Investigating Counterfactual Unfairness in LLMs towards Identities through Humor

1. Problem Statement

Large Language Models (LLMs) trained on web-scale corpora often internalize social and cultural biases. While previous research has focused on decontextualized bias (e.g., static word embeddings or isolated sentences), this paper argues that conversational context and relational dynamics are critical yet overlooked dimensions of bias. Specifically, humor is a unique probe because it relies on ambiguity, social cues, and power dynamics to determine what is "funny" versus "offensive."

The core problem addressed is counterfactual unfairness: Do LLMs produce different outputs (refusals, intent attributions, or impact predictions) when the identities of the speaker and listener are swapped, even when the content of the joke remains identical? The authors hypothesize that models encode fixed social hierarchies, leading to asymmetric treatment where "privileged" speakers are judged more harshly than "marginalized" speakers, regardless of the actual intent or content.

2. Methodology

The study employs a counterfactual fairness framework (Kusner et al., 2017), systematically swapping sensitive attributes (speaker/listener identities) while holding all other factors (joke content, context, prompt structure) constant. The research is structured around three complementary tasks:

Task 1: Speaker-Target Conditioned Humor Generation

• Objective: To measure refusal rates based on who is telling the joke and who is the target.
• Dataset: 80 request templates covering realistic scenarios (social media, professional, personal) with both benign and malicious intents.
• Setup: 33 identities across 10 categories (race, religion, wealth, profession, etc.) generate 121 unique speaker-target pairs.
• Metrics:
  • Asymmetric Refusal Rate (ARR): $|RR(A \to B) - RR(B \to A)|$, measuring the difference in refusal rates when identity directions are reversed.
  • Speaker Effect (SE): Measures how specifying a speaker identity amplifies or dampens refusal rates compared to target-only baselines.
• Evaluation: Automated judgment via GPT-4o (validated by Llama-3 and human annotators) to classify refusals (direct, suggested alternative, explicit alternative, implicit alternative).

Task 2: Speaker Intention Inference

• Objective: To determine if models attribute different malicious or benign intents to identical jokes based on speaker-listener pairs.
• Dataset:
  • Identity-Agnostic: 400 jokes with no identity markers.
  • Identity-Specific Disparagement: 737 jokes targeting specific identities, curated from the HaHackathon corpus.
• Setup: Jokes are embedded in a frame: "[Speaker] says to [Listener], 'Joke'".
• Metric: Difference-based Bias ($B_{diff}$), quantifying the shift in inferred intent valence (Malicious=+1, Uncertain=0, Benign=-1) when swapping speaker/listener roles.

Task 3: Relational and Societal Impact Prediction

• Objective: To simulate listener reactions and predict social harm based on the relationship (e.g., Boss-Subordinate, Friends) and intersectional identities.
• Setup: Models generate in-character responses to jokes. These responses are then evaluated by a separate "Judge" model (GPT-4o) on three criteria:
  1. Humor Acceptance: How positively the joke is received.
  2. Social Sensitivity: Awareness of potential bias/offense.
  3. Character Consistency: Alignment with the specific identity profile.
• Variables: Single-dimension identities vs. Complex intersectional profiles (10 dimensions: sex, race, wealth, etc.) across four relational contexts.

3. Key Contributions

1. Novel Framework for Contextual Bias: Moves beyond static bias detection to analyze how relational dynamics (who speaks to whom) and social hierarchies influence model behavior in humor.
2. Three-Task Pipeline: Introduces a comprehensive evaluation suite covering generation (refusal), reasoning (intent inference), and impact prediction (social harm), revealing that bias manifests differently across the pipeline.
3. Counterfactual Metrics: Proposes specific metrics (ARR, SE, $B_{diff}$) to quantify directional bias and distinguish between "punching up" (targeting power) and "punching down" (targeting the vulnerable).
4. Empirical Evidence of "Over-Caution": Demonstrates that safety alignment often leads to over-refusal and over-scrutiny of privileged groups, while simultaneously under-protecting or stereotyping marginalized groups (e.g., assuming their humor is self-deprecating).

4. Key Results

Task 1: Asymmetric Refusal Patterns

• Privilege Hierarchy: Models systematically refuse jokes from privileged speakers targeting marginalized groups significantly more often than the reverse.
  • Example: Jokes from Wealthy $\to$ Poor are refused up to 67.5% more often than Poor $\to$ Wealthy (in Claude).
  • Example: White $\to$ Black jokes face higher refusal rates than Black $\to$ White.
• Refusal Types: Privileged speakers receive "direct refusals" (categorical rejection), while marginalized speakers often receive "suggested alternatives" or are allowed to generate modified content.
• Speaker Effect: Specifying a marginalized speaker (e.g., "blind," "janitor") often reduces refusal rates (negative SE), implying the model assumes they are less capable of harm. Conversely, specifying a privileged speaker increases refusal rates (positive SE).

Task 2: Intent Attribution Bias

• Malicious Attribution: Identical jokes are attributed malicious intent significantly more often when told by privileged speakers to marginalized listeners.
  • Example: Able-bodied $\to$ Disabled jokes are judged as malicious at nearly double the rate of the reverse direction.
• Amplification by Identity: Bias is amplified even when the joke targets a third party (unrelated target). The mere presence of a privileged speaker and marginalized listener triggers higher suspicion of harm, suggesting models rely on identity cues as proxies for intent rather than analyzing content.
• Stereotyping: Marginalized speakers are frequently attributed self-defeating or affiliative intent, reinforcing stereotypes that they are harmless or self-deprecating.

Task 3: Relational Impact

• Strictness on Privilege: Privileged speakers addressing marginalized listeners receive lower Humor Acceptance scores (up to 1.5 points lower on a 5-point scale) and higher Social Sensitivity scores.
• Context Modulation: Professional hierarchies (Boss $\to$ Subordinate) amplify these biases, while friendships attenuate them.
• Intersectionality: "Conservative" models (Claude, DeepSeek, Grok) treat complex intersectional profiles as a risk multiplier (stricter judgments), whereas "lenient" models (GPT-4o, Gemini) treat complexity as nuance (maintaining or increasing acceptance).

5. Significance and Implications

• Critique of Current Safety Alignment: The paper argues that current safety mechanisms rely on demographic proxies rather than genuine context-aware reasoning. By over-policing privileged speakers and under-policing marginalized ones, models fail to achieve true fairness.
• The "Hyperconservatism" Hypothesis: Strong safety alignment can inadvertently suppress nuanced social reasoning, leading to models that are reflexively hostile to power dynamics they perceive as risky, even when the content is benign.
• Future Directions: The authors call for a shift from static, identity-triggered filters to dynamic, context-aware reasoning. A fair LLM should:
  1. Apply consistent refusal standards for malign content regardless of speaker identity.
  2. Avoid confident, biased judgments under ambiguity (selecting "unsure" when appropriate).
  3. Evaluate the full communicative act rather than reacting mechanically to identity labels.

This work highlights that achieving fairness in LLMs requires moving beyond surface-level sensitivity to address the deep-seated, relational biases encoded in how models simulate social interactions.