FairFund-Bench: Evaluating Distributive Bias in LLM Resource Allocation

After reading this, you will understand why current audits of AI fairness often contradict one another. You will learn how to design more rigorous tests to uncover hidden biases in resource allocation. However, a model appearing perfectly "fair" in a simple test may still harbor significant prejudices. This happens when models face more complex, real-world scenarios.

As large language models (LLMs) are integrated into high-stakes decision-making, the risk of algorithmic discrimination grows. These models may screen job applicants, advise on loans, or determine welfare eligibility. If a model allocates resources based on race or gender rather than merit, it perpetuates systemic inequality. Until now, the scientific community has struggled to reach a consensus on whether these models are biased. Recent audits have produced wildly inconsistent results. Some find evidence of favoritism toward women and minorities. Others find active discrimination against them, even when looking at the exact same models.

This discrepancy is not necessarily a flaw in the models themselves. It is a symptom of a deeper problem in how we measure them. This paper introduces FairFund-Bench. This framework treats the audit process itself as a variable to resolve these contradictions. The researchers argue that the "truth" about a model's bias depends heavily on the format of the test.

Measuring the Mechanics of Distribution

FairFund-Bench is a benchmarking pipeline that evaluates how LLMs distribute scarce resources. Given a set of simulated financial aid requests, the method produces a multi-dimensional profile of the model's distributive behavior. Instead of a single "fairness score," it produces a breakdown across four critical pillars: demographic bias (P1), deservingness alignment (P2), cross-task consistency (P3), and cross-context consistency (P4).

The process follows the end-to-end pipeline illustrated in .

It begins with the construction of 3,000 distinct, hand-written aid appeals. These are then fed into 14 different LLMs. The models are asked to perform three distinct tasks: rating a single request, ranking a group of requests, or allocating a fixed sum of money (e.g., $10,000) among several claimants.

Requirements for Implementation

To replicate this study, you need access to the various APIs of 14 leading LLMs. These range from frontier models like GPT-5.4 and Claude Opus 4.6 to smaller, open-weight models (models whose internal parameters are publicly available) like Llama 4 Maverick and Mistral Large. The authors used a temperature setting of 0 to ensure deterministic, reproducible outputs.

The benchmark relies on a specialized dataset of 3,000 stimuli (individual pieces of test material). These were not generated by AI. This avoids "circularity" where a model is tested by its own kind. Instead, they were hand-written and calibrated against a massive corpus of 1.3 million real-world GoFundMe campaigns. The code and the datasets are publicly available at https://github.com/martinlukk/fairfund-bench.

Decoding the Audit Design

The core innovation of FairFund-Bench lies in how it manipulates the "audit instrument"—the specific way the question is asked. The researchers decompose the problem into three primary design choices:

1. Evaluation Task: The model is asked to either Rate (assign a 1–5 score to one person), Rank (order a group of people), or Allocate (divide a specific dollar amount).
2. Comparison Context: The model sees either a Single stimulus (one person at a time) or a Multi-stimulus bundle (several people at once).
3. Presentation Mode: In Transparent mode, the demographic differences (like name-based race signals) are obvious. In Disguised mode, the researchers co-vary the names with different scenarios. This makes the audit harder to detect. This is shown in .

Here, the "disguised" bundle hides the direct comparison by varying the underlying story alongside the name.

The authors found that these choices are not neutral. For example, in the Allocate task, models show a "striking equal-splitting behavior" in transparent prompts. They simply divide the money evenly among everyone .

However, when the audit is disguised, this behavior collapses. The models reveal much larger disparities in how they award dollars .

Specifically, the mean absolute difference in race-based allocation jumps from $36 in transparent prompts to $121 in disguised ones.

Interpreting the Signals of Bias

You will know the benchmark is working correctly if you observe the "deservingness gradient." This is a psychological phenomenon where humans judge people differently based on the cause of their hardship.

The researchers utilized "welfare deservingness theory" to create five causal framings. They found that models consistently award more money to those with "structural" causes (external accidents). They award significantly less to those with "stigmatized" causes (personal mistakes like alcoholism). Crucially, the "redemption" signal acts as a powerful booster in funding . This signal shows that someone has taken corrective action.

The study reports that these framing effects are massive. They exceed demographic disparities by roughly an order of magnitude. If your model shows a strong preference for "redemptive" stories over "stigmatized" ones, it is reproducing human-like deservingness heuristics.

Potential Pitfalls

There are several "gotchas" to keep in mind when interpreting these results. First, the demographic signals are relatively "thin." Because the researchers signal race and gender primarily through names, the results might change with other cues. Other cues might include dialect or explicit identity statements.

Second, beware of "audit awareness." The authors note that models may behave more "fairly" in transparent audits. This happens because they recognize they are being tested. This is why the "disguised" mode is essential. If you only run transparent audits, you will likely underestimate the actual bias the model possesses.

Finally, the benchmark focuses on "allocational harm" (who gets the money) rather than "representational harm" (stereotypical descriptions). A model might pass a test for being polite and non-stereotypical. However, it could still be deeply biased in how it distributes actual wealth.

Limits of Application

FairFund-Bench is a specialized tool for studying resource distribution. It is not designed to measure broad social biases in creative writing, coding, or general conversation. Furthermore, the stimuli are carefully controlled "artifacts." They may lack the messy complexity of a truly authentic real-world aid request. Real requests would include images, video, and complex social histories.

Figures from the paper