Is your LLM an Agressive Narcissist with Depression? Analyzing Psychological Portraits of Large Language Models

Polina Fedorinina; Anna Fattakhova; Nadezhda Romanova; Ekaterina Lezina; Irina Nikishina

cs.AINov 2, 2026

Is your LLM an Agressive Narcissist with Depression? Analyzing Psychological Portraits of Large Language Models

Authors: Polina Fedorinina, Anna Fattakhova, Nadezhda Romanova, Ekaterina Lezina, Irina Nikishina

Organizations: HSE University · Moscow Institute of Psychoanalysis · MBZUAI

Abstract

Nowadays, interaction with AI requires caution not only regarding the factual reliability of generated content but also regarding personality markers that models may implicitly transmit during dialogue that can subtly influence users' perception of information. In this paper, we conduct a large-scale experimental study in which 22 large language models complete 11 psychological inventories. Furthermore, we investigate correlations between human and model responses to track the adjustments made for the users who interact with them on a daily basis. The results show that LLMs exhibit coherent and relatively stable psychological patterns. Personality traits remain largely consistent, whereas emotional and symptomatic dimensions are more variable, with anxiety and neuroticism emerging as dominant tendencies. Model explanations further suggest that LLMs generate internally consistent and safety-oriented profiles while also reflecting correlations with the users they regularly interact with.

Figures & tables

Figure 1: Comparison of normalized psychometric traits across API-based language models.

Figure 2: Overview of the experimental pipeline used in the study: four categories of psychological inventories were employed to construct psychological profiles of LLMs through API, conersational interfaces, and the humans, using such interfaces.

Category	Inventory	Alias	Original Source	Russian-language Source
Common mental disorders	Hospital Anxiety and Depression Scale	HADS	Zigmond and Snaith (1983)	Andriushchenko et al. (2003)
	Beck Anxiety Inventory	BAI	Beck et al. (1988)	Fedorenko and Kachay (2020)
	Beck Depression Inventory	BDI	Beck et al. (1961)	Tarabrina (2007)
	Generalized Anxiety Disorder Scale	GAD-7	Spitzer et al. (2006)	Zolotareva (2023)
	Symptom Checklist–90	SCL-90	Derogatis et al. (1973)	Tarabrina (2007)
Models of personality	Big Five Inventory	BFI-2	Soto and John (2016)	Kalugin et al. (2021)

Table 1: Psychological inventories included in the current study and the corresponding sources of their Russian-language versions. The Russian-language versions were obtained from published Russian sources describing their translation, adaptation, standardization, or psychometric evaluation, as applicable.

Figure 3: Stability of API-based LLM responses across psychometric assessments.

Figure 4: Top-5 most pronounced psychological traits for each model (API) based on normalized scores.

Figure 5: Examples of self-reported limitations identified by API-based LLMs during the generation of responses to psychological inventories.

Figure 6: Human ratings vs. model scores across selected psychological scales.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 13: Average distinction between interface-based and API-based LLM responses.

Figure 14: Heatmap of mean differences between model-generated and human scores across psychological scales. Positive values indicate higher scores generated by the model, whereas negative values indicate lower scores.

Figure 15: Personality traits for which interface-based models produce higher or lower scores relative to API-based models.

Figure 16: Deviation of API-based model and human test responses from interface-based model responses across the models.

Model	Default	0	0.1	1
alice_ai	✓	✓	✗	✓
deepseek-r1	✓	✓	✗	✓
deepseek-v3.1	✓	✓	✗	✓
deepseek-v4	✓	✓	✗	✓
GigaChat	✓	✓	✗	✓
gpt-4.1	✓	✓	✗	✓

Appendix

Table 2: Availability of temperature settings across model providers. A check mark (✓) indicates that the corresponding temperature was supported by the provider and included in our experiments, whereas a cross (✗) indicates that the temperature setting was unavailable and therefore was not tested.

Model	Variability	Max value	CV mean	CV max
deepseek-r1	$1.60_{\pm 1.86}$	5.58	$12.12_{\pm 38.30}$	223.61
gpt-oss-20b	$1.07_{\pm 1.16}$	4.83	$31.18_{\pm 27.81}$	125.00
depseek-v3.1	$0.75_{\pm 1.09}$	5.28	$28.88_{\pm 21.10}$	89.00
o4-mini	$0.66_{\pm 0.95}$	4.00	$14.35_{\pm 14.38}$	56.46
kimi-k2.5	$0.64_{\pm 0.91}$	3.78	$23.72_{\pm 24.00}$	131.34
qwen3	$0.63_{\pm 1.21}$	6.34	$17.94_{\pm 20.62}$	118.59

Appendix

Table 3: Model stability across temperature and cross-scale variability. Values are reported as mean

\pm

standard deviation where applicable.

Model	Missing responses
alice_ai	0
deepseek-r1	53
deepseek-v3.1	2
deepseek-v4	0
GigaChat	1
gpt-4.1	1

Appendix

Table 4: Number of missing responses for each evaluated LLM.

Trait / scale	GPT-5.4	DeepSeek-R1	Gemma-4	Kimi-K2.5	Qwen-3	Llama-3.3	Alice AI	GigaChat
Anxiety (HADS)
Depression (HADS)
Neuroticism (EPI)
Narcissism (SD3)
Obsessive–compulsive (SCL-90)
Extraversion (EPI)

Appendix

Table 5: Mean psychometric profile of eight representative API-based models across all available temperatures and repeated runs. Each score was normalized using the theoretical range of its questionnaire scale. Colors show seven equal-width bands from low to high normalized scores: , , , , , , and . A dash indicates that no valid score was available in the released results. For bipolar 16PF factors, labels run from the low-score pole to the high-score pole.

Figure 17: Representation of depression and anxiety scales across all API models in different questionnaires.

Figure 18: Results of the 16PF (Cattell) questionnaire across all scales for API-based LLMs.

Figure 19: Attachment style classification of API-based LLMs (based on ASQ results).

Figure 20: Temperament classification of API-based LLMs (based on EPI results).

Psychological inventories	Pearson correlation	$\Delta r$
all w/o BDHI	0.544	0.086
all w/o SCL-90	0.548	0.081
all w/o BFI-2	0.621	0.008
all w/o SD3	0.624	0.005
all w/o GAD-7	0.636	-0.007
all w/o ASQ	0.640	-0.010

Appendix

Table 6: Changes in Pearson correlation after leave-one-test-out analysis.

\Delta r=r_{\text{full}}-r_{\text{w/o test}}

. Positive values indicate that excluding a questionnaire decreases the overall correlation, whereas negative values indicate a slight increase.

Person	N	Pearson’s r	p	Strength
1	40	0.79	1.957491e-09	Strong
2	40	0.79	1.510057e-08	Strong
3	40	0.44	4.289652e-03	Moderate
4	40	0.83	2.152543e-11	Strong
5	40	0.84	1.148446e-11	Strong
6	40	0.88	6.756510e-14	Strong

Appendix

Table 7: Pearson correlations between human ratings and model scores for each participant. Pearson’s r indicates the direction and strength of the linear association between human ratings and model scores. The p-value indicates the statistical significance of the correlation. Correlation strength was classified as weak (

|r|<0.30

), moderate (

0.30\leq|r|<0.70

), or strong (

0.70\leq|r|<1.00

Figure 21: Human ratings and model scores across selected psychological scales for each participant. For each questionnaire, the scale with the strongest absolute Pearson correlation between human ratings and model scores was selected.

Figure 22: Systematic differences between interface-based model and human psychometric scores.

Provider	Models	Access date
Together AI	deepseek-ai/DeepSeek-R1, deepseek-ai/DeepSeek-V3.1, deepseek-ai/DeepSeek-V4, google/gemma-3n-E4B-it, Qwen/Qwen3-235B-A22B-Instruct-2507, openai/gpt-oss-20b, openai/gpt-oss-120b, llama-3.1-8b-instant, llama-3.3-70b-versatile, Gemma 3n E4B Instruct, Gemma 4 31B Instruct	Apr. 17, 2026; Sep. 28, 2026
OpenAI API	o4-mini, gpt-4.1, gpt-4.1-mini, gpt-5.2, gpt-5.2-chat-latest, gpt-5.3-chat-latest, gpt-5.4, gpt-5.4-mini, gpt-5.5, gpt-6-Luna	Apr. 15, 2026; Sep. 28, 2026
Kimi API	moonshotai/Kimi-K2.5	Apr. 19, 2026
Yandex Cloud	gpt://b1gro2jeni64i2m9qdcl/aliceai-llm	Mar. 27, 2026
GigaChat API	GigaChat	Mar. 27, 2026

Appendix

Table 8: LLM providers, accessed models, and access dates used in this study.

Personality Modeling in Language Models

Personality modeling in language models studies how traits are measured, represented, simulated, and controlled in model behavior.

See all 175 papers

Apparent Psychological Profiles of Large Language Models are Largely a Measurement Artifact
Jun 18, 2026Jelena Meyer, David Garcia, Dirk U. WulffLLM Evaluation Personality Modeling in Language Models
Max Planck Institute for Human Development. · University of Konstanz. · Barcelona Supercomputing Center. +1
Same Voice, Different Lab: On the Homogenization of Frontier LLM Personalities
Mar 20, 2026Avinash Krishna, Kalyana Chadalavada, Unso Eun Seo JoPersonality Modeling in Language Models Behavioral Consistency of LLM Agents
Independent · Anthropic · Cornell University
Human Psychometric Questionnaires Mischaracterize LLM Behavior
Sep 12, 2025Woojung Song, Dongmin Choi, Yoonah Park +3Personality Modeling in Language Models Psychometric Validation
1Graduate School of Data Science, Seoul National University · Department of Communication, Interdisciplinary Program in Artificial Intelligence, Seoul National University

Jun 18, 2026J/K move · Enter open · S save

Apparent Psychological Profiles of Large Language Models are Largely a Measurement Artifact

Jelena Meyer, David Garcia, Dirk U. Wulff

Max Planck Institute for Human Development. · University of Konstanz. · Barcelona Supercomputing Center.+1

LLM Evaluation Personality Modeling in Language Models Psychometric Validation

Open paper AskPDF arXiv

Psychological instruments designed for humans are increasingly used to assign large language models (LLMs) stable psychological profiles that affect their usability, safety assessment, and use as proxies for human participants in research. Using a formal psychometric framework, we show that these profiles are largely a measurement artifact. Administering a battery of personality and risk-preference instruments spanning self-reports and behavioral tasks to 56 instruction-tuned LLMs alongside large human reference samples, we report four findings. First, differences between models are driven not by the traits an instrument targets but by a directional response bias, a tendency to respond toward one end of the scale, or one labeled option, regardless of item content; a variance decomposition attributes 81-90% of between-model variation to this bias, against 9-16% in humans. Second, the bias declines with model capability but is not eliminated by it. Third, because bias rather than trait drives responding, an instrument's apparent reliability is almost entirely predicted by its response orthogonality, a term we coin for the proportion of items for which trait and bias point in opposite directions. Fourth, the profile a model appears to have shifts with the items used and can be manufactured through item selection. These results demonstrate that the apparent psychological profiles of LLMs are artifacts of the instrument used to measure them, not properties of the models themselves. As instruments borrowed from human psychology are rarely fully orthogonal and may inherently lack validity for LLMs, we call for dedicated assessments centered on response orthogonality.

Abstract

Personality Modeling in Language Models

Explore similar work

Apparent Psychological Profiles of Large Language Models are Largely a Measurement Artifact

Same Voice, Different Lab: On the Homogenization of Frontier LLM Personalities

Human Psychometric Questionnaires Mischaracterize LLM Behavior