Of course, you have no idea who Alice and Bob are, and probably have no interest in finding out. But if you are like most participants in our studies, you have already inferred that Alice is clever and Bob is cruel. Furthermore, you made these inferences without realizing it, and would likely deny them if asked.

The extensive research on such spontaneous trait inferences (STIs) and related phenomena supports some of Mitchell et al.'s arguments but challenges others. STI research does not use the framework or procedures of classical conditioning, so many of the terms in the target article (unconditioned stimulus [US], conditioned response [CR], blocking, overshadowing) are not applicable. But others are – for example, learning, propositions, associations, automatic and controlled processes, conscious and unconscious processes. So STI findings have direct bearing on the target article's more general claims.

Automatic propositional learning about others occurs. In a typical study of spontaneous inferences, participants read a series of trait-implying statements about other people, with an explicit goal other than forming impressions of them. The participants may be asked to memorize the material, or just familiarize themselves with it. Then they are tested in indirect ways to see whether they made trait inferences. Cued-recall, a favorite test in early research, also showed that participants were unaware of inferring anything about the targets (see Uleman et al. Reference Uleman, Newman, Moskowitz and Zanna1996b, for a review). That is, these inferences are unconscious. Although certainly aware of the stimuli, participants are not aware of their inference processes or outcomes.

The best evidence that STIs occur at encoding, rather than at retrieval, comes from on-line measures of concept activation. Zárate et al. (Reference Zárate, Uleman and Voils2001) had participants read trait-implying sentences for a subsequent memory test and simultaneously do a lexical decision task. Participants were quicker to identify as words implied traits that immediately followed trait-implying sentences than control sentences. Uleman et al. (Reference Uleman, Hon, Roman and Moskowitz1996a) used a slightly different task. Participants had to indicate, as quickly as possible, whether probe test words that immediately followed each sentence had literally been in the sentence. As predicted, participants were slower to correctly say “No” when sentences implied the trait probes than when they did not. So even though trait inferences interfered with optimal task performance, participants made them anyway. STIs are uncontrollable, as well as unintended.

Bargh (Reference Bargh, Wyer and Srull1994) identified efficiency as the fourth criterion of automaticity, and there is good evidence that STIs are highly efficient. Todorov and Uleman (Reference Todorov and Uleman2003) used a false recognition paradigm in which participants read 60 pairs of photos and sentences for a subsequent memory test. Then they judged whether particular traits had appeared in the sentences paired with particular photos. False recognition of implied traits was higher when traits were paired with actors' photos than with other photos, even when participants only saw each photo-sentence pair for two seconds, or viewed them to count the number of nouns in each sentence (not for a subsequent memory test), or viewed them while concurrently rehearsing a six-digit number. Thus, STIs are highly efficient, occurring under speeded, and shallow processing, and concurrent cognitive load conditions.

These false recognition results also show that STIs are about actors, and not just inferences about the behaviors. (See also Carlston & Skowronski Reference Carlston and Skowronski1994, for evidence from a different paradigm.) That is, STIs represent learned propositional knowledge, acquired relatively automatically (unconsciously, unintentionally, etc.). According to the target article (especially sect. 3), STIs cannot occur. Yet the evidence for spontaneous social inferences is extensive (Uleman et al. Reference Uleman, Saribay and Gonzalez2008).

Automatic associative learning about others also occurs. Suppose you read that Carol said Dan returned the lost wallet with all the money still in it, either for a subsequent memory test or merely to familiarize yourself with such information. When this information is paired with a photo of Carol but not Dan, you are likely to (unintentionally and unconsciously) associate honesty with Carol, the communicator (Skowronski et al. Reference Skowronski, Carlston, Mae and Crawford1998; Todorov & Uleman Reference Todorov and Uleman2004). Extensive research shows that this spontaneous trait transference (STT) does not occur because you confuse Carol and Dan, or intentionally draw conclusions about Carol. Instead, the activated trait is simply associated with the communicator, affecting subsequent trait ratings and other responses to her. Carlston and Skowronski (Reference Carlston and Skowronski2005) proposed that STTs represent mere associations between actors and traits, whereas STIs represent attributional inferences about actors.

Mitchell et al. emphasize the centrality of truth value to propositional knowledge. This suggests that if STI involves inferring propositions and STT does not, then making participants suspicious of the veracity of behavior descriptions paired with actors or communicators should interfere with STI but not with STT. This is exactly what Crawford et al. (Reference Crawford, Skowronski, Stiff and Scherer2007) found. These results not only support the propositional status of and relevance of truth value to STI, but they also provide another reason to distinguish between propositional and associative learning.

Explicit learning is dissociable from implicit learning about others. This dissociation challenges Mitchell et al.'s broad claims. Rydell et al. (Reference Rydell, McConnell, Mackie and Strain2006) asked participants to form an impression of Bob (so these are not spontaneous inferences). They saw 100 positive (or negative) statements describing Bob; then described how much they liked him; then saw 100 negative (or positive) statements; and then gave their explicit attitudes again. Uninterestingly, the participants liked Bob more (less) after positive (negative) descriptions, and changed their attitudes after reading inconsistent information. More interestingly, they were also subliminally primed on each trial with negative (or positive) words (e.g., love, cancer), and their implicit attitudes toward Bob were measured with a modified Implicit Association Test (IAT; Greenwald et al. Reference Greenwald, McGhee and Schwartz1998) after each series of 100 trials. The valence of the explicit and subliminal information always differed. After receiving positive explicit and negative subliminal information, the participants' explicit attitude was positive and their implicit attitude was negative. The second series of trials reversed these attitudes.

So, two kinds of evaluative learning occurred simultaneously about the same object (Bob): one intentional, conscious, and deliberate; the other, unintentional and unconscious. These dichotomies seem essential for describing the acquisition of explicit and implicit attitudes, even though Mitchell et al. claim that they are only useful for describing perceptual and performance processes.

These comments do not challenge the target article's claim that most, if not all, human conditioning is best understood as conscious, intentional, and effortful learning of propositions about the world. But they do challenge the authors' claim that all human learning can be subsumed under the rich and time-honored paradigms of conditioning. These comments also challenge the conflation of conscious, intentional, effortful, and controllable processes with each other and with all propositional learning – and of their “opposites” with each other and associative learning processes. There are varieties of both human propositional and associative learning that the target article neglects.