Detecting Deception Using MCI in Twitter

Introduction

Modified Cognitive Interviewing (MCI) has proven to be an effective method to detect deception [1, 2, 3, 4].

However, the increase in interactions via computer- mediated communication (CMC) is resulting in an interest in the dynamics of deception in online environments.

Over the past 10 years, numerous detecting deception studies have consistently demonstrated that an analysis of speech content – not voice stress - can consistently result in detecting deception rates at or above 82% - a rate that is significantly higher than that achieved by polygraph (i.e. 65%) or by chance (i.e. 50%) [1, 2, 5, 6, 7, 8, 9, 10].

These studies are important and demonstrate that speech content analysis – when derived through the use of a specific interviewing technique known as modified cognitive interviewing (MCI)–can be used to detect deception at rates significantly above other current approaches used by professionals.

At present, no studies have examined whether this MCI could effectively be used to discriminate truthful from denials or false claims in persons questioned in the computer-mediated environment of Twitter about the same task. This absence of research data means that our current understanding of how best to detect deceptive communications in social media is limited. Although one might assume that some of the language characteristics associated with deception that have been detected in traditional interviewing environments will also be manifested when people communicate in social media settings, at present there is a lack of research that might inform such an assumption. Thus, understanding whether alterations in speech content vary significantly between truthful claims, false claims or denials in social media would clarify not only whether features of deception generalize across communication environments, but also how well the method can be applied by professionals in the real world.

The present study was designed to examine A) whether MCI, performed in a limited text based environment (Twitter), derived speech content could be used to distinguish truthful accounts from false claims or denials in people who performed a cognitive activity; and B) the efficacy of MCI in Twitter compared to results previously derived from MCI in other formats.

Methods

Subjects: 44 students at the University of New Haven were the participants of this study. Each was given an oral briefing about the project and each provided written informed consent prior to participation in the study.

Design

This study design consisted of two phases: In Phase One participants engaged in, or only read about, a cognitive task; in Phase Two participants were interviewed about their claimed activity by interviewers through Twitter. When interviewing participants, the interviewers used a Modified Cognitive Interview (MCI) [1, 2, 3, 4]. The questions were adapted to fit into one Tweet each (140 characters). Transcripts from the Twitter interviews with participants were rated by human raters and also used to generate speech content variables for computer analysis.

Discussion

When exposed to the MCI in Twitter, genuine, false claimants and denial participants behaved in significantly different ways. Although the groups were indistinguishable from each other in the initial detailed prompt, they differed significantly when exposed to the second phase of the MCI (i.e., the mnemonic prompts). Truthful participants exhibited greater Unique Word counts and Response Lengths. The denial and false claim participants were unresponsive to the memory prompts and offered few additional details about their experience.

Within the context, it is reasonable to hypothesize that this reduced responsiveness to the mnemonic prompts could be due to the increased cognitive load associated with lying [11] and a desire to “tell their story” and stick to it so as to be believed, just as in Morgan CA, et al. [4].

Rater judgments were only slightly better than chance (54%). it is possible that raters were unable to assess the transcripts due to the short response length caused by the limited response possibility in Twitter. This finding is lower to that noted for raters in our previous study Morgan CA, et al. [4], where human raters performed significantly better (65%). Further studies may evaluate whether the rater accuracy can be increased. It bears emphasizing however, that as in many studies, the classification accuracies based on the speech content analysis by the computer were higher than those demonstrated by the human raters (78.5% vs. 54%, respectively) [12, 13].

As in previous studies, the application of the MCI method was effective at discriminating between genuine and false eyewitness accounts. As noted above, however, this field of inquiry is new and future research is needed in order to clarify the extent to which the current method and accuracies may apply to different types of experiences reported in social media. Although the current finding that MCI was effective for detecting deception in social media suggests this approach to detecting deception may have wider applicability in computer-mediated communications, it is important to acknowledge that social media communication platforms vary in the degree to which people text or video their observations. Therefore, a conservative interpretation of the present findings would be that they should be understood within the context of platforms in which people are sending text based communications.