Carruthers is right to reject the idea of a dedicated piece of cognitive architecture with the exclusive job of reading our own minds. But his mistake is in trying to explain introspection in terms of any one mindreading system. We understand ourselves in many different ways via many systems.

Carruthers claims that global workspace theory implies that sensory states, unlike propositional attitudes, are introspectible in a non-interpretative fashion. I argue that this claim is false, and defend a strong version of the “mindreading is prior” model of first-person access, according to which the self-ascription of all mental states, both propositional and sensory, is interpretative.

Judgments vary in the extent to which they are based on interpretation versus relatively direct access to mental contents. That is, a judgment might require a trivial amount of interpretation (e.g., translating one's immediately accessible “inner speech”) or a rather substantial amount of confabulation. Recognizing this continuum of interpretation underlying judgment could be more fruitful than debating a categorical introspection versus interpretation distinction.

Peter Carruthers correctly claims that metacognition in humans may involve self-directed interpretations (i.e., may use the conceptual interpretative resources of mindreading). He fails to show, however, that metacognition cannot rely exclusively on subjective experience. Focusing on self-directed mindreading can only bypass evolutionary considerations and obscure important functional differences.

Carruthers argues that there is no developmental or clinical evidence that metacognition is dissociable from mindreading, and hence there is no reason to think that metacognition is prior to mindreading. A closer look at the evidence, however, reveals that these conclusions are premature at best.

The engineer's look at how the mind works omits a central piece of the puzzle. It ignores the dynamic of motivations and the social context in which mindreading and metacognition evolved and developed in the first place. Mindreading and metacognition derive from a primacy of affective mindreading and meta-affectivity (e.g., secondary emotions such as shame or pride), both co-emerging in early development.

Mindreading in schizophrenia has been shown to be impaired in a multitude of studies. Furthermore, there is increasing evidence to suggest that metacognition is damaged as well. Lack of insight, or the inability to recognise one's own disorder, is an example of such a failure. We suggest that mindreading and metacognition are linked, but separable.

This commentary focuses on evidence from autism concerning the relation between metacognition and mindreading. We support Carruthers' rejection of models 1 (independent systems) and 3 (metacognition before mindreading), and provide evidence to strengthen his critique. However, we also present evidence from autism that we believe supports model 2 (one mechanism, two modes of access) over model 4 (mindreading is prior).

Defending first-person introspective access to own mental states, we argue against Carruthers' claim of mindreading being prior to meta-cognition and for a fundamental difference between how we understand our own and others' mental states. We conclude that a model based on one mechanism but involving two different kinds of access for self and other is sufficient and more consistent with the evidence.

This response defends the view that human metacognition results from us turning our mindreading capacities upon ourselves, and that our access to our own propositional attitudes is through interpretation rather than introspection. Relevant evidence is considered, including that deriving from studies of childhood development and other animal species. Also discussed are data suggesting dissociations between metacognitive and mindreading capacities, especially in autism and schizophrenia.

The past 50 years have seen an accumulation of evidence suggesting that associative learning depends on high-level cognitive processes that give rise to propositional knowledge. Yet, many learning theorists maintain a belief in a learning mechanism in which links between mental representations are formed automatically. We characterize and highlight the differences between the propositional and link approaches, and review the relevant empirical evidence. We conclude that learning is the consequence of propositional reasoning processes that cooperate with the unconscious processes involved in memory retrieval and perception. We argue that this new conceptual framework allows many of the important recent advances in associative learning research to be retained, but recast in a model that provides a firmer foundation for both immediate application and future research.

We discuss findings on evaluative conditioning (EC) that are problematic for the “conscious reasoning/propositional knowledge” account of learning, namely, dissociations between conscious beliefs and acquired (dis)liking. We next argue that, both for EC and for Pavlovian learning in general, conditioned responding cannot rationally be inferred from propositional knowledge type “CS refers to/signals US,” and that, therefore, performance cannot be explained.

Mitchell et al.'s claim, that their propositional theory is a single-process theory, is illusory because they relegate some learning to a secondary memory process. This renders the single-process theory untestable. The propositional account is not a process theory of learning, but rather, a heuristic that has led to interesting research.

In this commentary, we assess the propositional approach to associative learning not only in terms of veridicality and falsifiability, but also in heuristic value. We remark that it has furthered our knowledge and understanding of human, as well as animal, associative learning. At the same time, we maintain that models developed from the association formation tradition continue to bear great heuristic value as well.

We propose that learning has a top-down component, but not in the propositional terms described by Mitchell et al. Specifically, we propose that a host of learning processes, including associative learning, serve to imbue the representation of the conditioned stimulus (CS) with affective meaning.

The main conclusion to draw from Mitchell et al's article is that it is difficult to disentangle cognitive and learning processes in contingency and causal experiments. More compelling evidence for human associate learning comes from research where, because of the type of events involved, participants are unable or unlikely to think about the relationships between the events.

Mitchell et al. contemplate the possibility of rats being capable of propositional reasoning. We suggest that this is an unlikely and unsubstantiated possibility. Nonhuman animals and human infants do learn about the contingencies in the world; however, such learning seems not to be based on propositional reasoning, but on more elementary associative processes.

Mitchell et al. argue that conditioning phenomena may be better explained by high-level, rational processes, rather than by non-cognitive associative mechanisms. This commentary argues that this viewpoint is compatible with neuroscientific data, may extend to nonhuman animals, and casts computational models of reinforcement learning in a new light.

Additional support is presented for the necessity of awareness of the CS-US relation in human autonomic conditioning. However, possible limitations and exceptions regarding this general rule are discussed. Limitations include the lack of relationship between conditioned response (CR) strength and degree of awareness, and an important exception may be the finding of conditioning with backwardly masked CSs of a biologically prepared nature.

Mitchell et al. contend that there is no need to posit a contribution based on the formation of associative links to human learning. In order to sustain this argument, they have ignored evidence which is difficult to explain with propositional accounts; and they have mischaracterised the evidence they do cite by neglecting features of these experiments that contradict a propositional account.