he need never have made himself before she spoke. What she expects, rightly, is that her utterance will act as a prompt, making him recall parts of the book that he had previously forgotten, and construct the assumptions needed to understand the allusion. In both these examples Mary makes assumptions about what assumptions are, or will be, manifest to Peter. Peter trusts that the assumptions he spontaneously makes about the church and about Sense and Sensibility, which help him understand Mary’s utterances, are those she expected him to make. To communicate success-fully, Mary had to have some knowledge of Peter’s cognitive environment. As a result of their successful communication, their mutual cognitive environment is enlarged. Note that symmetrical co-ordination and mutual knowledge do not enter into the picture at all. The most fundamental reason for adopting the mutual-knowledge framework, as for adopting the code model, is the desire to show how successful communi-cation can be guaranteed, how there is some failsafe algorithm by which the hearer can reconstruct the speaker’s exact meaning. Within this framework the fact that communication often fails is explained in one of two ways: either the code mech-anism has been imperfectly implemented, or there has been some disruption due to ‘noise’. A noiseless, well-implemented code mechanism should guarantee per-fect communication. In rejecting the mutual-knowledge framework, we abandon the possibility of using a failsafe algorithm as a model of human communication. But since it is obvious that the communication process takes place at a risk, why assume that it is governed by a failsafe procedure? Moreover, if there is one conclusion to be drawn from work on artificial intelligence, it is that most cognitive processes are so complex that they must be modelled in terms of heuristics rather than failsafe algorithms. We assume, then, that communication is governed by a less-than-perfect heuristic. On this approach, failures in communication are to be expected: what is mysterious and requires explanation is not failure but success. As we have seen, the notion of mutual manifestness is not strong enough to salvage the code theory of communication. But then, this was never one of our aims. Instead of taking the code theory for granted and concluding that mutual knowledge must therefore exist, we prefer to look at what kind of assumptions people are actually in a position to make about each other’s assumptions, and then see what this implies for an account of communication. Sometimes, we have direct evidence about other people’s assumptions: for instance, when they tell us what they assume. More generally, because we mani-festly share cognitive environments with other people, we have direct evidence about what is manifest to them. When a cognitive environment we share with other people is mutual, we have evidence about what is mutually manifest to all of us. Note that this evidence can never be conclusive: the boundaries of cogni-tive environments cannot be precisely determined, if only because the threshold between very weakly manifest assumptions and inaccessible ones is unmarked. From assumptions about what is manifest to other people, and in particular about what is strongly manifest to them, we are in a position to derive further,

DOI link for he need never have made himself before she spoke. What she expects, rightly, is that her utterance will act as a prompt, making him recall parts of the book that he had previously forgotten, and construct the assumptions needed to understand the allusion. In both these examples Mary makes assumptions about what assumptions are, or will be, manifest to Peter. Peter trusts that the assumptions he spontaneously makes about the church and about Sense and Sensibility, which help him understand Mary’s utterances, are those she expected him to make. To communicate success-fully, Mary had to have some knowledge of Peter’s cognitive environment. As a result of their successful communication, their mutual cognitive environment is enlarged. Note that symmetrical co-ordination and mutual knowledge do not enter into the picture at all. The most fundamental reason for adopting the mutual-knowledge framework, as for adopting the code model, is the desire to show how successful communi-cation can be guaranteed, how there is some failsafe algorithm by which the hearer can reconstruct the speaker’s exact meaning. Within this framework the fact that communication often fails is explained in one of two ways: either the code mech-anism has been imperfectly implemented, or there has been some disruption due to ‘noise’. A noiseless, well-implemented code mechanism should guarantee per-fect communication. In rejecting the mutual-knowledge framework, we abandon the possibility of using a failsafe algorithm as a model of human communication. But since it is obvious that the communication process takes place at a risk, why assume that it is governed by a failsafe procedure? Moreover, if there is one conclusion to be drawn from work on artificial intelligence, it is that most cognitive processes are so complex that they must be modelled in terms of heuristics rather than failsafe algorithms. We assume, then, that communication is governed by a less-than-perfect heuristic. On this approach, failures in communication are to be expected: what is mysterious and requires explanation is not failure but success. As we have seen, the notion of mutual manifestness is not strong enough to salvage the code theory of communication. But then, this was never one of our aims. Instead of taking the code theory for granted and concluding that mutual knowledge must therefore exist, we prefer to look at what kind of assumptions people are actually in a position to make about each other’s assumptions, and then see what this implies for an account of communication. Sometimes, we have direct evidence about other people’s assumptions: for instance, when they tell us what they assume. More generally, because we mani-festly share cognitive environments with other people, we have direct evidence about what is manifest to them. When a cognitive environment we share with other people is mutual, we have evidence about what is mutually manifest to all of us. Note that this evidence can never be conclusive: the boundaries of cogni-tive environments cannot be precisely determined, if only because the threshold between very weakly manifest assumptions and inaccessible ones is unmarked. From assumptions about what is manifest to other people, and in particular about what is strongly manifest to them, we are in a position to derive further,

he need never have made himself before she spoke. What she expects, rightly, is that her utterance will act as a prompt, making him recall parts of the book that he had previously forgotten, and construct the assumptions needed to understand the allusion. In both these examples Mary makes assumptions about what assumptions are, or will be, manifest to Peter. Peter trusts that the assumptions he spontaneously makes about the church and about Sense and Sensibility, which help him understand Mary’s utterances, are those she expected him to make. To communicate success-fully, Mary had to have some knowledge of Peter’s cognitive environment. As a result of their successful communication, their mutual cognitive environment is enlarged. Note that symmetrical co-ordination and mutual knowledge do not enter into the picture at all. The most fundamental reason for adopting the mutual-knowledge framework, as for adopting the code model, is the desire to show how successful communi-cation can be guaranteed, how there is some failsafe algorithm by which the hearer can reconstruct the speaker’s exact meaning. Within this framework the fact that communication often fails is explained in one of two ways: either the code mech-anism has been imperfectly implemented, or there has been some disruption due to ‘noise’. A noiseless, well-implemented code mechanism should guarantee per-fect communication. In rejecting the mutual-knowledge framework, we abandon the possibility of using a failsafe algorithm as a model of human communication. But since it is obvious that the communication process takes place at a risk, why assume that it is governed by a failsafe procedure? Moreover, if there is one conclusion to be drawn from work on artificial intelligence, it is that most cognitive processes are so complex that they must be modelled in terms of heuristics rather than failsafe algorithms. We assume, then, that communication is governed by a less-than-perfect heuristic. On this approach, failures in communication are to be expected: what is mysterious and requires explanation is not failure but success. As we have seen, the notion of mutual manifestness is not strong enough to salvage the code theory of communication. But then, this was never one of our aims. Instead of taking the code theory for granted and concluding that mutual knowledge must therefore exist, we prefer to look at what kind of assumptions people are actually in a position to make about each other’s assumptions, and then see what this implies for an account of communication. Sometimes, we have direct evidence about other people’s assumptions: for instance, when they tell us what they assume. More generally, because we mani-festly share cognitive environments with other people, we have direct evidence about what is manifest to them. When a cognitive environment we share with other people is mutual, we have evidence about what is mutually manifest to all of us. Note that this evidence can never be conclusive: the boundaries of cogni-tive environments cannot be precisely determined, if only because the threshold between very weakly manifest assumptions and inaccessible ones is unmarked. From assumptions about what is manifest to other people, and in particular about what is strongly manifest to them, we are in a position to derive further,