For Mercier and Sperber (M&S), what appears as poor reasoning is actually appropriate argument–social dialogue facilitates reasoning by prompting agents to formulate arguments and defend them from objections. M&S propose a dual-process model with system 1 (S1) a consortium of inference mechanisms and system 2 (S2), an S1 apologist. We identify some features we think require clarification and provide alternative interpretations of phenomena used by M&S to support their model.
If S1 generates conclusions without revealing their derivation (modular-like), then where does S2 acquire the competence to support these arguments? What type of reasoning is required for it to construct these arguments, or does it run data back through S1 for a reasoned result? Related to this is the issue of argumentative contexts which trigger S2. These appear to be richer in information, creating a potential confound for the argumentative hypothesis: Is it the argumentative feature or the increased information that is critical?
The social psychology findings M&S adduce to support their view present a puzzle for it: How can truth win out amongst sophistical S2s committed not to discovering the facts but to defending S1's representation of them? Convergence-on-truth suggests there's more to S2 than defence of S1. One alternative views S2 as a dynamic, defeasible reasoner that sifts through S1 outputs, independently generating conclusions to be updated in the light of new information.
Presumably S1 must support probabilistic as well as deductive inferences. In which case, some regulatory role for S2 is inescapable. Suppose S1 has both deductive and probabilistic mechanisms and these produce compatible results with input X both deductively entailing and probabilistically supporting Y. Imagine new evidence E emerging that undermines Y so that X+E makes Y not probable. Nonetheless, E cannot affect the derivation of Y from X. So X+E still entails Y. Whence S2 has to decide whether to defend Y since it is derivable from X+E or surrender Y as X+E makes Y improbable. How would it make this decision?
Consider now M&S's views on confirmation bias. M&S deny confirmation bias is a flaw in reasoning. Yet if the aim of each agent's S2 is to persuade others, confirmation bias would just polarize views with no agent prepared to listen to another's arguments. Alternatively, if each S2 defends an agent's beliefs against objections, amassing evidence for those beliefs is important but anticipating likely objections and preparing a defence is no less so. Relative to aims of persuasion or defence, then, confirmation bias registers as a fault in reasoning.
Compare an M&S-styled S2-reasoner Aaron with a defeasible S2-reasoner Belle. Aaron is convinced the river mussels are good to eat since he's eaten them the past five days. Belle felt ill after eating them the day before. She advises Aaron to refrain. Aaron's S2 considers positive evidence and discounts negative evidence. So Aaron eats the mussels and falls ill. In contrast, Belle's S2 constructs fast generalizations on the fly. Having eaten them for four days, Belle inferred (G) the mussels are good to eat. But now her S2 enables Belle to adopt a position appropriate to the evolving evidence. The crucial difference between Aaron and Belle is this: Were they to swap roles, Belle would feel no internal pressure from her S2 to eat the mussels (unlike Aaron from his): Evidence someone else fell ill can prompt a defeasible reasoner to update (G) as disconfirming and confirming evidence are weighted equally. Whilst M&S's model allows S1 to update information, reasoning to a new conclusion (belief revision) appears anomalous.
Does the argumentative hypothesis yield the best explanation of reasoning performance? Take the Wason selection task. M&S claim that when agents are asked to assess the truth of (W) If there's a vowel on one side of a card, there's an even number on its other side for an E, K, 4, 7 array, their S1 matches cards to verbal cues, prompting them to choose the E card and 4 card. Their S2 then defends this choice against objections.
This matching hypothesis belies the semantic complexity of (W), which contains an indefinite “a card” nested within the scope of a conditional. Such constructions can vary in meaning. “If Ralph has a credit card, he buys tickets with it” is true if Ralph has four different credit cards but uses only one to buy tickets. However “if Sue has a pet, she feeds it” is false if Sue feeds her goldfish, starving her three kittens: “A pet” means every pet where “a credit card” means some credit card. Indefinites such as “a card” in (W) could be assigned a default existential reading (some) by an S1 semantic module. If contextual clues or background information don't subsequently override it, this (mistaken) interpretation could become fixed, leading subjects to look for some card with an even number on one side and a vowel on the other, as the majority do. Alternatively, if the semantic module defaults to a generic reading of (W), since generics (like the mussels are good) tolerate exceptions, the 7 card's role becomes opaque, rendering it non-salient once more.
For defeasible reasoners used to generic generalisations, figuring out the minimum number of occluded elements that need to be revealed to determine whether (W) is true is no trivial task. They fail not because they aren't called on to defend their view, as M&S contend, but because they have little or no experience testing exceptionless generalisations.
Why then do they succeed in group settings? A rival problem-solving model of reasoning says this is due to two factors:
-
1. Group settings are informationally rich with alternative hypotheses articulated, reasons for them explained and discussed, and
-
2. Subjects' semantic modules are specifically designed to compute the logical implications of others' statements to determine what those statements mean (Crain & Khlentzos Reference Crain and Khlentzos2010).
Further, since this model assumes subjects share a common goal of finding the correct solution, it is no mystery why subjects should converge on that solution.
For Mercier and Sperber (M&S), what appears as poor reasoning is actually appropriate argument–social dialogue facilitates reasoning by prompting agents to formulate arguments and defend them from objections. M&S propose a dual-process model with system 1 (S1) a consortium of inference mechanisms and system 2 (S2), an S1 apologist. We identify some features we think require clarification and provide alternative interpretations of phenomena used by M&S to support their model.
If S1 generates conclusions without revealing their derivation (modular-like), then where does S2 acquire the competence to support these arguments? What type of reasoning is required for it to construct these arguments, or does it run data back through S1 for a reasoned result? Related to this is the issue of argumentative contexts which trigger S2. These appear to be richer in information, creating a potential confound for the argumentative hypothesis: Is it the argumentative feature or the increased information that is critical?
The social psychology findings M&S adduce to support their view present a puzzle for it: How can truth win out amongst sophistical S2s committed not to discovering the facts but to defending S1's representation of them? Convergence-on-truth suggests there's more to S2 than defence of S1. One alternative views S2 as a dynamic, defeasible reasoner that sifts through S1 outputs, independently generating conclusions to be updated in the light of new information.
Presumably S1 must support probabilistic as well as deductive inferences. In which case, some regulatory role for S2 is inescapable. Suppose S1 has both deductive and probabilistic mechanisms and these produce compatible results with input X both deductively entailing and probabilistically supporting Y. Imagine new evidence E emerging that undermines Y so that X+E makes Y not probable. Nonetheless, E cannot affect the derivation of Y from X. So X+E still entails Y. Whence S2 has to decide whether to defend Y since it is derivable from X+E or surrender Y as X+E makes Y improbable. How would it make this decision?
Consider now M&S's views on confirmation bias. M&S deny confirmation bias is a flaw in reasoning. Yet if the aim of each agent's S2 is to persuade others, confirmation bias would just polarize views with no agent prepared to listen to another's arguments. Alternatively, if each S2 defends an agent's beliefs against objections, amassing evidence for those beliefs is important but anticipating likely objections and preparing a defence is no less so. Relative to aims of persuasion or defence, then, confirmation bias registers as a fault in reasoning.
Compare an M&S-styled S2-reasoner Aaron with a defeasible S2-reasoner Belle. Aaron is convinced the river mussels are good to eat since he's eaten them the past five days. Belle felt ill after eating them the day before. She advises Aaron to refrain. Aaron's S2 considers positive evidence and discounts negative evidence. So Aaron eats the mussels and falls ill. In contrast, Belle's S2 constructs fast generalizations on the fly. Having eaten them for four days, Belle inferred (G) the mussels are good to eat. But now her S2 enables Belle to adopt a position appropriate to the evolving evidence. The crucial difference between Aaron and Belle is this: Were they to swap roles, Belle would feel no internal pressure from her S2 to eat the mussels (unlike Aaron from his): Evidence someone else fell ill can prompt a defeasible reasoner to update (G) as disconfirming and confirming evidence are weighted equally. Whilst M&S's model allows S1 to update information, reasoning to a new conclusion (belief revision) appears anomalous.
Does the argumentative hypothesis yield the best explanation of reasoning performance? Take the Wason selection task. M&S claim that when agents are asked to assess the truth of (W) If there's a vowel on one side of a card, there's an even number on its other side for an E, K, 4, 7 array, their S1 matches cards to verbal cues, prompting them to choose the E card and 4 card. Their S2 then defends this choice against objections.
This matching hypothesis belies the semantic complexity of (W), which contains an indefinite “a card” nested within the scope of a conditional. Such constructions can vary in meaning. “If Ralph has a credit card, he buys tickets with it” is true if Ralph has four different credit cards but uses only one to buy tickets. However “if Sue has a pet, she feeds it” is false if Sue feeds her goldfish, starving her three kittens: “A pet” means every pet where “a credit card” means some credit card. Indefinites such as “a card” in (W) could be assigned a default existential reading (some) by an S1 semantic module. If contextual clues or background information don't subsequently override it, this (mistaken) interpretation could become fixed, leading subjects to look for some card with an even number on one side and a vowel on the other, as the majority do. Alternatively, if the semantic module defaults to a generic reading of (W), since generics (like the mussels are good) tolerate exceptions, the 7 card's role becomes opaque, rendering it non-salient once more.
For defeasible reasoners used to generic generalisations, figuring out the minimum number of occluded elements that need to be revealed to determine whether (W) is true is no trivial task. They fail not because they aren't called on to defend their view, as M&S contend, but because they have little or no experience testing exceptionless generalisations.
Why then do they succeed in group settings? A rival problem-solving model of reasoning says this is due to two factors:
1. Group settings are informationally rich with alternative hypotheses articulated, reasons for them explained and discussed, and
2. Subjects' semantic modules are specifically designed to compute the logical implications of others' statements to determine what those statements mean (Crain & Khlentzos Reference Crain and Khlentzos2010).
Further, since this model assumes subjects share a common goal of finding the correct solution, it is no mystery why subjects should converge on that solution.