The general public believes that it is possible to detect lying by observing nonverbal cues, yet these cues do not improve detection abilities (e.g., Bogaard et al., 2016). For these reasons, individuals who have not been trained to detect reliable cues (i.e., based on experimental evidence) typically detect lying at a rate only slightly higher than chance (i.e., 54%; see Bond and DePaulo, 2006, 2008; Hauch et al., 2017). After the meta-analysis ran by DePaulo and colleagues (2003, p. 95), an increase in the interface between language and cognition during deceit has emerged. More specifically, this meta-analysis showed that deceitful narratives differed mainly from truthful ones in terms of content, and slightly in terms of objective vocal and nonverbal cues. Therefore, the analysis of verbal lying behavior has gained interest in several domains: the forensic and legal framework (e.g., Vrij & Fisher, 2016), the insurance framework (e.g., Leal et al., 2015), or the human resources domain (e.g., Schneider et al., 2015) as meta-analyses showed that discourse analysis supported lie detection (e.g., DePaulo et al., 2003; Hauch et al., 2015; Oberlander et al., 2016). To provide an objective assessment of verbal deception, methods have been developed to provide structured analysis grids. Some of these methods have developed around research at the interface between psychology and linguistics. Common tools involve the Statement Validity Analysis (e.g., Amado et al., 2016, for a review), the Reality Monitoring Framework (e.g., Masip et al., 2005; Nahari, 2018a), or the Verifiability Approach (e.g., Nahari et al., 2014; see also Nahari, 2018a, for a review). The goal of these methods is to reduce the subjectivity of human judgments by providing objective indicators from research in cognitive science and psycholinguistics (Zuckerman et al., 1981; Strömwall et al., 2004). These methods tackling verbal deception correctly classify deceptive and authentic narratives about 70% of the time (Vrij, 2018). In this paper, we refer to verbal deception as a kind of deception that is expressed in words, thus including oral-and-transcribed deception and written deception in an interactional context. This type of deception is in contrast to non-verbal deception, which is expressed without words (e.g., through gestures such as head nods) or for which attempts at detection rely on analysis of non-verbal behavior. It is also in contrast with one-sided communication settings, such as online reviews, on which part of the literature that we will discuss here is based and provides interesting insights (see Rastogi and Mehrotra, 2017 for a review).

In 2018, several researchers and practitioners met for a 3-day workshop at Bar-Ilan University (Jérusalem, Israel), and provided an overview of the urgent issues and prospects in verbal deception detection (Nahari et al., 2019). Their goal was to answer the question “In your view, what is the most urgent, unsolved question/issue in verbal lie detection?” (p. 1). The consecutive article answering this question was structured in commentaries, each highlighting a matter that should further be investigated. Some commentaries insisted on the consideration of context and ecological validity of experimental studies (Commentaries 1, 3). Others required a better understanding and a fine-tuning of actual methods (Commentaries 2, 4, 5, 6, 10, 11). Finally, commentaries also regarded the communication between the field and research (Commentaries 7, 8, and 9). These commentaries covered the essential subjects regarding human judgements of deception and highlighted paramount issues on which researchers should focus. However, computational linguistic deception detection is barely mentioned in the publication of this 3-day workshop, while the interest for these tools and methods increases (e.g., Jupe et al., 2018; Kleinberg et al., 2018; Tomas et al., 2021c).

Similarly, in Vrij and Fisher (2016), deception detection tools are considered in regard to their potential application on the field. The article mentions various physiological methods (i.e., Control/comparison Question Test, Concealed Information Test ¹ ), nonverbal methods (i.e., Behavioral Analysis Interview ² ), and cognitive load methods (Imposing cognitive load, Asking unexpected questions, Encouraging the interviewees to say more, Strategic Use of Evidence ³ ). Only one verbal discourse analysis method, the Verifiability Approach, is mentioned (Nahari et al., 2014; Vrij et al., 2016). The Verifiability Approach suggests that liars construct their narratives strategically, and thus attempt to avoid verifiable information that would potentially confound them. No other information however is provided on verbal tools, nor on the extensive literature on computational verbal deception detection published at the time (e.g., Bond and Lee, 2005; Masip et al., 2012; Fornaciari and Poesio, 2013; Hauch et al., 2015; Rubin et al., 2015). The absence of the computational linguistic approach to detect deception from current reviews in the psychology literature is conspicuous as it is currently one of the most useful set of methods and tools when wishing to understand the interface between language and cognition (Jackson et al., 2020). Still, it has been suggested that research should determine whether the coding of verbal data can be operationalized by making computers and humans collaborate to improve deception detection (Biros et al., 2004; Nahari et al., 2019; Zhang et al., 2020).

In the present article, our aim is to inform the large variety of researchers from the fields of communication sciences, psychology, engineering, linguistics about prospects and issues regarding computational approaches to (deceptive) discourse analysis. We wish to provide some considerations regarding the computational approach to verbal deception detection and how it may provide answers and insights regarding the cooperation between humans and computers. To do so, we will rely on a set of scientific literature corpora that rarely mention each other: the theory-based approach and the classification-oriented approach. W e will make use of the theory-based literature, generally proposed by psychology and psycholinguistics researchers and relying on the notion of baselining and the Undeutsch hypothesis (Amado et al., 2015), and combine it with the computational classification -oriented approach supported by computational linguists and engineers. To the best of our knowledge, the combination of both these perspectives has not been recently proposed, and should provide interesting insights and perspectives on deception detection problems. Computational stylometric (i.e., statistical stylistics) measurements and methodologies might help answer some questions often considered in human deception detection methods, such as the standardization of coding commonly encountered in deception (i.e., inter-judge rating discrepancies), or intra-individual baseline-recall comparisons, as stylometry is deep-rooted in document comparison methods.

We thus want to shed light on the prospects and issues of deception detection relying on cue-based (e.g., LIWC) and model-based (e.g., word2vec, GloVe, BERT) computational stylometry, and the technical means provided by computational stylometry to solve certain issues in the deception detection research field. In this review, we shall only discuss the verbal content analysis methods and tools developed to detect deception. In other words, the data considered in this review is either written or transcribed. This is in line with the current research led on the topic indicating that using verbal content deception detection methods might be the most reliable and valid way (Aamodt and Custer, 2006; Hauch et al., 2015; Sternglanz et al., 2019). For this reason, this paper will not concern vocal analyses of deceitful speech, nor on nonverbal cues to deception. Despite their interest in the field, these tend to be less reliable than linguistic content approaches (see Sternglanz et al., 2019 for a review).

This article is divided into three sections. In the first one, we present and define stylometry as a subfield of authorship attribution and outline how it has been adapted to deal with the deception detection problem. This section also includes the main cues and model approaches used to detect deceit, and describes the accuracy of computational stylometric analyses of verbal deception. The second section is focused on the potential applicability to the field of stylometric deception detection methods. In the last section, we detail the advantages and limits of computational stylometric tools and methods for deception detection to finally outline crucial topics for further research.

Stylometry is considered as the statistical analysis of style in textual data (Chen, 2012; Chen et al., 2011; often used as a synonym for Natural Language Processing). It should be considered as a set methodologies and tools for the analysis of linguistic data (Daelemans, 2013). It emerged from the authorship attribution domain, in which an unidentified text is compared to a set of author-attributed groups of texts in order to determine by whom it was written (Love, 2002). Stylometry originates from the literary world but has soon evolved to be concerned with social matters, such as e-mail, text, or forum posts authorship identification (e.g., Afroz et al., 2014; Fatima et al., 2017), document obfuscation (e.g., Brennan et al., 2012), political discourse analysis (e.g., Barlow, 2013), or personality assessment (Verhoeven et al., 2016). Researchers and practitioners in the stylometric domain rely on the idiolect/stylome hypothesis: Every person has its own way to speak about a particular subject, and thus to express oneself on that matter (e.g., Baayen et al., 2002). This implies that while communicating verbally, every person chooses its own words in his/her available lexicon (Hallyday et al., 1964). Although the theoretical reasons for the existence of the idiolect have never been precisely clarified, one can assume an effect of frequency of word usage: The more a word is used by a speaker, the faster it can be accessed, and therefore prevails over other equally valid terms (e.g., O'Malley and Besner, 2008).

Relying on these idiosyncratic elements, stylometry aims to determine objective (i.e., determine the protagonists, the places mentioned, the time, etc.), subjective (i.e., opinions, values, etc.), or meta-knowledge (i.e., information about the author of the text in question, leading to a psychological and sociological profile of the author) from the verbal data analyzed (Daelemans, 2013). Importantly, the term profile is here taken directly from the literature on authorship attribution. Nonetheless, while stylometrics can better identify the psychological characteristics of the author, the notion of profiling has been debated significantly for several decades, and thus should be viewed with caution (see Fox and Farrington, 2018; for a meta-analysis on the issue). To identify the idiosyncrasies of the authors and analyze linguistic data, two approaches have been used to date: cue-based approaches and model-based approaches.

The first perspective, both in terms of historicity and importance, is the cue-based approach often represented by the classical Bag of Words (BoW) set of techniques. This approach considers words as individual atomic entities independently of any context, and covers the most basic semantic information extraction procedures (Zhang et al., 2010). In BoW approaches, one relies on taxonomic structures, and may use dictionaries and ontologies (e.g., Chung and Pennebaker, 2007), function and content words (e.g., Kestemont, 2014), n-grams (e.g., De Vel et al., 2001; Hernández Fulsilier et al., 2015), part-of-speech taggers (e.g., Hitschler et al., 2018), or parsers (e.g., Chen et al., 2015) to determine what every single word taken as an entity may indicate in terms of authorship. Note that it is not simply the word itself, but also its grammatical and syntactic characteristics (Rosso and Cagnina, 2017). But a major issue with these methods concerns the absence of context consideration. This problem can be easily exemplified with the example “I am not happy”, where the unit “happy” will be considered as a word denoting positive emotions when the general sentiment of the sentence is negative. Similarly, the word “bank” has ambiguous meaning, and ontologies working in an atomic perspective will have trouble distinguishing the institution from the side of a river.

To overcome this localist problem, the second perspective (i.e., model-based approaches) was developed. In this perspective, two approaches try to account for the context in which as word is presented. The first one is generally referred to as distributional representations or word embeddings. It relies on the distributional similarity theory according to which meaning is essentially conveyed by the context in which a unit of interest can be found, and thus that words found in a similar context have similar meaning (Harris, 1954; Mikolov et al., 2013). The objective of such models is to provide language models larger than the atomic perspective discussed previously. By commonly relying on recurrent or convolutional neural networks, one is capable of predicting in a sequential fashion words on the basis of context, or the context surrounding a given word (Raaijmakers, 2022). This is done by creating vector representations (also called word embeddings) of a given word from a large corpus, indicating the probability of a word occurring on the basis of its context. To do so, one may rely on training models such as word2vec, GloVe, or fastText (Nam et al., 2020). However, word embeddings have often been criticized for their interpretability problem despite some attempts being made to impart meaning on these vectors (Goodman and Flaxman, 2017; Şenel et al., 2018).

The second approach relies on deep learning and transformer-based language models, and most specifically on BERT (Bidirectional Encoder Representations from Transformers; Devlin et al., 2019). Transformers are algorithms that encode in parallel the input as a whole and combine it with the already produced output to determine the probability of the next word. A recent review regarding the properties of BERT shows that its embeddings incorporate syntactic tree-like structures, parts of speech, and semantic knowledge (Rogers et al., 2020). As for the vector representation approach, transformer-based models require a vast amount of data, and retain this black-box issue (Rogers et al., 2020).

By relying on these methods, stylometry can focus on the extraction of cues allowing to sketch a portrait of the author’s characteristics (i.e., certain socio-demographic characteristics, a potential mood state, a possible ideology, a spatio-temporal context, etc.), or at least suppose stylistic differences between one author’s characteristics and the others’ (Kocher and Savoy, 2018). These involve the evaluation of psychological and sociological traits of the author, and how psychological states may involve information processing, and thus variation in style. For instance, stylometry showed that one’s language was modified in a modified emotional or cognitive state (e.g., cognitive load, Khawaja et al., 2009; depression, Rude et al., 2004; Tackman et al., 2019; for a review, see Chung and Pennebaker, 2007). One can thus consider that this encompasses deception, as it implies for the sender a temporary psychological and emotional state that entails verbal strategies to create a false belief in the receiver (Walczyk et al., 2014). This is commonly expressed in other words by the Undeutsch hypothesis according to which a deceptive narrative will differ in form and content from a truthful narrative (and, by extension, its baseline; Amado et al., 2015). These strategies imply a potential deviation from the idiolect/stylome/baseline. In other words, one would be able to distinguish the deceptive narrative from truthful ones because they imply different states of minds, different verbal strategies, and thus different meta-knowledge cues in the account. To do so, the stylometric analyst may rely on a variety of tools, most of which have emerged for the authorship attribution domain.

To determine whether stylometry can be used in the field to reliably detect lying, we propose using a list of criteria developed by Vrij and Fisher (2016). Specifically, this list of criteria assesses 1) whether the items included in the previously presented methods were included as a result of a scientific approach and 2) whether their accuracy is supported by scientific studies.

To this end, this list extends the criteria of the Daubert case used as a benchmark for examining the scientific validity of evidence presented in court (Larreau, 2017). It includes 10 criteria in its final version (five criteria from Daubert and five criteria from Vrij and Fisher, 2016; see Table 1). The first five are used to determine whether evidence presented in court is scientifically admissible, and the last five are specific to lie detection.

To the first three criteria (i.e., “Is the scientific hypothesis testable?”, “Has the proposition been tested?”, and “Has the technique been peer-reviewed and published?”), the answer is yes for all of the stylometric methods included in this article. Indeed, lie detection from stylometric cues is a testable hypothesis: It is possible to determine, through human or algorithmic classification methods, whether the use of computational stylometry detects deception or not, in accordance with the falsifiability principle (Popper, 1959). Second, computational stylometry has been tested for its ability to detect deception with various tools and methods (e.g., for LIWC, see Ali and Levine, 2008; Fornaciari and Poesio, 2013; Newman et al., 2003; Tomas et al., 2021c; for named entity recognition, see Kleinberg et al., 2018; for morpho-syntactic labeling, see Banerjee and Chua, 2014; for n-grams, see Cagnina and Rosso, 2017; Hernández Fulsilier et al., 2015; Ott et al., 2013; for vector representations, see Nam et al., 2020; for BERT, see Barsever et al., 2020). Third, it has been the subject of over 20 peer-reviewed publications (e.g., Hauch et al., 2015; Forsyth and Anglim, 2020; Tomas et al., 2021a).

The fourth criterion from the Daubert list, regarding the known error rate of deception detection with stylometric cues, is complex. As we mentioned earlier, the accuracy rate seems to be at least as high as with the SVA or the RM methods for cue-based methods, and higher for model-based methods (see Table 1 for estimated error rates). However, to date, there is no systematic review providing an overview of the hits, misses, false alarms, and correct rejections when trying to classify deceptive narratives with stylometric cues. We thus argue that this criterion is to date not fulfilled because of the absence of general systematic review allowing for correct error rate assessment.

The fifth criterion questions whether the theory upon which the technique is based is globally supported by the scholar community. No clear definition has been given of what constitutes “acceptance in the scientific community”. Vrij and Fisher (2016) define this criterion as the amount of criticism the method has been subjected to. For instance, a lot of criticism has been uttered regarding the use of physiological cues in a polygraph context (e.g., Han, 2016; Vrij et al., 2016). If one considers the field of stylometry in a global perspective, including authorship analysis, there have been some critiques regarding the use of stylometry in court (Clark, 2011). However, while these critiques were accurate, they did not concern computational stylometry as an extraction procedure, but unfounded handmade stylometry, and its potential use of black box machine-learning algorithms for classification (Nortje and Tredoux, 2019). Computational stylometry as an extraction method has evidence-based background, robustness, and results that contradict these criticisms. Moreover, compared to the global literature regarding the stylometric assessment of authorship, the criticism is scarce. There seems to be little doubt regarding the global acceptance of its techniques by the scientific community (see Holmes, 1998, for a review). If one considers the field of forensic linguistics alone, the same global acceptance by the scientific community may be observed (Woolls, 2010). Moreover, the use of computational stylometric cues to detect deceit relies on various theoretical hypotheses commonly accepted by the community, such as the Reality Monitoring framework (e.g., Bond and Lee, 2005), or the Interpersonal Deception Theory combined with the Self-Presentation Perspective (e.g., Hancock et al., 2004). And as highlighted previously, the rationale behind stylometry is the one of the idiolect-baseline hypothesis, which has been supported by numerous studies (e.g., Barlow, 2013; Daelemans, 2013; Johnson and Wright, 2017; Kestemont, 2014; Stoop and van den Bosch, 2014; van Halteren et al., 2005; Wright, 2017). We thus argue that the theories underlying the use of the stylometric methodologies have generally been accepted by the scientific community.

Stylometry also ticks the box for the criteria 6–10. Indeed, it is easy to incorporate in a typical information-gathering setting (Criterion 6) as, up to now, it can only be used after the interviewing phase because of the transcription. This transcription phase is common to all current scientific methods. It tends more and more regularly to be part of the procedure in the case of audiovisual recordings for, for example, hearings of minors (e.g., art. 706-52 of the French code of criminal procedure, art. 4 of the ordinance n° 45-174 on delinquent children in France, articles 92-97 of the Belgian code of criminal instruction, art. 154, §4, let. d. of the Swiss code of criminal procedure) or adults (e.g., suspects in criminal cases, according to art. 715.01-715.2 of the Canadian Criminal Code; art. 64-1 of the French Code of Criminal Procedure; polygraphic credibility assessment according to art. 112ter, §1 of the Belgian Code of Criminal Procedure; Achieving Best Evidence in Criminal Proceedings, N8). Thanks to these transcripts, the use of stylometry during a second part of the hearing procedure allows the investigator to remain focused on the report writing phase and on gathering information during the interview.

Moreover, the use of post-interviewing stylometry will not affect the answer of a truthful interviewee (Criterion 7), and is easy to use (Criterion 8). Regarding Criterion 9, we argue that the automated extraction of stylometric cues protects the truth teller more than the use of any verbal deception detection grid, as the extraction itself is as free of biases as can be. Its analysis, however, may be questioned if one relies on deep learning methods, as they often lack transparency. Finally, regarding counter-measurements (Criterion 10), little is known in the public domain regarding stylometry for deception detection, leaving little room for counter-measurements.

If we follow the arguments developed by Vrij and Fisher (2016), stylometry should be allowed in the field when it comes to deception detection. However, there are a few issues that we wish to highlight to nuance this perspective. T hese criteria, although interesting in terms of scientific value, tend to avoid the main ethical concern behind deception detection, which is its accuracy. Evidently, scientific methods developed to tackle the detection of deceit provide better results than pure reliance on experience and intuition (DePaulo et al., 2003; Hauch et al., 2015; but see Sporer and Ulatowska, 2021 and Stel et al., 2020 for recent developments). Nevertheless, the performance and accuracy of these methods remains practically too dangerous to be put into practice, especially in risky situations such as employment, judicial events, or police interviews. Even in cases where impressive accuracy rates are reached (e.g., 93.6% in Nam et al., 2020), the chances to wrongly blame the sincere remain too high.

Despite this final argument regarding the current impossibility to rely on stylometry to detect deception in a definitie fashion, we still wish to develop the various advantages and caveats that set apart the stylometric way to detect deception from human methods in the next section.

As summarized in Table 2, computational stylometry has a few advantages and caveats when compared to human judgment methods to detect deceit.

We provided a review of common problems to deception detection and their potential solutions by relying on a majorly overlooked combination between theory-based and computational approaches to deception. To this day, the psychology and psycholinguistics research angle to deception detection commonly makes use of simple computational word-based stylometric tools while relying on theory-grounded approaches, while the computational approach proposed by the engineering literature relies almost exclusively on powerful algorithms without commonly mentioning the underlying theory explaining the differences between deceitful and sincere narratives. We argue for a combination of the strengths of both approaches (i.e., the understanding of cognitive and social mechanisms behind deception and the power of current algorithmic methods) for future research purposes.

Computational stylometry offers a set of tools that may help scholars and practitioners detect potentially deceptive verbal accounts. Relying on the power of computers and extraction algorithms, stylometry helps extract quickly the desired cues faster and more robustly than when humans use the actual grid to code textual data. The stylometric extracted data, when analyzed by algorithms, show above-chance discrimination rates between deceitful and authentic narratives while bringing interesting advantages for organizational concerns for legal practitioners. Little is known about how stylometric deception detection compares to human judgements, or even to the other methods of analysis of verbal statements, but the current studies seem to indicate at least similar results.

To the best of our knowledge, the present paper is also the first to make use of and extend the criteria developed by Vrij and Fisher (2016) regarding the applicability of stylometric deception detection on the field. Despite the fact that the computational stylometric approach seems to perform as well as the previous manual methods analyzed in the seminal study, and thus allegedly grant their use on the field, we nevertheless extended the list of criteria by considering ethical aspects. We argue that the lack of training datasets and current error rates, even in the best cases, remain insufficient to apply stylometric deception detection as a decision-making tool on the field.

This review also highlighted the limitations of current stylometric methods. To date, many studies in the psychology literature rely on the cue-based method to investigate the theoretical groudns of deception and make use of LIWC, while the other approaches remain explored (Hauch et al., 2015; Holtgraves and Jenkins, 2020; Schutte et al., 2021). It probably is the easiest stylometric tool to use to work on discourse analysis and deception detection. But this ease of use should not make scholars forget to look at other cues. As a matter of fact, a few studies have shown that LIWC involved certain issues that needed to be considered with care. Studies have highlighted for instance that LIWC was a word-based approach and consisted essentially of a list of individual words and could not take into account strings of words (Braun et al., 2019). For this reason, in sentences such as “I am happy” and “I am not happy”, the unigram happy will be considered each time as a sign of positive emotion, while it is not the case in the second sentence. LIWC thus omits context by relying solely on unique words fixed in a predefined ontology.

To counter this effect, a few studies have begun exploring new features to detect deception such as syntactic structure and part-of-speech tagging (Feng et al., 2012; Fornaciari et al., 2020) or Named-Entity Recognition (Kleinberg et al., 2018), and showed interesting insights in the content analysis of deceptive discourse. These encouraging paths need to be explored carefully, relying on evidence-supported linguistic and psychological models and interviewing procedures. But while stylometry shows a promising future for computational deception detection, there remain some methodological aspects that research should investigate.

Regarding the model-based approach, common machine-learning matters have rarely been tackled in the psychology and computational linguistics literature about deception detection. For instance, Kleinberg and colleagues (2019, 2021) explain how the methods developed to detect deceit computationally have rarely been tested on out-of-sample data. Most of the time, the division of the available data is favored to create a training sample (e.g., 20% of the available data) and a testing sample (e.g., 80% of the available data). Moreover, there are very few labeled datasets for deception detection purposes. We have mentioned Ott and colleagues’ dataset on deceptive opinion spam (Ott et al., 2013, Ott et al., 2012). Others have started to develop such as the one developed by Martens and Maalej (2019), the Paraphrased OPinion Spam (POPS; Kim et al., 2017), Ruspersonality (Litvinova et al., 2016), or the corpora made available by Amazon and Yelp. But as language is context-specific, the applications of the use of these datasets could not be transfered to computer-mediated communications, or even interactional deception detection (Sánchez-Junquera et al., 2020). Additionally, deception is, to the best of our knowledge, conceived as a binary variable in these datasets, where truthful and deceptive narratives are voluntarily antagonized. This means that even in the current datasets, a text is either deceptive or truthful. This does not reflect many cases where deception and truth are embedded in one another, representing an instance of the out-of-distribution generalization problem (Liu et al., 2020; Shen et al., 2021; Verigin et al., 2020a). Finally, the datasets artificially created for research purposes might not be reliable for training purposes (Fornaciari et al., 2020).

When considering the cue-based approach, the amount of cues investigated in linguistic deception detection could be widened if one considers the literature on authorship attribution and, to a bigger extent, on pragmatics. For instance, word length has rarely been considered, although it is supported by empirical evidence (Lewis and Frank, 2016). But as we explained above, the reliance on automatically extracted cues may cause an automation bias. We argue that it would be interesting to determine how this automation bias would appear in deception detection contexts, and how legal, human ressources, or insurance practitioners would consider a potentially automated decision-making system regarding deception detection. An adaptation of the aid/no-aid paradigm of Skitka et al. (1999) might be a first step to determining if a decision-supporting design may bias deception detection judgements. A second step would be to determine how trust in deception detection automation may influence decision-making when using computational deception detection methods.

We also wish to highlight how little is known about baselining in stylometric deception detection. In authorship attribution, from which stylometry originates, most studies rely on the comparison between an identified corpus of texts and statements to be attributed to an author. The determination of ground truth of the identified corpus in today’s authorship attribution problems is relatively reliable: Researchers rely often on tweets, blogs, or clearly signed data (Overdorf and Greenstadt, 2016). However such a signature does not exist in the field of deception detection. Still, we notice that little is known on how to adapt this baseline rationale to stylometric deception detection. For instance, could the linguistic style of the interviewer or the presentation of a model statement presented in Commentary 3 of the Nahari et al.‘s paper (2019) influence the language of the interviewee (see Richardson et al., 2014, for an example of Language Style Matching in police interrogation settings; see Porter et al., 2021 for a critical review of the model statement method)? This question bears significant importance, as the measurements used in stylometry and authorship attribution rely on similarity coefficients between identified and questioned documents, or in this case, authentic and deceptive texts. If the language of the interviewee changes and adopts the linguistic structures of the interviewer’s, there is a risk of obtaining inaccurate data, and thus a chance of biasing the samples used. Similarly, if a document guides the authentic or deceptive person’s narrative, it may introduce noise and disrupt the detection of linguistic signals that might indicate that a narrative has been manipulated. These factors should therefore be manipulated in experimental studies to determine their potential impact, but also to propose countermeasures if necessary.

There are other issues to consider when evaluating the baseline. For example, should one only rely on a single point of reference (as is now the case with baselining in deception detection), or use multiple linguistic sources to acquire the best overview of someone’s stylome (as is done in the authorship attribution domain)? Having a single point of reference may be problematic, as there are environmental and idiosyncratic factors that may impact the idiolect-baseline of the interviewee. Moreover, using a single reference point may have other disadvantages since its mere request may influence the subsequent narrative. Indeed, it has been shown that there is an interaction between the baseline and recall. Studies examining the change in order between baseline and recall have shown that the richness of detail and word count of the second statement was altered by the first (Verigin et al., 2020b). A recent study also highlighted this effect of the baseline on the second story (Tomas et al., 2021a).

But relying on a multiple points of reference causes other issues. For instance, ground truth may be harder to assess for each text. Moreover, the practical implications of multiple sources for baselining needs to be mentioned: Legal practitioners do not always have access to numerous documents, and if they do, the standardization of the documents may take a certain time that the legal field does not always have. These questions illustrate how little we actually know about verbal baselining. We argue that researchers should investigate the distinctiveness of verbal baseline establishment, and how to develop best practices combined with flexibility to be applied on the field.

Finally, as mentioned in the limits, modern verbal deception detection techniques rely on verbatim transcripts. Two problems arise: the time necessary for the transcription for oral data, and the method used for transcription. A few solutions should be investigated in order to tackle these limits. First, regarding the time-consuming approach of transcription, a human-in-the-loop approach might be of interest for this purpose ⁶ . Transcription softwares might provide a first transcript. Law enforcement officers or academic collaborators could then bring corrections to make it fit the above-mentioned guidelines. This interaction between humans and machines would provide an interesting solution to “meet the organizational time frame” mentioned in Nahari et al. (2019, p. 11). Second, as far as transcription methodology is concerned, we suggest that scholars focus on the emission of guidelines determining the rules to apply in transcribing the verbal discourse for deception detection purposes. More precisely, we suggest that researchers rely on the communication and linguistic domains to test experimentally what is best for verbal, and more particularly, computational stylometric cues extraction (e.g., Easton et al., 2000; Davidson, 2009).

These methodological issues prevent computational stylometry from being currently applied in the field. For this, we recommend that researchers in psychology, linguistics, and engineering cooperate and actively research these topics. This would be a way to expand each other’s knowledge: 1) for psychology scholars to provide evidence-based theory; 2) for linguists to provide the cues on which researchers may rely; and 3) for engineers to supply powerful natural linguistic processing algorithms to extract and analyze verbal data. The knowledge acquired from the coworking of these fields would be fine-tuned by discussing their practical application with legal practitioners to provide them with the best tools possible for verbal deception detection. These practical implications involve, for instance, deception detection in online forms allowing the filing of a report, the analysis of potentially deceptive e-mails, or the quest for truth in post-interview transcribed verbatims.

This tripartite collaboration will need to cover topics beyond the mere detection of deception. Deceptive verbal data can be found in many settings, ranging from insurance claims to opinions about hotels, including film reviews, investigative interviewing, malingering in a medical context. Some involve immediate interaction, while others may foster computer-mediated interaction, or no interaction at all. There is an increasing necessity to determine to what extent the context influences verbal deceit and challenges the idea behind the one-fit-for-all equation solving the deception detection problem (Sánchez-Junquera et al., 2020; Tomas et al., 2021b; Demarchi et al., 2021). The potential implications in the aforementioned contexts may involve computational methods beyond stylometry: If one relies on natural language processing, machine-learning, and artificial intelligence, there is increasing caution required on the ethics behind data collection, and the biases involved in the artificial intelligence literature. Data collected on the field, whether in insurance, investigative interviewing, medical, or in judicial context, are not collected for the purpose of deception detection: They are information gathered with the goal of making the best decision possible. If these private data are used for anything other than the purpose for which they were collected, it is ethically problematic to use them for deception detection, and more importantly so with the current detection rates. An essential discussion needs to happen regarding the treatment of verbal data for deception detection purposes.

Similarly (big) data evaluation through algorithmic and machine-learning means have also been pointed out for their biases. In machine-learning and artificial intelligence, bias refers to a priori knowledge and potential stereotypes that may lead to prejudice. Recent research has suggested that, like human beings, algorithms absorb stereotyped data and may taint, such as flowers and insects, with positive and negative emotions respectively because of the context in which they appear (Caliskan et al., 2017). If the flower-insect question seems trivial, studies have shown that natural processing trained on available data seemed to develop sexist representations of linguistic data, where men are identified as maestros and women as homemakers (Bolubasi et al., 2016; Caliskan et al., 2017; Papakyriakopoulos et al., 2019). Similarly, a race bias has also been observed (Manzini et al., 2019). It is thus extremely important to discuss these ethical issues in case of machine-learning based deception detection. Ethics are little discussed in the deception detection literature, but if, as developed by Vrij and Fisher (2016), stylometry or other verbal deception detection techniques may be used in the field, there is an urgent necessity to consider the ethics of data collection.

Computational stylometry offers a set of tools that can help academics and practitioners in insurance, justice, and human resources settings detect potentially deceptive verbal accounts. By relying on the power of computers and extraction algorithms, stylometry would allow the desired cues to be extracted more quickly and more robustly (agreement, reproducibility) than a solely manual work. Its development will constitute an important theoretical and practical advance in the field of deception detection. The more recent model-based approach, despite being considered as a black box, might provide better accuracy in terms of deception detection. Ethical questions regarding their use on the field need to be considered with absolute necessity. For this, we highly advocate for the combination of knowledge and skill from communication scientists, psychology and (psycho)linguistics researchers, philosophers, and computational scientists.