Logical Akrasia

Abstract

The aim of this paper is threefold. Firstly, sections 1 and 2 introduce the novel concept logical akrasia by analogy to epistemic akrasia. If successful, the initial sections will draw attention to an interesting akratic phenomenon which has not received much attention in the literature on akrasia (although it has been discussed by logicians in different terms). Secondly, sections 3 and 4 present a dilemma related to logical akrasia. From a case involving the consistency of Peano Arithmetic and Gödel's Second Incompleteness Theorem, it's shown that either we must be agnostic about the consistency of Peano Arithmetic or akratic in our arithmetical theorizing. If successful, these sections will underscore the pertinence and persistence of akrasia in arithmetic (by appeal to Gödel's seminal work). Thirdly, section 5 concludes by suggesting a way of translating the Dilemma of Arithmetical Akrasia into a case of regular epistemic akrasia; and further how one might try to escape the dilemma when it's framed this way.

1. Prologue

The Greek word “akrasia” translates literally as “lack of self-control,” but has come to be used as a general term for a weakness of will, i.e., a disposition to act contrary to one's own considered judgment. It will come as no surprise that such inability to act as one thinks right has interested ethicists since antiquity.

More surprising (perhaps) is the vast amount of attention that the analogous phenomenon epistemic akrasia has received lately.¹ A driving force behind this interest is the appealing thought that epistemic rationality requires coherence between: (A) an agent's doxastic attitudes in general, and (B) their specific beliefs about what doxastic attitudes are rational.² To illustrate, consider the medical resident Anandi who correctly figures that dosage 〈p〉 is appropriate for her patient; and thus believes that 〈p〉. Suppose she then learns she's been drugged herself, and further that the effects of the relevant drug very often lead to cognitive errors that are hard to detect from the inside. As a result, suppose she believes that 〈my belief that p is irrational〉 but that she maintains her belief that 〈p〉 nonetheless. Other things being equal, Anandi's doxastic state should strike us as irrational because she believes against her own standards of rationality, or as recent epistemological parlance will have it; because she believes akratically.³

Examples like Anandi's drug case have led some epistemologists to argue for a general anti-akrasia constraint on epistemic rationality (Feldman 2005; Littlejohn 2018; Smithies 2012; Titelbaum 2015):

The Akratic Principle. No [epistemic] situation rationally permits any overall [doxastic] state containing both an attitude A and the belief that A is rationally forbidden in one's situation. (Titelbaum 2019: 227)^{4, 5}

This principle is taken to imply that you should either have the attitudes you believe you ought to have, or stop believing that you ought to have those attitudes. Hence, in the name of rationality, the Akratic Principle forbids you to have certain combinations of attitudes such as not believing that 〈p〉 while believing that 〈believing p is rationally required in one's situation〉; or having credence(p) = 0.9 while believing that 〈having credence(p) = 0.9 is rationally forbidden in one's situation〉.

We'll return to the Akratic Principle in due course (cf. section 3), but for now let's consider a widely discussed case from the literature on epistemic akrasia to further guide our intuitions. The case concerns a sleep-deprived detective, Sam, who possesses misleading higher-order evidence (i.e., misleading evidence about what his first-order evidence supports):

Sleepy Detective. Sam is a police detective, working to identify a jewel thief. He knows he has good evidence – out of the many suspects, it will strongly support one of them. Late one night, after hours of cracking codes and scrutinizing photographs and letters, he finally comes to the conclusion that the thief was Lucy. Sam is quite confident that his evidence points to Lucy's guilt, and he is quite confident that Lucy committed the crime. In fact, he has accommodated his evidence correctly, and his beliefs are justified. He calls his partner, Alex. “I've gone through all the evidence,” Sam says, “and it all points to one person! I've found the thief!” But Alex is unimpressed. She replies: “I can tell you've been up all night working on this. Nine times out of the last ten, your late-night reasoning has been quite sloppy. You're always very confident that you've found the culprit, but you're almost always wrong about what the evidence supports. So your evidence probably doesn't support Lucy in this case.” Though Sam hadn't attended to his track record before, he rationally trusts Alex and believes that she is right – that he is usually wrong about what the evidence supports on occasions similar to this one. (Horowitz 2014: 719)

Provided the information of sleepy detective – and the background assumption that respecting one's total evidence is an important standard of epistemic rationality – what is the rational response to Sam's predicament? In other words, what doxastic attitude should he hold with respect to the identity of the thief? And what should he believe about what his first-order evidence supports?⁶

The literature is divided into three main camps. According to steadfast views, Sam should simply stick to his guns. That is to say, he should keep both his high confidence that 〈p〉 (i.e., 〈Lucy is the thief〉) and his belief that this is what his first-order evidence supports (Kelly 2005; Titelbaum 2015). A reason in favor of this response is that Sam actually got things right to begin with. So even though the later testimony from his partner Alex is higher-order evidence suggesting that his assessment of the first-order evidence is unreliable due to sleep deprivation, this is in fact misleading on the particular occasion.

In contrast, conciliatory views hold that Sam should reduce confidence both with respect to proposition 〈p〉 and the higher-order proposition stating that 〈my first-order evidence supports p〉 (Christensen 2007; Feldman 2005).⁷ A reason in favor of this position is that from Sam's first-person perspective the higher-order evidence constituted by Alex's testimony seems undefeated. Since Sam rationally trusts Alex to be right about his unreliable track record in relevantly similar circumstances, he should reduce his confidence at both first- and higher-order level.⁸

Notice that although steadfast and conciliatory views disagree about the rational response to cases like sleepy detective, they agree that Sam's confidence in 〈p〉 shouldn't conflict with his belief about what the first-order evidence supports. That is, both camps accept that any such level-incoherence is epistemically irrational.

Level-splitting views dispute this. According to the level-splitter, it can sometimes be epistemically rational to have a high confidence that 〈p〉 while also believing the higher-order proposition 〈my first-order evidence doesn’t support p〉. Imagine, for instance, a long deductive proof written on a whiteboard, and suppose that Beth thinks through the proof and comes to rationally believe a series of claims from which she competently deduces their conjunction, 〈p〉.⁹ Assume (quite plausibly) that Beth comes to rationally believe 〈p〉 by these means. Yet Beth knows that people like her – in similar situations involving long deductions – often make inferential errors. So, it may well be highly probable on her higher-order evidence that she has made an inferential error in the current situation, which suggests that 〈her first-order evidence doesn't support p〉 after all (even though we can stipulate that her belief in the truth of 〈p〉 is in fact correct).¹⁰ To be sure, the intended interpretation here is that the knowledge Beth possesses about people's shortcomings in situations relevantly similar to hers should be taken as higher-order evidence against her first-order attitude toward 〈p〉, but according to level-splitting views, what goes on at higher-order level need not affect the rationality of Beth's first-order attitudes.¹¹ Thus – by level-splitting lights – this is a scenario where Beth can have a high confidence in 〈p〉 while also believing the higher-order proposition 〈my first-order evidence doesn’t support p〉 and be rational nonetheless.

As with Beth's logic case, a level-splitting response to sleepy detective would have it that Sam should remain highly confident that 〈Lucy is the thief〉 and simultaneously believe that this isn't supported by his first-order evidence. Epistemologists such as Williamson (2011, 2014), Lasonen-Aarnio (2014, 2020), Wedgwood (2012), and Weatherson (2010), have all favored level-splitting views although their reasons for doing so diverge.

2. Logical akrasia

While epistemic akrasia is interesting in its own right, it will not be our main concern. Our primary focus will be on an analogous phenomenon in (formal) logic. The remaining sections aim to connect the discussion of epistemic akrasia from mainstream epistemology with another existing discussion in the philosophy of logic, which concerns the use of classical logic to prove metatheoretic results (such as soundness and completeness) about a weaker, non-classical logic. It will also be suggested that some level of akrasia is unavoidable in arithmetic because of Gödel's Second Incompleteness Theorem.

As a rough starting point, we'll take logical akrasia to consist of a mismatch between the deductive strength of the background logic one uses to prove metatheoretic results and the logical theory one believes (officially), i.e., a form of incoherence in logical theorizing akin to what we saw in the case of epistemic akrasia.¹² So, in other words, logical akrasia will occur when one explicitly appeals to (or at least implicitly commits to) a logical principle which is not endorsed by one's own theory.^{13, 14} At this point, we won't distinguish between meta-logic and metatheory, but we'll discuss the potential importance of drawing this distinction in section 2.1.

Now, to provide a concrete example of logical akrasia, consider the following passage from Beall and Restall:

The Intuitionist. What we have presented is a straightforward account of Tarski's model theory for classical predicate logic, and a simple account of truth conditions in a possible worlds semantics. We have claimed that such accounts deliver classical logic. Is this indeed the case? It is commonly thought that this is the case, but in present company we may have reason to question this thought. Upon an inspection of the usual soundness and completeness proofs, we shall see that the full power of classical logic is required to complete the proof. To show, for example, that in every model A∨¬A is satisfied, we need to show, for each model ${\rm {\frak M}}$, that ${\rm {\frak M}} {\Vdash} A$ or ${\rm {\frak M}{\nVdash} A}$. But this is an instance of the excluded middle! An intuitionist (for example) who rejects the law of the excluded middle will not endorse this reasoning. What can we say about this? (Beall and Restall 2006: 39)

One thing we could say – to answer Beall and Restall's query – is that the intuitionist is in a state of logical akrasia. The case of the intuitionist logician who, when doing the metatheory of intuitionistic logic, finds themselves using classical (nonconstructive) principles, is an illustrative example of logical akrasia as it involves a clear incoherence of the kind we are interested in. In sum: this logician happens to presuppose logical principles, when producing metatheoretic proofs, that are not endorsed by their own (intuitionistic) standards, and as was the case with the epistemic counterpart, logical akrasia seems self-undermining and irrational; or, to put this point more vividly: the intuitionist searching for an acceptable metatheory using a classical background logic seems akin to fixing a leaky roof by accustoming oneself to a wet floor.

2.1. Defining logical akrasia

So far, so good! Let's now define logical akrasia in a more regimented fashion:

Logical Akrasia. Subject S, believing logical theory T, is in a state of logical akrasia if and only if S commits to a logical principle that S's logical theory T fails to endorse as valid.¹⁵

As this definition is too coarse-grained to capture all relevant cases, we further distinguish between:

Weak. S, believing logical theory T, is weakly akratic if S commits to a logical principle that S's logical theory T fails to endorse as valid;

Strong. S, believing logical theory T, is strongly akratic if S commits to a logical principle that S's logical theory T rejects as invalid.

Based on this weak/strong distinction, we get two non-equivalent versions of logical akrasia. To appreciate this, consider a case where an intuitionist commits to an instance of the excluded middle in a restricted situation, and suppose that their theory doesn't endorse this as valid. Then, the intuitionist can extend their theory at a later stage such that the instance of the excluded middle becomes endorsed as valid (just stipulate that the situation is decidable) – e.g., it happens to be a case concerning a quantifier-free sentence of arithmetic like 2 + 2 = 4. Before the extension they were being weakly akratic, but not afterwards.¹⁶

Yet one should acknowledge that the tenability of the strong/weak distinction depends on the possible division: meta-logic/metatheory. There is, for instance, no difference between failing to endorse as valid and rejecting as invalid if one holds a formal and complete meta-logic rather than a non-formal and incomplete metatheory.¹⁷ If one's meta-logic is formal and complete over its domain, then the distinction between strong and weak logical akrasia collapses (since in that case everything which is not valid is simply invalid). When, say, an intuitionist holds a formal and complete meta-logic and the law of the excluded middle has counterexamples, then failing to endorse the excluded middle as valid and rejecting it as invalid amount to exactly the same.

2.2. Logical akrasia and incoherence

As we have seen above, states of logical akrasia seem to be incoherent. A point which is also frequently underscored in the literature:

If you take “logically valid” to obey a logic weaker than classical, you shouldn't ultimately be satisfied with developing your theory of that logic using inferences that are merely classically valid… (Field 2017: 14)

…what a strange approach to take, if one believes logic X is the correct logic. Why use an alien logic for one's metatheory – and if one does, why trust the result? (Read 2006: 208)

…it would be untoward in a logic to appeal in proof of its adequacy to principles in which the logic in question does not believe. (Meyer 1985: 13)

If he rejects classical logic for the object language, how is he entitled to rely on it for the metalanguage? (Williamson 2020: 6)

These quotes notwithstanding logical akrasia are deeply entrenched in our contemporary logical theorizing. For it is no secret that classical logic serves as the golden standard in evaluations of non-classical logics (Schurz 2021), i.e., it's common practice to take classical (first-order) logic as the “neutral” backdrop against which we evaluate non-classical logics. Examples are: Łukasiewicz's three-valued logic, Kleene's (strong) three-valued logic, Brouwer's intuitionistic logic, Priest's paraconsistent logic, etc.¹⁸

Where does this leave us? Is the current modus operandi of non-classical theorizing severely misguided? Well, insofar as we want reflective equilibrium (Resnik 1985, 1996, 2004) between our logical theories and considered judgments about logicality (i.e., validity, consistency, implication, equivalence, etc.), there is a sense in which the answer is affirmative. According to Michael Resnik, one's preferred logical theory and considered judgments about logicality are in a state of reflective equilibrium when:

…the theory rejects no argument that one is determined to preserve and countenances no argument that one is determined to reject… (Resnik 1996: 493)

Thus, we have at least two interpretations of reflective equilibrium, depending on how we read the term “reject”: (1) one's theory judges valid every argument one is determined to preserve; and (2) one's theory doesn't judge invalid any argument one is determined to preserve. For weak logical akrasia to violate reflective equilibrium, we need interpretation (1); not interpretation (2) (cf. section 2.1); or, as one may otherwise put it, for weak logical akrasia to violate reflective equilibrium, we need the strong interpretation of reflective equilibrium, not the weak one. The ideal of strong reflective equilibrium is incompatible with states of weak logical akrasia. Hence, the golden standard of non-classical logicians – committing themselves to classical principles in their metatheoretical pursuits – appears to be a standard of fool's gold in at least one sense.¹⁹

2.3. The analogy with epistemic akrasia

If we return to the analogy between epistemic and logical akrasia for a minute, we can now appreciate how both the weak and strong version of logical akrasia resemble the standard definitions of epistemic akrasia in various ways.

Recall first the appealing thought from section 1 stating that epistemic rationality requires coherence between: (A) an agent's doxastic attitudes in general, and (B) their specific beliefs about what doxastic attitudes are rational. Epistemic akrasia occurs whenever S adopts doxastic attitudes that don't live up to S's own standards of rationality. Logical akrasia, similarly, occurs when S commits to logical principles that don't live up to S's own standards of logic. So, in both cases, the problem is one of not meeting one's own ideal (rather than pursuing a spurious ideal).

Adding further to the analogy, we have no trouble imagining what steadfast and conciliatory responses to logical akrasia would look like. If one asserts that logical akrasia calls for a revision of one's logical commitments or theory, then it would count as a conciliatory view. If one, on the other hand, submits that no revision is required, it would be a steadfast view. One could also cook up a special level-splitting kind of steadfastness without noteworthy ingenuity. Assume, say, that one is a logical pluralist (of some sort), then one may argue that there are benign cases of logical akrasia. Given the pluralist's dictum that more than one logic can be correct, the (level-)incoherence of logical akrasia need not be problematic. Alternatively, one might be able to pull off a level-splitting response by appealing to logical instrumentalism (Haack 1974), i.e., the view that it doesn't make sense to think of logics as being correct or incorrect rather they are simply variously useful or not. It seems plausible to suggest that appealing to instrumentalism in some way could dissolve the tension in cases of logical akrasia to the level-splitter's satisfaction.^{20, 21}

3. Enter Gödel: rationality and akrasia

As advertised earlier, the two sections 3 and 4 make use of Gödel's (in)famous Second Incompleteness Theorem (Gödel 1931) to pose a dilemma related to logical akrasia. It should be noted up front that while the puzzle of arithmetic raised in these sections is related to logical akrasia, we'll leave it open what exactly the relation consists of, e.g., if second-order arithmetic is “logic” as per the neo-logicists, then the puzzle is just another case of logical akrasia; if Quine was right about bounds of logic, however, then it's not.

Gödel's theorem establishes that: assuming Peano Arithmetic (PA) is consistent, PA doesn’t derive Con _PA. In other words, if the theory PA is consistent, then PA cannot derive its own consistency.²²

Consider now the following akratic puzzle:

Akratic Peano Arithmetic. In using Peano Arithmetic, PA, subject S is at least implicitly committing to (and relying on) PA's consistency. After all, if the theory were inconsistent it's no help in sorting out truths from falsehoods. But if S's theory is PA, then the theory doesn't itself prove that PA is consistent (by Incompleteness). So, in that case S's theory doesn't prove the claim that S is committed to, and consequently S is at least weakly akratic (cf. section 2.1). Of course, S can easily extend S's theory (why not?), and then consider PA with Con _PA added as an axiom. This would provide S with a proof of Con _PA in a single line. But alas, now the issue of akrasia arises at the level of appeal to the consistency of that theory, i.e., PA + Con _PA. And so on ad infinitum…

In S's use of PA it turns out that S is not just committed to the theory PA itself, but also the stronger theory PA + Con _PA. Ergo, S is committed to a “logical” principle that their theory cannot prove, and thus S is in a state of akrasia.²³ By construction, there is no way for S to avoid committing to Con _PA (or $Con_{PA + Con_{PA}}$, or…) and escaping their akratic state (on the pain of triviality). While this doesn't necessarily show that S is irrational when using PA, it does entail that S is committed to something which goes beyond S's own theory in such situations.

Yet this is a positive kind of mismatch – i.e., S's background logic can prove something which S's believed theory cannot – rather than a violation of a negative rationality constraint such as the previously mentioned Akratic Principle. Recall the Akratic Principle stating that:

No [epistemic] situation rationally permits any overall [doxastic] state containing both an attitude A and the belief that A is rationally forbidden in one's situation. (cf. section 1)

As this is a negative principle in the sense that it involves an assertion about rationally forbidden states, the puzzle of akratic arithmetic is not in any obvious way a violation of it. Unless we are ready to grant that it's rationally forbidden to commit oneself to something which one cannot prove, of course, but this seems overly strong. Nobody in their right mind would suggest that provability is a plausible guide to rationality simpliciter.²⁴

Nonetheless, the puzzle of akratic arithmetic does illustrate a clash with our ideals concerning epistemic rationality insofar as reflective equilibrium is among them. As logically akratic states cannot be in reflective equilibrium – given interpretation (1) from section 2.2 – the case above does indeed suggest that S's doxastic state is epistemically irrational in a certain sense. How bad this sort of irrationality looks to the epistemologist does of course depend on the kind of good they take reflective equilibrium to achieve. On one interpretation, reflective equilibrium merely indicates that a reasoner has done what is rationally required of them relative to their initial data (e.g., a set of intuitions about certain logical inferences), but it would take further argument to show that a reasoner's doxastic attitudes are also likely to be true. Under this interpretation, reflective equilibrium is a rational ideal regarding the internal coherence of doxastic states rather than truth-conducive rationality.²⁵

4. The Dilemma of Arithmetical Akrasia

The upshot of section 3 is what we might call the Dilemma of Arithmetical Akrasia:

Insofar we take our theories to be appropriately formal and complete, then either:²⁶

1. (i) we must be agnostic about the consistency of PA (on the pain of triviality), which would be extremely odd at best;

2. (ii) or we must accept being arithmetically akratic, i.e., accept that we are trapped in an inescapable, infinite hierarchy of akratic states.

Notice that while taking the second horn of the dilemma doesn't rule out the existence of a rational fixpoint somewhere on the theoretical ladder, it does eliminate the possibility of an akrasia-free state which is accessible to us (since Gödel's incompleteness results range over all axiomatizable theories).

Further, the Dilemma of Arithmetical Akrasia is special in at least two ways. First, it involves an unsolvable case of akrasia while most cases of logical akrasia are clearly solvable, e.g., by converting to a fully classical theory. If the intuitionist we met in section 2 were willing to convert to a fully classical theory, then their akratic state would dissolve. But the case of PA is different as it looks more similar to an epistemic blindspot; where proposition 〈p〉 is an epistemic blindspot for subject S at time t if and only if 〈p〉 is consistent but unknowable by S at t (Sorensen 1988, 2020). Similarly, the consistency presumption is fundamental to our use of PA, but it just cannot be proved (and thus known) from within the bounds of the theory itself.

Second, unlike the akratic issues we focused on above (cf. section 2), the Dilemma of Arithmetical Akrasia cuts across the divide between classical and non-classical logicians. In the case of PA, it seems that we are all either agnostic (on the pain of triviality) or akratic!

So, in the end, taking the first horn doesn't sit well with our general intellectual outlook because we want to avoid being agnostic about the consistency of PA; on the other hand, going for the second horn is an unpleasant move as it reveals a boundary on arithmetical theorizing which seems to conflict with our rational ideals (e.g., reflective equilibrium).

5. Conclusion and perspectives: escaping the Dilemma of Arithmetical Akrasia

While the exact details remain to be worked out, this final section provides a quick and dirty proposal of how one can translate the Dilemma of Arithmetical Akrasia into a case of regular epistemic akrasia; and further how one might escape the dilemma when it's spelled out this way.

Let's first reformulate the dilemma such that it becomes a case of epistemic akrasia. Spelled out in terms of premises and conclusion(s), we get:

1. 1. S believes PA [by Indispensability].

2. 2. S believes Con _PA [by No-Miracles].

3. 3. S believes Con _PA is a logical principle [in absence of reasons to the contrary].

4. 4. S believes ${\nvdash}_{PA}Con_{PA}$ [by Incompleteness].

5. 5. S believes in the strong interpretation of reflective equilibrium with respect to PA: It's permissible for S to believe a logical principle only if PA proves it.

6. 6. Therefore: S believes Con _PA and believes that 〈it's forbidden to believe Con _PA〉.

7. 7. Ergo: S is epistemically akratic.

Now, it seems fair to suggest that we don't want to consider rejecting premises (1), (2), and (4); which leaves us with the possibility of rejecting one or both of (3) and (5) in order to escape the dilemma. That is to say:

1. 1. S believes PA [by Indispensability].

2. 2. S believes Con _PA [by No-Miracles].

3. 3. S believes Con _PA is a logical principle [in absence of reasons to the contrary].

4. 4. S believes ${{\nvdash}_{PA}}Con_{PA}$ [by Incompleteness].

5. 5. S believes in the strong interpretation of reflective equilibrium with respect to PA: It's permissible for S to believe a logical principle only if PA proves it.

6. 6. Therefore: S believes Con _PA and believes that 〈it's forbidden to believe Con _PA〉.

7. 7. Ergo: S is epistemically akratic.

Consider first the possibility of rejecting (3), i.e., the logicality of Con _PA. The principle stated by Con _PA is certainly a well-formed sentence of PA; and in that specific sense it is logical. It also expresses something essential to PA (or at least our use of PA). But perhaps Con _PA is still not logical in the right way for strong reflective equilibrium to apply to it. Is it really fair to expect strong reflective equilibrium to apply to paraconsistent logical theories, for example?

Consider next the possibility of rejecting (5) instead. While reflective equilibrium may be initially plausible when viewed as a philosophical method or as “the prima facie epistemology of logic” (Cohnitz and Estrada-González 2019: 137), it does seem like an overly demanding output to expect from applying the method in the context of logical theorizing and PA. Gödel's theorem already suggests that epistemic principles of that kind are hopeless, because every theory with the same expressive power as PA (or more) has a Gödel sentence. Why insist on something impossible? (This is an “ought-implies-can” violation perhaps). Moreover, what reason do we have to think that strong reflective equilibrium is a norm of belief within arithmetical theorizing? It seems that you'll need overly demanding bridge principles to establish the right connections between the formal sciences and epistemology in order to get this going (cf. MacFarlane 2004). Another way to put this point: the dilemma relies on it being an epistemic ideal that there is a reflective equilibrium between what one is committed to and what one's believed theory can prove (call it “RE”). An alternative, and perhaps more plausible ideal is that there be a reflective equilibrium between what one's committed to and what one's epistemic practice can justify (call it “RE*”). What speaks in favor of RE over RE*?

Supposing that at least one of the rough strategies outlined above is successful in letting us escape the dilemma – when it's framed in terms of epistemic akrasia – we are thus left to ask whether it's still a problem if S is arithmetically akratic after rejecting either of these premises and avoiding epistemic akrasia. Some theorists of the post-Gödel era may simply shrug their shoulders and bite the bullet here. In a way what Gödel's second incompleteness result tells us is that we can't both have consistency and syntactic completeness when it comes to theories with a certain amount of expressive power. So, perhaps some level of akrasia is just something that working logicians, mathematicians, computer scientists, etc., have come to live with in the aftermath of Gödel. They may also want to suggest that there is an important difference between the cases of logical akrasia exemplified by the intuitionist rejecting the excluded middle (cf. section 2) and the specific case of PA. In the former, the intuitionist can't combine their official theory and the background logic they are committing to into a jointwise consistent whole, whereas this is certainly possible in the latter – it's just that the background logic must be stronger than the theory PA itself.