Comparing human and AI emotional evaluations of images: GPT-4o performance across standard, persona, and language-specific prompting strategies Authors: Maiko Kobayashi, Akira Goto, Akihito Himuro Published: 2026 ABSTRACT BackgroundLarge language models, such as GPT-4o, have demonstrated strong performance in emotion recognition tasks, yet documented cultural biases raise concerns about cross-cultural validity. Prompt-engineering strategies, including persona prompting and native-language prompting, have been proposed as lightweight approaches to cultural adaptation, but their effectiveness in image-based emotion recognition remains underexamined, particularly in non-Anglophone European contexts. The present study examines whether language choice (Polish vs. English) and persona prompting alter GPT-4o emotional evaluation when benchmarked against a mixed European reference sample from the Nencki Affective Picture System (NAPS).MethodsGPT-4o (v.2024-08-06) evaluated 1,276 NAPS images on three emotional dimensions (valence, arousal, approach–avoidance) using a nine-point scale under three conditions: standard GPT-4o, persona-type GPT (Polish university student persona), and Polish-language GPT. Each image received 55 independent evaluations per condition, matching the original NAPS dataset (N = 204 evaluators; 60% Polish nationals, 40% European exchange students). Human–artificial intelligence (AI) agreement was evaluated using Pearson correlations, intraclass correlation coefficients (ICC[2,1]), Bayesian one-way ANOVA, and Anderson–Darling distributional tests.ResultsAll AI conditions positively correlated with human ratings across the three dimensions (r = 0.525–0.809, all p < 0.001). Neither persona nor Polish-language prompting consistently improved agreement over standard GPT-4o. The Polish-language condition aligned more closely with human ratings for arousal but diverged more strongly in approach–avoidance (d = 0.47). Valence showed the strongest correspondence (r = 0.784–0.809; BF₀₁ = 36.55 favoring equivalence), whereas arousal and approach–avoidance exhibited reliable mean-level and distributional divergence. Landscape stimuli produced the highest agreement (r up to 0.933) and faces the lowest (r as low as 0.253). Distributional differences persisted across all nine human–AI comparisons.ConclusionsWithin the constraints of a mixed European reference sample, persona-based and native-language adaptation strategies did not yield consistent improvements in human–AI agreement. Effects varied across dimensions and analytical methods, indicating that such strategies require dimension-specific empirical validation. AI evaluation aligned acceptably with human ratings for valence and scenic stimuli but diverged for arousal, approach–avoidance, and facial stimuli. These findings underscore the importance of jointly examining rank-order agreement, central tendency, and distributional structure in cross-cultural AI emotion recognition.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40359-026-05067-3. Background The rapid advancement of artificial intelligence (AI)-based emotion recognition has expanded opportunities for cross-cultural communication, mental health assessment, and human–computer interaction. Early studies demonstrated that machine learning could provide scalable, cost-efficient emotion assessment, shaping subsequent psychological research [1]. Building on this foundation, large language models (LLMs) such as OpenAI’s GPT-4o have shown strong performance in emotion evaluation tasks, combining multimodal processing with rapid response times (232–320 ms) [2]. Despite these advances, major concerns regarding cultural fairness and bias in AI-driven emotion recognition remain unresolved. Existing AI emotion-recognition systems exhibit well-documented cultural biases. GPT-4o, for example, shows systematic alignment with specific cultural value patterns across 107 countries [3]. These biases are commonly linked to imbalanced training data, as dominant datasets disproportionately reflect Western faces, expressions, and emotional norms, producing performance declines of 36–65% in non-Western populations [4]. Although ChatGPT can process Polish-language input without intermediate translation [5], whether this capability preserves culture-specific emotional interpretation remains unclear, raising concerns about equitable global deployment. Although prior research has examined cultural bias in LLM reasoning tasks [6] and nationality-based emotion attribution in text contexts [7], cultural adaptation strategies in visual emotion recognition remain poorly understood. Recent studies have benchmarked GPT-4 against human valence and arousal ratings using the Geneva Affective PicturE Database (GAPED) [8] and evaluated GPT-4 V across multiple emotion-recognition datasets [9]. However, these studies neither incorporated dimensional frameworks including approach–avoidance, such as the Nencki Affective Picture System (NAPS), nor tested interventions intended to reduce cultural bias in GPT-4o’s multimodal architecture. Research on persona prompting has similarly concentrated on symbolic reasoning domains, including mathematics and law [6, 10], rather than perceptual emotion evaluation. Existing persona- and nationality-based prompting studies have also remained confined to text-based scenarios [7]. Consequently, whether linguistic or persona-based adaptations meaningfully alter image-based emotion recognition remains an unresolved empirical question, particularly in Eastern European contexts that remain underrepresented in AI emotion research despite documented differences in emotional-expression norms [11, 12]. Russell’s circumplex model of affect [13], which organizes emotion along valence (pleasant–unpleasant) and arousal (activation–deactivation) dimensions, provides a robust framework for evaluating emotional judgment and generating theoretically grounded expectations for AI performance. Extensive neuroscientific and behavioral research has validated the model through strong correspondence with functional brain-imaging findings [14]. Combined with the approach–avoidance motivational system [15], the framework captures both affective experience and behavioral tendency, making it well suited for assessing whether AI systems reproduce human emotional evaluations across cultures. Barrett’s theory of constructed emotion further challenges universalist assumptions by arguing that emotions are actively constructed through cultural concepts, prior experience, and contextual interpretation [16]. From this perspective, AI systems trained predominantly on Western data may inadequately interpret emotions in non-Western settings, underscoring the importance of culture-specific emotional concepts in AI design. Effective emotion recognition may therefore depend less on presumed universal patterns than on sensitivity to cultural context. Cross-cultural research consistently demonstrates substantial variation in emotion recognition. A meta-analysis of 97 studies identified a 15–25% in-group advantage, whereby individuals more accurately recognize emotions expressed within their own cultural group [17]. This pattern parallels disparities observed in AI systems trained primarily on Western datasets. Eastern European emotional expression, in particular, differs from both Western European and East Asian norms [11], especially in conventions governing emotional directness and expression. Together, these findings indicate that cultural variation in emotion recognition is systematic rather than incidental and therefore requires explicit consideration in AI research. Although this study uses the NAPS, which is anchored in a Polish context, the human reference ratings derive from a mixed sample comprising 60% Polish nationals and 40% European exchange students of unspecified nationalities. Accordingly, the study should not be interpreted as a pure test of Polish emotional norms. Instead, it examines whether language choice (Polish vs. English) and persona prompting (Polish university student persona) alter AI emotional evaluation when benchmarked against a mixed European reference sample. This framing acknowledges the heterogeneous composition of the reference group while preserving the analytical value of testing cultural and linguistic adaptation strategies in a non-Anglophone European context. NAPS was selected for two methodological reasons. First, it provides a rare standardized affective dataset that includes the approach–avoidance dimension alongside valence and arousal. This three-dimensional structure is theoretically important because cultural differences in motivational orientation shape the information individuals attend to in their environment, whereas widely used alternatives such as International Affective Picture System (English) and Geneva Affective PicturE Database (GAPED; French) omit this dimension. GAPED has recently been used to benchmark GPT-4 valence and arousal ratings against human evaluations [8], offering a methodological precedent in a non-Polish European context. Second, NAPS provides large-scale standardized ratings (N = 204 evaluators across 1,356 images), enabling robust statistical comparison with AI-generated evaluations. The selection of a Polish linguistic and cultural context, despite the heterogeneous reference sample, was motivated by three considerations. First, Eastern European cultures remain underrepresented in AI emotion research despite documented differences in emotional-expression norms [11, 12]. The closest regional precedent involved administering a Russian-language Mayer–Salovey–Caruso Emotional Intelligence Test to GPT-4, which suggested cultural specificity in the model’s interpretation of complex emotions; however, that study was limited to verbal stimuli and did not examine image-based affective evaluation [18]. Second, Poland occupies a hybrid cultural position combining Western individualism with Eastern European collectivism [12], while also exhibiting norms of direct emotional expression and a socially accepted “culture of complaining” [12], offering analytical leverage beyond binary East–West comparisons. Third, Polish has been argued to encode complex meaning effectively in extended contexts [19], which may influence the interpretation of nuanced emotional content. Collectively, these factors make Poland a useful context for examining whether language choice and prompting strategies shape AI emotional evaluation, while recognizing that the mixed reference sample constrains strictly culture-specific conclusions. Prompt engineering and cultural adaptation strategies Prompt engineering (i.e., the modification of input instructions without retraining a model) has emerged as a lightweight approach for mitigating cultural bias in LLMs. Two widely studied strategies are persona prompting, which assigns culturally specific roles to LLMs [20], and native-language prompting, which replaces English with the target culture’s language. However, prior evidence suggests that the effectiveness of these strategies is highly task dependent [6, 10], and their role in image-based emotion recognition remains largely unexplored. Persona prompting shows particularly unstable task-dependent effects. Gupta et al. [6] demonstrated that persona assignment can expose latent biases in reasoning tasks, reducing performance in 7 of 12 evaluated datasets, with some personas causing declines exceeding 70%. Similarly, persona prompting can improve performance in some tasks while degrading it in others [10]. Existing research has focused almost exclusively on text-based domains such as mathematics, logic, and legal reasoning [6, 10], leaving its impact on perceptual emotion evaluation largely untested. Whether persona assignment meaningfully alters visual emotion recognition, a perceptual judgment process fundamentally distinct from symbolic reasoning. therefore remains unresolved. Native-language prompting introduces a related but equally complex trade-off. Although many LLMs achieve peak accuracy in English because of training-data concentration [21], native-language prompting may better preserve culture-specific emotional concepts [5]. English prompts tend to align outputs with North American cultural values even among non-Western users [3]. Similarly, mid-resource languages such as Polish remain underrepresented in training corpora, potentially reducing output reliability. Moreover, multilingual LLMs process languages as partially distinct systems rather than interchangeable variants [22], suggesting that language choice may shape not only expression but also the cultural framework through which emotional content is interpreted. This creates tension between technical optimization and cultural authenticity in emotion evaluation. Against this background, the present study empirically examines how these adaptation strategies operate in visual emotion recognition. Specifically, it provides: A systematic evaluation of GPT-4o using standardized affective stimuli from NAPS;A direct comparison of persona prompting and native-language prompting in image-based emotion evaluation;A focused analysis of a Polish/Eastern European cultural context underrepresented in prior AI emotion research; and.A multi-dimensional assessment spanning three emotional dimensions and five image categories. A systematic evaluation of GPT-4o using standardized affective stimuli from NAPS; A systematic evaluation of GPT-4o using standardized affective stimuli from NAPS; A direct comparison of persona prompting and native-language prompting in image-based emotion evaluation; A direct comparison of persona prompting and native-language prompting in image-based emotion evaluation; A focused analysis of a Polish/Eastern European cultural context underrepresented in prior AI emotion research; and. A focused analysis of a Polish/Eastern European cultural context underrepresented in prior AI emotion research; and. A multi-dimensional assessment spanning three emotional dimensions and five image categories. A multi-dimensional assessment spanning three emotional dimensions and five image categories. Whereas previous studies have examined GPT-4 V without adaptation strategies [9], persona prompting in text-based reasoning [6, 10], or broader cultural bias in LLMs [3], empirical evidence regarding adaptation strategies in image-based emotion recognition remains limited. Research hypotheses Three hypotheses guide this investigation. Hypothesis 1 predicts that persona prompting and Polish-language prompting will produce stronger agreement with human evaluators than standard GPT-4o, based on in-group advantage effects [17] and the possibility that native-language use preserves culture-specific emotional nuance [5].Hypothesis 2 predicts systematic variation in human–AI agreement across emotional dimensions, following the order valence → approach avoidance → arousal, reflecting differences in their cognitive and physiological foundations [13, 23].Hypothesis 3 predicts category-specific variation in agreement, following the order landscapes → animals/objects → people/faces, based on evolutionary universality [24], anthropomorphic projection mechanisms [25], and the culturally dependent nature of facial emotion recognition [26, 27]. Hypothesis 1 predicts that persona prompting and Polish-language prompting will produce stronger agreement with human evaluators than standard GPT-4o, based on in-group advantage effects [17] and the possibility that native-language use preserves culture-specific emotional nuance [5]. Hypothesis 1 predicts that persona prompting and Polish-language prompting will produce stronger agreement with human evaluators than standard GPT-4o, based on in-group advantage effects [17] and the possibility that native-language use preserves culture-specific emotional nuance [5]. Hypothesis 2 predicts systematic variation in human–AI agreement across emotional dimensions, following the order valence → approach avoidance → arousal, reflecting differences in their cognitive and physiological foundations [13, 23]. Hypothesis 2 predicts systematic variation in human–AI agreement across emotional dimensions, following the order valence → approach avoidance → arousal, reflecting differences in their cognitive and physiological foundations [13, 23]. Hypothesis 3 predicts category-specific variation in agreement, following the order landscapes → animals/objects → people/faces, based on evolutionary universality [24], anthropomorphic projection mechanisms [25], and the culturally dependent nature of facial emotion recognition [26, 27]. Hypothesis 3 predicts category-specific variation in agreement, following the order landscapes → animals/objects → people/faces, based on evolutionary universality [24], anthropomorphic projection mechanisms [25], and the culturally dependent nature of facial emotion recognition [26, 27]. Methods Human evaluation data reference Human evaluation data consisted of ratings from 204 healthy volunteers (119 females, 85 males; mean age = 23.9 years, SD = 3.4). Of the participants, 60% (N = 123) were Polish nationals, whereas the remainder were primarily European exchange students. In the original experimental design, participants evaluated subsets of 362 images, with each image independently rated by 55 individuals [28]. These ratings served as the human reference standard for assessing AI–human agreement. Image stimuli The study used stimuli data from NAPS [28], a standardized affective image database containing 1,356 high-quality images rated on valence, arousal, and approach–avoidance using a nine-point scale. The database includes five categories—animals, faces, landscapes, objects, and people—each spanning a broad emotional range. All 1,356 images were initially considered. However, some stimuli (e.g., images depicting corpses) could not be processed because of ChatGPT content-filter restrictions. Consequently, 1,276 images were retained for analysis (Table 1). The final dataset included animals in diverse emotional contexts (199 images), faces representing varied ethnic backgrounds (360 images), natural and urban landscapes (185 images), objects ranging from everyday items to abstract shapes (322 images), and full-body images of people in emotionally salient situations (210 images). All images were standardized to a resolution of 1,024 × 768 pixels. Table 1Distribution of Nencki Affective Picture System (NAPS) images by category before and after GPT-4o content filteringCategoryNAPS originalExcluded by GPT-4o content filterRetained foranalysisExclusion rate(%)People2504021016Animals2212219910Faces372123603.2Objects32863221.8Landscapes18501850Total1,356801,2765.9NAPS original = number of images in each category of the Nencki Affective Picture System. Content-filter exclusions were determined empirically during data collection: images blocked by GPT-4o (gpt-4o-2024-08-06) because of safety restrictions on sensitive content were removed from the analysis set Table 1Distribution of Nencki Affective Picture System (NAPS) images by category before and after GPT-4o content filteringCategoryNAPS originalExcluded by GPT-4o content filterRetained foranalysisExclusion rate(%)People2504021016Animals2212219910Faces372123603.2Objects32863221.8Landscapes18501850Total1,356801,2765.9NAPS original = number of images in each category of the Nencki Affective Picture System. Content-filter exclusions were determined empirically during data collection: images blocked by GPT-4o (gpt-4o-2024-08-06) because of safety restrictions on sensitive content were removed from the analysis set Distribution of Nencki Affective Picture System (NAPS) images by category before and after GPT-4o content filtering NAPS original = number of images in each category of the Nencki Affective Picture System. Content-filter exclusions were determined empirically during data collection: images blocked by GPT-4o (gpt-4o-2024-08-06) because of safety restrictions on sensitive content were removed from the analysis set Emotional evaluation was conducted across three dimensions—arousal, valence, and approach–avoidance—using a nine-point scale. Arousal ranged from 1 (very calm/relaxed) to 9 (very excited/activated), valence from 1 (very negative/unpleasant) to 9 (very positive/pleasant), and approach–avoidance from 1 (strong avoidance) to 9 (strong approach). AI experimental conditions Three experimental conditions—standard GPT-4o, persona-type GPT, and Polish-language GPT—were used to evaluate 1,276 images across three emotional dimensions: valence, arousal, and approach–avoidance. To maintain comparability with the human reference dataset, 55 independent evaluations were generated for every image, condition, and emotional dimension, matching the number of human ratings per stimulus. In the standard GPT-4o condition, anthropomorphic framing was introduced by instructing the model to “act like a human being and answer.” Each image was then evaluated using the prompt: “You are judging this image as … (from 1 = very negative to 9 = very positive, with 5 = neutral). Answer only with a score between 1 and 9.” Fig. 1 presents an example of the prompt structure and corresponding output. Fig. 1Example prompt and GPT-4o output in the standard condition. The left panel shows the prompt used in the standard condition, including anthropomorphic framing (“Act like a human being and answer”) and a 9-point valence scale (1 = very negative, 9 = very positive, 5 = neutral). The right panel shows the model’s single-integer output (rating = 7). The displayed image is from the Nencki Affective Picture System (NAPS) Example prompt and GPT-4o output in the standard condition. The left panel shows the prompt used in the standard condition, including anthropomorphic framing (“Act like a human being and answer”) and a 9-point valence scale (1 = very negative, 9 = very positive, 5 = neutral). The right panel shows the model’s single-integer output (rating = 7). The displayed image is from the Nencki Affective Picture System (NAPS) In the persona-type GPT condition, cultural role assignment was introduced by instructing the model to respond “as if you were a Polish university or graduate student.” Images were evaluated using the prompt: “My reaction to this image is … (from 1 = very negative to 9 = very positive, with 5 = neutral). Answer only with a score between 1 and 9.” In the Polish-language condition, all instructions were presented in Polish to examine the effect of language choice on emotional evaluation. Following the instruction Odpowiadaj, jak gdybyś był polskim studentem studiów licencjackich lub magisterskich, images were evaluated using the prompt: Oceniasz ten obraz jako… (od 1 = bardzo negatywny do 9 = bardzo pozytywny, przy 5 = neutralny). Odpowiedz tylko z wynikiem od 1 do 9. ChatGPT settings The study used ChatGPT v.gpt-4o-2024-08-06. GPT-4o was operated with the following hyperparameters: temperature = 1 (default); top_p = 1 (default); max_tokens = 1; presence_penalty = 0 (default); frequency_penalty = 0 (default); logit_bias = null (default); response_format = null (default); seed = null (default); and system_fingerprint = fp_e9b9b028d7. Temperature controlled the variance of the probability distribution used for token selection during generation. The choice of temperature = 1.0 was motivated by both methodological and practical considerations. Prior research reported no statistically significant differences in problem-solving performance across temperature values ranging from 0 to 1 [29]. In addition, widely used multimodal LLMs, including Claude and Gemini, employ a default maximum temperature of 1.0, making this setting representative of real-world deployment conditions. The max_tokens parameter defines the maximum number of generated tokens per response. Because the experimental design required only a single numeric rating between 1 and 9, max_tokens was restricted to 1 to prevent extraneous output and reduce computational cost. All remaining parameters retained their default settings. To eliminate carryover effects, conversational context was cleared after each evaluation, ensuring that every rating was generated from a fresh context window. Analytical method Effect-size thresholds were specified a priori based on prior empirical findings and practical significance. For cultural persona effects, partial eta squared (η²p) was set to 0.06, consistent with reported in-group advantage effects (d = 0.28–0.42) in cross-cultural emotion recognition [17]. For emotional-dimension differences, η²p was set to 0.10 to reflect theoretical distinctions in cognitive foundations. To evaluate the practical utility of AI emotion recognition, the minimum detectable correlation difference was defined as Δr = 0.15, given that correlations of r ≥ 0.3 are commonly considered meaningful in applied contexts. A priori power analysis conducted with G*Power 3.1.9.7 indicated statistical power greater than 0.99 for the specified effect size (f = 0.25) with a sample of 1,276 images. Correlation-based power calculations also confirmed adequate sensitivity to detect practically meaningful correlation differences (Δr = 0.15). All analyses were performed in R (v.4.3.1) using a multi-level statistical framework. Human–AI agreement was first evaluated using Pearson correlation coefficients and intraclass correlation coefficients (ICC(2,1)) with 95% confidence intervals [30]. ICC values were interpreted as follows: < 0.5 (poor), 0.5–0.75 (moderate), 0.75–0.9 (good), and > 0.9 (excellent). The primary inferential analysis used Bayesian factor analysis to assess equivalence across conditions (human, GPT, persona, and Polish) with the BayesFactor package [31]. For each emotional dimension—arousal, valence, and approach–avoidance—Bayes factors (BF₀₁) were calculated from one-way analysis of variance (ANOVA) models of mean ratings across the four conditions. BF₀₁ values greater than three indicated moderate evidence, whereas values greater than 10 indicated strong evidence in favor of the null hypothesis of no meaningful difference [32]. Cohen’s d and corresponding 95% confidence intervals were reported to quantify effect magnitude, with |d| < 0.2 interpreted as a small effect. Exploratory analysis In addition to the primary analyses, two exploratory analyses were performed. Category-specific Correlation analyses were conducted separately for five image categories—animals (n = 199), faces (n = 360), landscapes (n = 185), objects (n = 322), and people (n = 210)—to determine whether human–AI agreement varied by stimulus type. For each category, Pearson product–moment correlations were calculated between human evaluations and each AI condition (GPT, persona, and Polish) across the three emotional dimensions. 95% confidence intervals were reported to facilitate comparison across categories and AI conditions. Distributional differences Anderson–Darling tests were conducted for every human–AI comparison across all three emotional dimensions to determine whether observed differences reflected broader distributional discrepancies rather than shifts in central tendency alone. The Anderson–Darling test was selected because of its high sensitivity to deviations in distribution tails [33]. To control for multiple comparisons, the Benjamini–Hochberg procedure was applied with the false discovery rate fixed at 5% [34]. Results Descriptive statistics As shown in Table 2 and Supplementary Fig. 1, the mean ratings for arousal ranged from 4.439 (GPT) to 5.004 (human), with standard deviations spanning 1.001 (human) to 1.161 (Polish GPT). Human ratings also showed the narrowest observed range (2.043–8.046), whereas all AI conditions covered nearly the full scale. Valence ratings were higher overall, ranging from 5.387 (Polish GPT) to 5.574 (GPT), with standard deviations between 1.530 (human) and 1.812 (GPT). As in the arousal dimension, human ratings occupied a narrower range than AI-generated ratings. For approach–avoidance, mean ratings ranged from 4.952 (Polish GPT) to 5.511 (human), with standard deviations spanning 0.980 (GPT) to 1.363 (human). Unlike the arousal dimension, human ratings in approach–avoidance exhibited the greatest variability. Table 2Means, standard deviations, and ranges for three emotional dimensions by rating conditionVariable N MSDMinMaxArousalHuman1,2765.0041.0012.0438.046GPT1,2764.4391.0861.0368.855Persona GPT1,2764.4431.1171.0008.782Polish GPT1,2764.6811.1611.0189.000ValenceHuman1,2765.5471.5301.6968.540GPT1,2765.5741.8121.0009.000Persona GPT1,2765.5131.6291.0009.000Polish GPT1,2765.3871.7161.0009.000Approach-AvoidanceHuman1,2765.5111.3631.4678.460GPT1,2765.1500.9801.0008.964Persona GPT1,2765.0631.2491.0009.000Polish GPT1,2764.9520.9891.0009.000N number of images, M mean, SD standard deviation. Min and Max indicate the minimum and maximum mean ratings across images. Ratings were made on a 9-point scale. GPT = standard GPT-4o; Persona GPT = persona-type GPT; Polish GPT = Polish-language GPT Table 2Means, standard deviations, and ranges for three emotional dimensions by rating conditionVariable N MSDMinMaxArousalHuman1,2765.0041.0012.0438.046GPT1,2764.4391.0861.0368.855Persona GPT1,2764.4431.1171.0008.782Polish GPT1,2764.6811.1611.0189.000ValenceHuman1,2765.5471.5301.6968.540GPT1,2765.5741.8121.0009.000Persona GPT1,2765.5131.6291.0009.000Polish GPT1,2765.3871.7161.0009.000Approach-AvoidanceHuman1,2765.5111.3631.4678.460GPT1,2765.1500.9801.0008.964Persona GPT1,2765.0631.2491.0009.000Polish GPT1,2764.9520.9891.0009.000N number of images, M mean, SD standard deviation. Min and Max indicate the minimum and maximum mean ratings across images. Ratings were made on a 9-point scale. GPT = standard GPT-4o; Persona GPT = persona-type GPT; Polish GPT = Polish-language GPT Means, standard deviations, and ranges for three emotional dimensions by rating condition N number of images, M mean, SD standard deviation. Min and Max indicate the minimum and maximum mean ratings across images. Ratings were made on a 9-point scale. GPT = standard GPT-4o; Persona GPT = persona-type GPT; Polish GPT = Polish-language GPT Coefficients of variation (CVs) ranged from 0.190 to 0.325 across all conditions and dimensions, with human–AI differences remaining ≤ 0.047 within each dimension. For arousal, human ratings showed the lowest variability (CV = 0.200), whereas AI conditions produced slightly higher values (GPT = 0.245; persona GPT = 0.251; Polish GPT = 0.248). A similar pattern emerged for valence: human ratings (CV = 0.276) remained lower than all AI conditions (GPT = 0.325; persona GPT = 0.295; Polish GPT = 0.319). In contrast, approach–avoidance showed a mixed pattern. GPT (CV = 0.190) and Polish GPT (CV = 0.200) exhibited slightly lower variability than human ratings (CV = 0.247), whereas persona GPT matched the human level exactly (CV = 0.247). Correlation analysis between human and AI conditions Pearson correlation analyses showed consistently positive associations between human ratings and all AI conditions across the three emotional dimensions (Fig. 2; Supplement Table 1). For arousal, correlations ranged from r = 0.675 (human–GPT) to r = 0.703 (human–Polish), with all comparisons remaining significant after Bonferroni correction (p < 0.001). Confidence intervals were narrow and excluded zero (e.g., human–Polish: 95% CI [0.674, 0.729]), indicating stable estimates. Valence produced the strongest overall associations, ranging from r = 0.784 (human–Polish) to r = 0.809 (human–GPT), with all comparisons again reaching p < 0.001. For approach–avoidance, correlations ranged from r = 0.525 (human–Polish) to r = 0.680 (human–Persona), with human–GPT (r = 0.593) falling between these values. Fig. 2Heatmaps of Pearson correlation coefficients (left panel) and intraclass correlation coefficients (ICC[2,1]; right panel) between human and AI emotional ratings across three emotional dimensions. Each cell represents the association between human ratings and one AI condition (GPT, Persona, or Polish) across arousal, valence, and approach–avoidance (N = 1,276 images per comparison). Left panel: Pearson’s r (rank-order agreement). Right panel: ICC(2,1), two-way mixed-effects, absolute agreement, single measurement [30]. Darker cells indicate stronger association. All Pearson correlations remained significant after Bonferroni correction (p < 0.001). GPT = standard GPT-4o; Persona = persona-type GPT; Polish = Polish-language GPT Heatmaps of Pearson correlation coefficients (left panel) and intraclass correlation coefficients (ICC[2,1]; right panel) between human and AI emotional ratings across three emotional dimensions. Each cell represents the association between human ratings and one AI condition (GPT, Persona, or Polish) across arousal, valence, and approach–avoidance (N = 1,276 images per comparison). Left panel: Pearson’s r (rank-order agreement). Right panel: ICC(2,1), two-way mixed-effects, absolute agreement, single measurement [30]. Darker cells indicate stronger association. All Pearson correlations remained significant after Bonferroni correction (p < 0.001). GPT = standard GPT-4o; Persona = persona-type GPT; Polish = Polish-language GPT To complement rank-order correlations with a measure of absolute agreement, two-way mixed-effects intraclass correlation coefficients (ICC[2,1], absolute agreement, single measurement) were calculated and interpreted using confidence-interval-based guidelines [30]. Valence produced the highest ICCs, ranging from 0.775 (human–Polish) to 0.797 (human–GPT), with confidence intervals consistently within the good-agreement range (e.g., human–GPT: 95% CI [0.776, 0.816]). Arousal showed moderate agreement, with ICCs ranging from 0.587 (human–GPT) to 0.665 (human–Polish); corresponding confidence intervals also fell primarily within the moderate range (e.g., human–Polish: 95% CI [0.567, 0.737]). Approach–avoidance showed greater heterogeneity, with ICCs ranging from 0.450 (human–Polish) to 0.640 (human–Persona). The human–Polish comparison fell near the boundary between poor and moderate agreement (95% CI [0.300, 0.563]), whereas human–Persona remained clearly within the moderate range (95% CI [0.515, 0.726]). Bayesian results Bayesian analyses were conducted on 5,104 observations across the four conditions (Tables 3 and 4). For arousal, one-way Bayesian ANOVA strongly favored the alternative hypothesis (BF₀₁ < 0.001). Pairwise comparisons revealed a two-cluster structure. All human–AI comparisons strongly favored the alternative hypothesis (Human vs. GPT: BF₀₁ < 0.001, d = 0.54, 95% CI [0.46, 0.62]; Human vs. Persona: BF₀₁ < 0.001, d = 0.53, 95% CI [0.45, 0.61]; Human vs. Polish: BF₀₁ < 0.001, d = 0.30, 95% CI [0.22, 0.38]). Within AI conditions, GPT versus Persona strongly favored the null (BF₀₁ = 22.35, d = 0.00, 95% CI [− 0.08, 0.07]), whereas Polish versus GPT (BF₀₁ < 0.001, d = − 0.22, 95% CI [− 0.29, − 0.14]) and Polish versus Persona (BF₀₁ < 0.001, d = − 0.21, 95% CI [− 0.29, − 0.13]) strongly favored the alternative hypothesis. Human–AI effect sizes ranged from |d| = 0.30 (Human vs. Polish) to |d| = 0.54 (Human vs. GPT), with Human versus Persona at |d| = 0.53. According to Cohen’s criteria, the Human–GPT and Human–Persona comparisons reached approximately medium effect size (d ≈ 0.50), whereas Human–Polish fell within the small-to-medium range. Table 3Summary of Bayesian factor analysis results across emotional dimensionsDimensionOverall ANOVA BF₀₁InterpretationMax Effect Size |d|Arousal< 0.001Strong evidence for H₁0.54Valence36.55Very strong evidence for H₀0.11Approach-Avoidance< 0.001Strong evidence for H₁0.47BF₀₁ = Bayes factor comparing the null hypothesis (H₀: no difference between conditions) against the alternative hypothesis (H₁: presence of a difference). |d| = maximum absolute Cohen’s d across all pairwise contrasts within each dimension, including both human–AI and AI–AI comparisons. Cohen’s d values of 0.20, 0.50, and 0.80 represent small, medium, and large effects, respectively [35]. ANOVA = analysis of variance Table 3Summary of Bayesian factor analysis results across emotional dimensionsDimensionOverall ANOVA BF₀₁InterpretationMax Effect Size |d|Arousal< 0.001Strong evidence for H₁0.54Valence36.55Very strong evidence for H₀0.11Approach-Avoidance< 0.001Strong evidence for H₁0.47BF₀₁ = Bayes factor comparing the null hypothesis (H₀: no difference between conditions) against the alternative hypothesis (H₁: presence of a difference). |d| = maximum absolute Cohen’s d across all pairwise contrasts within each dimension, including both human–AI and AI–AI comparisons. Cohen’s d values of 0.20, 0.50, and 0.80 represent small, medium, and large effects, respectively [35]. ANOVA = analysis of variance Summary of Bayesian factor analysis results across emotional dimensions BF₀₁ = Bayes factor comparing the null hypothesis (H₀: no difference between conditions) against the alternative hypothesis (H₁: presence of a difference). |d| = maximum absolute Cohen’s d across all pairwise contrasts within each dimension, including both human–AI and AI–AI comparisons. Cohen’s d values of 0.20, 0.50, and 0.80 represent small, medium, and large effects, respectively [35]. ANOVA = analysis of variance Table 4Pairwise Bayesian comparisons across emotional dimensionsDimensionComparisonBF₀₁Evidence InterpretationCohen’s d95% CIArousalHuman vs. GPT< 0.001Strong for H₁0.54[0.46, 0.62]Human vs. Persona< 0.001Strong for H₁0.53[0.45, 0.61]Human vs. Polish< 0.001Strong for H₁0.30[0.22, 0.38]GPT vs. Persona22.35Strong for H₀0.00[–0.08, 0.07]GPT vs. Polish< 0.001Strong for H₁–0.22[–0.29, − 0.14]Persona vs. Polish< 0.001Strong for H₁–0.21[–0.29, − 0.13]ValenceHuman vs. GPT20.68Strong for H₀–0.02[–0.09, 0.06]Human vs. Persona19.40Strong for H₀0.02[–0.06, 0.10]Human vs. Polish1.05Weak for H₀0.10[0.02, 0.18]GPT vs. Persona15.10Strong for H₀0.04[–0.04, 0.11]GPT vs. Polish0.64Inconclusive0.11[0.03, 0.18]Persona vs. Polish3.73Moderate for H₀0.08[0.00, 0.15]Approach-avoidanceHuman vs. GPT< 0.001Strong for H₁0.30[0.23, 0.38]Human vs. Persona< 0.001Strong for H₁0.34[0.27, 0.42]Human vs. Polish< 0.001Strong for H₁0.47[0.39, 0.55]GPT vs. Persona3.38Moderate for H₀0.08[0.00, 0.16]GPT vs. Polish< 0.001Strong for H₁0.20[0.12, 0.28]Persona vs. Polish1.01Weak for H₀0.10[0.02, 0.18]N = 5,104 for all comparisons. BF₀₁ denotes the Bayes factor favoring the null hypothesis (H₀: no difference between conditions) over the alternative hypothesis (H₁: presence of a difference). Evidence thresholds were interpreted as follows: BF₀₁ > 10 = strong evidence for H₀; 3–10 = moderate evidence for H₀; 1–3 = weak evidence for H₀; 1/3–1 = weak evidence for H₁; 1/10–1/3 = moderate evidence for H₁; and < 1/10 = strong evidence for H₁. Bayes factors between 1/3 and 3 were classified as inconclusive. CI confidence interval Table 4Pairwise Bayesian comparisons across emotional dimensionsDimensionComparisonBF₀₁Evidence InterpretationCohen’s d95% CIArousalHuman vs. GPT< 0.001Strong for H₁0.54[0.46, 0.62]Human vs. Persona< 0.001Strong for H₁0.53[0.45, 0.61]Human vs. Polish< 0.001Strong for H₁0.30[0.22, 0.38]GPT vs. Persona22.35Strong for H₀0.00[–0.08, 0.07]GPT vs. Polish< 0.001Strong for H₁–0.22[–0.29, − 0.14]Persona vs. Polish< 0.001Strong for H₁–0.21[–0.29, − 0.13]ValenceHuman vs. GPT20.68Strong for H₀–0.02[–0.09, 0.06]Human vs. Persona19.40Strong for H₀0.02[–0.06, 0.10]Human vs. Polish1.05Weak for H₀0.10[0.02, 0.18]GPT vs. Persona15.10Strong for H₀0.04[–0.04, 0.11]GPT vs. Polish0.64Inconclusive0.11[0.03, 0.18]Persona vs. Polish3.73Moderate for H₀0.08[0.00, 0.15]Approach-avoidanceHuman vs. GPT< 0.001Strong for H₁0.30[0.23, 0.38]Human vs. Persona< 0.001Strong for H₁0.34[0.27, 0.42]Human vs. Polish< 0.001Strong for H₁0.47[0.39, 0.55]GPT vs. Persona3.38Moderate for H₀0.08[0.00, 0.16]GPT vs. Polish< 0.001Strong for H₁0.20[0.12, 0.28]Persona vs. Polish1.01Weak for H₀0.10[0.02, 0.18]N = 5,104 for all comparisons. BF₀₁ denotes the Bayes factor favoring the null hypothesis (H₀: no difference between conditions) over the alternative hypothesis (H₁: presence of a difference). Evidence thresholds were interpreted as follows: BF₀₁ > 10 = strong evidence for H₀; 3–10 = moderate evidence for H₀; 1–3 = weak evidence for H₀; 1/3–1 = weak evidence for H₁; 1/10–1/3 = moderate evidence for H₁; and < 1/10 = strong evidence for H₁. Bayes factors between 1/3 and 3 were classified as inconclusive. CI confidence interval Pairwise Bayesian comparisons across emotional dimensions N = 5,104 for all comparisons. BF₀₁ denotes the Bayes factor favoring the null hypothesis (H₀: no difference between conditions) over the alternative hypothesis (H₁: presence of a difference). Evidence thresholds were interpreted as follows: BF₀₁ > 10 = strong evidence for H₀; 3–10 = moderate evidence for H₀; 1–3 = weak evidence for H₀; 1/3–1 = weak evidence for H₁; 1/10–1/3 = moderate evidence for H₁; and < 1/10 = strong evidence for H₁. Bayes factors between 1/3 and 3 were classified as inconclusive. CI confidence interval For valence, the overall Bayesian ANOVA strongly favored the null hypothesis (BF₀₁ = 36.55). Pairwise comparisons largely supported this result. Human versus GPT (BF₀₁ = 20.68, d = − 0.02, 95% CI [− 0.09, 0.06]) and Human versus Persona (BF₀₁ = 19.40, d = 0.02, 95% CI [− 0.06, 0.10]) both strongly supported the null, whereas Human versus Polish provided only weak support for equivalence (BF₀₁ = 1.05, d = 0.10, 95% CI [0.02, 0.18]). Within AI conditions, GPT versus Persona strongly favored the null (BF₀₁ = 15.10, d = 0.04, 95% CI [− 0.04, 0.11]), Persona versus Polish moderately favored the null (BF₀₁ = 3.73, d = 0.08, 95% CI [0.00, 0.15]), and GPT versus Polish remained inconclusive (BF₀₁ = 0.64, d = 0.11, 95% CI [0.03, 0.18]). Across all human–AI valence comparisons, effect sizes ranged from |d| = 0.02 to 0.10, remaining below Cohen’s threshold for a small effect (d = 0.20). For approach–avoidance, the overall Bayesian ANOVA again strongly favored the alternative hypothesis (BF₀₁ < 0.001). All human–AI comparisons strongly supported the alternative hypothesis (Human vs. GPT: BF₀₁ < 0.001, d = 0.30, 95% CI [0.23, 0.38]; Human vs. Persona: BF₀₁ < 0.001, d = 0.34, 95% CI [0.27, 0.42]; Human vs. Polish: BF₀₁ < 0.001, d = 0.47, 95% CI [0.39, 0.55]), with the Polish condition showing the largest divergence from human ratings. Within AI conditions, GPT versus Persona moderately favored the null (BF₀₁ = 3.38, d = 0.08, 95% CI [0.00, 0.16]), Persona versus Polish showed only weak support for the null (BF₀₁ = 1.01, d = 0.10, 95% CI [0.02, 0.18]), whereas GPT versus Polish differed reliably (BF₀₁ < 0.001, d = 0.20, 95% CI [0.12, 0.28]). Across all human–AI comparisons, effect sizes ranged from |d| = 0.30 to 0.47, corresponding to small-to-medium effects by Cohen’s criteria [35]. Exploratory analyses were conducted separately across five image categories—animals (n = 199), faces (n = 360), landscapes (n = 185), objects (n = 322), and people (n = 210)—to determine whether AI–human agreement varied by stimulus content beyond the predictions of Hypothesis 3. All category-specific Pearson correlations remained significant after Bonferroni correction (p < 0.001; see Fig. 3 and Supplementary Table 2). Fig. 3Human–GPT agreement on emotional ratings across categories and conditions. Heatmaps display agreement between human ratings and three GPT conditions (GPT, Persona, and Polish) for arousal (top), valence (middle), and approach–avoidance (bottom) across five image categories. Left panels show Pearson’s r; right panels show ICC(2,1), two-way mixed-effects, absolute agreement, single measurement [30]. Darker cells indicate stronger agreement Human–GPT agreement on emotional ratings across categories and conditions. Heatmaps display agreement between human ratings and three GPT conditions (GPT, Persona, and Polish) for arousal (top), valence (middle), and approach–avoidance (bottom) across five image categories. Left panels show Pearson’s r; right panels show ICC(2,1), two-way mixed-effects, absolute agreement, single measurement [30]. Darker cells indicate stronger agreement Agreement varied substantially across categories. Landscapes produced the strongest AI–human agreement, with correlations ranging from r = 0.705, 95% CI [0.62, 0.77] (human–Polish, approach–avoidance), to r = 0.933, 95% CI [0.91, 0.95] (human–GPT, valence). Faces consistently produced the weakest agreement, with correlations ranging from r = 0.253, 95% CI [0.15, 0.35] (human–Polish, approach–avoidance), to r = 0.596, 95% CI [0.52, 0.66] (human–GPT, valence). A stable pattern also emerged across emotional dimensions. Valence produced the strongest correlations in all five categories, ranging from r = 0.521, 95% CI [0.44, 0.59] (faces, human–Polish), to r = 0.933, 95% CI [0.91, 0.95] (landscapes, human–GPT), while also showing the lowest cross-category variability (range = 0.41). Arousal showed substantially greater category sensitivity, with correlations spanning r = 0.264, 95% CI [0.17, 0.36] (faces, human–GPT), to r = 0.838, 95% CI [0.79, 0.88] (landscapes, human–Persona) (range = 0.57). Approach–avoidance showed comparable sensitivity (range = 0.59), with face correlations (r = 0.253–0.372) remaining lower than all other categories (r = 0.497–0.846). Intraclass correlation coefficients (ICC[2,1], absolute agreement, single measurement) reinforced these patterns. Using the confidence-interval-based criteria of Koo and Li (2016), valence achieved good-to-excellent agreement for landscapes (ICC = 0.880–0.910) and good agreement for objects (ICC = 0.828–0.845), people (ICC = 0.833–0.844), and animals (ICC = 0.805–0.828). Faces were the only category in which valence agreement remained moderate to poor (ICC = 0.437–0.569). For arousal, ICC values were moderate for landscapes (0.700–0.750), people (0.591–0.673), and animals (0.503–0.626), intermediate for objects (0.467–0.658), and poor for faces (0.180–0.273). Approach–avoidance again showed the weakest agreement for faces (ICC = 0.136–0.238, poor), whereas the remaining categories produced predominantly moderate-to-good agreement (ICC = 0.440–0.768). Across categories, the Polish condition generally produced comparable or slightly lower ICCs than GPT or Persona, particularly in the approach–avoidance dimension, mirroring the broader pattern observed in the overall analyses. Distributional differences between human and AI evaluations Anderson–Darling tests identified significant distributional differences between human and AI evaluations across all conditions and emotional dimensions (Table 5). Corresponding density distributions are shown in Fig. 4. Table 5Anderson–Darling test results for distributional differences between human and AI evaluationsDimensionComparisonAD Statistic p Adjusted pᵃArousalHuman vs. GPT113.190< 0.001< 0.001Human vs. Persona104.720< 0.001< 0.001Human vs. Polish64.612< 0.001< 0.001ValenceHuman vs. GPT22.435< 0.001< 0.001Human vs. Persona22.505< 0.001< 0.001Human vs. Polish29.591< 0.001< 0.001Approach-AvoidanceHuman vs. GPT131.27༐< 0.001< 0.001Human vs. Persona86.926< 0.001< 0.001Human vs. Polish207.81༐< 0.001< 0.001All comparisons remained significant after Benjamini–Hochberg correction for multiple testing. Larger Anderson–Darling (AD) statistics indicate greater distributional divergence between human and AI evaluations Table 5Anderson–Darling test results for distributional differences between human and AI evaluationsDimensionComparisonAD Statistic p Adjusted pᵃArousalHuman vs. GPT113.190< 0.001< 0.001Human vs. Persona104.720< 0.001< 0.001Human vs. Polish64.612< 0.001< 0.001ValenceHuman vs. GPT22.435< 0.001< 0.001Human vs. Persona22.505< 0.001< 0.001Human vs. Polish29.591< 0.001< 0.001Approach-AvoidanceHuman vs. GPT131.27༐< 0.001< 0.001Human vs. Persona86.926< 0.001< 0.001Human vs. Polish207.81༐< 0.001< 0.001All comparisons remained significant after Benjamini–Hochberg correction for multiple testing. Larger Anderson–Darling (AD) statistics indicate greater distributional divergence between human and AI evaluations Anderson–Darling test results for distributional differences between human and AI evaluations All comparisons remained significant after Benjamini–Hochberg correction for multiple testing. Larger Anderson–Darling (AD) statistics indicate greater distributional divergence between human and AI evaluations Fig. 4Anderson–Darling distributional analysis of human versus AI emotional evaluations. Kernel density estimates show mean ratings (1–9 scale) for each condition and emotional dimension. Human ratings were compared with each AI condition (GPT, Persona, and Polish) using two-sample Anderson–Darling tests with Benjamini–Hochberg-adjusted p-values Anderson–Darling distributional analysis of human versus AI emotional evaluations. Kernel density estimates show mean ratings (1–9 scale) for each condition and emotional dimension. Human ratings were compared with each AI condition (GPT, Persona, and Polish) using two-sample Anderson–Darling tests with Benjamini–Hochberg-adjusted p-values The magnitude of distributional divergence differed markedly across dimensions. Valence produced the smallest Anderson–Darling statistics, ranging from AD = 22.435 (human vs. GPT) to AD = 29.591 (human vs. Polish). Arousal showed substantially larger discrepancies, with AD values ranging from 64.612 (human vs. Polish) to 113.190 (human vs. GPT); the human versus Persona comparison produced AD = 104.720. The largest discrepancies emerged in the approach–avoidance dimension, where AD statistics ranged from 86.926 (human vs. Persona) to 207.810 (human vs. Polish). The latter represented the largest distributional divergence observed in the entire analysis. Condition-level patterns also differed by dimension. In arousal, the Polish condition showed the smallest divergence from human ratings (AD = 64.612), whereas GPT (AD = 113.190) and Persona (AD = 104.720) diverged more strongly. This pattern reversed in the approach–avoidance dimension: the Polish condition showed by far the greatest divergence (AD = 207.810), exceeding the Persona value (AD = 86.926) by more than twofold. The Polish AD statistic for approach–avoidance was the largest observed across all nine human–AI comparisons. Discussion This study examined whether cultural persona assignment or native-language prompting alters GPT-4o’s evaluation of emotional image content, using Poland as a representative Eastern European context that remains underrepresented in AI research. Using 1,276 NAPS images, three experimental conditions—standard GPT-4o, persona-type GPT, and Polish-language GPT—were compared against human reference ratings across valence, arousal, and approach–avoidance dimensions. The study contributes methodologically in two ways. First, it directly evaluates native-language prompting in visual emotion recognition, extending prior work focused primarily on text-based tasks [5, 22]. Second, it applies a multi-level analytical framework integrating correlation, Bayesian, and distributional analyses. Pearson correlations and intraclass correlation coefficients assessed whether AI and human ratings preserved relative stimulus ordering, Bayesian analyses assessed agreement in central tendency, and Anderson–Darling tests assessed agreement in distributional structure. These approaches yielded partially convergent but distinct conclusions. Valence showed strong rank-order and mean-level agreement despite persistent distributional differences, whereas arousal and approach–avoidance combined moderate rank-order agreement with reliable mean-level and distributional divergence. Taken together, the findings demonstrate that no single statistical approach adequately characterizes AI–human agreement in emotion evaluation. Positive correlations emerged across all conditions and dimensions (r = 0.525–0.809, all p < 0.001), indicating that AI systems broadly preserved the relative ordering of emotional stimuli. However, agreement strength varied by dimension (valence: r = 0.784–0.809; arousal: r = 0.675–0.703; approach–avoidance: r = 0.525–0.680). Bayesian analyses revealed dimension-specific patterns in central-tendency agreement. Human and AI ratings were highly similar for valence (overall BF₀₁ = 36.55; maximum human–AI |d| = 0.10) but diverged in arousal (overall BF₀₁ < 0.001; human–AI |d| = 0.30–0.54) and approach–avoidance (overall BF₀₁ < 0.001; human–AI |d| = 0.30–0.47). The Polish-language condition showed the largest divergence in approach–avoidance (d = 0.47). Anderson–Darling analyses further detected significant distributional differences across all nine human–AI comparisons (adjusted p < 0.001).Hypothesis 1 predicted that persona prompting and Polish-language prompting would improve agreement with human evaluations relative to standard GPT-4o. The results did not support this prediction consistently. Although the Polish-language condition showed closer agreement with human ratings in arousal, and the persona condition produced the strongest rank-order agreement in approach–avoidance, neither adaptation strategy improved agreement systematically across dimensions or analytical methods. Distributional differences from human ratings also persisted in every human–AI comparison regardless of prompting condition. Hypothesis 1 predicted that persona prompting and Polish-language prompting would improve agreement with human evaluations relative to standard GPT-4o. The results did not support this prediction consistently. Although the Polish-language condition showed closer agreement with human ratings in arousal, and the persona condition produced the strongest rank-order agreement in approach–avoidance, neither adaptation strategy improved agreement systematically across dimensions or analytical methods. Distributional differences from human ratings also persisted in every human–AI comparison regardless of prompting condition. Hypothesis 1 predicted that persona prompting and Polish-language prompting would improve agreement with human evaluations relative to standard GPT-4o. The results did not support this prediction consistently. Although the Polish-language condition showed closer agreement with human ratings in arousal, and the persona condition produced the strongest rank-order agreement in approach–avoidance, neither adaptation strategy improved agreement systematically across dimensions or analytical methods. Distributional differences from human ratings also persisted in every human–AI comparison regardless of prompting condition. In the valence dimension, correlation analyses showed strong rank-order agreement across all AI conditions (r = 0.784–0.809). Bayesian analyses likewise indicated minimal mean-level differences between human and AI ratings (|d| = 0.02–0.10). However, Anderson–Darling tests still identified significant distributional differences in all comparisons (AD = 22.435–29.591, adjusted p < 0.001). Across all three analytical levels, neither persona nor Polish-language prompting outperformed standard GPT-4o consistently. For arousal, correlations indicated moderate rank-order agreement across all AI conditions (r = 0.675–0.703). Bayesian analyses identified reliable human–AI differences in every condition (Human vs. GPT: d = 0.54; Human vs. Persona: d = 0.53; Human vs. Polish: d = 0.30). The Polish-language condition showed the smallest divergence from human ratings, whereas Persona and standard GPT produced nearly identical effect sizes. Anderson–Darling analyses confirmed distributional differences in all conditions, again with the Polish condition showing the smallest divergence (AD = 64.612) and standard GPT the largest (AD = 113.190). Although Polish-language prompting aligned more closely with human evaluations in this dimension, the effect was not sufficiently consistent to support broad claims of adaptation benefit. Approach–avoidance produced the weakest rank-order agreement overall (r = 0.525–0.680). Bayesian analyses again identified reliable human–AI differences across all conditions (Human vs. GPT: d = 0.30; Human vs. Persona: d = 0.34; Human vs. Polish: d = 0.47), with the Polish-language condition diverging most strongly from human ratings. Anderson–Darling analyses reinforced this result: the Polish condition produced the largest distributional discrepancy (AD = 207.810) across all nine human–AI comparisons. Thus, unlike the arousal dimension—where Polish-language prompting aligned most closely with human ratings—the same condition produced the greatest divergence in approach–avoidance across rank-order, mean-level, and distributional analyses. Prior research has proposed persona prompting and native-language prompting as mechanisms for adapting LLMs to specific cultural and linguistic contexts [3, 5, 6, 10, 20]. The present findings indicate that these strategies influenced AI emotion ratings, but their effects were inconsistent across emotional dimensions and analytical methods. In the arousal dimension, the Polish-language condition aligned more closely with human ratings than standard GPT in both rank-order agreement (r = 0.703 vs. 0.675) and central tendency (d = 0.30 vs. 0.54). In approach–avoidance, the persona condition produced the strongest rank-order agreement among AI conditions (r = 0.680), although it did not improve central-tendency agreement relative to standard GPT. In the valence dimension, neither persona nor Polish-language prompting outperformed standard GPT across any analytical framework. Despite these partial improvements, Anderson–Darling tests identified persistent distributional differences between AI and human ratings across all dimensions and conditions, with the Polish-language condition producing the largest divergence in approach–avoidance (AD = 207.810). Collectively, these findings suggest that persona and native-language prompting may produce localized improvements in human–AI agreement but do not provide robust or generalizable adaptation effects. Recent benchmarking has suggested that Polish can outperform English in long-context LLM tasks [19], which would predict a clearer advantage for the Polish-language condition. However, this short-prompt visual emotion task did not reproduce that pattern consistently. The dimension-specific and analysis-specific nature of the observed effects, combined with persistent distributional divergence across all comparisons, argues for cautious application of these prompting strategies in practice. Beyond comparisons between AI conditions, Hypothesis 2 (emotional dimension difficulty) examined whether human–AI agreement differed systematically across emotional dimensions. The hypothesis predicted the ordering valence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document} approach–avoidance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\to\:$$\end{document} arousal. Across analyses, the data showed a stable dimension-specific structure. Correlation analyses indicated that agreement with human ratings was strongest for valence (r = 0.784–0.809), intermediate for arousal (r = 0.675–0.703), and weakest for approach–avoidance (r = 0.525–0.680). ICC(2,1) values showed the same ordering (valence: 0.775–0.797; arousal: 0.587–0.665; approach–avoidance: 0.450–0.640). Bayesian analyses similarly showed that valence ratings were highly similar between AI and human conditions (overall BF₀₁ = 36.55; all human–AI |d| ≤ 0.10), whereas arousal (overall BF₀₁ < 0.001; human–AI |d| = 0.30–0.54) and approach–avoidance (overall BF₀₁ < 0.001; human–AI |d| = 0.30–0.47) both showed reliable divergence. The stronger correspondence observed for valence may reflect its status as the most fundamental dimension in Russell’s circumplex model of emotion [13]. Cross-cultural research has consistently identified valence as a stable cross-linguistic and cross-cultural regularity, including robust associations between visual properties such as brightness and positive evaluation [36, 37], as well as its broader role as a constraint within emotional meaning systems [38]. Because valence relies on relatively direct and widely shared affective cues, AI systems may encode it more reliably during training. By contrast, arousal and approach–avoidance judgments likely depend more heavily on embodied, motivational, and interoceptive processes that are less directly recoverable from training data. Anderson–Darling analyses reinforced this pattern. Distributional divergence was smallest for valence (AD = 22.435–29.591), intermediate for arousal (AD = 64.612–113.190), and largest for approach–avoidance (AD = 86.926–207.810). Across all analytical approaches, valence consistently showed the strongest human–AI agreement, whereas approach–avoidance showed the weakest. The predicted intermediate position of approach–avoidance was not supported, however. Although Bayesian analyses showed comparable mean-level divergence for arousal and approach–avoidance, correlation and ICC analyses demonstrated stronger agreement for arousal than for approach–avoidance, reversing the predicted ordering between these two dimensions. Hypothesis 2 was therefore only partially supported. Two mechanisms may help explain the divergence between AI and human ratings in arousal and approach–avoidance. First, AI systems are fundamentally disembodied. Human arousal experience depends on interoceptive and physiological signals—including heart rate, autonomic activation, and muscle tension—to which AI systems lack direct access [16]. Emotional understanding of this kind has been argued to depend on embodied social experience that cannot be fully approximated from text-based corpora alone [18]. Second, GPT-4o content filtering excluded 80 images (5.9% of the original NAPS dataset) depicting highly negative content, thereby removing many stimuli likely to evoke the strongest arousal responses in humans. This filtering may have contributed both to the narrower approach–avoidance distributions observed in some AI conditions and to the consistently lower mean arousal ratings generated by AI systems. The relative contribution of these mechanisms cannot be determined from the present data and remains for future investigation.Hypothesis 3 (image category difficulty) predicted higher human–AI agreement for landscapes than for objects and faces. The results partially supported this prediction. Landscapes consistently produced the strongest agreement (r = 0.705–0.933 across conditions and dimensions), whereas faces consistently produced the weakest (r = 0.253–0.596). Every landscape correlation exceeded every face correlation. Intermediate categories—animals, objects, and people—showed dimension-dependent variability rather than a stable ordinal ordering (see Supplementary Table 2). Thus, Hypothesis 3 was supported for the predicted positions of landscapes (highest agreement) and faces (lowest agreement), but not for the full ranking of intermediate categories. Hypothesis 3 (image category difficulty) predicted higher human–AI agreement for landscapes than for objects and faces. The results partially supported this prediction. Landscapes consistently produced the strongest agreement (r = 0.705–0.933 across conditions and dimensions), whereas faces consistently produced the weakest (r = 0.253–0.596). Every landscape correlation exceeded every face correlation. Intermediate categories—animals, objects, and people—showed dimension-dependent variability rather than a stable ordinal ordering (see Supplementary Table 2). Thus, Hypothesis 3 was supported for the predicted positions of landscapes (highest agreement) and faces (lowest agreement), but not for the full ranking of intermediate categories. Hypothesis 3 (image category difficulty) predicted higher human–AI agreement for landscapes than for objects and faces. The results partially supported this prediction. Landscapes consistently produced the strongest agreement (r = 0.705–0.933 across conditions and dimensions), whereas faces consistently produced the weakest (r = 0.253–0.596). Every landscape correlation exceeded every face correlation. Intermediate categories—animals, objects, and people—showed dimension-dependent variability rather than a stable ordinal ordering (see Supplementary Table 2). Thus, Hypothesis 3 was supported for the predicted positions of landscapes (highest agreement) and faces (lowest agreement), but not for the full ranking of intermediate categories. The highest agreement observed for landscape images is consistent with the relatively universal character of emotional responses to natural environments. Evolutionary frameworks such as Wilson’s biophilia hypothesis [24] propose an innate human sensitivity to natural settings, potentially reducing cultural variability in emotional interpretation. In addition, low-level visual features such as color and brightness—both linked to direct affective responses [39]—provide salient cues readily extractable by AI systems. Agreement for animal images varied across emotional dimensions. Extensions of the biophilia hypothesis suggest that emotional responses to animals also possess evolutionary foundations, although these responses are moderated by personal experience and culturally learned attitudes toward particular species [39]. Relative to landscapes, animal stimuli contain greater contextual and experiential variability, which may contribute to the lower agreement observed in this category. Object images likewise showed dimension-dependent agreement. This pattern is consistent with the interaction of low-level perceptual properties and higher-order cognitive interpretation. Shape-related features, including curved versus angular forms, have established associations with affective responses [40], while anthropomorphic projection mechanisms can influence emotional interpretation of inanimate objects [25]. Because emotional responses to objects are partly shaped by learned associations and individual experience, agreement levels may vary across dimensions. Images depicting people (full-body images) produced substantially higher agreement than face-only stimuli and showed agreement levels comparable to those of animals and objects. Full-body images likely provide richer contextual information (e.g., posture, clothing, and situational cues) that supplements facial information and enables more reliable AI interpretation than isolated faces alone. By contrast, face images produced the weakest agreement across all categories and dimensions (Pearson r = 0.253–0.596; arousal ICC values = 0.180–0.273). This pattern is consistent with the combined demands of specialized neural processing, including involvement of the fusiform face area [41], and culturally specific display rules governing emotional expression [26, 27]. Cross-cultural research further demonstrates that facial emotion recognition is strongly shaped by culturally learned norms [11], requiring integration of subtle and multidimensional social cues. The category-specific findings have broader implications for AI-based emotion recognition. Agreement was strongest in categories allowing AI systems to rely on salient visual or contextual information, including landscapes, objects, animals, and full-body images. In contrast, the weaker performance observed for face-only stimuli highlights the difficulty of interpreting culturally embedded and cognitively complex emotional signals. Improving facial emotion recognition may therefore require richer culture-sensitive training data and more advanced modeling strategies rather than reliance on generic visual features alone. This study also has several methodological limitations. First, although 60% of the human evaluators were Polish, the remaining participants were exchange students enrolled at Polish universities. The resulting reference sample may therefore not fully reflect Polish cultural norms, limiting the precision of culture-specific inference. Second, GPT-4o content filtering excluded 80 images (5.9% of the original NAPS dataset), primarily highly negative stimuli such as corpses. This exclusion narrowed the emotional range under analysis, potentially inflating correlations through range restriction while preventing assessment of contexts requiring sensitivity to extreme emotional content, including crisis intervention and trauma-related evaluation. Such applications are themselves under active investigation, including AI-integrated interventions for students experiencing psychological distress [42] and broader reviews of web-based applications in neuropsychiatric treatment [43]. Continued dialogue between basic research characterizing AI’s affective evaluation profile—including the boundaries imposed by content-filtering policies—and these applied research streams will be essential to translate current promise into clinically robust practice. Third, the study did not include a non-English, non-Polish linguistic comparison condition (e.g., Spanish). Consequently, even if stronger Polish-language effects had emerged, the design would still not permit clean separation of linguistic and cultural influences. The Polish-language condition should therefore be interpreted as a single-language probe rather than a definitive test of culture-specific adaptation. Several directions for future research follow from these limitations. Longitudinal analyses across successive model updates could determine whether native-language effects change as AI systems evolve. Comparative studies spanning multiple cultural regions are also needed to assess whether the present findings generalize beyond the Polish context. Finally, more detailed investigation into how cultural context is represented and operationalized within AI language processing may support the development of more precise and effective adaptation strategies. Conclusions The present study examined whether persona prompting and native-language prompting improve agreement between AI and human emotion evaluation using GPT-4o and the NAPS image database. Three findings have direct implications for the practical use of LLMs in emotion-evaluation tasks. First, persona prompting and native-language prompting did not produce generalized improvements in human–AI agreement. Where improvements emerged—Polish-language prompting in arousal and persona prompting in approach–avoidance rank-order agreement—they were restricted to specific dimensions and analytical frameworks, while distributional differences between AI and human ratings persisted in every comparison. These findings indicate that cultural-adaptation strategies do not operate uniformly and should be validated separately for each emotional dimension and application context before deployment. Second, agreement between AI and human ratings differed substantially across emotional dimensions. AI evaluation aligned closely with human ratings for valence but diverged for arousal and, most strongly, for approach–avoidance. Distributional differences nonetheless remained significant across all dimensions. Practical applications of AI-based emotion evaluation should therefore calibrate reliance on AI outputs according to the specific emotional dimension under consideration and incorporate human judgment when evaluating arousal or approach–avoidance. Third, agreement varied substantially across stimulus categories. AI evaluation showed the strongest agreement for landscapes and the weakest for faces, with animals, objects, and people occupying intermediate positions. Because facial emotion recognition depends on specialized perceptual and culturally conditioned processing, AI-based evaluation of facial emotion should be treated with particular caution, whereas evaluation of scenic or object-based stimuli may be comparatively more reliable. Several methodological limitations constrain the generalizability of these findings. The human reference sample consisted of 60% Polish nationals and 40% exchange students enrolled at Polish universities. Although acknowledged earlier as a limitation, this composition constrains interpretation of the Polish-language and persona conditions because divergence from the human reference cannot be attributed exclusively to mismatch with Polish cultural norms; part of the reference sample itself reflected non-Polish European perspectives. In addition, GPT-4o content filtering excluded 80 images (5.9% of the original NAPS dataset), disproportionately removing highly negative stimuli. The study also examined only a single linguistic and cultural context, limiting generalizability to other languages and regions. These findings do not invalidate cultural-adaptation strategies but indicate that such approaches require empirical validation on a dimension-specific basis. Future research should conduct simultaneous comparative analyses across multiple cultural regions, including East Asia, the Middle East, Africa, and South America; evaluate the influence of technical parameters such as temperature settings; and assess AI performance across the full emotional spectrum without content filtering. More broadly, the present findings underscore the importance of jointly examining rank-order agreement, central tendency, and distributional structure when evaluating cultural adaptation in AI systems. Supplementary Information Supplementary Material 1. Supplementary Material 1.