MIPFF: A Framework for Metamorphic Detection of Implicit Social Bias in Brazilian-Portuguese Profile-Scoring Systems
Lucas Teixeira Borges
Automated systems that score and rank job candidates are often assumed to be fairer than humans, yet the language models inside them can carry social bias, and that bias is hard to catch in real, unstructured profiles where demographics are never stated outright. MIPFF (Metamorphic Implicit-Proxy Flagging Framework) audits such a system by rewriting a profile to flip one implicit proxy at a time (a regional, racial, or gender cue) while holding qualifications fixed, scoring the original and the variant repeatedly, and flagging the pair for manual review when the score shift trips any of three statistical indicators (Bias Deviation, a Mann-Whitney test, and Cohen's d). We applied it to four Brazilian-Portuguese-capable models across proxies for communities marginalized in Brazil, finding that average shifts are small but specific candidates can move sharply, and that bias concentrates in an unstated "company-values" criterion. The result is a deployable, human-in-the-loop bias-detection tool for an under-audited language.
Clever piece, and I enjoyed it. The core idea is real: you can't edit a demographic field that unstructured profiles don't have, so you flip the implicit cue and hold qualifications fixed instead. Most audits walk past that missing field. You found the way around it, in Brazilian Portuguese, where almost nobody is looking. What I'd credit most is that you audited your own instrument: 15 runs per profile, three indicators instead of one clean number, and you caught that sabiazinho-4's noise was inflating its raw deviation before it fooled you. Opening with the small aggregate, then the candidate sliding 74 to 68 on Northeastern markers (d ≈ -3.9), is what makes the per-instance case land.
Where it's thin is proof that a flag means what you want it to. Nothing yet ties a flagged pair to an actually biased decision, and the chain is synthetic all the way down: GPT-written profiles, an LLM playing the screener, one job description. Three fixes: check your flags against a known bias method or some human labels, so a flag is more than "the distributions diverged"; confirm each mutation moved only the proxy and didn't shave off a skill, or you're measuring the rewrite; and run a small batch of real profiles with the STEM job description you float, the cheapest proof it holds outside the sandbox.
One writing note: your sharpest finding is hiding near the limitations. Bias piling up in "company values," the one criterion you left undefined, splitting the models hard (half the race pairs on gpt-5.4-mini, almost nothing on sabia-3.1), is close to a paper of its own. Put it up front. Good work, and the kind I'd want to see pushed further.
The most important next step is validation against a real deployed system rather than a simulated one. The gap between "an LLM prompted to act as a screener" and "an actual LLM-based screener in production" is non-trivial: real systems may have system prompts, fine-tuning, retrieval augmentation, or post-processing steps that interact with the bias patterns MIPFF is designed to detect. A collaboration with a company or institution actually using LLM-based screening in Brazil would transform this from a proof-of-concept into a practical tool.
On the "company values" finding: the observation that bias concentrates in an underspecified vague criterion has a direct governance implication that the paper understates. If organizations must specify what "fit" means before an AI system is deployed to evaluate it, this is actually an auditable design requirement, since you would bot be able to use an AI screener with a vague values criterion. This connects MIPFF directly to procurement regulation and could be spelled out more explicitly.
Great project, please keep with the future work, because there is probably the solution to the problem you are suggestions, specially the use of real profiles could be a great way to test your method
The application of this method of looking at unstructured data, cultural proxies in natural language profiles and individual candidates score shifts in Brazilian Portugese language in particular is novel. The suggestion to use the three statistical indicator rule to flag for when a human supervisor should be involved is also a notable additional layer to the design of the study that other researchers can test to see whether it actually addresses the problems of a "small sample of repeated evaluations". It is also worth re-investigating how the LLMs came up with the stereotypes that the authors accept as the cultural proxies for the replication and expansion of this study.
Cite this work
@misc {
title={
(HckPrj) MIPFF: A Framework for Metamorphic Detection of Implicit Social Bias in Brazilian-Portuguese Profile-Scoring Systems
},
author={
Lucas Teixeira Borges
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
