The fairness stakes — and why HR cannot treat AI like Marketing does

Two functions, two threat models

Marketing and HR look superficially similar from a distance — both produce a lot of text, both run campaigns, both use templates, both have started using AI for first drafts. The threat model is almost completely different.

Marketing’s worst plausible AI error is a clumsy line in a campaign, an off-brand image, a tagline that translates poorly, a generic email that goes to a list that deserved more thought. Embarrassing, sometimes costly, almost always recoverable. The audience is plural and largely anonymous. The fix is a corrected post and, occasionally, an apology.

HR’s worst plausible AI error is a single qualified candidate quietly filtered out of a shortlist, a performance rating that drops a woman a salary band she earned, a termination letter that misquotes the Labour Act, a screening tool that systematically prefers Kathmandu Valley CVs because that is what its training data over-represents. The audience is one person at a time, and that person is identifiable, present, and harmed. The fix may not exist; some of these errors only surface in a labour court six months later, by which time the harmed candidate has moved on, the firm’s reputation has moved on, and the only outcome on offer is a settlement, a fine, and a quiet rewrite of the policy.

Same technology. Two threat models. Borrowing Marketing’s risk posture for HR work is not “moving fast” — it is mispricing the downside by an order of magnitude.

What the law actually says

This is not a chapter on Nepali labour law, but three constraints have to sit on every HR professional’s desk before any AI conversation starts.

The Constitution, Article 18. The right to equality is a fundamental right. Sub-clauses prohibit discrimination by the State and by private actors on grounds of origin, religion, race, caste, tribe, sex, language, ideology, or “any other such grounds.” A hiring system that produces disparate outcomes by any of these — even unintentionally, even because the model inherited it — is a constitutional problem, not just an HR-policy problem. “The vendor said it was bias-free” is not a defence the Court has to accept.

The Labour Act 2074, with its rules and amendments. The Act sets out who must be hired in compliance with what, what notice and dues are owed at exit, what discrimination in the workplace looks like, and what an aggrieved worker’s remedies are. The relevant provisions for AI-era HR include — but are not limited to — the prohibition on discrimination in recruitment and employment conditions, the requirements around grievance handling, and the procedural rules for termination. A model-generated decision that bypasses required procedure is not a faster decision; it is a defective one.

Sector-specific rules. Banking, INGOs, government-contractor firms, BPOs serving GDPR-regulated clients — each carries an additional layer. NRB has views on bank hiring. Donor organisations have safeguarding requirements. International clients of Nepali BPOs increasingly ask whether AI tools used in HR decisions have been audited. None of this is satisfied by a chatbot that “seems fair.”

The Nepali context makes the stakes sharper

Borrowed AI playbooks usually arrive from US or European HR conversations, where the dominant fairness concerns are gender and ethnicity in a specific Western framing. Those concerns apply here too, but Nepal adds its own.

Gender bias in technical hiring. The pool of women applying for engineering and product roles in Nepal is smaller than the pool of men, for well-documented social reasons. A model trained on global hiring data learns that “successful engineer” correlates with male names, male-coded extracurriculars, and male-coded language. Run that model on a Nepali engineering shortlist and you do not get a neutral filter; you get the inherited bias of the training data, projected onto a pool where women already face more friction. The damage is silent and cumulative.

Caste-name screening risk. Nepali surnames carry caste information in a way that surnames in many other countries do not. A model that ranks CVs and has, somewhere in its training data, an association between certain surnames and lower-status occupations will reproduce that association. The model will not announce it is doing so. The recruiter will see a ranked list and trust it. This is exactly the silent harm Article 18 was written against.

Regional bias from Bagmati-centric data. Most digitised CVs, job postings, training data, and HR analytics in Nepal come from Kathmandu Valley. A strong programme-officer candidate from Surkhet with a Tribhuvan-University-affiliated college name the model has never seen, work experience at municipalities the model has never heard of, and references at Karnali-based NGOs the model has no record of, will be ranked lower than a Kathmandu candidate with the same competencies and a more recognisable paper trail. This is not a bug; it is the model doing exactly what it was trained to do. The fairness implication is severe in a country whose Constitution makes inclusive federal hiring an explicit obligation.

Linguistic and educational bias. A Nepali-medium degree from Far-West is, in a strict sense, equivalent to an English-medium degree from a Kathmandu private college for many roles. A model trained on global hiring patterns has no way to know this and tends to score the English-medium CV higher. Unless the firm is alert to this, the model is quietly making an educational-language bias decision the firm itself would never make if a human reviewer was reading the CVs.

The principle: fairness must be measured to be maintained

Sit with the three legal constraints and the four Nepali-context biases for a minute and one principle falls out. You cannot maintain fairness in an AI-assisted HR system by hoping it is fair. Fairness in any system that affects people’s livelihoods must be measured — periodically, with real data, against the same outcomes a non-AI process would have produced. If you do not measure, you do not know. If you do not know, you have not maintained fairness; you have just decided not to look.

This is the bridge to Chapter 3 of this course, which is about exactly that — what to measure, how often, who reviews the measurement, and what to do when the measurement shows the system has drifted. For now, the practical implication for Chapter 1 is simpler. Every AI use case in HR comes with an obligation the same use case in Marketing does not. Marketing’s tools can be evaluated on speed and cost. HR’s tools must additionally be evaluated on whether the decisions they shape are defensible against the Constitution, the Labour Act, and the simple test of whether you could explain, to the candidate’s face, why the system handled her the way it did.

The teams that get this right do not slow down. They go just as fast on the safe parts of the workflow — JDs, outreach, summarisation, translation, drafting — and they keep humans firmly in the decisional loop. The teams that get this wrong will, sooner or later, end up in a conversation with the Labour Court, a board investigation, or a journalist with a story about an algorithmic mis-hire. The professional choice is which conversation you would rather be in.

Check your understanding

Quick check

—

True or false: HR can safely treat AI tooling the same way Marketing does, because both functions produce a high volume of templated content and the underlying technology is the same.

1. ATrue — the technology is identical and content volume is similar, so the playbook should be too.
2. BTrue — HR is just a back-office function and the legal exposure is comparable to Marketing’s brand exposure.
3. CFalse — HR’s worst error is a specific identifiable harm to a person, on grounds the Constitution and Labour Act 2074 treat as serious, so HR’s defaults must be stricter than Marketing’s regardless of shared technology.
4. DFalse — the technology is actually different; HR uses entirely separate models from Marketing.

Check answer→

What comes next

Chapter 1 has set the frame: where AI fits in HR work, where in the monthly workflow the hours actually move, and why the fairness stakes punish carelessness in a way they do not in other functions. Chapter 2 turns to the first concrete pipeline most HR teams will touch — job descriptions and sourcing — and walks through the practical recipes for using AI on those tasks well, including the prompts, the review steps, and the failure modes to watch for as you scale up.