Blind screening feels like the responsible fix for hiring bias. The research shows models reconstruct race, gender, and class from the markers redaction leaves behind. The bias does not disappear. It goes underground.
There is a fix for hiring bias that sounds airtight. Strip the names. Hide the photos. Redact the dates and the addresses. Let the qualifications speak for themselves. Blind screening.
It feels like the responsible move, and plenty of teams have adopted it believing the problem is handled. It is not. The model reconstructs almost everything you tried to hide.
What survives the redactionProxies remain
A 2025 study set out to test exactly this: take resumes, remove the names and obvious identifiers, and see whether a language model still treats candidates differently. It does.
The models inferred race, gender, and socioeconomic class from the residue that anonymization leaves behind, the subtle sociocultural markers that no redaction tool flags: the languages a candidate lists, the activities and sports, the volunteering, the signal in a school or organization name. None of that is a protected characteristic on its face. All of it lets a model guess one (Small Changes, Big Impact, 2025).
So blind screening does not remove the bias. It hides the mechanism, which is worse, because now the discrimination runs through signals you are not even monitoring.
Why subtraction cannot get you to fairness
The instinct behind anonymization is that bias attaches to a few identifiable fields, and if you delete those fields, you delete the bias. But a resume is dense with correlated information. Almost any detail that tells you something useful about a candidate also leaks something about who they are.
You cannot redact your way around that without redacting the resume into uselessness. The proxies are not a bug in your anonymization, they are the document.
And the underlying model is already skewedBaseline bias
Remember what these models do before you even hand them an anonymized resume. In the largest audit of its kind, three production language models ranking real resumes preferred White-associated names in 85.1% of cases and female-associated names in only 11.1%. Names associated with Black men were never preferred over white-male names (Wilson & Caliskan, 2024).
Anonymization was supposed to neutralize that. Instead the model simply switches from the name it can no longer see to the proxies it still can. Same bias, quieter route.
What actually moves the needle
If you cannot subtract identity out of a resume, the answer is to stop betting the decision on the resume. A document is a record of claims and credentials, thick with the very markers that leak bias. It is not a record of what the person can do.
The durable move is to shift the weight of the decision onto evidence of demonstrated competence, what a candidate actually showed in a structured, human evaluation, rather than what a model inferred from a PDF. You are not trying to hide who someone is. You are trying to measure what they did.
So, does anonymizing resumes remove bias?
No. Models infer the traits you redacted from the sociocultural markers you cannot, so the bias goes underground rather than away, on top of a baseline that already favored some names over others.
You cannot subtract your way to fairness. You can only move the decision onto evidence of what the candidate demonstrated.
You cannot subtract identity out of a resume. Confeti moves the decision onto evidence of demonstrated competence from the interview, structured, comparable, and traceable, rather than signals inferred from a document.
See how it worksCommon questions
Does anonymizing resumes remove hiring bias?+
No. A 2025 study found models still infer race, gender, and class from sociocultural markers, languages, activities, volunteering, that survive removing names and identifiers.
What is a sociocultural proxy on a resume?+
A neutral-looking detail that correlates with a protected trait, such as listed languages, sports and activities, volunteering, or the name of a school or organization. Models use these to infer identity.
Isn't blind screening still better than nothing?+
Reducing obvious identifiers is a reasonable thin layer and removing a name won't make bias worse. The error is believing anonymization makes bias go away, when capable models see through it.
Why does the underlying model already matter?+
Because the base model is skewed before anonymization. In the largest audit, models preferred White-associated names in 85.1% of cases and never preferred Black-male names over white-male ones.
What actually reduces bias if anonymization doesn't?+
Shifting the decision weight onto evidence of demonstrated competence from a structured human evaluation, rather than traits inferred from a resume document.
References
- Small Changes, Big Impact: Demographic Bias in LLM-Based Hiring Through Subtle Sociocultural Markers in Anonymised Resumes (2025). arXiv. link
- Wilson, K., & Caliskan, A. (2024). Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), 1578–1590. link