Whose Values Are Generated? Measuring Visual Value Leakage in Multilingual Text-to-Image Generation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Whose Values Are Generated? Measuring Visual Value Leakage in Multilingual Text-to-Image Generation".
Jane: Here's the gist of the paper. Text-to-image (T2I) systems serve users across languages and cultural contexts, but evaluations of cultural representation often focus on recognizable elements such as landmarks and clothing while overlooking the cultural values that generated images The authors introduce VisionValueBench, a multilingual benchmark operationalizing six of Hofstede’s cultural value dimensions to investigate how T2I systems frame neutral social situations through a particular cultural orientation.
Tom: They find that value-neutral prompts yield images favoring particular value orientations, a phenomenon called visual value leakage.
Jane: The Core Finding on Value Leakage The central discovery is that even when prompts are value-neutral, the generated images consistently favor specific cultural orientations across all six dimensions.
Tom: Specifically, images tend to express low power distance, individualism, care and quality of life, long-term orientation, high uncertainty avoidance, and indulgence regardless of the T2I system used.
Title and authors: Tom: So, this paper introduces this VisionValueBench benchmark. It operationalizes those six Hofstede dimensions to systematically investigate how text-to-image systems frame neutral situations through different cultural orientations.
Jane: The key thing they’re showing is that these systems are not truly neutral; they have a built-in cultural bias when generating images, even when we try to keep the prompt open and value-neutral.
Lu: They set up this benchmark by converting contextualized questions into prompts describing social scenarios where the value orientation isn't specified, forcing the generator to express its own values.
Meng: The methodology involves using two ways to check for these orientations: Scenario-Anticipated Evaluation, or SAE, and Image-Adaptive Evaluation, or IAE.
Lalam: IAE is what they found was more reliable because it involves adaptively searching the generated image for visual evidence supporting each orientation; they saw that this achieved eighty-two point ninety two percent agreement with human reviewers on a subset of their data.
The paper's summary: Tom: The paper’s summary is that they looked at prompts varying in language and geographic cues across nine hundred sixty social scenarios and six text-to-image configurations to see what happened.
Jane: They found that while prompt language can have some influence, geographic cues are what really shift some dimensions toward associated cultural values, but the general direction of the image orientation stays pretty consistent.
Lu: For example, switching from English to Chinese prompts preserves orientations across all systems and dimensions entirely, which is a strong indicator of this widespread homogeneity in how these models operate.
Meng: But when you switch from US to China cues, they see larger changes primarily in the strength of the orientation rather than the direction itself.
Lalam: The paper suggests that multilingual text-to-image generation can expose users across different cultural contexts to these broadly shared value orientations without us necessarily recognizing them if this homogeneity isn't disclosed.
The paper's improvements: Tom: Looking at the improvements they suggest, they are focused on how we can get a better picture of what these systems are actually doing. They propose implementing an Image-Adaptive Evaluation strategy to identify visual evidence in each generated image without relying only on scenario anticipation.
Jane: That makes sense because IAE showed higher agreement with human reviewers under both SAE and IAE strategies, so they want us to use that approach when we can afford the computational cost.
Lu: They also suggest a multi-agent system for verification to check if the cultural evidence is actually grounded in the image and that its value interpretation fits the scenario, which should mitigate the bias of relying on just one model’s judgment.
Meng: From a production standpoint, I’d be interested in how they propose designing prompt engineering pipelines to explicitly control how different components like prompt refiners and diffusion models contribute to this visual value leakage.
Lalam: They also suggest integrating a scoring mechanism that evaluates each image on all six dimensions instead of just the scenario's source dimension, which allows for a more holistic assessment of system behavior across the board.
Conclusion: Tom: So wrapping up the paper "Whose Values Are Generated? Measuring Visual Value Leakage in Multilingual Text-to-Image Generation," they conclude there’s substantial homogeneity in the cultural values conveyed by text-to-image systems worldwide.
Jane: The main implication is that we need to start evaluating cultural responsiveness not just by what recognizable content the image shows, but also by the underlying values it expresses.
Lu: It suggests that users across different linguistic and cultural contexts might encounter particular value orientations without recognizing them if this homogeneity isn't disclosed in the systems themselves.
Meng: For deployment, they recommend using IAE when the budget allows because it has higher agreement with human-majority references, which is a practical piece of advice for anyone trying to build something that actually works on real traffic.
Lalam: They also flag a limitation: dependence on scenario distribution is noted because leaving value orientations unspecified doesn't guarantee that opposing orientations are equally easy to depict or recognize.
Tom: That’s the main point, this paper shows us that the cultural values leaking into these images are pretty consistent, which means we need better ways to measure and understand what's happening when we use these models.
University of Wisconsin–Madison · Microsoft Research Asia
cs.CV
Submitted: 2026-11-03
Updated: 2026-11-05
Comments: 36 pages
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: Text-to-image (T2I) systems serve users across languages and cultural contexts, but evaluations of cultural representation often focus on recognizable elements such as landmarks and clothing while…
Key concepts
- Hofstede's Cultural Dimensions
- These are six established frameworks used to measure different cultural values, such as power distance and individualism. The researchers operationalized these abstract concepts into concrete visual criteria to test how AI images reflect these underlying cultural orientations in real-world scenarios.
- Visual Value Leakage
- This phenomenon occurs when text prompts that are intentionally neutral still cause the generated images to lean toward specific, pre-determined cultural values. The research found that regardless of the prompt's neutrality, images consistently favored certain orientations across all six dimensions.
- Image-Adaptive Evaluation (IAE)
- This is a method used to score AI-generated images by adaptively searching the image for visual evidence supporting each cultural orientation. IAE was found to be more reliable than other methods because it achieved higher agreement with human reviewers when assessing value expression.
- Prompt and Cue Sensitivity
- The research analyzed how changes in prompt language (e.g., English to Chinese) or geographic cues affect the resulting image values. Findings indicated that language has limited influence, but geographic cues can shift orientation strengths while largely preserving the prevailing cultural orientations.
Terminology
Summary
Text-to-image (T2I) systems serve users across languages and cultural contexts, but evaluations of cultural representation often focus on recognizable elements such as landmarks and clothing while overlooking the cultural values that generated images The authors introduce VisionValueBench, a multilingual benchmark operationalizing six of Hofstede’s cultural value dimensions to investigate how T2I systems frame neutral social situations through a particular cultural orientation. p. 0 edu
ABSTRACT
Text-to-image (T2I) systems serve users across languages and cultural contexts. However, evaluations of cultural representation focus mainly on recognizable elements such as landmarks, clothing, and activities, overlooking the cultural values that generated images convey. , individualism versus collectivism) to investigate how T2I
systems frame neutral social situations through a particular cultural orientation. We find that value-neutral prompts, which admit diverse cultural interpretations,
nonetheless yield images that favor particular value orientations, a phenomenon
we call visual value leakage.Where the paper says it
Where the paper says it
The Core Finding on Value Leakage
The central discovery is that even when prompts are value-neutral, the generated images consistently favor specific cultural orientations across all six dimensions. Specifically, images tend to express low power distance, individualism, care and quality of life, long-term orientation, high uncertainty avoidance, and indulgence regardless of the T2I system used. p. 2 92% agreement with the human reviewers. Across six T2I systems including both commercial and open-weight models, we observe a striking degree of cross-system value consistency: images generated by all systems tend to express
low power distance, individualism, care and quality of life, long-term orientation, high uncertainty
avoidance, and indulgence, shown in Figure 1 (bottom). Prompt language has limited influence on these orientations, whereas geographic
cues shift some dimensions toward the associated cultural values while largely
preserving the prevailing orientations. These findings reveal substantial homogeneity in the cultural values conveyed by T2I systems worldwide.Where the paper says it
Where the paper says it
Benchmark Construction and Methodology
VisionValueBench was constructed grounded in Hofstede’s cultural dimensions, covering six pairs of contrasting value orientations in four languages. The benchmark involves converting contextualized questions into prompts describing value-neutral social scenarios, filtering out those that prescribe a value orientation and leaving the orientation open. p. 3 Value-relevant context
Translate
Figure 2: VisionValueBench construction pipeline. Building on CDEval, a text-based benchmark
grounded in Hofstede’s cultural dimensions, we (1) convert contextualized questions into prompts
describing value-neutral social scenarios, (2) filter out scenarios that constrain or preclude certain
cultural orientations, and (3) translate the resulting English prompts into other languages. This benchmark allows us to not only
compare value orientations across generators, but also investigate how value orientation shifts with
prompt language and geographic cues that specify the country in which the scene takes place. To reliably infer orientations of a generated image, we develop two complementary evaluation strategies: Scenario-Anticipated Evaluation (SAE) and Image-Adaptive Evaluation (IAE) (Figure 3). To reliably infer orientations of a generated image, we develop two complementary evaluation strategies: Scenario-Anticipated Evaluation (SAE) and Image-Adaptive Evaluation (IAE) (Figure 3). SAE anticipates how a value might be expressed in the given scenario, while IAE adaptively searches
the generated image for visual evidence supporting each orientation.Where the paper says it
Where the paper says it
Where the paper says it
Prompt and Cue Sensitivity Analysis
The study investigates how prompt language and geographic cues influence these orientations. Prompt language has limited influence on these orientations, whereas geographic cues shift some dimensions toward associated cultural values while largely preserving the prevailing orientations. p. 0 Generated
images consistently favor the same value orientations across all six dimensions,
for example, equality over hierarchy (low power distance), individual autonomy
over group obligation (individualism), and enjoyment over restraint (indulgence). Prompt language has limited influence on these orientations, whereas geographic
cues shift some dimensions toward the associated cultural values while largely
preserving the prevailing orientations. Purely switching from English to Chinese preserves orientations across all systems and dimensions. Switching from US to China cues
produces larger changes, primarily in orientation strength, while largely preserving the prevailing
directions.Where the paper says it
Where the paper says it
Evaluation and Robustness
The evaluation framework involves multiple models independently verifying visual grounding and value interpretation using SAE and IAE strategies. Human validation showed that IAE achieves higher agreement with human reviewers, achieving 82. p. 6 (a) Hypothetical distributions of 100 supported images: each of the last three rows yields ∆S = +0. 1 relative to the first,
which corresponds to a net shift of ten images per 100 supported images; (b) IAE achieves higher
agreement with both human-majority references, SAE-H and IAE-H. On a human-annotated
subset, our IAE orientation decisions achieve 82. 92% agreement with the human reviewers. Across 960 scenarios and
six T2I systems, generated images exhibit broadly homogeneous value orientations. Our experiments show that prompt language has limited influence on these orientations, while geographic cues
primarily change their strength but still largely preserve the prevailing directions. Extending the comparison to Hindi and Thai prompts and cues on the three commercial
systems shows the same pattern: strength varies while direction largely persists (Appendix D. These findings suggest that multilingual T2I generation can expose users across cultural contexts
to broadly shared value orientations, with limited control over which orientations prevail through
prompt language and geographic cues.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Implications
The paper concludes that VisionValueBench reveals substantial homogeneity in the cultural values conveyed by T2I systems worldwide. These findings motivate evaluating cultural responsiveness through expressed values alongside recognizable cultural content. p. 9 When this homogeneity is not disclosed, users may encounter particular value
orientations without recognizing them. These findings motivate evaluating cultural responsiveness
through expressed values alongside recognizable cultural content and developing ways to make the
cultural orientations depicted in generated images visible to users. When this homogeneity is not disclosed, users may encounter particular value
orientations without recognizing them. These findings motivate evaluating cultural responsiveness
through expressed values alongside recognizable cultural content and developing ways to make the
cultural orientations depicted in generated images visible to users. Human review shows that IAE
recovers UAI evidence that SAE misses, supporting its advantage in identifying image-specific evidence on this dimension. For practitioners assessing the value orientations of their own T2I systems,
we recommend IAE when the evaluation budget permits. This homogeneity highlights the importance of evaluating
expressed value orientations alongside recognizable cultural content. ACKNOWLEDGMENTS
This work is supported by Microsoft Research Fellowship Program (The Challenge of Grounding
AI in Global Majority Languages, Cultures, and Contexts).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 36 ---
Preprint
G LIMITATIONS AND FUTURE DIRECTIONS
From cultural constructs to visual evidence. Our evaluation operationalizes cultural values through Hofstede’s six dimensions and their associated visual criteria. p. 0 However, evaluations of cultural representation focus mainly on recognizable elements such as landmarks, clothing, and activities, overlooking the cultural values that generated images convey. To bridge this gap, we introduce VisionValueBench, a multilingual benchmark that operationalizes six of Hofstede’s cultural
value dimensions (e. For
example, all six systems favor low power distance and indulgence under both English and Chinese
prompts (Figure 6). For culturally diverse users, switching T2I systems, changing prompt language,
or specifying a country may therefore provide limited control over the value orientations expressed
in generated images. We find that value-neutral prompts, which admit diverse cultural interpretations,
nonetheless yield images that favor particular value orientations, a phenomenon
we call visual value leakage. We evaluate prompts varying in language and geographic cues across 960 social scenarios and six T2I configurations. When this homogeneity is not disclosed, users may encounter particular value
orientations without recognizing them. These findings motivate evaluating cultural responsiveness
through expressed values alongside recognizable cultural content and developing ways to make the
cultural orientations depicted in generated images visible to users. This is why we hold the scenario fixed
in VisionValueBench and compare value orientations across languages and geographic cues, rather
than varying the requested culture itself. Cultural values in text-to-image generation. In Figure 2 (Scenario), the teacher and students
form the focal subject, and their classroom communication provides the value-relevant context. How
they communicate remains unspecified, allowing different power-distance orientations to emerge.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 34 ---
Preprint
F.3 SCENARIO FILTERING PROMPT
The screening prompt is one user message with the task instructions, the attribution rules and decision cards for all six dimensions, an output contract and the scenario. The model decides whether the sentence already commits the image to one orientation, so that any image faithful to the sentence has to express it. p. 2 CulText2I
uses short concepts rather than situations that put a value orientations at stake, and its judgments are
not tied to verifiable visual evidence. We instead pose each value questions as a scenario with its
orientation left open, tie every judgment to identified visual evidence, and record whether an image
supports one orientation, both, or neither. CulText2I
uses short concepts rather than situations that put a value orientations at stake, and its judgments are
not tied to verifiable visual evidence. We instead pose each value questions as a scenario with its
orientation left open, tie every judgment to identified visual evidence, and record whether an image
supports one orientation, both, or neither.Where the paper says it
Where the paper says it
--- Page 33 ---
Preprint
F.4 EVALUATION PROMPTS
All four prompts use JSON schemas and run independently on GPT-5.6 Luna, Gemini 3. p. 5 ≥ 2/3
Frozen
Per-image proposals
Dynamic
Per-scenario proposals
Does any evidence item have
both a valid observation and
a justified interpretation? 6-luna Gemini-3. ≥ 2/3
Frozen
Per-image proposals
Dynamic
Per-scenario proposals
Does any evidence item have
both a valid observation and
a justified interpretation? 6-luna Gemini-3. 6-luna Gemini-3. 7-flash Muse spark 1. 6-luna Gemini-3. 7-flash Muse spark 1. Each scenario specifies a value-relevant situation while
leaving its orientation open. This section describes how we derive scenarios from existing text-based
value questions, filter out scenarios that prescribe a value orientation, validate that retained scenarios
leave value orientations open, and vary prompt language and geographic cues. IAE (lower row) proposes evidence from each
image and allows verifiers to add missed evidence. Within each strategy, candidates from three
models are pooled without proposer identities.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 32 ---
Table 20: Condensed decision cards for automatic annotation
Dim. Defining issue Orientation criteria Insufficient alone PDI Authority or control over decisions, resources or participation, and how unequal power is treated High: higher standing is treated as entitling control, deference, or limits on others’ voice. Low: their standing to contribute, question or seek accountability is acknowledged or asserted, even if a leader keeps final authority. p. 4 In the classroom example in Figure 3, the relevant content is how students’
opportunities to contribute are treated within the teacher–student relationship. A teacher listening
attentively to a speaking student may support low power distance because the student’s contribution is acknowledged. In the classroom example in Figure 3, the relevant content is how students’
opportunities to contribute are treated within the teacher–student relationship. A teacher listening
attentively to a speaking student may support low power distance because the student’s contribution is acknowledged.Where the paper says it
Where the paper says it
--- Page 31 ---
Table 16: Pooled orientation scores under the six shared conditions
Method System PDI IDV MAS UAI LTO IVR.
--- Page 30 ---
Figure 20: Paired scores when every image is scored on all six dimensions
Same contrasts and layout as Figure 7.
--- Page 29 ---
Table 17: Paired score changes
(a) Language EN → ZH; no cue.
--- Page 33 ---
Figure 20: Paired scores when every image is scored on all six dimensions Same contrasts and layout as Figure 7.
--- Page 34 ---
Table 19: Prompt rendering templates
English Generate an image:.
--- Page 32 ---
F.2 CONSTRUCTION, TRANSLATION, and Rendering Prompts
Both construction calls use DeepSeek-V4-Flash at temperature 0; their rationales are never sent to image generators. Translation uses Gemini 3. p. 2 , 2026), and multimodal judges for visualization quality (Xie et al. CulText2I
uses short concepts rather than situations that put a value orientations at stake, and its judgments are
not tied to verifiable visual evidence. Drift would be indistinguishable from a
genuine language effect. Our translation rubric asks for natural target-language wording that retains
the original meaning, detail, and intentional omissions, while excluding cultural adaptation, which
could supply the orientation the scenario is meant to leave open.Where the paper says it
Where the paper says it
--- Page 32 ---
F.4 EVALUATION PROMPTS
All four prompts use JSON schemas and run independently on GPT-5.6 Luna, Gemini 3.7 Flash and Muse Spark 1.3 with medium reasoning effort.
--- Page 32 ---
F.4 EVALUATION PROMPTS
Prompt 1: SAE proposal summary: scenario only Identify the focal situation, what the scenario establishes and what it leaves open.
--- Page 34 ---
F.4 EVALUATION PROMPTS
Prompt 2: IAE proposal summary: image + scenario Rank relevance to the depicted focal situation, then propose candidates grounded in the image.
--- Page 35 ---
F.5 DECISION CARDS FOR AUTOMATIC ANNOTATION
The scope is the scenario supplies context such as alternatives, the object of a response and an arrangement’s purpose, while the image decides any open outcome and prevails if the two conflict. Evidence connects the depiction, its role in the focal situation and the card’s criterion; readable text counts in context, but nationality, ethnicity, gender and cultural markers alone never do. p. 4 Each candidate connects an observable depiction to a value orientation through its role in the focal situation: the central interaction, choice, or arrangement being
depicted. Recognizing visual content alone does not establish its value orientation, as VLMs may attend to irrelevant details or misinterpret a situation despite recognizing its objects (Bae et al.Where the paper says it
--- Page 34 ---
F.6 DECISION CARDS FOR HUMAN REVIEW
Reviewers use plain-language versions of the same six cards on the review page, each with a worked two-step example. The PDI example shows how to interpret Low PDI. p. 8 For example, a festival
screen asks attendees to trust a young, inexperienced management team while an attendee’s cup
reads “We’ll see,” so the same image supports both orientations (Appendix D. −1
0
1
−1
0
1
−1
0
1
Orientation score S PDI IDV MAS UAI LTO IVR
(a) Language
EN → ZH
(no cue)
(b) Geographic
cue · EN
US cue →
China cue
(c) Geographic
cue · ZH
US cue →
China cue
MAI-2.Where the paper says it
--- Page 34 ---
Table 20: Condensed decision cards for automatic annotation
Dim. Defining issue Orientation criteria Insufficient alone PDI Authority or control over decisions, resources or participation, and how unequal power is treated High: higher standing is treated as entitling control, deference, or limits on others’ voice. Low: their standing to contribute, question or seek accountability is acknowledged or asserted, even if a leader keeps final authority.
--- Page 34 ---
Table 20: Condensed decision cards for automatic annotation
Dim. Defining issue Orientation criteria Insufficient alone PDI Authority or control over decisions, resources or participation, and how unequal power is treated High: higher standing is treated as entitling control, deference, or limits on others’ voice. Low: their standing to contribute, question or seek accountability is acknowledged or asserted even if a leader keeps final authority. p. 4 In the classroom example in Figure 3, the relevant content is how students’
opportunities to contribute are treated within the teacher–student relationship. A teacher listening
attentively to a speaking student may support low power distance because the student’s contribution is acknowledged.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Implement an Image-Adaptive Evaluation (IAE) strategy to identify visual evidence in each generated image without relying solely on scenario anticipation, as IAE achieves higher agreement with human majority references under both SAE and IAE.
-
Develop a multi-agent system for verification that checks if cultural evidence is grounded in the image and that its value interpretation fits the scenario, mitigating the bias of relying on a single model’s judgment.
-
Design prompt engineering pipelines to explicitly control value orientation expression by characterizing how different components—such as prompt refiners and diffusion models—contribute to visual value leakage.
-
Integrate a scoring mechanism that evaluates each image on all six dimensions rather than only the scenario's source dimension, allowing for a more holistic assessment of system behavior.
-
Create a method to disentangle the separate contributions of prompt processing, the image-generation model, and both from shaping value orientations in generated images.
Abstract
Text-to-image (T2I) systems serve users across languages and cultural contexts. However, evaluations of cultural representation focus mainly on recognizable elements such as landmarks, clothing, and activities, overlooking the cultural values that generated images convey. To bridge this gap, we introduce VisionValueBench, a multilingual benchmark that operationalizes six of Hofstede's cultural value dimensions (e.g., individualism versus collectivism) to investigate how T2I systems frame neutral social situations through a particular cultural orientation. We find that value-neutral prompts, which admit diverse cultural interpretations, nonetheless yield images that favor particular value orientations, a phenomenon we call visual value leakage. We evaluate prompts varying in language and geographic cues across 960 social scenarios and six T2I configurations. Generated images consistently favor the same value orientations across all six dimensions, for example, equality over hierarchy (low power distance), individual autonomy over group obligation (individualism), and enjoyment over restraint (indulgence). Prompt language has limited influence on these orientations, whereas geographic cues shift some dimensions toward the associated cultural values while largely preserving the prevailing orientations. These findings reveal substantial homogeneity in the cultural values conveyed by T2I systems worldwide.
Sources
- Text-to-Image Models and Their Representation of People from Different Nationalities Engaging in Activities
- Understanding the Effects of Distractors on Reasoning Vision-Language Models
- Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values
- Beyond Aesthetics: Cultural Competence in Text-to-Image Models
- On the Cultural Gap in Text-to-Image Generation
- Self-improving data synthesis can scale embedding-model adaptation across domains and tasks
- Multilingual Conceptual Coverage in Text-to-Image Models
- Bidirectional Human-AI Alignment in Education for Trustworthy Learning Environments
- Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems
- CDEval: A Benchmark for Measuring the Cultural Dimensions of Large Language Models
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs