According to well established at least since Kant belief, the human practice, language and thought depend on an ability to follow rules. However, it is not clear what it means that we follow rules in practice, including using words and concepts. Commonsensical belief says that in order to follow a rule we need to be aware of the rule. However, if we agree on that, then we embark on a regress, for example, to be able to use rules, concepts and words we need another rules that would establish standards of correctness for the application of the first rules, the concepts and the words. To be able to apply the rules we require to follow further rules for the correct application of the rules. Therefore, following any rule would require following a further rule and so on. For instance, suppose the relevant rule for ‘green’ was formulated in the sentence: “Apply word ‘green’ when pointing at green objects”. Then, the words in that rule such as ‘object’, ‘apply’ etc. would require other rules for their correct application. Wilfrid Sellars (1954 p. 204) sketches the regress as follows:

Thesis. Learning to use a language (L) is learning to obey the rules of L.

But, a rule which enjoins the doing of an action (A) is a sentence in a language which contains an expression for A.

Hence, a rule which enjoins the using of a linguistic expression (E) is a sentence in a language which contains an expression for E—in other words a sentence in a metalanguage.

Consequently, learning to obey the rules for L presupposes the ability to use the metalanguage (ML) in which the rules for L are formulated.

So that learning to use a language (L) presupposes having learned to use a language (ML). And by the same token, having learned to use ML presupposes having learned to use a meta-metalanguage (MML) and so on.

But this is an impossible (a vicious) regress.

Therefore, the thesis is absurd and must be rejected.

To put it roughly, using a language L is obeying rules of L; to obey rules of L we need metarules of ML for the correct application of the rules of L; the metarules for the correct application of the rules need further meta-metarules of MML and so on. The purport of Sellars’s argument is the rejection of the thesis that learning to use language L is learning to obey the rules of L. One remark is necessary: the argument refers not only to language games but extends to the practice as such. Language games are an instance of practice in general.

The argument implies that, first, we can learn to participate in a rule-governed (language) practice without necessarily learn to obey rules of the (language) practice. Second, it suggests that we can partake in a in a rule-governed (language) practice without obeying the rules of the (language) practice. However, it does not mean that the (language) practice is not rule-governed. Just the opposite, if we want to speak sensibly about a practice, we have to speak about the practice in terms of rules of the practice. Without rules there is no practice at all. Therefore, it is not correct to say that a (language) practice does not establish what we must or must not or may or may not do. Partaking in a practice means that we have an operative knowledge what behavior counts as correct and incorrect behavior. If we do not know whether an application of the word ‘green’ when pointing at green objects is the correct application of the word ‘green’ or not, then we are not competent English speakers and we do not fully partake in the language practice. Nevertheless, if partaking in a rule-governed (language) practice does not depend on obeying rules of the practice, then it seems to follow that there is nothing we could recognize as a correct partaking in the practice.

In order to solve the difficulty and stop the regress it is tempting to distinguish, following Sellars, between learning to obey a rule and learning to conform to a rule. In Some Reflections on Language Games (1954 p. 205) Sellars writes:

[T]here is a simple and straightforward way of preserving the essential claim of the thesis [that learning a language is learning to obey its rules] while freeing it from the refutation. It consists in substituting the phrase ‘learning to conform to the rules… for ‘learning to obey the rules…’ where ‘conforming to a rule enjoining the doing of A in circumstances C’ is to be equated simply with ‘doing A when the circumstances are C’—regardless of how one comes to do it.

The basic idea is that we can think about rule-following in terms of one’s disposition or propensity to do what the rule R requires in the circumstances C. For example, small children can be said to follow the rule of the language grammar, so long as they have a disposition to speak in accordance with the grammar. However, conforming to a rule does not involve any conceptualizing of R and therefore it is contrasted with obedience to a rule that requires having the rule in mind and conscious intending to follow the rule. In short, rule-following is a matter of becoming conditioned to conform to a rule R but not yet to conscious obeying the rule R. As a consequence resulting behavior is rule-governed but it does not mean that it is rule-obeying.

The purport of the distinction is to include into rule-following considerations human activities for which there are rules and yet in which agents participate without being able to formulate the rules they conform to and consequently, where obeying the rules is not essential to participating in practice as such. The main advantage is that by the distinction between rule-conformity and rule-obedience we avoid the regress described above. If conforming to a rule is an instance of following a rule, then one can follow a rule without the necessity of having it in mind, where the rule-conformity means that one exhibits regularity in one’s actions, i.e. a one has an inclination to do the right things in the right circumstances. Therefore, one conforms to a rule so long as one acts in accordance with what the rule says to do, even if one’s intention is not to do so because of what the rule says.

However, after closer look this regress-stopping solution, at least at this stage, says too little. Conformity with a rule or a regularity of one’s actions does not mean that one is following a rule. Our actions exhibit a variety of regularities we are not even aware of. One’s action could be consistent with an infinite number of incompatible patterns. As a consequence, first, one is subject to infinite number of rules; second, if there is no single pattern that one’s performance exhibits, then we cannot determine which rule one is following. If that is so, then we cannot determine what is and is not an error. Additionally, if that is true, then we cannot say that one is following a rule at all, because in order to say that one is following a rule we need a possibility of breaking the rule.

Similar point is made by Wittgenstein (2001). Let’s take a sequence of numbers: 1, 2, 3. If we ask a student what number should follow the sequence, then the answer “4” is correct only if we apply a rule “add 1”. However, a student could equally apply a rule “give a next prime number”, and if that is so, then the correct answer is not “4” but “5”. It could go on indefinitely, because one can find an adequate rule for every answer. Of course the issue concerns not only mathematical sequences. The same we can say about phenomenal properties. If the subject s possesses a mental image of a linden then this image is in accord with the representation of a linden as well as Berlin regardless what all past experiences of the subject s. Therefore one has no criterion to decide whether one actually possesses a representation of a linden or Berlin. As a consequence, the regularity of the sequence does not decide what rule one follows and what rule one ought to follow. At the end we are left, as McDowell (1998 p. 242) famously stressed it, between Scylla and Charybdis.

[The] problem is to steer a course between Scylla and Charybdis. Scylla is the idea that understanding is always interpretation […]. We can avoid Scylla by stressing that, say, calling something ‘green’ can be like crying ‘Help!’ when one is drowning—simply how one has learned to react to this situation. But then we risk steering onto Charybdis, the picture of a basic level at which there are no norms.

The problem, which is also called regularism-problem, is captured by Brandom (1994, p. 28). Regularism presupposes that exhibiting a regularity can count as following a rule, but without some kind of a supplement there is nothing that counts as an irregularity, hence nothing counts as an error. Nothing helps the reference to dispositions instead of one’s past behaviors. Even if we constrain the infinite set of cases under consideration to finite set of one’s dispositions, the problem remains. According to dispositionalist view, if one is disposed to add 1 to each predecessor, then one is following the rule to add 1 to each predecessor. On this basis one can exclude other incompatible rules by saying that one does not have a disposition to act in other way, and so following a rule is doing what one is disposed to do and breaking the rule is doing not what one is disposed to do.

The main problem with the dispositionalist view, what was most famously stressed by Kripke (1982), is that the dispositions as such do not say what one ought to do. Exhibiting a disposition does not say what is correct and incorrect. It is possible to imagine that one intends to do A in C, even if it is not correct to do A in C. The dispositions do not settle what should be done and what not, and so there is no necessary connection between what one is disposed to do and what is correct to do.