PaperlayerView exact revision on arXiv

Source-linked AI summary

How Humans and LLMs Read Gender into "Gender-Neutral" Physical Descriptions

Yingjia Wan, Lin Lin, Elisa Kreiss

arXiv:2609.16366v2cs.CLcs.AIcs.CYcs.HC

TL;DR

Whether physical descriptions are meaningfully gender-neutral has remained an empirical question. The paper builds GAPA from human ratings of physical attributes and evaluates LLMs against those judgments. It finds that descriptions carry gender associations and that models only partially recover human patterns, with systematic misalignment.

  • Problem

    Whether physical descriptions carry gender associations among readers, despite being used as a seemingly neutral alternative to identity labels, remains underexplored.

    Source: §1 · p. 2

  • Method

    The authors compile 316 physical descriptors into GAPA, collect gender-association ratings, compare 16 LLMs with human judgments, and train a proxy predictor.

    Source: §1 · pp. 2–3, §4 · p. 4

  • Results

    Physical descriptions carry structured, graded gender associations; 53% of attributes are significantly more associated with one gender, while models partially recover patterns with compressed ratings and weaker alignment for men.

    Source: §1 · p. 3, §4.3 · p. 5, §7 · p. 10

  • Takeaways & Limitations

    Physical descriptions are not necessarily gender-neutral in human interpretation, challenging their use as substitutes for explicit identity labels.

    Source: §1 · p. 2, §7 · p. 10

    Ratings concern socially shared associations, not inherent gendering, and isolated attributes omit discourse context, so findings are not a complete account of gendered meaning.

    Source: Limitations · p. 10

Abstract

from arXiv · show

When foundation models describe people, recent work in AI fairness, accessibility, and ethics recommends avoiding inferred identity labels (e.g., "she", "his") in favor of seemingly "objective" physical descriptions (e.g., "short hair", "a defined jawline"). Yet whether such descriptive language achieves gender-neutral communication remains an open empirical question. To study this, we introduce GAPA (Gender Associations of Physical Attributes), a dataset of 316 common physical attributes drawn from diverse sources, paired with 14,706 gender-association ratings from 304 US-based annotators. Results show that physical descriptions carry structured and graded gender associations among readers, with more consistent and distinctive associations for women and men than for non-binary identities. Next, we evaluate 16 LLMs across model families, sizes, and post-training variants against human ratings. The models partially recover human associations but exhibit systematic alignment biases, including compressed rating distributions, weaker alignment for associations with men, and asymmetric abstention that disproportionately targets the non-binary category. Finally, we release the best-performing proxy model trained to predict humans' gender associations of descriptive language and demonstrate its utility through a sociolinguistic analysis of character descriptions in LitBank. Together, our findings provide the first empirical evidence that seemingly "objective" physical descriptions can retain systematic gender associations in human interpretation, and uncover systematic patterns of model-human misalignment. This challenges the assumption that replacing explicit gender labels with physical descriptions necessarily yields gender-neutral communication, and highlights downstream challenges in using such descriptions to communicate subjective identity categories in human-AI interaction.

1 Introduction

The paper questions whether physical descriptions are meaningfully gender-neutral, since readers may associate them with gender even when they replace explicit identity labels. It introduces GAPA and evaluates how well LLMs represent those human associations.

1 Introduction

  • Physical descriptions may carry systematic gender associations, challenging the assumption that they provide a gender-neutral alternative to identity labels.The paper asks whether isolated descriptors have consistent associations among readers and whether language models reproduce or distort them.Source: §1 · p. 2
  • GAPA contains 316 physical-attribute descriptions and 14,706 ratings from 304 U.S.-based participants on associations with women, men, and non-binary people.The attributes come from LLM generation, human elicitation, and contemporary fiction.Source: §1 · p. 2
  • Sixteen LLMs partially recover human associations, while showing systematic distortions across multiple dimensions.A proxy predictor trained on GAPA achieves r=0.76 against held-out human ratings and is used to analyze character descriptions in LitBank.Source: §1 · p. 3

2 Related Work

Prior research examines gender inference from visual and acoustic cues and has developed NLP benchmarks for gender bias. These benchmarks largely focus on labels, leaving gender associations in descriptive vocabulary underexplored.

2 Related Work

  • Social-science research treats gender as graded and context-dependent, and documents gender inference from cues including voice, facial morphology, and motion.Research also finds that these associations are systematic and context-sensitive.Source: §2 · p. 3
  • NLP research has developed gender-bias benchmarks and mitigation methods, while more recent work reports model difficulties with gender-neutral language and non-binary categories.Examples include struggles with gender-neutral pronouns, lower performance on non-binary categories, and inconsistent generation of gender-neutral language.Source: §2 · p. 3
  • Prior NLP work has focused mainly on gender associations with labels rather than descriptive vocabulary, leaving physical descriptions underexplored as a possible source of gender associations.The paper investigates this vocabulary in both human interpretation and language models.Source: §2 · p. 3

3 GAPA: A Collection of Physical Attributes

GAPA assembles physical attributes from LLM-generated descriptions, human-written terms, and novels to capture a range of attribute types and styles.

3 GAPA: A Collection of Physical Attributes

  • The collection draws on LLM-generated attributes for broad categorical coverage (n=170), human-written attributes for high-frequency terms (n=50), and novel-derived attributes for unusual references (n=96).The novel-derived attributes come from six novels.Source: §3 · pp. 3–4
  • Attribute selection procedures are described in Appendix A.The collection combines sources chosen to provide broad coverage, frequent terms, and unusual or stylized references.Source: §3 · pp. 3–4
  • These three sources yield a collection of 316 physical attributes.The paper directs readers to Appendix A for procedural specifications on attribute selection.Source: §3 · p. 4

4 Physical Attributes Carry Systematic Gender Associations for Humans

A human-rating study finds that physical attributes carry structured gender associations rather than functioning as generally gender-neutral descriptions. These associations are less consistent for non-binary targets, and their patterns vary across attributes and sources.

4.1 Experiment Details

  • The study measured associations in ratings from 304 participants using a 7-point scale, with each participant rating an attribute for only one target gender.Participants rated the likelihood that someone would say a woman, man, or non-binary person had an attribute; the design averaged 17 ratings per attribute and gender category.Source: §4 · p. 4, §4.1 · p. 4

4.2 Metrics

  • Gender Association is the mean rating for an attribute–gender pair, while Gender Distinctiveness is the gap between the highest and second-highest category means.Together, the metrics separate strength of association from how clearly an attribute differentiates among categories.Source: §4.2 · p. 4

4.3 Results

  • Across GAPA, physical attributes were associated more with women than men and least with non-binary people, and were not perceived as generally gender-neutral.The authors note that the overall pattern could reflect either a feminine skew in sampled attributes or broader biases in language use.Source: §4.3 · pp. 4–5
  • 53% of attributes significantly favored one gender; conventional woman–man ranking patterns were more frequent and significant (68%–73%) than patterns ranking non-binary people highest.Non-binary-highest patterns were the rarest and had much lower significance rates.Source: Figure 2 · p. 5, §4.3 · p. 5
  • Man–woman associations were negatively correlated overall (r = −.16, p = .004), but source-specific correlations ranged from r = −0.61 for human-sourced attributes to r = 0.27 for novel-extracted attributes.Non-binary associations with both women and men were significantly positively correlated, so an attribute being common for a gender did not necessarily make it gender-diagnostic.Source: §4.3 · pp. 5–6
  • Human ratings were most reliable for women, intermediate for men, and least reliable for non-binary targets across attribute sources.The authors interpret the lower non-binary reliability as less consistency across raters.Source: Figure 4 · p. 6, §4.3 · p. 6

5 Physical Attributes Carry Systematic Gender Associations for LLMs

Across 16 LLMs, ratings only partially recover human gender associations for physical attributes: distributions are compressed toward the midpoint, and alignment is weaker for the man category. The evaluation also finds model-specific differences and non-binary-targeted abstention.

5.1 Method

  • The zero-shot evaluation prompts 16 diverse models to give a single 1–7 rating for each physical attribute, using the direct-generation responses for reported results.Top-logit token extraction was also tested, with results consistent with direct generation.Source: §5.1 · p. 6, §5 · p. 6

5.2 Evaluation Metric & Baselines

  • Alignment is measured primarily with Pearson r against averaged human ratings, with RMSE as an auxiliary metric and human reliability measures as reference baselines.The leave-one-rater-out correlation approximates a typical rater; ICC(1, k) approximates the maximum correlation attainable against aggregated judgments.Source: §5.2 · pp. 6–7

5.3 Results

  • LLMs show moderate alignment overall, with higher correlations for woman and non-binary person than for man, centering around or exceeding the typical-rater baseline.The comparison uses Pearson correlation with averaged human ratings and human reliability baselines; Figure 5 shows rating distributions and model–human correlations.Source: §5.3 · p. 7, Figure 5 · p. 7
  • Models concentrate ratings in the mid-range (3–5) and use extreme scores (6–7) less often than human annotators.This pattern indicates more conservative, less differentiated ratings than human judgments.Source: §5.3 · p. 7, Figure 5 · p. 7
  • Claude-Opus-4.6 has the strongest average Pearson correlation (r = 0.669), while GPT-OSS-20B has the weakest overall (r = 0.173).Proprietary models show only a modest collective advantage over open-source models, with substantial performance overlap.Source: §5.3 · p. 7, Figure 6 · p. 8
  • Instruction-tuned open-source models generally improve over their base counterparts in RMSE, but show no consistent Pearson r gains.The reported pattern suggests lower absolute prediction divergence without substantially improved rank-order alignment.Source: §5.3 · p. 8, Figure 6 · p. 8
  • Abstentions disproportionately target the non-binary category and are concentrated in Gemini-3-flash and instruct-tuned open-source models.Corresponding base models show little comparable abstention; Figure 6c reports abstention counts.Source: §5.3 · p. 8, Figure 6 · p. 8

6 Scalable Socio-Linguistic Analyses Using a Trained Proxy Predictor

The authors train a proxy predictor on GAPA and apply it to physical descriptions in LitBank. The analyses find associations with character and author gender, while their interpretation is bounded by the contemporary U.S. reader perceptions represented in GAPA.

6.1 Training a Proxy Model for Predicting Humans’ Gender Associations

  • The proxy is trained on GAPA to predict aggregated human ratings for woman, man, and non-binary person from physical-attribute descriptions.Models are fine-tuned with a regression head; the study uses a 60/15/25 train/validation/test split and a standardized hyperparameter search.Source: §6.1 · p. 9
  • The olmo2_7b_base proxy achieves the highest test-set correlation (r = 0.764) and lowest RMSE (0.633) among the evaluated candidates.The authors propose it as a simulator of human-like gender-association interpretations for unseen items.Source: §6.1 · p. 9, Figure 7 · p. 9

6.2 Scaled-Up Analyses on LitBank (Bamman et al., 2019)

  • The LitBank analyses reflect contemporary U.S. reader perceptions because the GAPA participants determine the proxy’s ratings.This bounds what the analyses establish about descriptions in older literary texts.Source: §6.2 · p. 9
  • In LitBank, descriptions of male and female characters differ in the direction of their gender associations, while their absolute distinctiveness is comparable.Attributes describing male characters are rated more associated with man than woman, with the reverse ordering for female characters.Source: Figure 8 · p. 9, §6.2 · p. 10
  • Author gender is associated with feminine ratings for descriptions of male characters: male authors’ descriptions have weaker feminine associations than female authors’ (4.06 vs. 4.22; β = −0.12, p = .03).Ratings of association with man are largely invariant to author gender, and female characters elicit similar descriptions from both author groups.Source: Figure 8 · p. 9, §6.2 · p. 10

7 Conclusion

The conclusion argues that physical descriptions do not ensure gender-neutral communication: readers associate them with gender, and LLMs only partially reproduce those associations. It presents a proxy predictor as a way to analyze such patterns in naturalistic text.

7 Conclusion

  • Physical descriptions carry structured, graded gender associations among human readers, and the 16 evaluated LLMs partially recover those patterns while showing systematic misalignment.The conclusion identifies midpoint-compressed ratings, weaker alignment for man, and non-binary-targeted asymmetric abstention as patterns of misalignment.Source: §7 · p. 10

Limitations

The authors qualify the human norming results: ratings capture socially shared associations, not inherent gendering, and the isolated, heterogeneous attributes do not provide a complete account of gendered meaning.

Limitations

  • Ratings may reflect stereotype knowledge, perceived prevalence, and expectations about language use as well as semantic association.The task measures socially shared associations between attributes and gender categories, not whether attributes are inherently gendered.Source: Limitations · p. 10
  • Attributes were rated in isolation despite a heterogeneous set, and non-binary appearance prototypes may be less culturally shared, so findings are not a complete account of gendered meaning.The authors note that discourse context can shape interpretation and that lower agreement for non-binary person may reflect greater conceptual heterogeneity.Source: Limitations · p. 10

Ethics Statement

The study analyzes socially shared associations in language and perception, not individuals’ gender from appearance, and frames its findings descriptively. It calls for safeguards against using the dataset or models for inference, profiling, or decisions about real people.

Ethics Statement

  • The analysis concerns aggregate patterns in language and perception, not defining or inferring an individual’s gender from appearance.The authors caution against using the dataset or models for gender inference, profiling, or decision-making about real individuals.Source: Ethics Statement · p. 11
  • Human-subject data were collected with informed consent and fair compensation, without personally identifiable information beyond basic demographics.Source: Ethics Statement · p. 11
  • The released dataset and models are intended for research, such as auditing implicit bias and improving AI transparency, with safeguards and contextualization.Source: Ethics Statement · p. 11

A Supplementary Information of Dataset Collection for GAPA

GAPA attributes were assembled from LLM generation, human elicitation, and contemporary fiction. Participant recruitment and the final sample are described alongside these collection procedures.

A Supplementary Information of Dataset Collection for GAPA

  • Physical attributes were gathered from three distinct source domains to support broad coverage, cognitively prominent concepts, and rare or stylized descriptors, using LLM-generated, human-written, and novel-extracted attributes.For the novel-extracted set, 16 attributes were randomly sampled from each of six contemporary novels.Source: A Supplementary Information of Dataset… · p. 14

B.1 Participant Demographics

  • The study recruited 320 U.S. participants through Prolific, requiring native English speakers, U.S. residence, a 95% approval rate, and passing an attention check.Participants were 19–83 years old (M = 42.70).Source: §B.1 · p. 14

B.2 Descriptive Results

The descriptive results compare human gender-association ratings across attribute sources, while the accompanying material explains reliability ceilings and the novel-extraction procedure. The prompt operationalizes extracted attributes as described body parts and applies textual criteria to extraction and gender labeling.

B.2 Descriptive Results

  • Across human-written, LLM-generated, and novel-derived attributes, ratings were highest for women, followed by men, and lowest for non-binary targets.The same ordering held within most individual sources and in the overall aggregates.Source: §B.2 · p. 14, Table 2 · p. 16

B.3 Correlation Ceilings from Human Uncertainty

  • LOO estimates a typical rater’s agreement with other raters, while ICC(1, k) estimates the reliability of averaged ratings and a stricter ideal-model correlation ceiling.Because the human targets are averages of noisy judgments, the authors use both ceilings as references for interpreting model correlations.Source: §B.3 · p. 17
  • Finite-sample evaluation noise can make empirical model correlations occasionally exceed the ICC(1, k) ceiling reference.Source: §B.3 · p. 17

C Extraction Pipeline; C.1 Extraction Pipeline Process

  • The novel-extraction pipeline filters likely physical-description sentences, uses an LLM to extract attributes, and postprocesses the results, including pronoun normalization.The filtering stage uses physical-attribute noun and adjective lexicons; extraction requires a body part and a characterizing modifier.Source: §C.1 · p. 18

C.2 Prompt for LLM-Based Extraction

  • The prompt requests complete noun phrases describing a body part, excludes bare body parts or descriptions without a body part, and labels gender only when the text provides solid evidence.Source: §C.2 · p. 18

D.2 Method Details

The section describes how LLM ratings are elicited and compared with human ratings, then outlines the proxy-model training and tuning setup. It also bounds the LitBank analysis to associations held by recently recruited US participants.

D.2 Method Details

  • Direct-generation and top-logit extraction nearly perfectly agree except for Mistral-7b; directly generated ratings are reported as primary for accessibility.The two strategies capture verbalized responses and first-token rating preferences, respectively.Source: §D.2 · p. 20

D.3 Complete Results

  • Figure 6a presents ratings across genders for all models against human ratings, while Table 3 reports per-gender RMSE and Pearson r statistics.Source: §D.3 · p. 20, Table 3 · p. 21

E.1 Training Details

  • Training uses LLM-generated GAPA for 60/15/25 train/validation/test splits, with novel-derived and human-written attributes reserved for out-of-domain evaluation.A linear head predicts aggregated human ratings as continuous scores; MSE training encourages shared annotator signal while attenuating idiosyncratic noise, with equal tuning budgets and consistent cross-validation and multi-seed settings.Source: §E.1 · pp. 20–21

E.2 Full Configurations for Hyperparameter Search

  • The hyperparameter configuration aims for controlled comparisons across model family, scale, and instruction-tuning variants.Source: §E.2 · p. 21, Table 4 · p. 22

F How to Interpret the Exploratory LitBank Analyses

  • The LitBank analysis reflects gender associations held by recently recruited US participants, not a direct view of authors’ absolute biases decades earlier.The passage frames contemporary reader interpretation and comparisons between author groups as a basis for hypotheses about potentially persistent perceptions.Source: F How to Interpret the Exploratory… · p. 22
Loading 2609.16366v2…