment of doing so is still far in the future, but when the time of engaging in debate comes closer at hand, suddenly its costs appear disproportionally high in consideration of its delayed expected benefit, and thus the agent may balk at it.

To sum up, costs of arguing include both direct consumption of valuable resources (time, cognitive effort, physical energy, etc.) and loss of opportunities to achieve alternative goals relevant to the arguer. Both categories of costs are bound to increase over time, while the benefits of arguing in most cases will either remain stable or increase at a lower rate: this implies that the utility function of arguing is decreasing over time, so that argumentation, if prolonged for more than a certain duration, will become a liability, regardless of the arguer's success in achieving his goal. These facts are (or should be) factored in the decision to argue or not, depending crucially on the expected time horizon of the discussion. Time not only happens to multiply the costs of arguing, but also acts as a filter in assessing the utility of costs vs. benefits, since the former are typically immediate and certain, whereas the latter are delayed and uncertain. Due to hyperbolic temporal discounting, this may lead the subject to overestimate costs and downplay benefits of arguing, as well as generating some temporal inconsistencies in his argumentative plan.

Here we use the label ‘dangers’ to refer to the negative utility produced by a side effect or consequence of arguing, such that it jeopardises some goals of the agent. This is a technical and somehow restrictive use of the term, since in loose talk people consider as dangers also things that do not qualify under our definition: for instance, most notably, the ‘danger’ of losing the argument. However, this restrictive connotation is useful to distinguish dangers from costs, as discussed in Section 3. Moreover, we are especially interested in the role of dangers in the decision to argue: this means that here we are considering dangers that are deliberately accepted (or rejected) by the agent, not just unforeseen events that happen to create negative utility. As a consequence, these argumentative dangers entail a certain level of personal responsibility: e.g. if things go badly, the arguer may regret it or blame himself for making a poor decision, and these negative self-evaluations constitute themselves a liability of arguing.

Some dangers are specific of a given instance of argumentation, since they refer to specific goals of the agent that arguing might put in jeopardy: e.g. arguing with one's in-laws may involve the danger of straining their relationship with the arguer's spouse, and this is likely to be an unwanted consequence of the debate and a (defeasible) reason not to get involved in it. Clearly, specific dangers of arguing need to be treated on a case-by-case basis. In contrast, here we are interested to briefly discuss generic dangers of arguing: that is, types of risks that are frequently associated with argumentation and thus recur in the arguer's decision-making across different contexts and situations.

First of all, an argument may ‘backfire’, to use Cohen's (2005) apt terminology, in the sense of undermining the credibility of the conclusion it aimed to prove. When an argument fails to prove its conclusions, this should not be taken to indicate that the audience is left with the same views on the matter they had before being exposed to the argument. Quite often, what happens is that a failed argument is taken as evidence to the contrary: the very fact that an argument for p failed is easily interpreted as an argument for ¬p – which is clearly a highly unsatisfactory result for the arguer. Even more dramatically, once argumentative failure has occurred, the more effort the arguer put into the original argument, the more likely it is that the audience will consider this failure as good proof of the falsity of the intended claim. After all, if a determined and competent arguer, after putting so much effort into arguing for p, still failed to prove it, the most likely reason for that failure is that p is false, since there is no question as to the competence and motivation of the arguer. Like in a nightmarish subversion of all standard values, backfiring arguments may happen to retort against the arguer his very best efforts. This also implies that the dangers of arguing, in terms of negative effects on the credibility of the conclusion, increases as a function of duration, efforts, and quality of the argumentative process.9

A second relevant danger concerns the possibility that argumentation will be conducive of disagreement between the parties, rather than helping to solve it (Paglieri 2009). With very few exceptions, in argumentation theories it is taken for granted that agreement, partial or complete, is the desirable end-state of several types of dialogue (Walton and Krabbe 1995; Walton 1998), or even of argumentation in general (Gilbert 1997; Johnson 2000; van Eemeren and Grootendorst 2004). But argumentation, even when explicitly aimed at increasing the level of agreement between the parties, may fail spectacularly to do so. When this happens, it is not true (contra argumentative optimism; see Section 2.1) that people are left where they started: in many circumstances, their predicament will be worsened by having argued, in the sense of heightening their differences of opinion rather than solving them. Moreover, this happens because, and not in spite, of the nature of their argumentation: it is not due to a rational or procedural mistake made by one or both parties, but rather a natural consequence of their (correct) ways of arguing.

Four main factors contribute to make escalation of disagreement a potential danger of perfectly rational arguments (for further details on this point, see also Paglieri 2009). From an epistemological point of view, divergence of opinion per se is considered legitimate and possibly faultless, whereas keeping wrong beliefs in the face of (supposedly) valid arguments is regarded as a stubborn and irrational attitude: as a consequence, any disagreement that survives arguing, especially after long-winded discussion is likely to be less tolerated than any pre-argumentative disagreement between the parties. Hence arguing may turn a legitimate dissensus into an irrational refusal to listen to reason, from the standpoint of both parties, thus worsening the quality of their disagreement. Second, the decreasing utility of argument over time (see Section 5) implies that the longer we argue, the more I stand to lose – and nobody likes losing more rather than less, especially is the chances of ever convincing the counterpart seem to become smaller and smaller the longer we talk. These economical pressures contribute to make us increasingly short-tempered and impatient with our interlocutors in prolonged debate, and this in turn increases the likelihood of escalation of disagreement. Third, arguing leads to explore in further detail a matter over which we disagreed from the onset, and this may well multiply the potential issues of dispute between us: if we disagree on whether Obama or McCain should have been elected as President and start arguing to settle the matter, it is likely that we will discover that we disagree also on many other things, some of them even more important than the original point of discussion, e.g. abortion rights and their extent. Finally, taking at face value what is said is, in most social contexts, the ‘right’ thing to do, i.e. what social conventions demand of people: if someone doubts your word on a given matter, the skeptic has better to produce some reason to back up that challenge, otherwise he will be labelled as unfriendly, rude, blunt, uncouth, opinionated, etc. Conversely, arguing, either to persuade the opponent of some claim he does not yet endorse, or to refute the reasons provided by the proponent, challenges this default trust, and thus requires some justification, whereas the same is not true for lack of argument (e.g. directly assenting or keeping silent).10

The damages associated with argumentation, and thus its dangers, can also concern the emotional well-being of the arguers. Arguing is rarely a neutral activity (Gilbert 1997), and the confrontational setting that it often implies may put a significant strain on people's mood: the arguers will typically feel compelled to succeed, while at same time being exposed to the critical cross-examination of the counterpart. If winning an argument is likely to cost you a friendship, or even your own peace of mind, you should (and will) think twice before entering that particular debate. Indeed, considerations of emotional risks enter the decision-making process of arguers: this is why we are very reluctant to argue about topics that are sensitive for ourselves or the counterpart, or both. For instance, only the most foolhardy husband will engage his wife in a debate on what is best to do during childbirth. Since breaching certain topics can be painful, and since this fact is often known to both parties, arguing about such topics appears as a deliberate aggression, cruelly aimed at the counterpart's ‘soft spot’. This is likely to make the parties deaf to reason and turn even the most amicable suggestion into a fight. This is especially true with family and friends, and help explaining why some of the bitterest quarrels happen with people who are very dear to us. The feelings of friends and relatives are more easily hurt in a debate because venting open disagreement, as required by arguing, could be construed as a betrayal of legitimate social expectations and not because arguing in itself is especially hurtful for them. I was expecting your approval, and the fact that it is denied to me breaks this expectation, thus producing a certain amount of pain, disquiet, or discomfort. A stranger, lacking any specific expectation on whether you share or care for his views, is much less vulnerable to your argumentative attitude.

A further danger of arguing concerns drawing unwanted attention from the counterpart on topics that it would be in the arguer's best interest to keep hidden or out of focus. Argumentation works by raising certain issues and making explicit their connections with facts already proven, or at least provable: the thematic focus of the discussion shifts during the discussion and can be manipulated to try serving the best interest of each party – an operation known in pragma-dialectics as strategic manoeuvring (van Eemeren and Houtlosser 2002; van Eemeren 2009). The danger inherent in such process is that, if mismanaged, it can end up undermining the arguer's own case: this happens when the arguer's inept manoeuvring makes the opponent notice things that are either damaging for the conclusion the arguer wants to prove (e.g. fuelling the opponent's counterarguments) or counterproductive for some other extra-dialogical goals of the arguer or both. Just to mention a very disastrous example, imagine a discussion where Adam tries to persuade Bob to lend him some money; when Adam argues that ‘You should lend me the money, since you already did last year’, Bob retorts with ‘This reminds me – you never paid me back! Not only I will not lend you any new money, now I also demand immediate payment of your standing debt to me!’. Clearly, Adam should have considered more carefully all the dangers associated with invoking precedents in a discussion on debts.

Finally, argumentation is risky also inasmuch as a failure may well have negative effects on the arguer's reputation, not only as a skilled arguer but also as a rational thinker. Albeit reputation (Conte and Paolucci 2002) and trust (Castelfranchi and Falcone 2010) are important research topics in social psychology and distributed AI (for a review, see Sabater and Sierra 2005), they have rarely been focused in the context of argumentation theories – and even when they were, their role was mainly confined to specific argumentation schemes, such as argument from expert opinion or from testimony (Walton, Reed, and Macagno 2008). In contrast, reputation and trust have a much more pervasive impact on argumentation, since we frequently trust (or distrust) the arguments of a speaker mainly or even solely on the ground of his reputation. Moreover, a good reputation as an arguer entails benefits that go beyond argumentation: being right and being able to prove it is considered indicative of many positive qualities, such as lucidity of judgment, strength of character, clarity of reasoning, inventiveness, and fluency. On the contrary, argumentative failure is taken to show lack of these desirable features, and this is something we definitely do not want the general public to witness. Indeed, we would be happy to showcase our argumentative triumphs, if only there was no danger of making our dialectical debacles equally manifest. But there is such a danger, so caution suggests keeping our disagreements as private as possible, to avoid devastating effects on our reputation as rational agents (see also Paglieri 2009).

To sum up, aside from any specific risk that might be associated with a given piece of argument, arguing in general entails certain dangers, of which arguers are aware and that factor in their decision on whether and how to argue. These dangers include undermining one's own case, escalating the disagreement with the counterpart, worsening the emotional state of both parties, attracting the opponent's attention on unwanted aspects of the subject matter, and ruining the arguer's reputation due to failures in proving one's point. Not all these aspects need to matter for computational agents as they do for human arguers, yet many of them are absolutely crucial – as we shall discuss in the next section.

It is time to ponder to what extent consideration of benefits, costs, and dangers is crucial for technological applications of argumentation theories. Argumentation technologies constitute a vibrant field of research in AI, as witnessed by the recent inauguration of this journal, as well as by several scientific meetings (the workshop series ArgMAS, CMNA, Persuasive Technologies, and since 2006 the large-scale biennial conference COMMA), many other publications (aside from the proceedings of the conferences just mentioned, see also Reed and Norman 2004; Walton 2005; Rahwan and Simari 2009; and the special issue of Artificial Intelligence edited by Bench-Capon and Dunne (2007)), and some large-scale research projects (e.g. ASPIC – Argument Service Platform with Integrated Component, http://www.argumentation.org/ and the support COST action on Agreement Technologies, http://www.agreement-technologies.org/).

Broadly speaking, the relevance of argumentation theories for AI is threefold, spanning both programming issues, human–computer interaction (HCI), and computer-mediated communication (CMC). Concerning computer programs and protocols, especially in distributed systems, argumentation techniques are used to guarantee a good balance between flexibility and stability (Rahwan and Simari 2009): each software agent is endowed with more or less sophisticated argumentative skills to interact with other agents and solve specific problems, e.g. information retrieval, automated decision-making, negotiation, etc. As far as interfaces are concerned, argumentation theories provide inspiration on how to design dialogue systems that human users will perceive as natural and convincing, and that will be able to engage in conversations of non-trivial complexity, e.g. providing appropriate replies, asking pertinent questions, offering suggestions in light of the user's needs and tastes, etc. (Guerini, Stock, and Zancanaro 2007; Mazzotta, de Rosis, and Carofiglio 2007). As for CMC applications, argumentation theories offer insights on how to support online interactions between end-users, both in synchronous and asynchronous modalities and with different levels of tutoring, in order to favour an orderly and productive development of the debate (Sillince and Saeedi 1999; Andriessen, Baker, and Suthers 2003), and also help defining criteria for qualitative comparison between online discussion and face-to-face interaction (Weinberger and Fischer 2006; Joiner, Jones, and Doherty 2008).

As discussed above, the assumption ‘more argument, better results’ is mistaken as a general rule: in addition, it may also be very misleading, when it comes to argumentation technologies. Let us start from the application of argument theories to communities of artificial agents interacting with each other. These agents typically have no special concern for their personal emotional welfare, and they do not care for social image and for being in good terms with their fellows: indeed, they have no feelings that can be hurt, so there is nothing personal at stake in their interactions. Nonetheless, their decision to engage in argumentation should still be guided by (i) energetic constraints, so that they should not pursue argumentation if the expected benefits are lower than the likely costs and (ii) some measure of foresight on the extra-dialogical consequences of arguing, so as to weigh benefits and dangers of argumentation in deciding whether to engage in it. In contrast, building artificial arguers that are blind to such elementary precautions may result in computational inefficiency, either because the agents lose themselves in the subtleties of prolonged debate, thus slowing or sidetracking the performance of the system as a whole (something both programmers and users are keen to avoid) or because the agents fail to consider the risks of arguing and therefore produce a result which is below standard (e.g. obstinate agents that keep using poor arguments may well undermine the credibility of the claim they intend to support, due to backfiring effects). The right rule in this case is to incorporate strategic considerations in the agent's decision to either argue or let the matter rest, and these considerations should somehow approximate those discussed above.

A cost–benefit analysis of argumentation is even more important when it comes to design users’ interfaces based on argumentation technologies or software for supporting computer-mediated arguments. In HCI and CMC, it is paramount that the software interface is capable of eliciting agreement in its users, both as a baseline to ensure smooth interaction in general (users should not get mad at the program or at each other), and concerning specific functionalities of the system (users should be willing to accept or at least consider the system's suggestions). The former type of agreement is instrumental to the latter, especially in those cases where (i) the software is supposed to make decisions or suggestions in the best interest of the user and (ii) this may not always coincide with what the user would do, left to his own devices. Expert systems capable of supporting human decision-making constitute very promising applications in AI, but they need to be designed keeping in mind the users’ argumentative attitudes. In particular, users should not be subjected to questioning or correction unless this is strictly necessary – and here ‘necessary’ means not only that the user may be about to make a mistake, but also that such a mistake is relevant for the user's current goals. In these applications, it is of the utmost importance to include an ‘agreement strategic manager’ as part of the artificial system, to ensure that the user is not bothered all the time with puny arguments or fastidious suggestions, and he is instead best prepared to keep an open-minded attitude towards the system's indications when these are truly relevant. In short, a computer should not be designed to be quarrelsome or pedantic, lest it ends up like the Talking Cricket in The Adventures of Pinocchio, with an infuriated user throwing a mallet at it.

As we discussed, a cost–benefit analysis of argumentation involves a rich variety of considerations: which ones are to be significant for a given argumentation technology depend on the specifics of that technology, including its intended context of application. For instance, an argumentation-based protocol for negotiation among autonomous artificial agents needs to be concerned with the reputational effects of arguing only to the extent that it incorporates some kind of agent's reputation (for arguments on why it should, see Conte and Paolucci 2002); similarly, an expert system with the sole purpose of advising end-users on the manifest structure of their arguments need not fear to hurt their feelings – albeit, it would still need to have some inkling on the users’ goals, as to prioritise the advice given to them. However, even if different technologies have different needs with respect to the assessment of benefits, costs, and dangers of arguing, some general considerations may be distilled from the analysis of human argumentation, and offered to computer scientists as a sort of (preliminary and incomplete) checklist in designing future argumentation technologies:12

The take-home message of this paper is that arguing is not always the brightest solution for the agent's predicament, be it a human or an artificial arguer. In everyday discourse, arguers are all keenly aware of this fact, and thus they treat the decision to argue as a strategic problem. We suggest more attention should be paid in argumentation theories to this strategic dimension of arguing and that lessons learned from studying the decision-making process of human arguers will provide useful guidance for the development of more effective argumentation technologies. In this paper, we moved some preliminary steps in that direction, by providing the bare outline of the strategic considerations that enter the decision to argue and tentatively discussing their implications for argument-based computational systems. The main points we considered include the need to put every argumentative act in the broader context of the agent's practical reasoning (typically, we do not argue just for the sake of it, but rather to foster further ends, and it is with respect to such extra-dialogical goals that the arguer's decision-making is to be understood), the possibility that argumentation might produce disastrous results (in the precise sense of worsening the arguers’ initial predicament according to some standard of quality, e.g. credibility of the conclusion, level of agreement between the parties, etc.) and the threefold concerns that shape the arguer's decision – namely, expected benefits, likely costs, and foreseen dangers.

While the cost–benefit analysis of argumentation is without doubt still in its infancy, we hope to have shown that it has much potential to blossom into a rich, relevant, and highly interdisciplinary domain of inquiry. There is much virgin territory to be explored and charted, and we believe future studies should cover all levels of analysis. Taking the liberty of acting as pathfinders, here we suggest some promising ‘tracks’ that we see opening in front of argumentation scholars:

Theoretical track: Efforts should be made to develop the somehow haphazardly observations presented in this paper and elsewhere in the literature (see Section 1) into a coherent and comprehensive theory of argumentative decision-making. Moreover, the significance of argumentative decision-making should be discussed and clarified in the broader context of argumentation theories, that so far paid only lip service to this topic.

As mentioned, the road ahead is long, unexplored, and full of bifurcations. Nevertheless, the benefits of taking it far exceed its costs and dangers. Studying argumentation without considering the decision to argue is like reverse engineering a bicycle without understanding its purpose – a hopeless endeavour at its worst, a mere intellectual exercise at its best. Argumentation theories should have a higher aim, and they will need to keep the arguer's decisions in the loop to successfully hit the target.