Why "we removed the name" does not solve bias

The mechanism — proxies are how protected information travels

A CV is not a database with separate fields for “name” and “merit.” It is a piece of writing in which every choice — what school is named, what neighbourhood appears in the address, what extracurriculars are listed, what English-language patterns the candidate uses — carries information about who the candidate is and where they come from. A machine-learning model is, in essence, a very fast pattern-finder rewarded for predicting an outcome. If the outcome it is asked to predict correlates even slightly with gender, caste, or region — and in any pipeline shaped by past human decisions, it does — the model will find the patterns in the remaining columns that approximate the redacted ones. It does not need the name. It has the address, the school, the surname mentioned in a project description, the phrasing of the personal statement, the list of organisations the candidate has volunteered with. Any of these can stand in.

This is not a failure of careful engineering. It is what the model is for. A model rewarded for predicting outcomes is also rewarded for finding shortcuts to that prediction. If the shortcut runs through a proxy for a protected attribute, the model will take the shortcut, and the shortcut will produce decisions that look statistically similar to having the protected attribute itself. The mechanism is mathematically inescapable so long as the proxies remain and the model is asked to predict outcomes that were historically shaped by the attribute. Removing the column does not remove the information; it only removes the audit trail.

A simple test. Take a real CV from your last screening cycle. Black out the name. Ask a human reviewer — not a model, a person — whether they can guess the candidate’s likely gender, likely caste group, likely class background, likely province of origin, from what remains. In a Nepali CV, the honest answer is almost always yes, often within a confidence range that would distress the people who designed the redaction. If a competent human can guess from the residue, so can a large language model — only faster, more confidently, and across more CVs.

The Nepal-specific proxy map

The proxy problem is universal; the specific proxies are local. A defensible HR team needs to know which features in a Nepali CV carry which protected information, because those are the features your audit needs to watch and your rubric needs to handle carefully.

Surname carries caste. This is the most direct proxy and the hardest to redact. Even if you blank the name field, surnames appear in project descriptions, reference contacts, family-business mentions, and quoted publications. A pattern-finding model that has seen the Nepali surname distribution — and every major model has — can categorise a surname into a likely caste group with high accuracy. The same is true for many ethnic surnames carrying region and language community information.

Neighbourhood and address carry caste, class, and region together. A Kathmandu address in Baluwatar means something different from one in Balaju, and both mean something different from a Janakpur or Dhangadhi address. The municipality, the ward, the postal code, the school listed nearby — each is a fingerprint of the candidate’s economic and social background. A model that has read enough Nepali web text knows the geography well enough to use it as a feature.

School and university name carry class and language background. An English-medium A-Level school in Kathmandu marks a candidate differently from a community Nepali-medium school in Surkhet. A degree from Kathmandu University is read by the model differently from a degree from a Tribhuvan-affiliated rural campus, regardless of the candidate’s actual performance. The school is not the candidate’s competence; it is a marker of the family resources that opened a particular door at age fifteen.

Extracurriculars and clubs carry gender and class. Cricket and football are more often boys’ activities in Nepali school culture; dance and debate cluster differently. Rotary Interact suggests one kind of school; a volunteer year with a rural NGO suggests another. None of these are deterministic; all of them shift the model’s prediction in a direction that correlates with the protected attribute.

Language patterns in the CV itself carry class and education background. Sentence length, formality register, frequency of certain Nepali-English code-switches, the kind of project verbs used — implemented versus built versus helped with — all map onto the candidate’s English-language exposure, which maps onto schooling, which maps onto class. The model does not need to be told the candidate’s class; it can read it in the prose.

What actually helps — and what does not

If redaction does not work, what does? The honest answer is a system of practices, none of which is sufficient on its own, all of which are necessary in combination. They are the practices the last two sections built up to, and the proxy problem is the underlying reason for each.

The outcome audit (Section 2) is the load-bearing intervention. Because protected information travels through proxies you cannot fully remove, the only measurement of fairness that holds up is the measurement of outcomes by group. The audit does not need to know how the bias got in; it only needs to detect that the outcomes are uneven, and then trigger the investigation. This is why the audit is the highest-leverage habit in the chapter.

The written rubric (Section 1) constrains what the model is allowed to weight. A rubric that says, in writing, we are scoring on these three specific must-haves, with these specific evidence types, narrows the model’s discretion. It cannot stop the model from using proxies, but it can stop the model from inventing new criteria on the fly that happen to correlate with school or surname. The rubric is the first ring of defence; the audit is the second.

The sample-from-bottom review (Section 1) catches systematic proxy-driven errors at the rejected stage. If the model is silently dropping a category of candidate — say, every CV in mixed Nepali-English phrasing — the sampled rejections will reveal it within a cycle or two. Without the sampling, the systematic error is invisible.

Calibration of human reviewers is the fourth piece, and it deserves more attention than it gets. Humans use the same proxies the model does. A two-hour calibration session where reviewers score the same set of CVs independently and discuss the divergences typically reveals reviewer-level patterns the team did not know it had. The model is not the only source of bias in the funnel; the audit will not always find the model at fault, and the calibration is how the reviewer-level fix gets made.

What does not help: redaction alone, vendor fairness certifications without an audit you can re-run, “blind” screening that strips only the obvious fields, and the assumption that a model trained on a global pool will be neutral on a Nepali one. Each of these is a comfort, not a control.

The honest summary

The proxy problem is not a temporary engineering limitation that better tools will solve in two years. It is a structural feature of how predictive models work in any domain where the outcome being predicted has a history shaped by the attributes you are trying to be fair on. In Nepali hiring — where caste, gender, region, and class have shaped opportunity for generations — those attributes have left their signature across every column of the candidate’s resume that the model can read. The signature does not come off because you blanked one field.

This is not a counsel of despair. It is a counsel of seriousness. The path to defensible AI-assisted screening is not the search for a clever redaction scheme that finally works. It is the unglamorous discipline of writing the rubric, running the cycle with humans on every rejection, sampling the rejected pile, and auditing the outcomes every quarter. That discipline does not make bias impossible; nothing does. It makes bias detectable, addressable, and, when challenged, answerable. That is the honest professional standard, and it is the one the next few years of Nepali HR practice will be measured against.

Check your understanding

Quick check

—

True or false: removing names, photos, and addresses from CVs before AI screening makes the screening bias-free.

1. ATrue — without those fields, the model cannot know the candidate’s identity.
2. BFalse — protected attributes leak through proxies such as surname mentions in project text, school and university names, neighbourhood-fingerprinted addresses, extracurricular patterns, and the writing style of the CV itself; the model follows the proxies regardless of which columns are removed.
3. CTrue, but only if photos are also removed.
4. DFalse — but only because addresses are too hard to fully remove.

Check answer→

What comes next

That closes Chapter 3, the highest-risk chapter of the course. You now have a workflow for a single cycle, an audit habit for the pipeline over time, and a clear mechanism-level understanding of why the most popular bias defence does not work. Chapter 4 moves from screening to interviewing — where AI is finding new uses in question generation, transcript summarisation, and structured scoring, and where a different set of risks (recorded conversations, candidate-side AI use, asymmetric power in the room) needs the same kind of honest treatment.