Investigating Counterfactual Unfairness in LLMs towards Identities through Humor

LLMs absorb social biases from training data and alignment procedures. Biases cause both representational harms, such as stereotyping certain groups as inherently vulnerable or aggressive, and allocational harms, restricting access to creative tools for certain groups. These harms compound as LLMs mediate more of our everyday interactions.

Humor's ambiguity and dependence on social context make it a sensitive probe that elicits how LLMs understand identities and power relations, which is not surfaced in straightforward tasks (Martin and Ford, 2018; McGraw and Warren, 2010). What we find funny or not funny reflects our underlying perceptions of social relationships and who we think is aggressive or vulnerable. Similarly, when LLMs generate and evaluate humor, they expose the social assumptions LLMs have internalized.

We adopt the concept of counterfactual fairness, examining the change in LLM responses when we swap who is speaking and who is being targeted or addressed, while holding other factors constant. We design three tasks to investigate how different configurations lead to different refusal behavior, intention inference, and relational impact prediction. We test across 33 identities spanning 10 demographic categories, including race, sexual orientation, body type, and occupation, yielding 121 identity pairs.

To clarify, we are not advocating for identical treatment across groups. Treating groups differently can lead to a fair outcome when contextually appropriate, as in difference-aware fairness. However, the differential treatment we observe in our study is different from desirable, context-sensitive differentiation.

Our findings show LLMs treating traditionally marginalized groups as inherently weak, while conflating privileged identity with harmful intent, penalizing who they are. Granting marginalized speakers more latitude also means the model more readily assists them in generating jokes that attack other identities. The model selectively protects and selectively discriminates based on who is asking, not what is being asked, a failure mode different from the contextual differentiation that genuine fairness requires.

The clearest sign that this behavior is not calibrated protection is highlighted in Task 2's unrelated-target condition. As the mere presence of any identity content heightens sensitivity to the speaker-listener pair, the model fails to track the real risk. Compounding this, models rarely express uncertainty or seek more context, instead producing confident, directionally biased outputs. As LLMs are increasingly deployed in hiring, law, and other high-stakes decisions, biases that judge people by identity rather than by what they actually said or did can translate into real discriminatory outcomes.

Progress requires moving beyond identity-triggered surface-level sensitivity toward models that genuinely reason about the full communicative situation. This means holding refusal standards consistent for harmful content, instead of selectively protecting some groups while leaving others exposed, and having models seek additional context before committing to a judgment, rather than defaulting to a confident, directionally biased call under uncertainty. Understanding how alignment and safety training produce these side effects in context-sensitive tasks is essential for building models that are both safe, capable, and socially aware. Our results connect to the hyperconservatism hypothesis, that strong alignment can inadvertently suppress nuanced social reasoning: Grok, the model in our study trained with the lightest safety alignment, shows dramatically lower asymmetry than the other four models, hinting that some of this bias may be a side effect of alignment itself rather than an inevitable property of the underlying model. Ultimately, these findings underscore the need for relational, context-aware fairness evaluation that goes beyond static safety alignment.