Key findings
- ≤0.25Jaccard similarity with the peak-career configuration at the stage of crossing lines of loyalty, in all 13 participants
- p = .005configurational distance from the peak-career configuration was greater at crossing the lines of loyalty than at Crisis 1 and at Crisis 2 (adjusted p)
- 69.2%of participants had Anxiousness in their configuration at crossing the lines, against none at peak career
- 93.7%agreement when the same narratives were coded a second time, across 1,300 facet-level comparisons (Cohen's κ = 0.78)
Abstract
- Background
- Life narratives contain rich information about how psychological characteristics are expressed across different stages of a person’s life. Translating this material into systematic psychological representations that can be compared within individuals across biographical stages remains a methodological challenge.
- Objective
- This study examined an artificial intelligence-supported method for identifying psychological configurations in life narratives and comparing them across different stages within the same individual.
- Methods
- We analyzed in-depth interviews with 13 former members of criminal organizations who subsequently became state witnesses. Each narrative was divided into 10 biographical stages, which were coded independently by a large language model using a predefined theoretical framework comprising the 25 personality facets of the Personality Inventory for the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (PID-5). The model was not given access to configurations generated from other stages. Configurations were compared within individuals across the peak of their criminal careers, subsequent crises, and the stage of crossing lines of loyalty.
- Results
- Overlap between the peak-career configuration and configurations from earlier crisis stages varied considerably across participants. In contrast, at the stage of crossing lines of loyalty, all 13 participants showed low overlap with their peak-career configuration (Jaccard similarity coefficient ≤0.25). Configurational composition also differed across stages. Grandiosity and Risk Taking were more frequently represented at the peak of the criminal career, whereas Anxiousness, Depressivity, and Suspiciousness were more frequently represented at the crossing stage.
- Conclusion
- The findings demonstrate that a large language model operating within a predefined psychological framework can systematically code life narratives and generate multidimensional configurations for comparison across biographical stages. This approach extends artificial intelligence-supported psychological analysis beyond individual characteristics assessed at a single point in time to configurations represented across a biographical sequence. The findings reflect psychological characteristics as represented and coded in the narratives rather than directly measured changes in personality, providing a methodological basis for studying intraindividual variation in complex biographical processes.
Introduction
Betrayal involves a violation of a previously established relationship of trust and loyalty and may be accompanied by loss of trust, relational disruption, and substantial changes in the experience of the person who has been betrayed.1 Among state witnesses from criminal organizations, crossing lines of loyalty has a particular significance. A person who was previously embedded in a group organized around loyalty, secrecy, and mutual dependence begins to act against its members. The importance of trust and loyalty in criminal organizations has also been documented in research on criminal networks, where trust plays a central role in maintaining and coordinating relationships among group members.2,3, The psychological question is therefore not only why a person violates an established loyalty, but how a process develops through which an action that was previously unthinkable becomes possible.
Early explanations of loyalty violations focused primarily on characteristics of the individual. Research has linked betrayal and deviations from loyalty to personality characteristics, particularly antagonistic features such as narcissism, manipulativeness, and psychopathy.4 At the same time, research on espionage and insider threat has moved beyond the search for a particular “type” of person who may become a threat toward models that describe a process in which individual vulnerabilities interact with stressors, life circumstances, and decision-making over time.5,6 Recent research similarly suggests that personality characteristics and situational factors are associated with willingness to report insider-threat behavior.7
This shift from traits to processes, however, leaves an important question unanswered: what changes within the individual over the course of such a process? If certain characteristics have been present for years, their presence alone cannot explain why a crossing of loyalty boundaries occurs at one point rather than another. Likewise, a crisis by itself may not provide a sufficient explanation when the same individual has experienced earlier crises without crossing those boundaries. The question therefore shifts from identifying characteristics that distinguish people to examining how multiple psychological characteristics are expressed and organized within the same person over time.
This perspective is consistent with dynamic approaches to personality. Whole Trait Theory describes traits in terms of distributions of momentary states, distinguishing relative trait stability from variation in how traits are expressed across situations and contexts.8 From this perspective, changes observed during a crisis do not necessarily imply a lasting change in personality structure. They may instead reflect changes in how psychological characteristics are expressed and in the pattern in which they occur together.
From Traits to Configurations
This distinction is consistent with dimensional models of personality and psychopathology. Krueger et al.9 developed the Personality Inventory for the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), known as the PID-5, which assesses 25 maladaptive personality traits organized into five broad domains. The facet-level structure allows multiple emotional, interpersonal, and behavioral characteristics to be represented simultaneously rather than reducing an individual to a single category or trait.
When the research question concerns change over time, however, examining each characteristic separately may not be sufficient. Person-centered approaches to personality research examine profiles formed by combinations of multiple characteristics. A systematic review of this literature found that personality profiles can differ both in the levels of individual characteristics and in their relative patterning, with distinct configurations of personality characteristics emerging across profiles.10 This supports the use of a psychological configuration as a unit of analysis: the set of characteristics represented together at a given point in the narrative.
A configurational approach becomes particularly relevant when applied within the same individual over time. Rather than asking only which characteristics are present, it allows examination of whether their organization is maintained or changes across different stages. In the context of crossing lines of loyalty, the configuration represented at that stage can be compared with configurations represented earlier in the same person's narrative, particularly during the peak of criminal involvement and preceding crises. The individual therefore serves as an internal reference point, allowing the analysis to focus on intraindividual change.
In the present study, the 25 PID-5 facets were used to represent multidimensional psychological configurations and to compare their similarity and distance across biographical stages within the same individual.
From Life Narratives to Psychological Coding
Examining intraindividual change across a biographical sequence requires information about different stages of a person's life. Life narratives provide a rich source of such information because they contain accounts of events, relationships, emotions, behaviors, and meanings attributed to personal experiences. Research on language and computerized text analysis has shown that the way people use language can provide systematic information about psychological and social processes.11,12
Research on computational personality assessment has further demonstrated that naturally occurring language can provide information about psychological characteristics. Park et al.13 showed that personality traits could be estimated from language use and evaluated against established measures of personality. Their findings provide an empirical foundation for examining personality-related information in natural language, beyond the identification of isolated words or expressions.
The methodological challenge is to transform complex narrative material into a systematic psychological representation. Psychological characteristics are not always expressed explicitly, and their meaning may depend on the context in which they appear. The challenge becomes greater when several psychological dimensions must be identified and the same coding procedure applied consistently across different stages of an individual's life.
Grimmer and Stewart14 emphasized that automated text analysis requires careful alignment between theoretical concepts and computational measurement, together with validation appropriate to the research question and the context of application. This perspective underscores the importance of predefined conceptual frameworks and cautious interpretation of representations derived from textual data.
Computational approaches to text analysis provide one way of addressing this problem. Neuman and Cohen proposed a computational approach to personality assessment in which psychological constructs provide the basis for representing and measuring personality characteristics from text.15 This approach is relevant to the present study because it separates the psychological model from the computational procedure: the theoretical framework specifies the constructs to be examined, while the computational method is used to identify their expression in text.
The development of large language models (LLMs) has expanded the possibilities for this type of analysis. Demszky et al.16 reviewed applications of LLMs in psychological research, including the analysis and generation of language data, while emphasizing the need for systematic evaluation of their reliability and limitations. Related studies have also demonstrated the use of predefined psychological knowledge in computational analyses of social phenomena and personality-related language.17,18 These studies support the integration of psychological theory with computational analysis while underscoring the need to establish the validity and stability of the resulting coding.
More recently, Brickman et al.19 examined the use of large language models for psychological assessment, emphasizing the importance of clearly defined psychological constructs, appropriate analytic procedures, and systematic evaluation of measurement validity. These considerations are particularly relevant when language models are used to identify expressions of psychological characteristics in naturalistic text within a predefined theoretical framework.
Recent research has also extended the application of generative language models to the qualitative coding of complex textual material. Than et al.20 demonstrated their potential to identify multidimensional concepts in extended texts, including concepts that are expressed implicitly and require contextual interpretation rather than simple keyword identification. These findings are particularly relevant to life narratives, in which psychological characteristics may be expressed through descriptions of events, behaviors, and interpersonal relationships rather than through explicit statements about personality.
In psychological research, Bunt et al.21 further demonstrated the potential of large language models to classify psychologically defined phenomena in text, while emphasizing the importance of semantic, predictive, and content validity in evaluating the resulting classifications. Their findings highlight the distinction between the consistency of computational coding and the validity of the psychological constructs that the coding is intended to represent.
A particularly relevant example is the work of Neuman and Hazdai, who used an LLM within a predefined theoretical framework to analyze attachment patterns in life narratives across biographical stages.22 In that approach, attachment theory defined the constructs to be identified, while the language model was used to code their expression in the narrative. The present study applies a similar principle to the PID-5. The LLM was not asked to diagnose participants or to infer their personality without constraints. Instead, the 25 PID-5 facets provided a fixed theoretical framework for coding psychological expressions in the narratives. The resulting scores are not psychometric PID-5 scores and do not constitute clinical diagnoses. This distinction is particularly important given the growing use of LLMs in mental health research and the need to distinguish computational performance from the validity of psychological measurement.23–25
The Present Study
The present sample provides an opportunity to examine this approach in a distinctive population. All participants eventually crossed lines of loyalty, but their life narratives also included earlier stages and crises. This structure makes it possible to use each individual as an internal reference and to examine how psychological configurations represented in the narratives vary across the biographical sequence leading to crossing lines of loyalty.
The present study examined whether theory-guided LLM analysis could identify multidimensional configurations of psychological characteristics in naturalistic life narratives and examine how these configurations varied across biographical stages. Rather than focusing on individual facets or an overall psychological classification, the study represented psychological expression in the narrative as configurations of PID-5 facets and compared these configurations within individuals across stages.
Two research questions were examined: (1) whether the psychological configurations represented in the narratives varied across different biographical stages within the same individual; and (2) what differences in configuration composition emerged between the peak of criminal involvement, subsequent crises, and the stage of crossing lines of loyalty.
We hypothesized that the configurations represented in the narratives would vary across the biographical sequence. Specifically, we hypothesized that the configuration at the stage of crossing lines of loyalty would show less overlap with the peak-career configuration than the configurations observed during the preceding crises.
Methods
Participants
The study was based on semi-structured, in-depth interviews with 13 former senior members of organized crime groups who crossed the lines of loyalty and testified against members of their own organizations. The interviews were conducted in Israel between 2018 and 2020. The participants represented approximately 80% of those formally recognized as state witnesses by the competent authorities during the study period; the remaining 20% were not accessible for participation. Participants were recruited with the assistance of law enforcement agencies and the Witness Protection Authority. Recruitment and data collection were conducted with the approval and oversight of the Ministry of National Security.
All participants provided written informed consent. Institutional review board approval was obtained from the governmental authority responsible for witness protection (IRB identifier: 08384917). The interview protocol was also approved by the Ethics Committee of Bar-Ilan University on February 1, 2018.
Interview procedure
The study used a narrative approach centered on the chronology of participants' lives. Individual interviews were conducted by two experienced clinicians. Participants were asked to describe their life histories from early childhood through adolescence, entry into criminal activity, integration into the criminal organization, subsequent crises, and the events and processes that ultimately led to crossing the lines of loyalty and cooperating with law enforcement authorities.
Interviews lasted approximately 2.5 to 6 hours, with breaks provided according to participants' needs. All interviews were recorded and transcribed in full.
Division of narratives into biographical stages
Following a holistic reading of the interviews, the two researchers divided the narratives into 10 common biographical stages. The division was based on the chronology of each life history and on the sequence of events, crises, and resolutions described in the narratives. The same stage structure was applied across participants to permit within-person comparison across corresponding points in the life course.
The 10 stages were: (1) childhood; (2) early adolescence; (3) adolescence and the early stages of the criminal career; (4) peak criminal career; (5) Crisis 1; (6) resolution of Crisis 1; (7) Crisis 2; (8) resolution of Crisis 2; (9) the final crisis preceding the crossing of the lines of loyalty; and (10) crossing the lines of loyalty, defined as the stage in which the participant's transition from loyalty to the criminal organization to cooperation with law enforcement was established.
The large language model was not used to determine the number, meaning, or boundaries of the biographical stages. Stage definitions were established before the psychological coding procedure.
Large Language Model–Based Psychological Coding
Each biographical unit was analyzed using ChatGPT with the GPT-5.6 Luna model (OpenAI), applying the same standardized coding prompt to all narrative units. The pretrained language model was used as a structured coding tool for identifying psychological characteristics expressed in the narratives. It was not trained or fine-tuned on the study data, nor was it used to make clinical diagnoses. Rather, its application was intended to examine the applicability of a theory-guided approach to mapping psychological configurations across successive biographical stages.
Psychological coding was based on the 25 personality facets of the Personality Inventory for DSM-5 (PID-5).9 The definitions of the 25 facets were specified in advance in the coding prompt. For each facet, the model was instructed to identify supporting evidence in the narrative and assign an intensity score from 0 to 4 according to the predefined coding instructions.
The coding focused on the psychological expression of a facet rather than on the severity or nature of the event or behavior described. The model was specifically instructed not to infer a personality characteristic solely from the presence of criminal, violent, dangerous, or otherwise high-risk behavior. Each coding decision was required to be supported by evidence contained in the narrative.
Each biographical unit was analyzed independently. When coding a given unit, the model received only the text belonging to that unit and was not provided with configurations generated from other stages of the same participant.
The complete definitions of the 25 PID-5 facets used in the coding procedure are provided in Appendix 1.
Coding Prompt
The same standardized prompt was used for all biographical units:
Personality Inventory for DSM-5 Facet Coding Prompt. Read and analyze the following narrative text using the dimensional personality-trait framework of the PID-5 (Appendix 1). The framework defines 25 personality facets representing patterns of emotional experience, cognition, self- and interpersonal perception, and behavior.
For the purposes of this analysis, identify evidence of each PID-5 facet in the narrative and rate the intensity of its expression on a 0–4 scale according to the predefined coding instructions.
Step 1: Identify relevant material. Identify direct or indirect descriptions that may provide evidence of a personality facet.
Step 2: Classify and rate. For each relevant expression:
- Assign it to the PID-5 facet it most directly represents.
- Quote the specific sentence or passage supporting the classification.
- Determine the degree to which the evidence corresponds to the facet definition.
- Assign an intensity score from 0 to 4 according to the predefined coding instructions.
Core instruction: Rate the psychological expression of the facet, not the severity of the event or behavior described. Do not infer a personality facet solely from the presence of criminal, violent, dangerous, or otherwise high-risk behavior. The coding decision must be supported by evidence contained in the narrative.
Analyze each biographical unit independently. Do not use configurations generated from other stages of the same participant when coding the current unit.
The complete definitions of the 25 PID-5 facets are provided in Appendix 1.
Configurational Similarity and Distance
Configurational similarity was assessed using the Jaccard similarity coefficient. For each participant, the Crisis 1, Crisis 2, and crossing-of-the-lines configurations were compared with the participant's peak-career configuration. The Jaccard coefficient was calculated as:
J(A, B) = |A ∩ B| / |A ∪ B|
where A and B represent the sets of included PID-5 facets in the two configurations. Values range from 0, indicating no shared facets, to 1, indicating identical configurations.
Configurational distance was calculated separately using the original 0–4 facet scores. For each comparison stage, distance was defined as the sum of the absolute differences across the 25 PID-5 facets:
D = ∑j=125 |xj − xj,peak|
where xj is the facet score at the comparison stage and xj,peak is the corresponding score at peak career. Higher values indicate greater configurational distance from the peak-career profile.
Repeatability of the Coding Procedure
To examine the repeatability of the coding procedure, the same narrative material was processed a second time using the same standardized coding protocol. The repeated analysis used the same biographical stages, facet definitions, scoring instructions, and configuration threshold as the primary analysis.
Example of the Coding Procedure
Table 1 presents a shortened example from the peak criminal-career stage of one participant (pseudonym: “Eric”) and illustrates the coding process from textual evidence to facet identification and intensity rating.
Table 1.
| PID-5 facet | Intensity score | Evidence from narrative | Coding rationale |
|---|---|---|---|
| Grandiosity | 4 | “It brought me to a position... almost like an arbitrator, perhaps the youngest there was”; “I was the head of the company.” | Self-description emphasizing elevated status, authority, and an exceptional position relative to others. |
| Hostility | 4 | “The violent part, the part that requires force, leave that to me.” | Direct expression of willingness to use force within the group. |
| Rigid Perfectionism | 3 | “Look at every group I joined. I integrated into the hierarchy at a very high level. I am a perfectionist.” | Explicit reference to perfectionism and demanding standards concerning conduct and performance. |
This example illustrates that coding was based on specific textual evidence evaluated against the predefined facet definitions rather than on the participant's involvement in criminal activity itself.
Results
4.1 Configurational similarity and distance relative to the peak criminal-career configuration
Configurational distance differed significantly across the three stages, Friedman χ2(2) = 13.51, p = .001. In post hoc Wilcoxon tests with Benjamini-Hochberg correction, configurational distance at crossing the lines of loyalty was greater than at Crisis 1 (adjusted p = .005) and Crisis 2 (adjusted p = .005). No significant difference was observed between Crisis 1 and Crisis 2 (adjusted p = .212).
Table 2. Configurational Similarity Relative to the Peak Criminal-Career Configuration
| Participant | Peak career (stage 4) | Crisis 1 (stage 5) | Crisis 2 (stage 7) | Crossing the lines (stage 10) |
|---|---|---|---|---|
| 1 | 1.00 | 0.36 | 0.50 | 0.14 |
| 2 | 1.00 | 0.23 | 0.82 | 0.13 |
| 3 | 1.00 | 0.40 | 0.00 | 0.22 |
| 4 | 1.00 | 0.50 | 0.57 | 0.25 |
| 5 | 1.00 | 0.40 | 0.00 | 0.25 |
| 6 | 1.00 | 0.38 | 0.20 | 0.22 |
| 7 | 1.00 | 0.00 | 0.14 | 0.00 |
| 8 | 1.00 | 0.33 | 0.33 | 0.20 |
| 9 | 1.00 | 0.17 | 0.17 | 0.17 |
| 10 | 1.00 | 0.14 | 0.00 | 0.14 |
| 11 | 1.00 | 0.00 | 0.00 | 0.25 |
| 12 | 1.00 | 1.00 | 0.00 | 0.14 |
| 13 | 1.00 | 0.50 | 0.50 | 0.25 |
Note. Values represent Jaccard similarity coefficients relative to each participant's peak criminal-career configuration. The value of 1.00 in the peak-career column represents the reference configuration and is not an independent comparison.
For Crisis 1, the median similarity with the peak-career configuration was 0.36 (IQR = 0.17–0.40; range = 0–1.00). For Crisis 2, the median similarity was 0.17 (IQR = 0–0.50; range = 0–0.82). At the stage of crossing the lines of loyalty, Jaccard similarity values ranged from 0 to 0.25.
4.2 Composition of psychological configurations across stages
The frequency of the 25 PID-5 facets included in the configurations at the four focal stages is presented in Table 3.
Table 3. Frequency of PID-5 Facets Included in the Configurations by Stage, n (%)
| PID-5 facet | Peak career (stage 4) | Crisis 1 (stage 5) | Crisis 2 (stage 7) | Crossing the lines (stage 10) |
|---|---|---|---|---|
| Anhedonia | 0 (0.0%) | 0 (0.0%) | 1 (7.7%) | 1 (7.7%) |
| Anxiousness | 0 (0.0%) | 5 (38.5%) | 5 (38.5%) | 9 (69.2%) |
| Attention Seeking | 5 (38.5%) | 1 (7.7%) | 1 (7.7%) | 0 (0.0%) |
| Callousness | 7 (53.8%) | 2 (15.4%) | 3 (23.1%) | 1 (7.7%) |
| Deceitfulness | 1 (7.7%) | 2 (15.4%) | 0 (0.0%) | 1 (7.7%) |
| Depressivity | 0 (0.0%) | 0 (0.0%) | 4 (30.8%) | 8 (61.5%) |
| Distractibility | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Eccentricity | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Emotional Lability | 0 (0.0%) | 2 (15.4%) | 2 (15.4%) | 5 (38.5%) |
| Grandiosity | 10 (76.9%) | 3 (23.1%) | 3 (23.1%) | 0 (0.0%) |
| Hostility | 8 (61.5%) | 8 (61.5%) | 7 (53.8%) | 7 (53.8%) |
| Impulsivity | 4 (30.8%) | 2 (15.4%) | 2 (15.4%) | 1 (7.7%) |
| Intimacy Avoidance | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Irresponsibility | 3 (23.1%) | 2 (15.4%) | 1 (7.7%) | 0 (0.0%) |
| Manipulativeness | 5 (38.5%) | 2 (15.4%) | 2 (15.4%) | 1 (7.7%) |
| Perceptual Dysregulation | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Perseveration | 11 (84.6%) | 10 (76.9%) | 9 (69.2%) | 13 (100%) |
| Restricted Affectivity | 2 (15.4%) | 0 (0.0%) | 1 (7.7%) | 0 (0.0%) |
| Rigid Perfectionism | 3 (23.1%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Risk Taking | 11 (84.6%) | 6 (46.2%) | 5 (38.5%) | 1 (7.7%) |
| Separation Insecurity | 1 (7.7%) | 2 (15.4%) | 0 (0.0%) | 4 (30.8%) |
| Submissiveness | 1 (7.7%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Suspiciousness | 0 (0.0%) | 8 (61.5%) | 7 (53.8%) | 7 (53.8%) |
| Unusual Beliefs and Experiences | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Withdrawal | 0 (0.0%) | 0 (0.0%) | 2 (15.4%) | 2 (15.4%) |
Note. Each cell shows the number and percentage of participants (N = 13) for whom the facet received an intensity score of ≥2 and was therefore included in the configuration at that stage.
The composition of the configurations differed between the peak-career and crossing stages. At the peak-career stage, Perseveration, Risk Taking, and Grandiosity were among the most frequently represented facets. At crossing, Anxiousness, Depressivity, and Suspiciousness were more frequently represented, while Perseveration remained highly prevalent across both stages.
4.3 Sensitivity analysis: narrative length
No association was observed between narrative-unit length and the number of PID-5 facets included in the configuration (Spearman's ρ = -0.03, n = 52).
Stage-specific correlations were small: peak career, ρ = -0.08; Crisis 1, ρ = -0.25; Crisis 2, ρ = 0.08; and crossing the lines of loyalty, ρ = -0.18.
4.4 Reproducibility of the coding procedure
The reproducibility of the coding procedure was assessed by independently reprocessing the same narrative material using the identical coding protocol, including the same 25 PID-5 facets, 0–4 scoring scale, and ≥2 inclusion threshold. Across 1,300 facet-level coding comparisons, the repeated coding showed 93.7% overall agreement, with Cohen's κ = 0.78. The mean Jaccard similarity between the configurations generated in the two coding runs was 0.73.
Discussion
The present study examined whether a large language model (LLM), operating within a predefined theoretical framework based on the 25 PID-5 facets, could be used to systematically analyze psychological configurations derived from life narratives. The PID-5 defined the set of characteristics examined, while the LLM was used to code expressions of these characteristics within their narrative context. This approach made it possible to derive multidimensional configurations from different parts of the narratives and to examine the relationships among them using quantitative measures.
The main finding was a marked reorganization of the psychological configurations represented in the narratives around the crossing of lines of loyalty. The greatest departure from the peak-career configuration occurred at the crossing stage, accompanied by a shift in the relative representation of several personality facets. This pattern suggests a change in the narrative representation and relative organization of psychological characteristics, rather than evidence of a change in underlying personality structure or the prominence of a single characteristic.
Implications for research on crossing the lines of loyalty
In the context of crossing the lines of loyalty, the present findings add an intra-individual perspective to our understanding of the transition from loyalty to a group to the violation of that loyalty. Previous research has emphasized the interplay between personality vulnerabilities, situational factors, and processes unfolding on the path toward insider threat behavior,5,6 while recent research has identified associations between personality characteristics and situational conditions in reported willingness to engage in such behavior.7 The present study adds to this literature by examining changes in the psychological composition represented in the narrative around critical turning points in the process.
Whole Trait Theory distinguishes between the relative stability of traits and variation in how they are expressed across situations and contexts.8 This distinction provides a useful framework for understanding the present findings: the relative prominence of psychological characteristics may change across the course of a process even when the underlying traits themselves are relatively stable. In the present study, the configuration provides a way of describing this change at the level of narrative representation. It captures which characteristics appear together, which become more or less prominent, and how this pattern changes around the turning points associated with crossing the lines of loyalty.
This perspective broadens the research question beyond identifying individual characteristics associated with betrayal or a breach of loyalty. It makes it possible to examine how several characteristics operate together within a given life course and how their relative patterning is expressed during periods of relative stability, crisis, and transition. In this sense, the configuration provides a useful unit of analysis for processes in which psychological meaning emerges from the combination of several characteristics rather than from any single characteristic alone.
Implications for psychological research and mental health
The broader significance of this approach lies in the possibility of using natural language data as a source of multidimensional psychological representations. Narratives, interviews, and self-reports contain information about emotions, relationships, behaviors, and perceptions, making them a rich source of psychological information. Previous research has shown that language can provide information about psychological processes and aspects of personality, and that such information can be analyzed systematically using computational methods.11–13
In mental health research, natural language processing has been used to analyze information arising in therapeutic processes and to examine processes associated with psychological interventions.23 The use of LLMs extends this possibility through their ability to process open-ended text while taking into account the context in which it occurs. Combining this capability with systematic evaluation of validity, stability, bias, and safety is an important requirement for developing applications in mental health, particularly when such systems are intended to support assessment or treatment.16,24,25
From a research perspective, this approach makes it possible to examine rich narrative material at a finer level of resolution than a measure based on a single time point. Rather than producing an overall estimate of a personality characteristic, separate configurations can be derived from different parts of a narrative and their relationships examined. The same principle may be relevant to other narrative sources, such as repeated interviews, diaries, and therapy transcripts, where psychological characteristics can be tracked as they are represented over time.
Methodological contribution
The methodological contribution of the present study lies in linking a predefined psychological framework with the systematic analysis of open-ended narrative text. Neuman and Cohen demonstrated a computational approach to personality assessment based on semantic representations of text, in which the psychological construct defines the space of computational analysis.15 Subsequent work by Neuman and Cohen extended this principle to other forms of computational analysis, including the use of psychologically grounded models to identify social norm violations and the construction of AI-generated language data linked to theoretically defined personality dimensions.17,18 Demszky et al. likewise highlighted the potential of LLMs for psychological research while emphasizing the need for systematic evaluation of their validity, reliability, and limitations.16
In the present study, the PID-5 framework defined the coding space in advance, and the LLM applied this framework to separate narrative units. This made it possible to transform complex narrative material into multidimensional representations with a consistent structure that could be examined quantitatively. The PID-5 comprises 25 personality facets organized into five broad domains, allowing multiple psychological dimensions to be considered simultaneously.9 Person-centered approaches to personality research likewise provide a basis for examining patterns formed by combinations of characteristics and the relationships among them.10
A related approach was demonstrated in earlier work by Neuman and Hazdai, in which an LLM was used within a predefined theoretical framework to analyze attachment patterns in life narratives.22 The present study extends this principle from attachment styles to a multidimensional framework comprising 25 personality facets and adds the possibility of quantitatively assessing the similarity and distance between the resulting configurations.
The methodological value of this approach therefore lies in producing repeated, multidimensional, theory-based representations of narrative material and then examining the relationships among them systematically. The procedure links psychological theory, narrative text, computational coding, and quantitative analysis, providing a framework that can be tested and replicated in future research. Recent research on qualitative coding with generative language models further supports the relevance of examining complex, context-dependent concepts in extended textual material, while highlighting the need for careful methodological evaluation.20
The repeated-coding analysis provides additional support for this aspect of the method, as the procedure produced consistent representations when the same material was analyzed again under the same conditions. This finding provides a basis for using the coding protocol as a structured procedure for deriving psychological representations from narratives and highlights the importance of further testing the approach across different models, coding conditions, and samples.
An important distinction concerns the reproducibility of the coding procedure and the validity of the psychological representations it produces. Although repeated coding yielded consistent configurations under the same conditions, this consistency does not, by itself, establish that the identified facets correspond to independently assessed personality characteristics. Recent methodological research emphasizes that the evidentiary requirements for using language models as text-classification tools differ from those required to support broader claims about psychological constructs.19,21,26 Accordingly, the present findings provide initial evidence of procedural reproducibility, while independent validation is needed to establish the accuracy and construct validity of the resulting psychological classifications.
Limitations and Future Directions
The study represents an initial demonstration of the approach in a small and distinctive sample. The 13 participants constituted the empirical sample used to evaluate the proposed coding procedure; their interviews were not used to train the language model, which was pretrained and subsequently applied to the narrative material using predefined coding instructions. The sample therefore limits the generalizability of the observed configurational patterns and does not provide a basis for population-level inference. Future research should examine the method in larger and more diverse samples and in populations with different biographical and clinical characteristics, thereby providing a broader basis for evaluating the validity and generalizability of the approach.
The analysis was based on retrospective narratives. Accordingly, the configurations reflect how psychological characteristics were represented and coded in the narratives at the different stages. The findings therefore describe changes in the narrative representation of these characteristics across the life course. The PID-5 served as a theoretical framework for coding, and the resulting representations reflect the outcome of the coding process within this framework rather than psychometric PID-5 scores. Consequently, the findings should not be interpreted as direct evidence of changes in participants' underlying personality traits over time.
Coding results may be influenced by the model, the coding instructions, and the way the threshold for including a characteristic in a configuration is defined. Future studies could examine the stability of the findings across different models, alternative coding protocols, and different inclusion thresholds. Incorporating prospective and longitudinal data would make it possible to evaluate the approach in settings where psychological information is collected over time and closer to the occurrence of the relevant events.
A further limitation concerns the potential for systematic bias in language-model-based coding. Ashwin et al.27 demonstrated that classification errors in the automated analysis of qualitative interviews may be associated with characteristics of the interviewees rather than occurring randomly. This concern is relevant to the present study because the analysis was based on retrospective narratives from a distinctive population. Future research should therefore examine the correspondence between model-generated classifications and independent human coding, as well as the stability of the resulting configurations across alternative models, coding instructions, and narrative contexts.
Conclusion
The study proposes a theory-guided framework for deriving psychological configurations from life narratives and examining changes in their composition across the life course. The findings indicate a meaningful change in the narrative representation and composition of the configurations around the crossing of the lines of loyalty, alongside the continued presence of certain characteristics across the stages examined.
The central contribution of the study lies in combining an intra-individual perspective on the process of crossing the lines of loyalty with a multidimensional, theory-based representation of personality characteristics. This approach provides a way to examine how several psychological characteristics are expressed together and how their relative prominence changes across complex biographical processes. Methodologically, the study demonstrates how a predefined psychological framework can be combined with an LLM to derive structured representations from narrative text and examine the relationships among those representations quantitatively.
Declarations
- Conflict of Interest
- The authors declare no conflicts of interest.
- Funding
- No funding was received for this study.
Appendix 1. Definitions of the 25 PID-5 Facets
- Anhedonia: Reduced capacity to experience pleasure, interest, or satisfaction.
- Anxiousness: A tendency toward worry, tension, fear, heightened vigilance, and anticipation of negative outcomes.
- Attention Seeking: Behavior aimed at attracting attention and making oneself the focus of others’ attention.
- Callousness: Limited concern for the feelings or well-being of others, accompanied by reduced guilt, remorse, or empathy.
- Deceitfulness: A tendency to lie, mislead, distort information, or present information falsely.
- Depressivity: A tendency toward low mood, hopelessness, pessimism, shame, or feelings of worthlessness.
- Distractibility: Difficulty maintaining attention and goal-directed behavior because of distractions.
- Eccentricity: Unusual or unconventional behavior, thoughts, appearance, or speech.
- Emotional Lability: Emotional instability and frequent or rapid shifts in emotional state.
- Grandiosity: A sense of superiority, self-importance, or entitlement, together with an exaggerated view of one’s abilities or status.
- Hostility: A tendency toward anger, irritability, resentment, antagonism, or vindictiveness.
- Impulsivity: Acting on immediate urges without sufficient planning or consideration of the consequences.
- Intimacy Avoidance: Avoidance of closeness and deep intimate or emotional relationships.
- Irresponsibility: Failure to meet duties, obligations, agreements, or responsibilities toward others.
- Manipulativeness: The use of interpersonal strategies to influence or control others in pursuit of personal goals.
- Perceptual Dysregulation: Unusual experiences or disruptions in the perception and experience of oneself or the surrounding environment.
- Perseveration: Continuing a thought, task, or course of action despite a need or reason to change it.
- Restricted Affectivity: A limited range of emotional experience and expression.
- Rigid Perfectionism: A demand for high and inflexible standards, accompanied by difficulty adapting or completing a task when those standards are not met.
- Risk Taking: Engagement in dangerous activities with limited consideration of their potential negative consequences.
- Separation Insecurity: Fear of rejection, loss, or separation from significant others, together with a need to maintain closeness to them.
- Submissiveness: Adapting one’s behavior to the wishes of others, even at the expense of one’s own needs or preferences.
- Suspiciousness: Expectation of harm, exploitation, or deception, together with a tendency to attribute negative intentions to others.
- Unusual Beliefs and Experiences: Unusual beliefs, thoughts, or experiences concerning reality.
- Withdrawal: A preference for keeping away from others, avoidance of social interaction, and reduced interpersonal contact.
References
- Rachman S. Betrayal: a psychological analysis. Behav Res Ther. 2010;48(4):304-311. doi:10.1016/j.brat.2009.12.002
- Gambetta D. Codes of the Underworld: How Criminals Communicate. Princeton University Press; 2009.
- von Lampe K, Johansen PO. Organized crime and trust: on the conceptualization and empirical relevance of trust in the context of criminal networks. Global Crime. 2004;6(2):159-184. doi:10.1080/17440570500096734
- Paulhus DL, Williams KM. The dark triad of personality: narcissism, Machiavellianism, and psychopathy. J Res Pers. 2002;36(6):556-563. doi:10.1016/S0092-6566(02)00505-6
- Shaw ED, Ruby KG, Post JM. The insider threat to information systems: the psychology of the dangerous insider. Security Awareness Bulletin. 1998;(2-98).
- Shaw ED, Sellers L. Application of the critical-path method to evaluate insider risks. Stud Intell. 2015;59(2):41-48.
- Forsyth L, Ewens T, Anglim J. Personality and situational predictors of insider threat: a vignette study. Aust J Psychol. 2026;78(1):2628384. doi:10.1080/00049530.2026.2628384
- Fleeson W, Jayawickreme E. Whole Trait Theory. J Res Pers. 2015;56:82-92. doi:10.1016/j.jrp.2014.10.009
- Krueger RF, Derringer J, Markon KE, Watson D, Skodol AE. Initial construction of a maladaptive personality trait model and inventory for DSM-5. Psychol Med. 2012;42(9):1879-1890. doi:10.1017/S0033291711002674
- Yin K, Lee P, Sheldon OJ, Li C, Zhao J. Personality profiles based on the FFM: a systematic review with a person-centered approach. Pers Individ Dif. 2021;180:110996. doi:10.1016/j.paid.2021.110996
- Pennebaker JW, Mehl MR, Niederhoffer KG. Psychological aspects of natural language use: our words, our selves. Annu Rev Psychol. 2003;54:547-577. doi:10.1146/annurev.psych.54.101601.145041
- Tausczik YR, Pennebaker JW. The psychological meaning of words: LIWC and computerized text analysis methods. J Lang Soc Psychol. 2010;29(1):24-54. doi:10.1177/0261927X09351676
- Park G, Schwartz HA, Eichstaedt JC, et al. Automatic personality assessment through social media language. J Pers Soc Psychol. 2015;108(6):934-952. doi:10.1037/pspp0000020
- Grimmer J, Stewart BM. Text as data: the promise and pitfalls of automatic content analysis methods for political texts. Polit Anal. 2013;21(3):267-297. doi:10.1093/pan/mps028
- Neuman Y, Cohen Y. A vectorial semantics approach to personality assessment. Sci Rep. 2014;4:4761. doi:10.1038/srep04761
- Demszky D, Yang D, Yeager DS, et al. Using large language models in psychology. Nat Rev Psychol. 2023;2:688-701. doi:10.1038/s44159-023-00241-5
- Neuman Y, Cohen Y. AI for identifying social norm violation. Sci Rep. 2023;13:8103. doi:10.1038/s41598-023-35350-x
- Neuman Y, Cohen Y. A data set of synthetic utterances for computational personality analysis. Sci Data. 2024;11:623. doi:10.1038/s41597-024-03488-6
- Brickman J, Gupta M, Oltmanns JR. Large language models for psychological assessment: a comprehensive overview. Adv Methods Pract Psychol Sci. 2025;8(3):25152459251343582. doi:10.1177/25152459251343582
- Than N, Fan L, Law T, Nelson LK, McCall L. Updating “the future of coding”: qualitative coding with generative large language models. Sociol Methods Res. 2025;54(3):849-888. doi:10.1177/00491241251339188
- Bunt HL, Goddard A, Reader TW, Gillespie A. Validating the use of large language models for psychological text classification. Front Soc Psychol. 2025;3:1460277. doi:10.3389/frsps.2025.1460277
- Neuman Y, Hazdai K. An AI-supported methodology for identifying attachment styles. medRxiv. Preprint posted September 4, 2025. doi:10.1101/2025.08.30.25334439
- Malgaroli M, Hull TD, Zech JM, Althoff T. Natural language processing for mental health interventions: a systematic review and research framework. Transl Psychiatry. 2023;13:309. doi:10.1038/s41398-023-02592-2
- Stade EC, Stirman SW, Ungar LH, et al. Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation. NPJ Ment Health Res. 2024;3:12. doi:10.1038/s44184-024-00056-z
- Hua Y, Na H, Li Z, et al. A scoping review of large language models for generative tasks in mental health care. NPJ Digit Med. 2025;8:230. doi:10.1038/s41746-025-01611-4
- Lin Z. From prompts to constructs: a dual-validity framework for large language model research in psychology. Annu Rev Psychol. Published online August 10, 2026. doi:10.1146/annurev-psych-100925-034807
- Ashwin J, Chhabra A, Rao V. Using large language models for qualitative analysis can introduce serious bias. Sociol Methods Res. 2026;55(3):795-839. doi:10.1177/00491241251338246
Article information
Authors
- Keren HazdaiRuppin Academic Center, IsraelCorresponding author · [email protected]
How to cite
Hazdai, K., 2026. From Traits to Configurations: An Artificial Intelligence Supported Methodology for Mapping Psychological Configurations Across Life Narratives. Medical Research Archives, [online] 14(10). https://doi.org/10.18103/mra.2026.0726
Record
- Journal
- Medical Research Archives
- Issue
- Vol 14 No 10 (2026)
- Section
- Research Articles
- Published
- 31 October 2026
- ISSN
- 2375-1924
- Downloads
- 68
Copyright and licence
Published by the European Society of Medicine under a Creative Commons Attribution 4.0 International licence (CC BY 4.0), which permits unrestricted use, distribution and reproduction in any medium, provided the original authors and source are credited.