Evaluating Multimodal AI for Greenwashing Detection: A Comparative Analysis of ChatGPT, Claude, and Gemini in ESG Reports

Abstract

The rapid expansion of sustainability reporting under the EU Corporate Sustainability Reporting Directive (CSRD) has intensified concerns about greenwashing, particularly in visual communication within ESG reports. Recent advances in multimodal artificial intelligence offer new possibilities for automated detection, yet their reliability in non-English corporate reporting contexts remains unclear. This study evaluates the greenwashing detection capabilities of three leading multimodal AI systems—ChatGPT 5.1, Claude 4.5 Sonnet, and Gemini 2.5 Flash—using a purposively selected sample of 20 Polish ESG reports benchmarked against ESRS-aligned performance scores from the national “Ranking ESG”. A standardized auditing prompt was applied across all tools to generate comparable assessments of visual greenwashing. Contrary to theoretical expectations and all four hypotheses, the models did not demonstrate negative correlations between performance and AI-detected greenwashing; instead, high-performing firms frequently received higher greenwashing scores. Dimensional analyses showed inconsistent and often contradictory evaluations across Environmental, Social, and Governance pillars, while inter-tool reliability proved extremely low (Krippendorff’s α ≈ 0). These findings indicate that current multimodal AI systems conflate communication sophistication with deceptive intent and lack sufficient contextual understanding for ESG assurance. The study highlights significant methodological limitations and outlines directions for developing domain-specific, ESRS-aligned AI tools for greenwashing detection.

1. Introduction

The transition toward sustainable business practices mandates robust Environmental, Social, and Governance (ESG) reporting for corporate accountability. The European Union’s Corporate Sustainability Reporting Directive (CSRD), implemented alongside the European Sustainability Reporting Standards (ESRS) in July 2023 [1,2,3], represents a paradigmatic shift, expanding mandatory disclosure from approximately 12,000 to an estimated 50,000 companies. However, this proliferation of reporting has exacerbated concerns regarding greenwashing—making misleading claims—with recent data indicating an increased severity of incidents [4].

The advent of Artificial Intelligence (AI) and multimodal machine learning offers a scalable solution for systematic greenwashing detection. AI-driven approaches, particularly those leveraging Natural Language Processing (NLP), Retrieval-Augmented Generation (RAG), and Large Language Models (LLMs), can analyze both textual and complex visual elements in corporate communications to identify deceptive practices and assign trustworthiness scores [5,6].

The Polish context presents a particularly compelling research environment for investigating automated greenwashing detection. On 13 December 2024, Poland implemented the CSRD into national law through amendments to the Act on Accounting, Statutory Auditors, Audit Firms, and Public Oversight [7]. Polish companies are now navigating the transition from voluntary sustainability disclosure to mandatory ESRS-compliant reporting, with large public-interest entities with more than 500 employees required to report for fiscal year 2024 [8]. Research indicates that Polish companies demonstrate varying levels of preparedness for ESRS-compliant reporting, with particular challenges in environmental disclosures, the translation of EU regulations into reduced environmental pressure, and the strategic demands of industrial decarbonisation [9,10,11]. The existence of established ESG ranking systems, such as “Ranking ESG. Odpowiedzialne Zarządzanie”, which aligns with European Sustainability Reporting Standards, provides a benchmark for validating AI-based detection methods [12].

Despite these technological advances and regulatory developments, a critical research gap exists in the systematic evaluation of leading multimodal AI tools’ capabilities to detect greenwashing in non-English language contexts, particularly for visual greenwashing patterns in corporate reports. While previous studies have explored AI applications in ESG analysis [13] and developed greenwashing detection frameworks [14], comprehensive comparative evaluations of commercially available AI systems against validated ESG performance benchmarks remain scarce. Furthermore, the specific challenges of detecting visual greenwashing—through misleading graphs, selective imagery, and manipulative infographics—have received limited attention in the automated detection literature.

This study addresses these gaps by conducting a comparative evaluation of three leading multimodal AI systems (ChatGPT Plus 5.1, Claude 4.5 Sonnet, and Gemini 2.5 Flash) in detecting greenwashing patterns within Polish ESG reports. The research employs a novel methodology that combines visual rhetoric analysis with ESRS-aligned performance benchmarking, utilizing a purposively stratified sample of 20 Polish companies with documented ESG performance variations. By focusing on the intersection of AI capabilities, regulatory compliance requirements, and linguistic-cultural specificity, this study contributes to both the theoretical understanding of automated greenwashing detection and the practical implementation of AI-assisted ESG verification systems.

This study addresses the critical research gap in systematically validating multimodal AI tools for detecting visual greenwashing in non-English, mandatory reporting contexts.

The primary aim of this research is to systematically evaluate and comparatively validate the capability, accuracy, and limitations of leading multimodal AI systems in detecting visual and textual greenwashing patterns within Polish corporate ESG reports against established, ESRS-aligned performance benchmarks. This aim is operationalized through the following research questions:

RQ1 (Aggregate): To what extent do the AI-generated greenwashing scores (textual and visual) correlate with the external, ESRS-aligned ESG performance of Polish companies?

RQ2 (Dimensional variation): Does the strength of the correlation between AI-detected greenwashing scores and external performance vary significantly across the Environmental, Social, and Governance (ESG) dimensions?

RQ3 (Compensatory communication): For companies exhibiting significant performance gaps between their strongest and weakest ESG dimensions, do the multimodal AI systems detect a significantly higher intensity of greenwashing (visual and textual) in the weaker dimension?

RQ4 (Inter-tool agreement): What is the level of inter-rater agreement (reliability) among the three leading multimodal AI systems (ChatGPT Plus 5.1, Claude 4.5 Sonnet, and Gemini 2.5 Flash) when assigning greenwashing scores and classifications?

RQ5 (Feasibility and limitations): What are the practical limitations and implications of using accessible multimodal AI systems for greenwashing detection in supporting regulatory compliance, auditing, and stakeholder decision-making under the new CSRD/ESRS framework?

The findings have significant implications for regulators, investors, auditors, and sustainability practitioners navigating the evolving landscape of mandatory ESG disclosure in the European Union.

2. Literature Review and Theoretical Framework

2.1. Greenwashing: Conceptual Evolution and Dimensionality

The concept of greenwashing emerged in the 1980s, coined by Jay Westerveld to criticize the hotel industry’s misleading environmental claims regarding towel reuse. While early scholarship focused primarily on product advertising, the concept has evolved substantially alongside intensifying stakeholder pressure for corporate accountability. A seminal shift occurred with Delmas and Burbano [15], who advanced a rigorous definition characterizing greenwashing as “the intersection of two firm behaviors: poor environmental performance and positive communication about environmental performance”. This definition established the critical principle that greenwashing acts as a discrepancy between disclosure and reality—a principle central to our research design.

Recent scholarship has expanded the conceptualization beyond environmental claims to encompass the broader ESG framework. Lyon and Montgomery [16] synthesized diverse forms of decoupling, distinguishing between selective disclosure (omitting negative information) and symbolic actions lacking substantive implementation. This evolution reflects the transition of corporate reporting from a narrow environmental focus to comprehensive ESG integration mandated by frameworks such as the GRI and the European Sustainability Reporting Standards (ESRS).

The dimensionality of greenwashing has also received increasing scrutiny. While Seele and Gatti [17] redefined the phenomenon as a legitimacy crisis triggered by accusations, other scholars have focused on the specific channels of deception. Notably, de Freitas Netto et al. [18] and Parguel et al. [19] differentiate between claim greenwashing (textual assertions) and executional greenwashing (misleading visual or aesthetic presentation). This distinction highlights that greenwashing operates multimodally, combining textual claims with visual rhetoric to manage stakeholder impressions.

2.2. Visual Greenwashing and Sustainability Report Design

Visual communication in sustainability reports has emerged as a critical vector for greenwashing, yet it remains significantly under-researched compared to textual analysis. Building on Impression Management theory, Merkl-Davies and Brennan [20] suggest that visual elements are often used to obfuscate poor performance. Hrasky [21] empirically demonstrated the visual rhetoric compensation hypothesis, finding that companies with lower environmental performance rely more heavily on photographs and visual metaphors to distract from unfavorable data.

Scholars have identified several specific visual tactics employed in this context [22,23]:

• Nature imagery: the use of pristine landscapes (e.g., forests, blue skies) disconnected from actual business operations to prime positive associations.
• Aesthetic manipulation: utilizing specific color palettes (e.g., greenery, earth tones) to trigger psychological heuristics associated with sustainability.
• Data visualization bias: designing charts that obscure negative trends or emphasize minor favorable metrics through scale manipulation.
• Visual complexity: creating overwhelming layouts to obscure key performance indicators (KPIs).

These tactics exploit cognitive heuristics, allowing companies to convey a green impression regardless of textual accuracy. Crucially, experimental research by Parguel et al. [19] confirmed that nature imagery significantly influences consumer perceptions of environmental performance independent of actual attributes. Similarly, Cho et al. [24] demonstrated that the visual sophistication of reports can influence stakeholder trust, often masking the substance of disclosures.

2.3. Theoretical Framework: Legitimacy Theory and Visual Rhetoric

Our framework integrates legitimacy theory with visual rhetoric analysis. Legitimacy theory posits that organizations seek to align their activities with societal norms to ensure survival [25]. When a legitimacy gap emerges—a disconnect between performance and expectations—companies deploy symbolic management strategies to restore perceived legitimacy.

Cho et al. [26] demonstrated that companies facing legitimacy threats respond through enhanced disclosure sophistication rather than substantive performance improvement. Visual elements serve as powerful legitimacy tools because they convey emotional meanings that bypass rational evaluation [27]. We integrate this with Birdsell and Groarke’s [28] theory of visual argument, which posits that images function as persuasive arguments through implicit association.

Building on the theoretical framework of legitimacy theory and visual rhetoric in sustainability communication, we advance four hypotheses regarding AI-based greenwashing detection in Polish ESG reports

H1 (Aggregate validation): We hypothesize that AI-generated greenwashing scores will exhibit a moderate negative correlation (r ∈ [−0.40, −0.60], p < 0.05) with external ESG performance benchmark. We advance the compensatory communication hypothesis: companies with weak ESG performance will compensate through enhanced visual rhetoric. Therefore, we should observe a negative correlation between actual ESG performance and AI-detected greenwashing signals. This predicted moderate strength acknowledges that visual greenwashing represents only one component of a company’s total reporting strategy. We anticipate moderate rather than strong correlations because some high performers may employ aggressive visual communication to highlight achievements, while some poor performers may adopt conservative disclosure approaches to avoid scrutiny.

H2 (Dimensional variation): The strength of the negative correlation between AI greenwashing scores and performance will vary systematically across the three ESG dimensions, with differences tested via Fisher’s r-to-z transformation. We hypothesize the strongest negative correlation for Environmental (r < −0.50) due to the higher objectivity and quantifiability of environmental metrics; the weakest correlation for Social (r ∈ [−0.20, −0.35]) due to greater subjectivity and cultural contextualization; and intermediate correlation for Governance (r ∈ [−0.35, −0.45]). This pattern reflects the differential nature of ESG dimensions, where environmental metrics allow clearer performance-disclosure comparison, while social performance involves cultural interpretation that may complicate AI assessment.

H3 (Dimensional gap effect): Companies exhibiting performance gaps exceeding 30 percentage points between ESG dimensions (operationalized as |max(E,S,G) − min(E,S,G)| > 30) will demonstrate significantly higher AI-detected greenwashing intensity in their weakest dimension compared to their strongest dimension (Wilcoxon signed-rank test, p < 0.05, effect size r > 0.50). This hypothesis directly tests compensatory communication strategy as predicted by legitimacy theory, where companies attempt to visually offset poor performance in critical areas through enhanced visual rhetoric.

H4 (Inter-tool variation): The three multimodal AI tools (ChatGPT Plus 5.1, Claude 3.5 Sonnet, and Gemini 1.5 Pro) will demonstrate substantial but incomplete inter-rater agreement (Krippendorff’s α ∈ [0.65, 0.75]) in assigning greenwashing scores. This incomplete agreement suggests complementary rather than redundant capabilities, reflecting differences in training architectures and optimization objectives.

3. Materials and Methods

3.1. Research Design

This study employs a quantitative comparative evaluation design using correlation analysis to assess multimodal AI tools’ ability to detect visual greenwashing in ESG reports. The research design addresses the core question: Can multimodal AI tools accurately detect visual greenwashing in ESG reports when validated against external performance benchmarks? We operationalize “accuracy” through correlation between AI-generated greenwashing assessments and actual ESG performance scores, following the definitional principle that greenwashing exists where poor performance coincides with positive communication.

The research proceeded through four sequential phases: (1) Sample selection and data collection, (2) Standardized AI testing, (3) Quantitative analysis, (4) Synthesis and interpretation. The design deliberately prioritizes external validity through real-world corporate reports over experimental control through artificial stimuli, accepting measurement noise in exchange for practical relevance to CSRD implementation contexts.

3.2. Validation Benchmark: Ranking ESG

We selected “Ranking ESG. Odpowiedzialne Zarządzanie” as our primary validation benchmark based on four critical criteria: ESRS alignment, dimensional breakdown, methodological rigor, and Polish market focus. Organized by Koźmiński University Business Hub in partnership with Deloitte Poland, this ranking evaluates companies against European Sustainability Reporting Standards (ESRS) categories, providing separate percentage scores for Environmental (E), Social (S), and Governance (G) performance dimensions—a structure essential for dimension-specific validation unavailable in aggregate ESG ratings. While acknowledging limitations inherent in any single benchmark (potential measurement error, possible self-reporting bias, temporal lag between assessment and report publication), the ranking’s ESRS alignment makes it particularly relevant for CSRD implementation research. Alternative international benchmarks (MSCI ESG, Sustainalytics, Bloomberg ESG) were considered but rejected due to lack of dimensional breakdown, or limited Polish market coverage.

3.3. Sample Selection

We employed purposive stratified sampling to select 20 companies from the Ranking ESG participant pool, maximizing variance across ESG performance levels and sectoral representation.

Table 1. Sample companies with dimensional ESG performance scores.

Company	Sector	E%	S%	G%	Total Points
Santander Bank Polska	Banking, financial and insurance sector	100	100	100	100
Coca-Cola HBC	Consumer goods	100	97	100	99
Ergo Hestia	Banking, financial and insurance sector	100	94	100	98
Żywiec Zdrój	Consumer goods	100	94	100	98
Orange Polska	Telecommunications, technology, media	100	100	93	98
Castorama	Trade	89	95	86	87
PSE	Energy, fuels, mining	82	78	87	82
Holcim	Industrial and chemical production	85	69	93	82
mBank	Banking, financial and insurance sector	88	64	100	82
Bogdanka	Energy, fuels, mining	74	72	90	78
Polpharma	Pharmaceuticals and medicine	82	69	77	76
Unimot	Transport and logistics	53	81	93	75
Hegelmann Transporte	Transport and logistics	76	64	83	74
Edge1 Solution	Telecommunications, technology, media	57	75	80	74
Grupa Luxmed	Pharmaceuticals and medicine	88	53	83	71
CD Projekt	Telecommunications, technology, media	69	31	57	50
PKO BP	Banking, financial sector and insurance	100	25	60	49
Ikano Bank	Banking, financial sector and insurance	85	36	43	47
GS1 Polska	Telecommunications, technology, media	38	19	70	41
Amica	Industrial and chemical production	46	39	37	40

Note: E% = Environmental performance, S% = Social performance, G% = Governance performance.

presents the complete sample with dimensional performance scores.

Sample stratification achieved three objectives. First, performance distribution spans the full range, with 5 companies (25%) scoring top of the ranking, 10 companies (50%) in the medium range, and 5 companies (25%) with the lowest score in the ranking, ensuring variance necessary for correlation analysis. Second, sectoral diversity includes 8 distinct industry categories: banking/financial services/insurance (n = 5), telecommunications/technology/media (n = 4), consumer goods (n = 2), energy/fuels/mining (n = 2), industrial and chemical production (n = 2), transport and logistics (n = 2), pharmaceuticals and medicine (n = 2), and trade (n = 1), enabling examination of industry-specific greenwashing patterns. Third, ownership structure variation encompasses international subsidiaries (e.g., Santander, Orange Polska, Coca-Cola HBC), domestic private entities (e.g., CD Projekt, Polpharma), and mixed ownership structures, addressing potential differences in reporting sophistication and stakeholder pressure.

3.4. ESG Report Collection

For sustainability report (reporting years 2023–2024) [29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48]. Reports were sourced from corporate websites, investor relations pages, or parent company disclosures for subsidiaries. Document characteristics were systematically recorded: total pages, report type (standalone sustainability report, integrated annual report, dedicated ESG section), language(s), and format (PDF). In case of the integrated annual reports ESG sections were extracted for further analysis

The sample exhibits substantial variance in report sophistication: page counts range from 21 to 291 pages (mean = 91.95, SD = 69.62), reflecting dramatically different approaches to ESG disclosure. The distribution shows notable extremes: PSE’s 291-page report and mBank’s 217-page document contrast sharply with concise reports from Żywiec Zdrój (21 pages), Holcim (23 pages), and Unimot (23 pages). This variance—with a coefficient of variation of 75.71%—suggests no standardized reporting format in the Polish market. Report length clusters emerge: compact reports under 50 pages (n = 8, 40%), moderate reports 50–100 pages (n = 4, 20%), substantial reports 100–150 pages (n = 6, 30%), and extensive reports exceeding 150 pages (n = 2, 10%).

3.5. AI Tool Selection and Testing Protocol

We selected three leading multimodal AI systems based on market adoption, documented capabilities, and practical accessibility: OpenAI’s ChatGPT Plus 5.1, Anthropic’s Claude 4.5 Sonnet, and Google’s Gemini 2.5 Flash. These systems represent the current advances in multimodal AI, each employing different architectural approaches and training methodologies. Selection excluded open source alternatives due to inferior performance on complex document analysis tasks.

We developed a standardized prompting protocol applied identically across all tools and reports to ensure comparability. The prompt underwent iterative refinement through pilot testing, optimizing for: (1) clear task definition, (2) structured output format enabling quantitative analysis, (3) dimensional breakdown aligned with ESRS categories, (4) specific greenwashing technique identification, (5) confidence level assessment enabling reliability evaluation. The finalized prompt instructs each AI to function as an “ESG auditor detecting greenwashing”, analyze the uploaded report, and provide outputs in a standardized format.

The prompt requests five components in each response: (1) Numerical scores (0–10 scale where 0 = authentic, 10 = severe greenwashing) for overall greenwashing, environmental dimension (E), social dimension (S), and governance dimension (G), with scores summed to total (0–40 range); (2) Categorical classification (authentic: 0–12 total, mixed: 13–24, greenwashing: 25–40) to enable agreement analysis; (3) Specific evidence: 5–7 concrete examples with page numbers and descriptions; (4) Dimension analysis: 2–3 sentence assessment for each E/S/G dimension explaining scoring rationale; (5) Confidence level (low/medium/high) with 1–2 sentence justification, enabling assessment of tool certainty. The complete prompt, which defines the scoring scale, output structure, and the targeted greenwashing techniques, is provided in Appendix A for replication purposes.

Testing procedure followed strict protocol to ensure consistency and minimize confounds. Each report was uploaded to each of the three AI tools as a separate task in randomized order (preventing sequence effects). The standardized prompt was submitted without modification. Testing occurred within a one-week period to minimize temporal effects from potential model updates. All AI responses were captured in full-text format, with structured data extracted into a standardized coding template for quantitative analysis.

Cost and accessibility were crucial factors in selecting the AI tools for this study. We deliberately prioritized publicly available large language models (LLMs) to ensure the findings are relevant and practically accessible to a wider audience, including academic researchers and small- and medium-sized ESG consulting firms. This approach contrasts with high-barrier enterprise alternatives or complex API-based solutions that require specialized data science infrastructure and advanced technical implementation expertise. By focusing on tools deployable by practitioners without specialized resources, this accessible AI strategy directly reflects the study’s applied objective: evaluating systems ready for immediate use in real-world ESG verification.

3.6. Analytical Approach

We employ a multi-stage quantitative analytical strategy to rigorously test Hypotheses H1 through H4. The primary validation analysis examines the linear relationship between AI-generated greenwashing scores and external ESG performance benchmarks.

• Aggregate validation (H1): We will calculate the Pearson correlation coefficient between the AI Overall Greenwashing Score and the validated Ranking ESG Composite Score.
• Dimensional variation (H2): Dimension-specific correlations will be calculated for each ESG pillar. The hypothesized difference in correlation strength across dimensions will be tested directly using Fisher’s r-to-z transformation on paired correlation coefficients.
• Dimensional gap effect (H3): This hypothesis is tested using a two-part approach. First, we will use Wilcoxon signed-rank tests (appropriate for non-normally distributed paired data) to compare the AI-detected greenwashing scores between the weakest and strongest dimensions for companies with a performance gap exceeding 30 points. Second, a correlation analysis will test if larger dimensional performance gaps predict larger corresponding AI greenwashing gaps.
• Inter-tool variation (H5): Inter-rater reliability will be assessed using Krippendorff’s Alpha for the categorical classifications and the Intraclass Correlation Coefficient (ICC) for absolute agreement on the continuous Total Score.

4. Results

4.1. Performance Assessment

The comparative evaluation of ChatGPT, Claude 3.5 Sonnet, and Gemini 2.5 Flash reveals substantial variations in their ability to detect greenwashing in Polish ESG reports.

Table 2. Overall visual greenwashing performance scores.

Company	Benchmark—Total Score	GPT	Claude	Gemini
Amica	40	2	6	7
Bogdanka	78	4	4	8
Castorama	87	4	6	7
CD Projekt	50	4	3	1
Coca-Cola HBC	99	7	4	6
Edge1 Solution	74	6	6	4
Ergo Hestia	98	4	3	7
Grupa Luxmed	71	6	3	7
GS1 Polska	41	3	5	4
Hegelmann Transporte	74	6	7	5
Holcim	82	6	5	4
Ikano Bank	47	4	6	4
mBank	82	4	3	3
Orange Polska	98	6	6	5
PKO BP	49	6	4	3
Polpharma	76	5	5	7
PSE	82	4	2	5
Santander Bank Polska	100	5	4	6
Unimot	75	4	6	7
Żywiec Zdrój	98	6	7	4

presents the aggregate performance scores across all 20 companies evaluated. The three LLMs demonstrated considerable inconsistency in greenwashing detection, with high inter-model score variations. Surprisingly, companies with exemplary ESG performance (Santander Bank Polska = 100, Coca-Cola HBC = 99) received moderate greenwashing scores (4–7 range) comparable to those of lower performers, suggesting that AI models may conflate legitimate sustainability communication with greenwashing tactics regardless of actual ESG achievement. The mean greenwashing scores were similar across models (GPT: M = 4.80, SD = 1.25; Claude: M = 4.75, SD = 1.44; Gemini: M = 5.20, SD = 1.78). Overall visual greenwashing performance scores are presented in Table 2.

Analysis of the environmental dimension reveals even more pronounced inconsistencies in AI-based greenwashing detection. Claude 4.5 Sonnet demonstrated the most balanced approach (M = 5.35, SD = 1.59). Gemini 2.5 Flash showed the highest average scores but with greater variability (M = 6.05, SD = 1.77), suggesting potential oversensitivity to environmental claims. ChatGPT exhibited the most conservative pattern (M = 5.7, SD = 1.31), consistently scoring companies in a narrow range. While GPT demonstrated a strong positive correlation with environmental benchmark scores (r = 0.608, p = 0.004), Claude showed a negative trend (r = −0.271 p = 0.24) and Gemini showed no relationship (r = −0.045 p = 0.85), resulting in an overall null correlation (r = 0.095, p = 0.690). Notably, the six companies with perfect environmental performance scores (100) still received moderate to high greenwashing assessments, with Coca-Cola HBC receiving the highest greenwashing score from GPT (8) despite its perfect environmental performance.

The environmental dimension shows higher greenwashing detection scores across all three LLMs compared to general ESG assessment (average increase of 0.78 points), possibly indicating that environmental claims trigger more skepticism from AI models or that companies use more promotional language when discussing environmental achievements, regardless of their actual performance level. The environmental performance scores are presented in

Table 3. Environmental performance scores.

Company	Benchmark—Environmental Score	GPT	Claude	Gemini
Amica	46	3	7	8
Bogdanka	74	5	5	9
Castorama	89	5	7	8
CD Projekt	69	5	4	2
Coca-Cola HBC	100	8	5	7
Edge1 Solution	57	7	7	5
Ergo Hestia	100	5	4	7
Grupa Luxmed	88	7	4	8
GS1 Polska	38	3	5	5
Hegelmann Transporte	76	7	8	6
Holcim	85	7	6	5
Ikano Bank	85	5	6	5
mBank	88	5	2	3
Orange Polska	100	7	5	6
PKO BP	100	6	5	4
Polpharma	82	6	6	6
PSE	82	5	2	7
Santander Bank Polska	100	6	5	7
Unimot	53	5	7	8
Żywiec Zdrój	100	7	7	5

.

The dimensional analysis reveals striking differences in AI greenwashing detection patterns across ESG components. Social dimension scores were systematically lower across all three LLMs (M = 3.73) compared to environmental (M = 5.70) and general ESG (M = 4.85) assessments, despite social performance having the lowest benchmark scores (M = 67.8) and highest variability (SD = 26.2). Unlike the Environmental category, none of the models (GPT, Claude, Gemini) achieved a statistically significant correlation with the official benchmark, indicating a failure to accurately reflect the true social standing of the companies. Furthermore, their evaluations are inconsistent, mutually contradictory, and effectively random in this context, rendering them currently unsuitable for professional ESG auditing in the social domain. Most notably, companies with perfect social performance scores (Orange Polska and Santander Bank Polska, both at 100) received higher greenwashing assessments (M = 4.00) than companies with poor social performance below 40 (M = 3.13), suggesting that LLMs may interpret legitimate social achievements as promotional rhetoric while overlooking potential greenwashing in poorly performing companies. This suggests that current LLMs lack adequate frameworks for evaluating social sustainability claims. This finding is particularly concerning given that social washing may be more subtle and culturally contextual than environmental greenwashing, requiring nuanced understanding of labor practices, community relations, and social impact metrics that appear beyond current AI capabilities. The social performance scores are presented in

Table 4. Social performance scores.

Company	Benchmark—Social Score	GPT	Claude	Gemini
Amica	39	2	5	5
Bogdanka	72	3	3	6
Castorama	95	3	5	6
CD Projekt	31	3	2	1
Coca-Cola HBC	97	5	3	5
Edge1 Solution	75	5	5	3
Ergo Hestia	94	3	2	6
Grupa Luxmed	53	4	2	6
GS1 Polska	19	2	3	3
Hegelmann Transporte	64	4	6	4
Holcim	69	5	4	3
Ikano Bank	36	3	4	3
mBank	64	3	3	2
Orange Polska	100	4	5	4
PKO BP	25	5	3	3
Polpharma	69	4	3	4
PSE	78	3	1	5
Santander Bank Polska	100	4	3	4
Unimot	81	3	5	5
Żywiec Zdrój	94	4	4	3

.

Governance assessment showed the most pronounced differences between tools. Chat GPT demonstrated limited discrimination capability (M = 2.90, SD = 0.97), while Claude 4.5 Sonnet showed moderate performance (M = 3.15, SD = 1.42). Gemini 2.5 Flash exhibited the greatest variability (M = 3.40, SD = 2.11), including extreme scores (0 for CD Projekt and 8 for Bogdanka). Correlation analysis between benchmark governance scores and AI performance ratings revealed consistently negligible and statistically insignificant relationships across all three tools: GPT (r = 0.140, p = 0.557), Claude (r = −0.017, p = 0.944), and Gemini (r = −0.024, p = 0.921). Notable outliers emerged in the analysis, particularly among companies with perfect governance scores. For instance, Żywiec Zdrój (governance score: 100) received dramatically different assessments, with Claude scoring eight while Gemini scored zero, exemplifying the challenges in establishing consistent automated detection protocols for governance-related visual greenwashing. The complete absence of correlation in governance dimension, combined with extreme scoring anomalies suggests fundamental limitations in current AI models’ ability to evaluate non-environmental ESG claims. The social governance scores are presented in

Table 5. Governance performance scores.

Company	Benchmark—Governance Score	GPT	Claude	Gemini
Amica	37	2	5	7
Bogdanka	90	3	3	8
Castorama	86	2	5	4
CD Projekt	57	3	2	0
Coca-Cola HBC	100	4	2	4
Edge1 Solution	80	4	4	4
Ergo Hestia	100	2	2	5
Grupa Luxmed	83	3	2	4
GS1 Polska	70	2	2	4
Hegelmann Transporte	83	5	4	4
Holcim	93	4	3	4
Ikano Bank	43	2	3	2
mBank	100	2	2	2
Orange Polska	93	3	3	2
PKO BP	60	4	2	2
Polpharma	77	3	4	3
PSE	87	2	1	4
Santander Bank Polska	100	3	2	2
Unimot	93	2	4	5
Żywiec Zdrój	100	3	8	0

.

4.2. Hypothesis Testing

H1. 

Aggregate Validation.

Hypothesis H1 predicted a moderate negative correlation (r in [−0.40, −0.60]) between AI greenwashing scores and external ESG performance, assuming that poor performers would engage in more deceptive visual practices. This hypothesis was rejected. Contrary to prediction, the analysis revealed a positive correlation between external performance and AI-detected greenwashing, which reached statistical significance (significant positive correlation) for Chat GPT (r = 0.50, p = 0.025). Gemini 2.5 showed a weak positive trend (r = 0.32, p = 0.166) and Claude 4.5 showed no correlation (r = −0.06, p = 0.802). The positive correlation indicates that companies with higher ESG performance benchmarks (e.g., Santander, Coca-Cola HBC) received higher greenwashing scores from the AI tools than low performers (e.g., Amica, CD Projekt). This greenwashing paradox suggests that AI models may be conflating visual sophistication (common in high-maturity reports) with manipulation, or conversely, that high-performing firms are indeed deploying more aggressive impression management strategies to maintain their legitimacy.

H2. 

Dimensional variation.

H2 predicted that the strength of negative correlations between AI greenwashing scores and ESG performance would vary systematically across dimensions, with Environmental showing the strongest negative correlation (r < −0.50), Social the weakest (r ∈ [−0.20, −0.35]), and Governance intermediate (r ∈ [−0.35, −0.45]). This hypothesis was not supported by the data. For the Environmental dimension, correlations ranged from r = −0.271 (Claude, p = 0.248) to r = 0.608 (GPT 5.1, p = 0.005), with GPT showing a significant positive rather than negative relationship. Social dimension correlations ranged from r = 0.14 (Claude, p = 0.556) to r = 0.444 (Gemini, p = 0.05), all positive rather than negative as hypothesized. Governance correlations were near-zero across all tools (GPT 5.1: r = 0.14, p = 0.557; Claude: r = −0.017, p = 0.944; Gemini: r = −0.024, p = 0.921). Fisher’s r-to-z transformation tests revealed no significant differences between dimensional correlations within any individual AI tool (all p > 0.10), indicating that the hypothesized systematic variation across ESG dimensions was not observed. These results suggest that the differential nature of ESG dimensions does not translate into predictable patterns in AI-detected visual greenwashing as theorized.

H3. 

Dimensional gap effect.

H3 predicted that companies with performance gaps exceeding 30 percentage points between ESG dimensions would demonstrate significantly higher AI-detected greenwashing in their weakest dimension compared to their strongest dimension (Wilcoxon signed-rank test, p < 0.05, effect size r > 0.50). Seven companies met the gap criterion: PKO BP (gap = 75 points), GS1 Polska (51), Ikano Bank (49), Unimot (40), CD Projekt (38), mBank (36), and Grupa Luxmed (35).

We tested H3 separately for each AI tool by extracting dimensional greenwashing scores for companies’ strongest- and weakest-performing ESG dimensions. Results were remarkably consistent across tools, all contradicting the hypothesis. For Chat GPT, mean greenwashing scores in strongest dimensions (M = 4.14, SD = 2.12) exceeded scores in weakest dimensions (M = 3.57, SD = 1.13), yielding a mean difference of −0.57 points opposite to predictions. Only 28.6% (2/7) of companies showed higher greenwashing in their weakest dimension. A Wilcoxon signed-rank test revealed no significant difference (W = 7.0, p = 0.781, one-tailed, r = 0.29).

Gemini produced nearly identical patterns: mean strongest = 4.29 (SD = 2.11), mean weakest = 3.71 (SD = 2.21), mean difference = −0.57 (W = 6.0, p = 0.859, r = 0.41). Only 14.3% (1/7) of companies exhibited the predicted pattern. Claude showed similar results: mean strongest = 3.86 (SD = 1.86), mean weakest = 3.43 (SD = 1.13), mean difference = −0.43 (W = 10.0, p = 0.773, r = 0.28), with 42.9% (3/7) showing the predicted direction. None of the three tools approached conventional statistical significance or the hypothesized effect size threshold.

Intertool consistency revealed that six of seven companies showed unanimous directional agreement across all three tools: four companies (CD Projekt, Grupa Luxmed, Ikano Bank, PKO BP) unanimously exhibited lower greenwashing in their weakest dimensions, contradicting H3; one company (Unimot) unanimously showed higher greenwashing in its weakest dimension, supporting H3; and two companies (GS1 Polska, mBank) produced mixed patterns. This high level of inter-tool agreement on the reverse pattern suggests that the absence of compensatory communication effects is not attributable to tool-specific limitations but rather reflects genuine characteristics of the reports or fundamental challenges in detecting dimensional strategic behavior through visual rhetoric analysis.

H4. 

Inter-tool variation.

H4 predicted that the three multimodal AI tools would demonstrate substantial but incomplete inter-rater agreement (Krippendorff’s α ∈ [0.65, 0.75]), reflecting complementary rather than redundant capabilities. This prediction was strongly rejected. Krippendorff’s alpha for overall greenwashing scores was α = 0.021, indicating near-zero inter-rater reliability and falling far outside the predicted range. This value is well below the conventional threshold for tentative conclusions (α ≥ 0.67) and suggests that the three tools are essentially measuring different constructs rather than providing different perspectives on the same phenomenon.

Dimensional analyses revealed similarly low reliability across all ESG dimensions: Environmental α = 0.104, Social α = −0.018, Governance α = −0.039. Negative alpha values for Social and Governance dimensions indicate that agreement was worse than would be expected by chance, suggesting systematic differences in how tools approach these dimensions.

Pairwise Pearson correlations confirmed the absence of inter-tool agreement. GPT 5.1 and Gemini showed a weak negative correlation (r = −0.140, p = 0.557), GPT 5.1 and Claude showed a weak positive correlation (r = 0.111, p = 0.642), and Gemini and Claude showed a weak positive correlation (r = 0.078, p = 0.744). None approached statistical significance, and the mean pairwise correlation was r = 0.016, effectively zero. Similar patterns emerged across dimensional scores, with correlations ranging from r = −0.141 to r = 0.224, all non-significant.

Despite this profound disagreement in scoring, the three tools did not differ significantly in their mean score levels (ANOVA: F(2,57) = 0.509, p = 0.604). Mean overall scores were GPT 5.1 = 4.80 (SD = 1.28), Gemini = 5.20 (SD = 1.82), and Claude = 4.75 (SD = 1.48). This suggests that the disagreement reflected not systematic bias in stringency but rather fundamental differences in which companies and which aspects of reports triggered higher greenwashing assessments.

Case-level analysis illustrated the extent of inter-tool divergence. Five companies showed inter-tool score ranges exceeding 3 points despite the 1–10 scale: Amica (GPT 5.1 = 2, Gemini = 7, Claude = 6, SD = 2.65), Bogdanka (4, 8, 4, SD = 2.31), Grupa Luxmed (6, 7, 3, SD = 2.08), Ergo Hestia (4, 7, 3, SD = 2.08), and CD Projekt (4, 1, 3, SD = 1.53). Only two companies showed high agreement with score ranges ≤1: mBank and Orange Polska. The mean standard deviation across all 20 companies was 1.42, indicating that substantial disagreement was the norm rather than the exception.

5. Discussion

Delmas and Burbano [15] defined greenwashing as the intersection of poor performance and positive communication, implying a negative correlation between ESG achievement and deceptive disclosure. Our findings contradict this foundational premise: ChatGPT showed a significant positive correlation (r = 0.50, p = 0.025), with high performers like Santander Bank Polska receiving higher greenwashing scores than low performers like Amica. This greenwashing paradox may reflect what Lyon and Montgomery [16] termed symbolic corporate environmentalism—sophisticated communication that triggers AI detection regardless of authenticity. Alternatively, following Seele and Gatti’s [17] reconceptualization of greenwashing as accusation-based legitimacy judgment, AI tools may function as accusation-generating mechanisms responding to communication intensity rather than performance-disclosure gaps. Most likely, however, AI systems conflate the visual sophistication characteristic of mature ESG reports—what Cho et al. [24] found influences stakeholder trust—with manipulative intent, penalizing high performers for professional presentation.

Hrasky’s [21] visual rhetoric compensation hypothesis predicted that companies with weak performance would deploy enhanced visual rhetoric to distract from unfavorable data. Our H3 test contradicted this: companies with large dimensional performance gaps showed lower, not higher, greenwashing scores in their weakest dimensions. The impression management mechanisms identified by Merkl-Davies and Brennan [20]—selective emphasis, strategic omission, rhetorical framing—may operate too subtly for AI pattern recognition. The distinction between claim and executional greenwashing proposed by Parguel et al. [19] and de Freitas Netto et al. [18] suggests that AI may detect executional elements (nature imagery, green aesthetics) without connecting them to dimensional performance gaps.

Dimensional analysis revealed no systematic patterns matching theoretical expectations. Environmental correlations ranged from r = 0.608 (ChatGPT) to r = −0.271 (Claude), contradicting assumptions about metric objectivity. Boiral’s [23] concept of sustainability reports as simulacra may explain why even quantifiable environmental disclosures resist consistent AI assessment. The Social dimension showed inverse relationships: companies with perfect scores received higher greenwashing flags than poor performers, suggesting that AI interprets legitimate achievements as promotional rhetoric—precisely the cultural contextuality that Parguel et al. [19] warned complicates interpretation. Governance correlations near zero likely reflect that visual rhetoric theory [27,28] emphasizes emotional and aesthetic communication, while governance disclosures are more technical and visually understated.

The near-zero inter-rater reliability (Krippendorff’s α = 0.021) represents the most consequential finding. Rather than the predicted complementary capabilities (α ∈ [0.65, 0.75]), the three AI tools essentially measure different constructs. Seele and Gatti’s [17] emphasis on greenwashing as fundamentally accusation-based gains support: different AI architectures may embody different implicit accusation thresholds based on training data. Cho et al.’s [24] finding that visual sophistication influences trust suggests each tool weights sophistication signals differently, producing systematic disagreement about which companies exhibit concerning patterns.

Legitimacy theory [25] may require refinement for automated detection contexts. Cho and Patten’s [26] finding that companies facing legitimacy threats respond through enhanced disclosure sophistication explains the positive correlations: AI detects legitimacy management behaviors without assessing whether they constitute deception or confident achievement communication. For CSRD implementation in Poland [7,8], where companies are transitioning to mandatory ESRS-compliant reporting [9,10], these findings counsel caution: AI tools may penalize the disclosure improvements that regulators seek to encourage. Until domain-specific training datasets and improved explainability emerge, AI should function as preliminary screening requiring expert validation rather than authoritative assessment.

6. Conclusions

This study evaluated three multimodal AI systems for detecting visual greenwashing in Polish ESG reports against ESRS-aligned benchmarks. All four hypotheses were rejected. Contrary to Delmas and Burbano’s [15] performance-disclosure gap principle and legitimacy theory predictions [25,26], AI tools exhibited positive correlations between performance and detected greenwashing. Hrasky’s [21] visual rhetoric compensation hypothesis was contradicted: companies showed lower greenwashing scores in weak dimensions. Inter-tool reliability proved observably low (α = 0.021), indicating the tools measure fundamentally different constructs.

These findings demonstrate boundary conditions of established theoretical frameworks when applied to automated detection. The Delmas and Burbano [15] definition, while conceptually sound, does not translate into AI-detectable patterns. Hrasky’s [21] compensation hypothesis, validated through expert coding, does not manifest in ways multimodal AI can identify. Seele and Gatti’s [17] emphasis on greenwashing as accusation-based legitimacy judgment gains support from inter-tool divergence reflecting different implicit accusation thresholds.

For practitioners, the near-zero inter-tool reliability undermines confidence in AI-assisted assessment. Current consumer-grade multimodal AI is not ready for professional ESG auditing. Future research should develop domain-specific ESRS-aligned training datasets, investigate hybrid human-AI evaluation frameworks, and conduct cross-national comparisons to establish whether observed patterns reflect universal challenges or context-specific factors. Until reliable domain-specific AI systems emerge, greenwashing detection must remain fundamentally a human-expert endeavor.