LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits

George Rahul, Ronit Jahilal, Bhaskara Hanuma Vedula, Ponnurangam Kumaraguru, Pushkar ShuklaarXiv 2026 · arXiv

TL;DR

The paper introduces ADDIS (Additive Intersectional Screen), a statistical framework for auditing intersectional bias in Large Language Models (LLMs) that distinguishes true identity interactions from simple additive effects. By applying ADDIS to a new benchmark of 10,000 Indian personas, the authors demonstrate that most "intersectional" biases reported by traditional methods are actually just the sum of individual attribute biases, though a significant core of genuine interactions remains.

What Problem Does This Paper Address?

Intersectionality theory posits that the harm faced by individuals with multiple marginalized identities is not merely the sum of the harms of each identity. However, current algorithmic audits often fail to operationalize this correctly. They typically flag a subgroup (e.g., "Muslim woman") as biased if it fares worse than its constituent groups ("Muslim" or "woman") individually.

The authors argue this is a form of "Gerrymandering" in bias auditing. If a model is biased against women and independently biased against Muslims, a Muslim woman will naturally face a higher total bias. This is a purely additive effect, not a unique intersectional interaction. A true intersectional effect is a departure from this expected sum—what Kimberlé Crenshaw described as an experience "greater than the sum of racism and sexism." Without a proper additive baseline, current audits suffer from high false-positive rates, misidentifying independent effects as complex interactions.

Fig. 1. The single-attribute criterion cannot tell addition from interaction. Relative to a Hindu man, woman adds 3 points of attributed anxiety and Muslim 4 of depression. Points beyond the dashed arc are flagged, and red marks Ard 𝛿from their sum (open circle). (a) Without interaction, the Muslim woman is still flagged. (b, c) The same interaction raises the excess more along the sum than at a right angle. (d) If Muslim lowers anxiety instead, the shifts nearly cancel and a real interaction stays inside the arc.

Methods

The authors propose ADDIS, which moves away from simple group comparisons toward a regression-based approach:

  1. Additive Baseline: ADDIS fits a linear regression model over all identity attributes (gender, age, caste, etc.) across the entire dataset. This model predicts the "expected" response for any combination of identities, assuming no interactions exist.
  2. Additive Residual Distance (ARD): The metric $\delta$ measures the distance between a subgroup's observed mean response profile and the additive model's prediction.
  3. Freedman–Lane Null Hypothesis: To determine if a measured $\delta$ is statistically significant, ADDIS uses a sign-flipping procedure on residuals. This preserves the main effects of each identity while destroying interaction effects, creating a robust "null" distribution for comparison.
  4. Family-Wise Error Rate (FWER) Control: Because audits test hundreds of potential identity pairings, ADDIS uses a max-$z$ rule to control the probability of false discoveries across the entire search space.

Datasets

  • Synthetic Benchmark: 150 datasets generated with known additive effects and zero interactions, plus datasets with "planted" interactions of varying sizes to validate recall and precision.
  • IndiPersona: A newly introduced dataset of 10,000 synthetic Indian personas crossing six identity attributes: gender identity, age, sexual orientation, socioeconomic status (SES), caste, and religion. Each persona is described in one of 50 everyday scenes.
  • Counterfactual Quartets: A targeted dataset used to isolate causal interactions by changing only two labels (e.g., caste and class) while keeping all other persona details identical.

Key Insights and Ideas

  • Additive Bias is not Intersectional Bias: Most current audits "double count" bias. If a model is consistently biased against two groups, it will naturally be more biased against their intersection. ADDIS shows that this is often just consistent single-attribute bias, not a unique intersectional failure.
  • The "Right-Angle" Interaction: Intersectionality is often about a combination doing something different than its parts. In high-dimensional response spaces, most interactions point in directions perpendicular to the main effects. Traditional metrics often miss these "right-angle" interactions entirely.
  • Rounding Creates Spurious Interactions: When LLMs provide answers on a Likert scale (e.g., 1–5), the act of rounding a continuous internal representation can create mathematical interactions where none exist in the underlying logic. ADDIS is more robust to this than previous methods but highlights it as a persistent challenge for scale-based audits.
  • Caste $\times$ Class Interactions: While observational audits often lack the data density to find rare subgroups (like "Dalit $\times$ ultra-rich"), counterfactual testing revealed that caste and class interact significantly in almost all audited models.

Key Results

  • False Positive Reduction: On synthetic data with no interactions, the standard "single-attribute" criterion flagged 66.1 false subgroups per dataset on average. ADDIS flagged only 0.09.
  • Improved Recall: ADDIS recovered 84% of planted interactions in synthetic benchmarks, compared to only 32% for the standard criterion.
  • LLM Audit Findings: Auditing 13 open models (including Llama 3, Gemma, and Qwen) and DeepSeek V4.1, ADDIS found that only ~3% of subgroups were truly intersectional on scale scores.
  • Primary Interaction Hub: The most consistent and largest intersectional interaction across all models was gender identity $\times$ sexual orientation, particularly affecting transgender and queer personas.

Limitations and Future Work

  • Synthetic Personas: The audit relies on synthetic personas, which may not perfectly reflect how models treat real people or how bias manifests in real-world Indian social contexts.
  • Pairwise Focus: The current implementation of ADDIS focuses on pairs of attributes. The authors suggest extending this to three-way or higher-order intersections in future work.
  • Harm vs. Attribution: The paper measures how models attribute traits to personas (e.g., "this person is anxious"), which is a proxy for bias but not a direct measure of downstream harm in specific applications like hiring or medical triage.
  • Likert Sensitivity: While superior to alternatives, ADDIS still sees an increased false-positive rate (up to 20%) on rounded Likert data, suggesting a need for ordinal-aware models in future iterations.

Summary updated

Abstract

Intersectional bias arises when a model treats people who hold several marginalised identities at once differently from what their separate identities would predict. Auditing such intersectional subgroups remains a challenge: their number grows quadratically, and there is no agreed baseline for what independent attributes would predict. Without such a baseline, many existing audits flag a subgroup when it exceeds each attribute alone, a condition that attributes acting independently already satisfy. We test this standard criterion on 150 synthetic datasets built with no interaction. With the shuffle null, it flags 66.1 false subgroups per dataset on average, and at least one in every dataset. When we plant interactions of known size into these datasets, a repaired null lets the criterion recover only 32% of them. We introduce ADDIS (Additive Intersectional Screen), which fits an additive model over all attributes, scores each subgroup by its Additive Residual Distance (ARD) from that model's prediction, and tests it against a Freedman-Lane null. On our synthetic benchmark, ADDIS flags 0.09 false subgroups per dataset and recovers 84% of the same planted interactions. We then build IndiPersona, 10,000 synthetic Indian personas with six identity attributes, and audit 13 open language models. The standard criterion flags 7.1% of the 738 tested subgroups on average; ADDIS flags 3.0%. Gender identity crossed with sexual orientation is the largest pairing in every model, and counterfactual personas confirm it and additionally reveal a caste x class interaction in all but one model that the observational test lacked the data to detect. Much of what audits report as intersectional bias can thus arise from independent single-attribute effects alone, yet a small core of interactions survives.
BibTeX
@misc{rahul2026llm,
  author = {George Rahul and Ronit Jahilal and Bhaskara Hanuma Vedula and Ponnurangam Kumaraguru and Pushkar Shukla},
  title = {{LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits}},
  year = {2026},
  eprint = {2611.26538},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26538},
}
Figure