Levels of Intentionality

We discuss philosophical work on speech acts with two objectives. First, we consider what the computer language use of speech acts can learn from the extensive work by philosophers. Second, considering speech acts as we want computers to do them sheds light on the philosohical problems. The two aspects of the philosohical treatments are so interrelated that we discuss them together. Here are some remarks.

1. Philosophers treat speech acts as natural phenomena to be studied. However, they propose not to treat them from the point of view of anthropology or linguistics. Instead they study their essential characteristics in terms of what they accomplish. This makes their work more relevant to computer science and artificial intelligence than linguistic or anthropological work would be.
2. My view of speech acts is that they are necessary in the common sense informatic situation in which people interact with each other to achieve their goals. The most important features of this informatic situation are independent of the fact that we are humans. Martians or robots with independent knowledge and goals would also require speech acts, and many of these would have similar characteristics to human speech acts.
3. The point of this paper is that speech acts are valuable when we design computer systems to interact with humans and with each other.
4. However, only some of the characteristics that philosophers have ascribed to speech acts are valuable for our purposes. Which ones they are will depend on the purposes.
5. Besides the speech acts that are common in human society, it is convenient to invent others. Indeed human institutions often involve the invention of speech acts. An example is the airplane reservation discussed in this paper.
6. A particular kind of speech act is an entity in an approximate theory in the sense of (McCarthy 1979a). For this reason attempts at precise definitions, e.g. of an airplane reservation, are likely to be beside the point. Instead we will have nonmonotonic axioms (McCarthy 1986) that partially characterize them.
7. Regarding speech acts as events of execution of program statements may be useful for philosophers also.
8. Performatives that are not really speech acts because they don't result in external output are also useful. Our main example is the commitment. When a program makes a commitment, its correctness requires the fulfillment of the commitment.
9. A key question we share with philosophers is that of what must be true in order that a speech act of a given kind be successfully performed.
10. This paper plays some role in the controversy between John Searle (1984) and the artificial intelligence community about whether computers can really believe and know. As I understand Searle's position, a natural extension would say that computers can't really promise either. Our position is that Elephant 2000 programs can perform some kinds of speech acts as genuinely as do humans. The difference between what Elephant 2000 programs do and what some humans to in this respect will be similar to the differences among humans. We expect that the programming community will for a long time be interested in speech acts of a more limited sort than Searle has discussed. However, human speech acts when performed in certain institutional settings also have a limited character. For example, giving a reservation usually does not involve an opinion that it will benefit the person to whom it is given.

    It's not clear that a difference of opinion on this point has practical consequences for programming.

11. It will often be possible to regard a program not written in Elephant as though it were by regarding its inputs as questions and requests and its outputs as promises, etc. This is the Elephant analog of Newell's (1982) logic level or my (1979a) about ascribing mental qualities to machines.
12. Austin and Searle distinguish illocutionary from perlocutionary speech acts. An example is that ordering someone to do something is illocutionary, but getting him to do it is perlocutionary. The same sentence may serve as both, but the conditions for successful perlocutionary acts don't just involve what the speaker says; the involve its effect on the hearer. Both philosophers mention difficulties in making the distinction precise, but for the purposes of Elephant it's easy.

    The correctness conditions for an illocutionary act involve the state of the program and its inputs and outputs. The correctness conditions for a perlocutionary act depends also on events in the world. An airline reservation program may reasonably be specified in terms of the illocutionary acts it performs, i.e. by its inputs and outputs, whereas the correctness of an air traffic control program is essentially perlocutionary, because stating the full correctness of the latter involves stating that it prevents the airplanes from colliding.

    In this connection it may be worthwhile to go beyond philosophical usage and apply the term perlocutionary to inputs as well as outputs. Namely, perlocutionary conditions on the inputs state that they give facts about the world, e.g. the locations of the airplanes. The correctness of an air traffic control program depends on assumptions about the correctness of the inputs as well as on assumptions about the obedience of the pilots and the physics of airplane flight.