Lying, Deceptive Implicatures, and Commitment

6.1. Participants

In this experiment, participants were recruited on Prolific Academic and completed an online survey implemented in Unipark. All participants were required to be at least 18 years old, English native speakers, and to have an approval rate from previous studies on the platform of at least 90%. 402 participants started the survey, and the data of 371 were included in the analysis. 31 participants were excluded for finishing the experiment in less than three minutes, failing the attention check, or indicating severe technical issues when doing the survey. The mean age was 32.3 years, 43% were male, 56% female, and 1% were non-binary or preferred not to disclose their gender. Participants received £1.50 for an estimated 15 minutes of their time (£6/h).

6.2. Design, Procedure and Materials

We implemented a 4 (kind of deception: Lie vs. Deceptive Implicature_a vs. Deceptive Implicature_b vs. Truth) × 14 (vignettes: fourteen short stories modeled after popular and different kinds of false implicature) mixed design. The kind of deception was tested as a between-subjects factor, while the fourteen different vignettes were tested within subjects. Participants were randomly assigned to one of the four between-subjects conditions. In the Lie condition, participants evaluated prototypical cases of lying; in the Truth condition, they evaluated prototypical cases of telling the truth. The two Deceptive Implicature conditions differed in the order in which the relevant test questions were asked (see below for further details).

Participants first read general instructions to familiarize themselves with the task and the response formats. They were then presented with the fourteen scenarios in randomized order. Afterwards, they were presented with an attention check in the form of a vignette featuring a person who accidentally said something that turned out to be false, although the agent had very good reason to believe what he said to be true (we excluded participants who judged this agent to be lying). On the final page, participants were asked about their gender, age, and whether any technical or other issues occurred.

The fourteen vignettes were modeled after cases of conversational implicatures discussed in the pertinent literature (e.g., Grice 1989b and Levinson 2000; see Meibauer 2006 for an overview). We can classify these cases (i) along the GCI/PCI (generalized conversational implicature/particularized conversational implicature) distinction and (ii) by the Gricean maxims or principles involved.

Here is one of the fourteen cases we employed (for the wording of the remaining vignettes, see the appendix):

1. Quantity—GCI—scalar <his, $a$> (Grice 1989b: 37–38)

   Ben and Pia have been married for 10 years and are on their way to a party in Paris. In their hometown, word on the street is that their marriage is in a bad place. Coincidentally, Robert is in Paris for business at the same time. Robert detests Ben, as Ben got elected mayor of their town: a job Robert wanted to have. Robert happens to meet Ben and Pia one evening in the hotel bar. When Robert returns to their home town, he sees an opportunity to get Ben in trouble. So he says to his friends in the tennis club, “In the hotel bar I saw Ben kissing a woman!” In fact, Ben was kissing his wife Pia, and Robert knew that.

[Test prompt:]

• Robert lied to his friend.

• Please indicate the extent to which you disagree or agree with this claim.

• completely disagree (1)-(2)-(3)-(4)-(5)-(6)-(7) completely agree

In the Lie condition, the deceptive implicatures were made explicit to transform the utterance into a prototypical lie. Hence, in the Lie condition the speaker in vignette (1) says:

“In the hotel bar I saw Ben kissing a woman—and the woman was not his wife!”

and in the Truth condition he tells his friends:

“In the hotel bar I saw Ben kissing a woman—it was Pia, his wife!”

In the two Deceptive Implicature and the Lie conditions, participants were presented with five additional questions:

• Commitment: To what extent do you disagree or agree with the following claim?

  • Robert committed himself to the claim that Ben was kissing a woman who was not his wife.

• Suggestion: To what extent do you disagree or agree with the following claim?

  • Robert suggested that Ben was kissing a woman who was not his wife.

• Probability: After hearing what Robert said, how likely is it that Robert’s friends will come to believe that Ben kissed a woman who was not his wife?

• Falsity: To what extent do you agree with the following claim? Robert’s statement was factually false.

• Morality: How would you morally evaluate Roberts’s behavior in the story you just read?

These five questions were added in the two Deceptive Implicature conditions, to test which of these five measures would best predict lie ratings for the deceptive implicatures. As already hinted at in Section 4, we expected commitment ratings to do best. The other proxy measures were included for the following reasons. The Suggestion question is, so to say, a weaker version of the Commitment question and would track lie ratings more closely if lying only required that a speaker’s utterance suggested—rather committed herself to—a believed-false claim. The Probability question takes into account that causing a false belief in the addressee seems to be a main function of lying. However, even though intended, the speaker cannot know for sure that they will cause the false belief. The Falsity question tracks the idea that cases of deceptive implicatures might amount to lying because people consider a literally true utterance carrying a deceptive implicature as overall false. The Morality question was included because morality has been shown to influence a wide range of judgments (see Knobe 2010 for an overview). It is plausible that people are more likely to judge those utterances to be lies that they consider morally wrong.

The two Deceptive Implicatures conditions (a & b) differed only with respect to the order in which the lie-question and the five proxy measure-questions were presented. In Deceptive Implicatures_a (and in Truth and Lie), participants first saw the vignette together with the lie-question. After they provided their lie rating, the vignette was again presented together with the five proxy measure-questions. In Deceptive Implicatures_b, participants first saw the vignette together with five proxy measure questions (order randomized), and after they provided their judgments to the proxy measures questions, they saw the vignette together with the lie-question on the next page.

We implemented these two orders of presentation because we could not rule out that the participants’ proxy measure-ratings would influence their lie ratings (and vice versa). For instance, judging that a certain assertion was factually false might increase the probability of considering this assertion to be a lie; or first judging that a behavior was morally good might decrease the probability of judging the corresponding assertion to be a lie. Alternatively, once participants have judged the assertion to be a lie, this might increase the probability of the five proxy measure-questions being answered in support of the lying judgment.

6.3.1. Deceptive Implicature_a versus Deceptive Implicature_b

We first tested, by means of a mixed ANOVA, whether lie ratings in the two Deceptive Implicature conditions were affected by the order in which the lie-question and the proxy measure-questions were presented, with the order of presentation as a between-subjects factor, and the vignettes as a within-subject factor. This was not the case. There was neither a main effect of order of presentation $\text{(}F\text{(}1,185\text{)}=1.94;p=.0165;{\eta }_p^2=.011\text{)}$, nor any interaction of order of presentation and vignette $\text{(}F\text{(}13,2379\text{)}=1.61;p=.074;{\eta }_p^2=.009\text{)}$ . Hence, we collapsed the data for the two Deceptive Implicature conditions (combined n = 185) in order to analyze the lie ratings.⁶

6.3.2. Lie ratings

The lie ratings for the three conditions and the fourteen cases are summarized in Figure 1. Unsurprisingly, lie ratings in the Lie condition were all significantly above the midpoint of the scale (all p < .0001) and close to the upper limit of the scale. Lie ratings in the Truth condition were significantly below the midpoint of the scale (all p < .0001) and close to the lower limit of the scale. The lie ratings for each of the fourteen deceptive implicatures lay between their corresponding prototypical variants. Ten out of the fourteen Deceptive Implicatures vignettes received mean lie ratings significantly higher than the midpoint of the scale (4), three were not significantly different from the midpoint, and one was significantly below it (see Table 2). Looking at the median ratings instead of the means, the result pattern (see Figure 2) is similar, with eleven medians greater than the midpoint, two being equal to the midpoint, and one being below it. Furthermore, in ten out of the fourteen cases the most frequently chosen option was the one indicating highest agreement with the claim that the agent lied (see Figure 2).

Figure 1. Lie judgments for the fourteen vignettes in the three conditions. Error bars represent 95% CI. The scale ranged from 1 to 7, with “1” indicating strong disagreement with the claim that the agent lied and “7” indicating strong agreement.

Figure 2. Violin plot of the lie ratings for the fourteen vignettes in the Deceptive Implicature condition. White circles represent medians, black circles represent modes.

6.3.3. Proxy measures

To assess how strongly the mean lying ratings for the fourteen scenarios are associated with the mean ratings of each of the five proxy measures for the fourteen scenarios, we correlated the mean lying ratings from the condition in which the lie-question was always asked first (Deceptive Implicature_a; n=93) with the mean ratings of the five proxy measures in the condition in which the proxy measure-questions were always asked first (Deceptive Implicature_b, n=92), so that the resulting correlation was based on two independent samples, that is, each participant either contributed a lying rating or a proxy measure rating. The mean ratings for both the five proxy measures and the lying ratings from these two conditions are summarized in Table 3. All except for the Falsity proxy measure exhibit a significant correlation with the lying ratings. The strongest correlation was found for Commitment ratings (see Figure 3), the agreement or disagreement with the claim that the agent committed herself to the implicature (r = .87, p < .001); followed by Suggestion (r = .79), Probability (r = .68), Morality (r = -.58) and Falsity (r = .46). Furthermore, the mean squared error (MSE) was lowest for Commitment (MSE = 0.41), followed by Morality (MSE = 0.60; inverted ratings), Probability (MSE = 0.93), Suggestion (1.00), and Falsity (4.48).⁷ Unsurprisingly, commitment ratings for the cases of straightforward lies (i.e., believed-false assertions) were—similarly to lie ratings—consistently close to the ceiling (never below 6).

Figure 3. Lie and Commitment ratings for the fourteen scenarios (r=.87). The lie-statement had the form ‘[Speaker] lied to [addressee]’ and the commitment-statement ‘[Speaker] committed herself to the claim that [implicature]’.

To further assess the robustness of commitment as the best proxy measure, we zoomed in to the participant’s level.⁸ Commitment was confirmed to have the strongest relationship with the lying ratings (r = .74), followed by Suggestion (r = .62), Morality (r = -.55), Falsity (r = .52), and Probability (r = .36). Moreover, using—mutatis mutandis—the same method at the vignette level, Commitment was again the best proxy measure (r = .5), followed by Falsity (r = 0.47), Morality and Suggestion (r = -.37 and r = .37, respectively), and Probability (r = .16).

6.3.4. Individual response pattern

An exploratory analysis of lying ratings at the level of individual participants reveals that agreement ratings differed quite a lot across the fourteen cases for most participants. In fact, not a single participant provided the same rating for all cases. On the contrary, the difference between the highest and the lowest rating (henceforth: range) was 6 for many participants (78 out of 185, 42%), that is, they strongly agreed (“7”) and strongly disagreed (“1”) at least once. Moreover, 28% (51/185) had a range of 5, and for 19% (36/185) the range was 4. By contrast, only 11% (20/185) had a range of 3 or lower (6% (12/185) had a range of 3, 4% (7/185) had a range of 2, and only one participant had a range 1).⁹

7.1. Limitations and Objections

Before we turn to the implications of our study for the debate on the definition of lying, we would like to address some potential objections to and limitations of our study. First, in our experiment, as in most experimental approaches, rating scales were used. However, it is perhaps unclear whether a 7-point scale is the right device for representing participants’ intuitions about lying. A scale seems to suggest or even presuppose that lying comes in degrees, rather than being a yes/no decision between “The utterance was a lie” and “The utterance was not a lie” (Rutschmann & Wiegmann 2017). Since there are arguments for the scalar nature of lying (Marsili 2014; 2019; see also Coleman & Kay 1981, who consider lying to be a prototype concept), a scalar measure seems appropriate. Moreover, a recent study (Viebahn, Wiegmann, Engelmann, & Willemsen 2020) found that binary and degree-measures of lying judgments correlate virtually perfectly, suggesting that they both measure the same aspects in an experimental study.

Second, one might worry that the participants who judged a speaker using a deceptive implicature to be lying only did so because we did not offer—in addition to “The speaker lied to the addressee”—more fine-grained response options such as “The speaker misled the addressee”. This is a reasonable worry that we cannot refute with the data we collected in the current experiment. Fortunately, though, data from a study (Reins & Wiegmann 2021) that included such options still found that most cases of deceptive implicatures were considered to be cases of lying (see Viebahn et al. 2020, who found that including this option slightly lowered lie attributions).¹⁰

Third, the vignettes that we used all represent cases discussed in the Gricean, neo-Gricean and post-Gricean literature in which they serve as examples for various kinds of implicatures. However, it is not clear whether, or to what extent, these single vignettes are representative of an implicature phenomenon. We can safely say, however, that for a certain cover story that belongs to a certain class of implicatures, we find a specific effect. Nevertheless, it would be premature to conclude that any other cover story belonging to that class would yield the same results.

7.2. Implications for the definition of lying

The results of our experiment suggest that the traditional definition of lying is too narrow to capture ordinary speakers’ judgments that classify some deceptive implicatures as cases of lying. If one agrees that a crucial aim of the definition of lying is to capture the folk concept, one possible reaction to such findings would be to extend the traditional definition to include all deceptive implicatures (cf. Meibauer 2005; 2014b: 125). This would lead us to the following broader definition of lying (the “or” is inclusive):

A lies to B if and only if

there is a proposition p such that

1. A asserts that p to B, and

2. A believes that p is false

or there is an implicature q such that

1. A implicates that q to B, and

2. A believes that q is false

However, this definition might be too broad, because of the empirical finding that some cases of deceptive implicatures were not judged to be instances of lying by laypeople. Furthermore, it would not be able to account for a salient feature of our results. While participants overall tended to agree that speakers can lie by means of deceptive implicatures, the agreement ratings were always significantly higher for the corresponding deceptive assertions. This gap is mirrored by the degree to which participants perceived the speaker to have committed herself to the believed-false claim. As briefly mentioned above, when the speaker lied in a straightforward way (i.e., asserted a believed-false claim), commitment ratings were—as one would expect—consistently close to the ceiling of the rating scale and always higher than in corresponding cases of deceptive implicatures. Furthermore, commitment ratings also tracked the finding that the lying ratings for the fourteen cases of deceptive implicatures differed to quite some extent, as indicated by the result that commitment exhibited the highest correlation with the lying ratings and the lowest MSE. In light of these findings, understanding lying as commitment to a claim that is believed to be false seems promising. Perhaps a definition along the lines of Viebahn’s (2021) is worth considering:

A lies to B if and only if there is a proposition p such that:

1. A performs a communicative act C with content p;

2. with C, A intends to communicate p to B;

3. with C, A commits herself to p; and

4. A believes that p is false.

Obviously, in order to establish a commitment-based definition of lying, several issues need to be addressed. The two most pressing questions are the following. What exactly does it mean to be committed to the truth of a proposition? And what are the features of an implicature that decide whether or not a speaker is committed to its content? The first question has been extensively discussed in the context of assertion (see Pagin 2016 for an overview), and has been the focus of recent discussions (see the special issue in Theoretical Linguistics, “Communication as Commitment Sharing” 2019). Given our empirical approach, it will not come as a surprise that we believe that answering the question of what commitment means in this context requires further empirical studies, which is why we will refrain for now from proposing a specific account.

Nevertheless, it is probably worth considering how commitment has been spelled out in the literature.¹¹ We have chosen to present Viebahn’s (2021) recently proposed account for the following reasons. First, it is a variant of a view endorsed by several authors, namely the idea that commitment to p entails a certain justificatory responsibility to defend p if challenged (see Pagin 2016, for an overview of such accounts). Moreover, since Viebahn’s account has been developed in the course of defending a commitment-based definition of lying that includes certain implicatures, it might help us to address the second question (about the features of an utterance that determine whether the speaker is committed to the implicated content). This account is described as follows:

The notion of commitment […] is to be understood as the commitment in assertion […]. By requiring liars to commit themselves to a proposition they believe to be false, […] [this notion of commitment] requires them to take on a justificatory responsibility to defend that they know the proposition put forward. And this justificatory responsibility, in turn, is based on the absence of consistent dismissals to knowledge-challenges regarding the proposition in question.

Viebahn’s (2021) account not only provides a characterization of what it means to be committed to p, but additionally offers a criterion for determining whether a speaker committed herself to a certain proposition.

The Commitment in Lying

In performing a communicative act C with a proposition p as its content, a speaker A commits herself to p (in the sense relevant for the lying-misleading distinction) iff A cannot consistently dismiss an audience challenge in response to C to defend (or justify) that she knows that p.

While it is beyond the scope of this paper to closely examine Viebahn’s (2021) criterion, it seems fair to note two things about it. First, it seems to make correct predictions for several kinds of implicatures. For instance, it seems plausible that a speaker cannot consistently dismiss an audience challenge in response to C to defend (or justify) that she knows that p in cases of substitute implicatures, such as saying I’ve got tomatoes coming out of my ears to communicate that one has obtained a good crop (Viebahn 2017; 2021). Second, it seems that the question of whether the speaker can consistently dismiss the challenge is not easy to answer. For instance, let us consider again the case of Robert saying In the hotel bar I saw Ben kissing a woman! and assume that someone challenges Robert by asking Do you really know that he did not kiss his wife? Can Robert consistently dismiss this challenge by saying something like I did not claim that he kissed someone who wasn’t his wife. I merely claimed that he kissed a woman? The answer to this question does not seem to be straightforward (and Viebahn himself does not deal with GCIs). This need not be a bad thing, because it might reflect the finding that participants’ judgments for deceptive implicatures are not as clear as they are for the corresponding assertions. However, there might be easier-to-identify features of an utterance that determine whether the speaker committed herself to an implicature. Again, further empirical studies seem the way to go, but let us make some educated guesses here. First, as discussed in Section 4, one kind of case in which a speaker is considered to have committed herself to the truth of an implicature might be if her utterance would not be relevant at all without deriving the implicature. Without deriving the implicature, replying I have to work to the question Are you coming to the party? would not be relevant to the question at all (Party case). The same holds for the answer Evelyn was here yesterday, and she is a smoker to the question of who the cigarettes belong to (Cigarette case). By contrast, consider case (3). Here, answering I believe that he has a girlfriend seems relevant to the question of how chances are for getting together with the man in question, even without deriving the implicature that the speaker does not know it. Second, a speaker might commit herself to the truth of an implicature if the implicature corresponds to a strong default interpretation. For instance, the empirical evidence (in our study but also in Weissman & Terkourafi 2018) that deceptive cardinals are cases of lying might be explained by the fact that, in contexts where the exact number of something is important, utterances like have two children seem to be naturally understood as exactly two (cf. relevance theorists, Carston 2002; Sperber & Wilson 1986, who consider GCIs as explicatures and belonging to the explicit content expressed). A further candidate for when speakers are committed to implicatures might be cases in which the implicature is a substitute for what is said, as in metaphors and hyperbole (cf. Viebahn 2017). For instance, Carl saying, in (11), that the kids were bouncing and bouncing and bouncing seems to be just a substitute for saying that the kids were bouncing for a long time.

Let us conclude the discussion with addressing a potential worry that could arise at this point. One might be skeptical that without specifying commitment in the context of lying, any real progress has been made by the current study. However, such an assessment would neglect important findings and features of the study. First, the high number of participants and vignettes in the current study arguably provides the strongest evidence that ordinary people consider some cases of deceptive implicatures as cases of lying. Second, concerning the relations of lying and several proxy measures, the finding that commitment best tracks lying ratings is not trivial or obvious, even though the view that commitment is important for lying can already be found in the literature (Meibauer 2014a; Viebahn 2017; 2020; 2021). The other proxy measures we investigated were arguably not any less plausible from the start. For instance, if falsity had been the best proxy measure, there would have been a very natural explanation of why one can lie with deceptive implicatures: they can affect truth evaluations—and the more strongly they render an utterance factually false, the more likely the utterance is to be considered a case of lying. Moreover, assuming that the point of lying is to cause a false belief in the addressee, it would not be surprising if the probability of doing so would be the best proxy measure of lying. And taking into account the seminal work in experimental philosophy done by Knobe (2003; 2010), it would not have been shocking if our findings had revealed that moral evaluations most closely track lie judgments. Hence, although there is still a long way to go, the findings of the current study might point in the right direction.

To sum up, the current study provides strong evidence that ordinary speakers consider a wide range of deceptive implicatures to be cases of lying and suggests a link between lying and being perceived to be committed to a believed false claim. The latter finding suggests that defining lying in terms of commitment might be promising. However, before a commitment-based definition of lying can be established further theoretical and empirical research is needed to clarify the concept and its determents.