acceptodds
Under review as a conference paper at ICLR 2027

Multimodal Bias in AI-Mediated Markets: Benchmarking, Preference Optimization, and Human Valuation

Abstract

Generative models mediate economic decisions by selecting the information people see. Demographic associations in that selection process can change consumer valuations even when the underlying product and reviews remain fixed. We investigate this pathway through multimodal benchmarking, paired review-summary generation, Direct Preference Optimization (DPO), and an incentivized human marketplace experiment. In a multimodal benchmark adapted from a standard human implicit association test (IAT), the unadapted instruction-tuned Gemma 3 12B checkpoint (baseline Gemma) exhibits implicit demographic bias in negative-descriptor selection. Across 839 paired review sets, Black-cue review summaries score 12.6% lower in sentiment relative to the White-cue mean under the study’s generation procedure. DPO fine-tuning with low-rank adaptation uses 1,000 preference pairs. The reported sentiment gap narrows from −1.061 to −0.007 score points. An ethics-approved, preregistered, incentive-compatible marketplace study includes 976 eligible completions and 15,616 product evaluations. Under randomized assignment of review summaries and seller photos, review summaries from baseline Gemma yield an estimated Black-cue versus White-cue willingness-to-pay gap of −18.7% in the study’s reference condition, compared with −0.6% for summaries from DPO-tuned Gemma. The estimated willingness-to-pay gap therefore narrows by 18.1 percentage points (95% CI: 4.4, 31.8; p = .0097). The findings connect measurable model behavior to human–AI interaction: generated market information changes incentivized human valuation, and DPO-tuned review summaries yield a smaller downstream valuation gap.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
vision-language models, implicit bias, direct preference optimization, human-AI interaction, review summarization, willingness to pay

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3pUyOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2(+1)
Soundness
1
Read review
Paper summary

The paper links demographic cues in Gemma-generated summaries to human valuation. Across 839 review pairs, Black-cue sentiment is 1.061 points lower. DPO uses 800 training and 200 development pairs. Among 775 focal survey completions, the reference-cell WTP gap changes across baseline and DPO summary conditions.

Critical strengths
  1. The demographic-cue–text–valuation pathway is a meaningful contribution.

  2. Paired generation holds products and reviews fixed; within-condition randomization separates summary effects from displayed-photo effects.

  3. Broader face-panel checks and limitations of the generator comparison are disclosed.

Critical weaknesses
  1. Gemma's 32.9-point gap on 12 faces shrinks to 1.3 points on the 400-image panel (q = .623). This limits generalization of the focal diagnostic, not the separate summary and valuation findings.

  2. The approximately-zero diagnostic gap lacks precise uncertainty. Baseline group means are reported, but post-DPO absolute scores, source faithfulness, and capability retention are needed to distinguish reduced disparity from degraded summaries. Teacher identity and training hyperparameters also require clarification.

  3. As Appendix G acknowledges, generator conditions do not isolate weight changes from cohort and generation-run differences. The additional, non-preregistered interaction supports attenuation for these fixed summary sets, not a general causal effect of DPO.

  4. Only 16 products are used. Product selection and marginal effects across disclosure conditions should be reported; broader product-level claims require stimulus-level uncertainty.

Questions for authors
  1. What are the post-DPO gap intervals, absolute scores, and faithfulness results?

  2. How were generator cohorts and the 16 products selected?

  3. Do marginal effects and product-level sensitivity analyses support the reference-cell result?

1 comment
AuthorsOpenReview ↗

What the face benchmarks and summary comparisons show, and how buyer valuation gaps change with tuned summaries

Thank you for your time and detailed comments. We will upload an updated pdf soon.

1. The 12-face and UTKFace results. The focal diagnostic uses standardized, neutral-expression faces from Project Implicit’s Race-IAT materials, with gender balanced across groups and the same prompt repeated 50 times per face. Gemma’s Black-White gap was 32.9 percentage points on those 12 faces (95% CI [15.7, 48.4]). As you pointed out, on UTKFace, the gap was 1.3 points in the same direction. We looked further into this and it is because UTKFace has additional variance in age, expression, pose, and lighting, and its age and gender labels are estimated (Zhang et al., 2017). We will describe the UTKFace finding more narrowly in the revision. FairFace served a separate purpose. We matched the FairFace images before constructing the fine-tuning pairs to reduce noise in age, gender, background, and expression.

2. The near-zero gap result. In the 12-face run on tuned Gemma, every recorded descriptor was negative for both groups. The Black–White gap was zero and the between-group test was undefined. In a separate 12-face follow-up, ORPO fine-tuning yielded a gap of about -1 percentage point. Our main contribution comes from the buyer experiment. Randomly assigning summaries showed that demographic differences in generated text can change what buyers will pay for the same product. The smaller valuation gap for the tuned summary set shows how changing the text buyers receive can reduce that disparity in the tested setting. We used rank-8 LoRA. Two independent undergraduate research assistants manually reviewed separate batches of 500 preference pairs each.

3. Review-summary quality and faithfulness. For the texts buyers saw, mean sentiment on the 1-10 scale was as follows.

ConditionBlack-cue meanWhite-cue meanGap (95% paired CI)
Untuned5.199.19−4.00 [−4.85, −3.15]
DPO summary8.639.06−0.44 [−0.71, −0.16]

In a blind assessment of 16 summary pairs, two independent undergraduate raters unanimously agreed that every Black-cue version was more negative and verified that both versions drew evidence from the original reviews, just picked different words.

4. Survey groups and buyer valuation. English-speaking U.S. adults recruited on Prolific evaluated the same 16 products using either the untuned summary set (532 completions) or the DPO-tuned-model summary set (243 completions). Within each survey, we randomized the summary version and the displayed seller photo independently. With a human source label, no representativeness statement, and a White seller photo displayed, the Black-minus-White willingness-to-pay effect was −USD 0.688 for untuned summaries and −USD 0.022 for the DPO-tuned-model summary set. The gap narrowed by USD 0.666 (95% CI [0.162, 1.171], p<.01p < .01). This additional analysis compares the two fixed summary sets across survey groups. Within-survey randomization establishes that the generated summary can change what buyers will pay for the same product. The valuation gap was smaller for the DPO summary set in a marketplace task with real payoffs.

5. Product selection and marginal effects. Before the buyer study, three independent raters selected everyday items where possible and prioritized review summary pairs with visibly different summary wording. A significant average sentiment difference does not guarantee a noticeable difference for each product. The budget limited us to 16 products and the USD 6-15 product price range for the purchase lottery. The products were glass jars, cord organizers, a kitchen towel set, reusable hooks, push hooks, labels with a marker, magnets, measuring cups and spoons, a meat thermometer, wall décor, a microwave cover and silicone mat, motion-sensor lights, a table runner, a second kitchen towel set, trash bags, and an electric wine opener.

Averaging over the same disclosure and displayed-photo conditions in both groups, the Black-minus-White willingness-to-pay effect was −USD 1.247 for untuned summaries and −USD 0.056 for the DPO summary set. That is a USD 1.191 narrowing of the gap (95% CI [0.985, 1.397]). The gap also narrowed in each of the six disclosure conditions. We refit the regression after dropping each possible pair of products (120 combinations) for cross-validation. Comparing the untuned summary results to DPO summaries shifted the Black-minus-White willingness-to-pay gap by +USD 0.957 to +USD 1.331 (all p<.01p < .01).

Zhang, Zhifei, Yang Song, and Hairong Qi. 2017. “Age Progression/Regression by Conditional Adversarial Autoencoder.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.

Reviewer_Q8ZrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

My understanding of this paper is as follows: the authors argue that in the marketplace setting, specifically with online product reviews, AI models are being used to summarize reviews. These reviews are biased as measured by their Claude rater which rates the credibility and sentiment. The authors argue that the bias is due to the demographic information of the user which was included as a visual signal along with the review. They then show that if they intervene by fine tuning the AI model that does that summarization, the bias is mitigated. I think it is an interesting domain with practical applications, but I’m confused by the experimental setup.

Critical strengths

The use case is clear and motivates studying the specific instance based type of bias it could induce. However, I'm not sure that the experimental setup actually reflects the real-world use case.

Critical weaknesses

My understanding of this paper is as follows: the authors argue that in the marketplace setting, specifically with online product reviews, AI models are being used to summarize reviews. These reviews are biased as measured by their Claude rater which rates the credibility and sentiment. The authors argue that the bias is due to the demographic information of the user which was included as a visual signal along with the review. They then show that if they intervene by fine tuning the AI model that does that summarization, the bias is mitigated. I think it is an interesting domain with practical applications, but I’m confused by the experimental setup. Why would you provide the reviewer’s user profile image when summarizing a review? The photo of the user has nothing to do with the substance in the review and I don’t understand under what condition someone would actually do this. This seems to be a fundamental flaw, because prior work has already demonstrated that visual demographic information causes models to behave in stereotypical ways. Where is the experiment without this visual information? As in, demographics are implicitly or explicitly stated in the review. Your current setup is inducing the bias by including the demographically varied images of the users.

Authors state: “The application uses a locally assembled Amazon corpus with 10–20 customer reviews per product. For each product, baseline Gemma generates a review summary under a Black seller image and a matched White image from SocialCounterfactuals (Howard et al., 2024). Product title, review text, model, decoding settings, and instruction are fixed within the pair. The reported study protocol uses a neutral request for a two-sentence summary of the customer reviews, approximately 50 words, with the seller image supplied as context. The same request is used for base and DPO-tuned Gemma, as well as for both image conditions.”

Can the authors clarify this?

The writing is hard to follow, and sometimes a bit unprofessional. For example, “The term describes measured model behavior, not a claim that models have human mental states.” -> . No one is reading your previous paragraph as a claim that models have human mental states…and this sounds out of place and hurts your writing. Moreover, authors use terms like willingness-to-pay without a clear introduction and explanation of this concept. I appreciate the figure 1 overview but it can be made clearer such that on a skim we can understand your other experiments, especially about WTP.

Also section “Association and generative-bias measurement” under related work is missing a lot of bias/fairness/eval work that actually use the exact datasets you cite, but in the last few years 2026,2025, 2024. Please review the works that cite those datasets and do similar work as you and include them in your related work.

Questions for authors

Please clarify about the experimental setup, WTP, provide updated related works, etc. and address weaknesses listed above.

1 comment
AuthorsOpenReview ↗

How seller photos enter review generation and how independently randomized summaries affect downstream evaluations

Thank you for your time and detailed comments. We will upload an updated PDF soon.

1. Why we supplied a seller photo. We gave Gemma a Black or matched White seller photo alongside the same customer reviews to test whether the photo changes the summary. This resembles real world usage. From their official documents, Yelp Assistant uses reviews and photos (Yelp Inc., 2025), and eBay generates listing text from product photos (eBay News Team, 2023). Im et al. (2021) combine reviews and images for opinion summarization.

2. How we generated and scored the summaries. We used 10-20 Amazon customer reviews for each of 839 products. For each product, we generated one summary with a Black seller photo and one with a matched White seller photo from SocialCounterfactuals (Howard et al., 2024). For those two summaries, the product title, raw reviews, and instructions were identical. The prompt requested two sentences of about 50 words in total. An automated rater scored sentiment, clarity, detail, and first-hand experience. Two undergraduate research assistants each rated a separate set of 210 summaries as a human cross-check.

3. Effects of the summary and displayed photo. Buyers saw product information, a review summary, and a seller photo. We randomized the summary and displayed photo independently. Buyers did not know which photo Gemma had received. With a human source label, no representativeness statement, and a White photo on screen, Black-cue summaries lowered willingness to pay by USD 0.688 relative to White-cue summaries in the untuned condition (18.7%, p < .001). The corresponding estimate in the comparison summary set was −USD 0.022. Holding the White-cue summary and disclosures fixed, switching the displayed photo from White to Black had no statistically detectable effect (−USD 0.014, p = .917).

4. No-seller-photo control. A separate group saw the same marketplace interface and summaries with seller photos hidden for all products. Comparing products shown with Black- versus White-cue summaries, the raw WTP gaps were −USD 1.96 (untuned, N=103) and −USD 0.97 (comparison, N=41).

5. Model check without a face cue. With every face replaced by the same gray image, untuned Gemma showed a 0.00-percentage-point gap in negative-descriptor selection across the original photo groups. We will document this check in the supplement.

6. How we measure willingness to pay. We use a Becker–DeGroot–Marschak-style multiple price list (Becker et al., 1964, and Wertenbroch and Skiera, 2002). At each price from USD 0 to USD 15, buyers choose between USD 15 in cash and the product plus the cash left after subtracting that price. The list allowed one switch from product to cash as prices rose. Each buyer had a 1-in-20 chance of having one randomly chosen row implemented. Because that row’s price was fixed, buyers had an incentive to choose the option they preferred. If a buyer chooses the product through USD 6 and cash from USD 7, we record USD 6 as their willingness to pay.

7. We will remove the sentence about mental states, clarify Figure 1, and link to the exact summaries buyers saw. Howard et al. (2025), Cheng et al. (2025), and Lan et al. (2026) use SocialCounterfactuals to audit or mitigate model bias. Our experiment extends this work by randomly assigning summaries to buyers and measuring how the summaries change downstream consumer response for the same product.

Becker, G. M., DeGroot, M. H., and Marschak, J. (1964). “Measuring Utility by a Single-Response Sequential Method.” Behavioral Science 9(3), 226–232.

Cheng, H., Guo, Y., Guo, Q., Yang, M., Gan, T., Guan, W., and Nie, L. (2025). “Social Debiasing for Fair Multi-modal LLMs.” ICCV, 1740–1750.

eBay News Team (2023). “‘Magical’ Listing Tool Harnesses the Power of AI to Make Selling on eBay Faster, Easier, and More Accurate.” eBay Innovation.

Howard, P., Madasu, A., Le, T., Moreno, G. L., Bhiwandiwalla, A., and Lal, V. (2024). “SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples.” CVPR, 11975–11985.

Howard, P., Fraser, K. C., Bhiwandiwalla, A., and Kiritchenko, S. (2025). “Uncovering Bias in Large Vision-Language Models at Scale with Counterfactuals.” NAACL, 5946–5991.

Im, J., Kim, M., Lee, H., Cho, H., and Chung, S. (2021). “Self-Supervised Multimodal Opinion Summarization.” ACL-IJCNLP, 388–403.

Lan, J., Schlegel, U., Hannan, T., Zhang, G., Chen, H., and Seidl, T. (2026). “Unveiling the ‘Fairness Seesaw’: Discovering and Mitigating Gender and Race Bias in Vision-Language Models.” arXiv:2505.23798v2.

Wertenbroch, K., and Skiera, B. (2002). “Measuring Consumers’ Willingness to Pay at the Point of Purchase.” Journal of Marketing Research 39(2), 228–241.

Yelp Inc. (2025). “Yelp Expands AI Features to Make Local Discovery More Conversational, Visual and Seamless.” Press release, October 21.

Reviewer_aKGhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether demographic cues in multimodal generation can affect economic decisions through the information presented to consumers. It combines a face-to-descriptor diagnostic, paired summaries of 839 fixed product-review sets, DPO-based mitigation of Gemma 3 12B, and an incentivized marketplace experiment. Baseline summaries generated under Black seller cues receive lower sentiment scores and produce lower willingness to pay in the study’s reference condition. These gaps are substantially smaller for the evaluated DPO-generated summaries. The main contribution is the connection between demographic disparities in generated information and consequential human valuation.

Critical strengths
  1. The problem is practically important. The paper studies how demographic cues may influence consumers indirectly through AI-generated information, rather than only through explicit discriminatory statements.
  2. The human experiment adds substantial value beyond model-level bias metrics. Rather than only static benchmark results, this paper observes real-human incentivized valuation.
Critical weaknesses
  1. The control of demographic and other visual features needs further justification. (1) The focal face-to-word diagnostic in Section 3.1 uses only 12 images. Although gender is balanced and facial expressions are standardized, such a small image sample raises concerns about whether the measured gap is driven by particular faces rather than a stable demographic association. This concern is reinforced by the much smaller Gemma gap on the larger matched UTKFace panel. (2) Sections 3.1 and 4.1 should clarify how other visual differences are controlled before attributing the results specifically to Black–White demographic cues.

  2. The rationale connecting the DPO preference construction to bias mitigation needs to be clearer. The paper describes two types of preference pairs, but it remains unclear how the chosen–rejected labels are determined and why these preferences should specifically correct unjustified Black–White differences

  3. The reduction in demographic gaps is not accompanied by sufficient validation of summary quality. The post-tuning evaluation mainly reports group differences, leaving absolute quality and faithfulness to the original reviews insufficiently assessed. DPO might reduce sentiment gaps by omitting genuine criticisms or adding unsupported positive claims. Factuality, fluency, and coverage should also be checked.

  4. The presentation of some results could be improved. For example, Section 4.2 reports several numerical comparisons in prose.

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Visual input controls, DPO preference pairs, and the sentiment and faithfulness of review summaries

Thank you for your time and for identifying where the methods and results need more detail. We will upload a revised pdf soon.

1. Face panels and visual controls. The 12-image panel uses standardized, neutral-expression Project Implicit Race-IAT faces, with gender balanced within each racial group. The human test is widely used. Morehouse and Banaji (2024) report a pro-White association for 65% of 3.3 million U.S. respondents from 2007 to 2020. In our model diagnostic, we ran 50 queries per face and computed uncertainty at the image level. Gemma’s Black–White gap in negative descriptor selection was 32.9 percentage points (95% CI [15.7,48.4][15.7, 48.4]). On the broader UTKFace panel, the estimate was in the same direction but smaller (+1.3 points). Differences in age, expression, pose, and lighting, as well as uncertainty in estimated labels, could contribute to the difference between panels. We will describe the UTKFace finding more narrowly in the revision.

For review generation, we used paired synthetic Black and White seller portraits from SocialCounterfactuals (Howard et al., 2024) to limit other visual differences. The product title, reviews, and instructions were identical within each pair. We then randomized the summary and the seller photo shown to buyers independently. This lets us estimate how the resulting text changes what buyers will pay for the same product.

2. How we labeled the DPO pairs. We matched FairFace portraits before constructing the preference pairs. For 700 profile pairs, the preferred response was a teacher rewrite grounded in visible evidence, paired with the untuned Gemma response to the same image. For 300 descriptor pairs, we favored words supported by the visible expression over alternatives containing more unsupported negative terms. Preference judges saw no race labels, and two independent undergraduate assistants each audited a separate set of 500 pairs.

3. Absolute scores and faithfulness. For the summaries buyers saw, the table reports mean sentiment on a 1–10 scale. The pp-values come from two-sided paired tt-tests across the 16 products.

Survey conditionBlack cueWhite cueBlack − White (95% CI)pp
Untuned Gemma5.199.19−4.00-4.00 ([−4.85,−3.15][-4.85, -3.15])<.001< .001
DPO summary set8.639.06−0.44-0.44 ([−0.71,−0.16][-0.71, -0.16]).004.004

The gap narrowed by 3.56 points (95% CI [2.61,4.52][2.61, 4.52], p<.001p < .001), mainly because Black-cue summaries scored higher.

In a blind assessment of 16 pairs, two independent undergraduate raters agreed that every Black-cue version was more negative and traced the information in both versions to the original reviews. We also checked specific criticisms in the texts buyers saw in the DPO condition. Both kitchen-towel summaries reported that the towels were thin and better suited to decoration. Both trash-bag summaries reported tearing. We will include these examples alongside the scoring rubric for clarity, detail, and firsthand experience.

In a separate 839-product comparison, we scored summaries from untuned Gemma and tuned Gemma. Each cell lists Black-cue / White-cue results. The two-sided paired tests compare within the same products, separately for each racial cue. Even if the magnitudes were small, the tuned Gemma actually achived higher scores for clarity, detailness, and firsthand experience when generating review summaries.

MeasureUntunedDPO TunedPaired pp (Black / White)
Clarity8.87 / 8.969.00 / 9.00<.001< .001 / <.001< .001
Detail5.38 / 5.495.67 / 5.66<.001< .001 / <.001< .001
Firsthand experience3.62 / 4.044.69 / 4.77<.001< .001 / <.001< .001

4. Presentation. Thank you for pointing out that the comparisons in Section 4.2 are hard to follow in prose. We will put the group means, gaps, and confidence intervals in a single table.

References

Morehouse, K. N., and M. R. Banaji. 2024. “The Science of Implicit Race Bias: Evidence from the Implicit Association Test.” Dædalus 153(1), 21–50.

Howard, P., A. Madasu, T. Le, G. L. Moreno, A. Bhiwandiwalla, and V. Lal. 2024. “SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples.” Proceedings of CVPR, 11975–11985.

Zhang, Z., Y. Song, and H. Qi. 2017. “Age Progression/Regression by Conditional Adversarial Autoencoder.” Proceedings of CVPR, 4352–4360.

Thank you for the responses. This has addressed some of my concerns, except for the limited number of samples in the chosen dataset and only one dataset is applied.

AuthorsOpenReview ↗

Additional out-of-sample data evidence and the limitation of human experiment

Thank you for your follow-up. We incorporate additional evidence into the revised PDF.

1. Additional matched-face dataset. To address generalization beyond the original 12 faces, we evaluated 72 new Black–White pairs from FairFace, a large, racially balanced dataset (Karkkainen and Joo, 2021). Pairs matched gender, age band, expression, and image quality.

For the 24 negative-expression pairs, five queries per portrait produced 240 responses and 720 descriptor selections. Negative-word rates were 12.50% for Black faces and 1.67% for White faces: +10.83 pp (95% paired-bootstrap CI [−0.83, 25.00]). Black faces again received more negative descriptors than matched White faces, although the difference was not statistically significant.

2. Human sample and product scope. Our incentivized design and participant scale follow behavioral and human–AI experiments (Logg et al., 2019; Noy and Zhang, 2023). We sought varied products and broad U.S. participation through Prolific, obtaining 12,400 valuations from 775 survey completions across 16 products priced USD 6–15. What we could do was restricted by how human recruitment and actual purchases impose costs beyond model-only evaluations. Product-omission checks support within-sample robustness. Our contribution shows that demographic differences in AI-generated information affect valuations of the same product.

References

Karkkainen, K., and Joo, J. (2021). “FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age for Bias Measurement and Mitigation.” WACV, 1548–1558.

Logg, J. M., Minson, J. A., and Moore, D. A. (2019). “Algorithm Appreciation: People Prefer Algorithmic to Human Judgment.” Organizational Behavior and Human Decision Processes 151, 90–103.

Noy, S., and Zhang, W. (2023). “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.” Science 381, 187–192.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Seller-photo controls, out-of-sample evidence, summary faithfulness and capability tests, and buyer-valuation effects

Thank you to reviewers Q8Zr, aKGh, and 3pUy for your detailed feedback. This comment summarizes the revised paper and supplementary materials, with section references for each change. The central finding is that demographic differences in generated information change buyers’ valuations of the same product.

1. Generation and buyer design (Q8Zr; §4.1, §6.1–6.2, Figure 1). Gemma receives a seller photo alongside 10–20 reviews for each of 839 products. Product title, reviews, and a neutral request for a two-sentence, approximately 50-word summary are identical within each Black/White SocialCounterfactuals portrait pair. Yelp, eBay, and multimodal opinion-summarization research motivate this image-and-text setting.

Buyers see product information, a summary, and an independently randomized seller photo. Each participant sees one model’s summaries throughout to minimize cross-model carryover. Our monotone price list offers cash versus the product plus remaining cash. A 1-in-20 lottery implements one randomly chosen row, providing incentives to reveal WTP. Three independent raters prioritized everyday products and noticeably different summary wording. The focal analysis covers 775 completions and 12,400 valuations across 16 products priced USD 6–15.

2. Buyer outcomes and sensitivity (Q8Zr, 3pUy; §6.3, Appendix F). With a human label, no representativeness statement, and a White displayed photo, the Black-minus-White WTP effect is −USD 0.688 (−18.7%, p<.001) for untuned summaries and −USD 0.022 for tuned summaries. The gap narrows by USD 0.666 (95% CI [0.162, 1.171], p<.01). The displayed-photo effect is −USD 0.014 (p=.917). Independent randomization separates the effects of assigned summaries and displayed seller photos.

Averaging over identical disclosure/photo conditions gives gaps of −USD 1.247 and −USD 0.056, a USD 1.191 reduction (95% CI [0.985, 1.397]). Attenuation appears in all six disclosure conditions and all 120 two-product omissions (USD 0.957–1.331, all p<.01). Without the generator-by-summary-cue interaction, Black-cue summaries still reduce WTP by USD 0.478 (95% CI [−0.730, −0.226], p<.001). AI-source and representative-label effects remain positive and significant. Cross-survey attenuation is identified as exploratory. No-photo controls show raw gaps of −USD 1.96 untuned (N=103) and −USD 0.97 tuned (N=41).

3. Face diagnostics and visual controls (aKGh, 3pUy; §3, Appendices C–D). The 12 standardized, gender-balanced Race-IAT faces receive 50 queries each. Untuned negative selections are 353/900 for Black and 57/900 for White faces: +32.9 pp (95% CI [15.7, 48.4]). We narrowed the claim for noisy UTKFace. Among 72 new FairFace pairs matched on gender, age, expression, and image quality, 24 negative-expression pairs yield 240 responses and 720 selections: 12.50% versus 1.67%, +10.83 pp (95% CI [−0.83, 25.00]). This adds a same-direction comparison on new faces. Gemma endorses none of ten classifiable explicit stereotypes. Replacing faces with one gray image gives a 0.00-pp gap. The reported DPO diagnostic has 900/900 negative selections in both groups, explaining its zero gap and undefined within-tuned t-test. Summary and buyer outcomes provide the consequential mitigation tests.

4. Preference construction (aKGh, 3pUy; §5.1, Appendix D). FairFace portraits were matched before constructing 1,000 preferences: 700 pairs favor grounded teacher rewrites over untuned Gemma responses, and 300 favor expression-supported descriptors over alternatives containing unsupported negative terms. Judges saw no race labels. Two assistants audited nonoverlapping sets of 500 pairs. We clarify GPT/Claude roles and report rank-8 LoRA with all parameters. These details explain how preferences target unsupported demographic differences.

5. Summary quality and capability (aKGh, 3pUy; §5.2, Appendix E). Across 839 products, sentiment gaps fall from −1.061 to −0.007 (tuned 95% CI [−0.036, 0.022]); tuned Black/White means are 8.644/8.651. Clarity, detail, and firsthand-experience scores improve (paired p<.001). Buyer-text sentiment gaps fall from −4.00 to −0.44; two independent blind raters traced both versions to source reviews.

We evaluate all 3,356 summaries using claim support, source-complaint retention, VADER, and BERTScore. The VADER gap shrinks 99.3%; BERTScore precision changes from .85683 to .85623 (adjusted p=.329). Overall claim support falls from 86.12% to 80.32%, and selected criticism coverage from 52.05% to 33.67%; tables report both pooled and racial-gap changes. On 1,000 TextVQA questions, visual-language accuracy is 66.5% untuned and 66.2% tuned (−0.30 pp, 95% CI [−0.74, 0.12], adjusted p=.393). These tests assess quality alongside the smaller summary and valuation gaps.

We also revised Figure 1, explained WTP, removed the mental-states sentence, added recent counterfactual-bias literature (§2), and included the exact buyer summaries and TextVQA records.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.