Same Sentence, Different Construct: Identity-Swapped Rationales Reveal Measurement Non-Invariance in LLM Hate Speech Classifiers

Wenchang Wang, Yue SongarXiv 2026 · arXiv

This paper investigates a critical vulnerability in Large Language Models (LLMs) used for content moderation: measurement non-invariance. While modern LLMs achieve stellar aggregate accuracy on hate speech benchmarks, the authors demonstrate that these models often apply fundamentally different rules depending on which social group is being targeted. Specifically, models frequently fail to recognize hateful statements directed at women, recasting them as "personal dislikes" while correctly identifying identical statements directed at other protected groups as hate speech.

TL;DR

Even LLMs with >95% accuracy exhibit "measurement non-invariance," applying different evidentiary standards for hate speech depending on the target identity. By swapping identity terms in identical sentence templates, researchers found that models systematically miss misogynistic abuse by rationalizing it as interpersonal conflict rather than group-based hate.

What Problem Does This Paper Address?

As LLMs are increasingly deployed as "surrogate annotators" for social science research and automated moderation, we typically validate them using aggregate metrics like balanced accuracy. However, the authors argue that aggregate agreement is insufficient. A scientifically sound instrument must satisfy measurement invariance: it must map the latent construct (e.g., "hate speech") onto observed labels symmetrically across all subpopulations.

If a model flags "I hate [Group A]" but ignores "I hate [Group B]," it isn't just making a random error; it is operating under a "construct shift." This paper aims to diagnose these shifts by looking past the final label and auditing the model's stated rationales.

Methods

The researchers employed an identity-controlled audit framework (illustrated in Figure 1) to isolate the causal impact of identity terms.

Fig. 1. Identity-controlled audit framework, illustrated on a HateCheck template using Hy4 Preview generations. (1) Syntactic templates fix the surrounding lexical context and ground-truth label; only the identity designation is systematically permuted across seven protected classes. (2) Each prompt variant is evaluated under an identical protocol to extract discrete verdicts and qualitative rationales. (3) Estimand E2 benchmarks target-specific error rates against the six-group baseline on the identical template; E3 identifies criterion shifts by comparing rationales across divergent and convergent classifications within the same template.

Key techniques included:

  • Identity-Swap Testing: Using templates where every word is identical except for the target group (e.g., "women" vs. "trans people").
  • Excess Error Estimands: Measuring how much more likely a model is to miss hate speech for one specific group compared to the average of all other groups on the exact same sentence structures.
  • Rationale Contrast Analysis: Comparing the model-generated justifications for divergent classifications. They coded for "protected-status invocation" (recognizing group targeting) vs. "denial or personalization" (framing the attack as an individual grievance).
  • Definitional Ablation: Testing if providing the model with an explicit, authoritative definition of hate speech (including "sex" as a protected category) fixes the invariance.

Datasets

  • HateCheck: A functional diagnostic suite containing 3,728 cases. The authors focused on 421 templates instantiated across seven protected groups: women, trans people, gay people, Black people, disabled people, Muslims, and immigrants.
  • Counterfactually Augmented Sentiment Data (CAD): A set of movie reviews minimally edited to flip sentiment. This served as a non-identity-based baseline to distinguish systematic bias from general classification noise.

Key Insights and Ideas

  • The "Personalization" Loophole: When models miss hateful statements about women, they don't just "fail"—they invent a reason. In 22 of 30 missed templates for one model, the rationale explicitly reframed misogyny as a "personal dislike toward a specific individual," even when the phrasing was identical to cases where it correctly identified group-level hate for others.
  • Definitions are Double-Edged Swords: Adding an explicit definition that included "sex" did not close the coverage gap for women. However, it did cause models to stop flagging hate speech against immigrants because "immigrants" were not explicitly listed in the provided text. Models seem to adhere strictly to negative exclusions (what is left out) while resisting explicit inclusions (what is added).
  • False Positives vs. False Negatives: The paper reveals a crucial distinction. Disparities in false negatives (missing hate) are usually due to model construct shifts. Disparities in false positives (over-flagging) are often "compositional artifacts"—they vanish under identity control and are usually caused by the presence of reclaimed slurs in the dataset for specific groups.

Key Results

The audit of four state-of-the-art models (DeepSeek Flash, GLM-5.3-Flash, Hy4 Preview, and MiniMax M3.1 Flash) yielded several quantitative "aha" moments:

  • High Aggregate Accuracy: All models exceeded 0.95 balanced accuracy on HateCheck, appearing "safe" by traditional standards.
  • The Misogyny Gap: Despite high accuracy, Hy4 and DeepSeek showed excess false-negative rates for women of 8.3 and 4.4 percentage points, respectively (see Figure 2).
  • Rationale Divergence: As shown in Table 2 (Figure 5), a model might label "I really can't stand women like you" as a "personal dislike" while labeling "I really can't stand trans people like you" as "hate speech targeting a protected characteristic."
Fig. 2. Within-template excess error across target groups and models (percentage points, 95% template-bootstrap CI). Top: hateful templates (positive values indicate elevated false-negative rates relative to identical counterfactual statements). Bottom: non-hateful templates (positive values indicate elevated over-flagging). Women are highlighted.
Table 2. Representative paired rationales from identical syntactic templates (verbatim outputs, truncated). Lexical framing is held constant while permuting solely the protected target.

Limitations and Future Work

  • Synthetic vs. Naturalistic Text: The study uses synthetic templates (HateCheck) to isolate causal effects. While effective for auditing, real-world social media text involves more complex pragmatics and sarcasm.
  • Geographic/Cultural Alignment: All audited endpoints were from China-based providers. The authors suggest future work should investigate if Western-developed models exhibit the same structural persistence of misogyny gaps.
  • Prompt Engineering: The study tested standard prompts and definitional prompts; exploring Chain-of-Thought (CoT) or few-shot demonstrations remains a promising direction for mitigation.

Ultimately, the authors argue that we must move beyond aggregate accuracy. To build truly fair classifiers, we need standardized reporting that includes identity-controlled tests of measurement invariance.

Summary updated

Abstract

Large language models are increasingly used as measurement instruments for socially consequential constructs such as hate speech, and are typically validated by their aggregate agreement with human labels. A valid instrument must also apply the same construct to every group it measures. We test this requirement using 421 \hc{} templates that are instantiated identically for seven protected groups, so that only the identity term varies. Four LLM classifiers each exceed 0.95 balanced accuracy, yet within identical templates two of them miss hateful statements about women at markedly higher rates than the same statements about other groups (excess false-negative rates of 8.3 and 4.4 percentage points), a gap first reported for fine-tuned classifiers in 2021. We show how the construct shifts. Comparing rationales within each template, we find that models invoke protected-group status for other groups but recast the identical statement about women as personal dislike or as aimed at an individual, in 22 of 30 missed templates for one model. Adding the benchmark's own definition, which lists sex as a protected characteristic, does not close the gap, and the contrast persists. Apparent false-positive disparities, by contrast, largely vanish under identity control and trace to reclaimed slurs present only for some groups. Errors shared by all models concentrate where model and benchmark definitions diverge, and pairwise counterfactual accuracy adds little beyond endpoint accuracy. We argue that validating LLM measurement requires identity-controlled tests of measurement invariance that examine both decisions and the criteria models state for them.
BibTeX
@misc{wang2026same,
  author = {Wenchang Wang and Yue Song},
  title = {{Same Sentence, Different Construct: Identity-Swapped Rationales Reveal Measurement Non-Invariance in LLM Hate Speech Classifiers}},
  year = {2026},
  eprint = {2611.02491},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02491},
}
Figure