Chapter 06 · Section II · 16 min read
Societal implications of AI
Work, fairness, language, and sovereignty — the four questions Nepal will have to answer in the next decade.
The interesting questions about AI are not technical. They are about who benefits, who is harmed, and who decides. For a country like Nepal, four big questions are coming whether we are ready or not: work, fairness, language, and sovereignty. This section takes them in turn, with one Nepali example for each.
Work: what the day looks like
A common claim is “AI will eliminate jobs.” A more accurate claim is “AI will shift what jobs spend time on.” The difference is not pedantic. Jobs that disappear entirely tend to be narrow tasks easy to automate end to end. Jobs that change are the more common case — the human is still needed, but for different things than they used to do.
Case: a junior lawyer in Kathmandu. In 2024, much of a junior lawyer’s day is contract drafting — taking a template and adapting it to a specific client. A model that drafts contracts well changes this work without erasing it. The lawyer’s day shifts from drafting (now mostly the model’s job) to reviewing the model’s output, advising the client on what to push back on, and handling the cases where the model fails.
The total number of junior lawyers may shrink. The remaining ones are doing higher-judgement work. Their pay may rise; the path to becoming one may narrow. None of this is a “robots take over” story — it is a story about the texture of work changing.
The same pattern repeats across industries that produce a lot of draft outputs: marketing, journalism, accounting, translation, customer support. In each, the human is increasingly an editor rather than a producer. The skills that gain value are judgement and taste. The skills that lose value are routine production.
Fairness: the model only knows what it was shown
Every model is a snapshot of the data it was trained on. If the data systematically under-represents some group, the model will be systematically worse for that group. This is not a flaw; it is a feature of how supervised learning works. But the consequences are very real.
Case: facial recognition at an airport. A vendor sells the government a facial recognition system trained mostly on Indian and Chinese face datasets. Deployed at Tribhuvan International Airport, it works passably on most travellers — and significantly worse on travellers from the Eastern hills and Terai-Madhes communities whose facial features differ from the training-set average. False alarms hit those communities disproportionately. The system is “97% accurate” overall and yet produces a pattern of discrimination on real users.
This is algorithmic bias. It is not because anyone meant to discriminate; it is because the data did not include the deployment population. Nepal, with its profound ethnic and linguistic diversity, is unusually exposed to this failure mode. Any deployed AI system here needs to be audited against Nepal’s actual population, not against the vendor’s average benchmarks.
The cost of getting this wrong is borne by people who already carry historical disadvantage. AI literacy in Nepal includes the right to demand that systems used on people here are evaluated on people here.
Language: a policy choice nobody made explicitly
A government that adopts English-only AI is, in effect, conducting policy in English. The language in which a system operates is a policy choice — even when no policymaker realised they were making it.
Case: a government chatbot. A municipality contracts a vendor to build a chatbot that answers citizen queries. The vendor builds it on a foundation model that works best in English and adds a thin Nepali layer. The Nepali responses are passable. The Maithili responses are not handled at all. The Bhojpuri ones are passed through Hindi translation as an approximation.
Nothing in the procurement said “this system is for English-and-passable-Nepali speakers only.” But that is what the system is. Speakers of other Nepali languages are quietly excluded. The vendor did not intend this. The municipality did not intend this. It happened because nobody insisted, at procurement time, that the system meet a language coverage standard.
Multilingual deployment is a hard problem and it is expensive. But pretending it is solved when it is not is the more expensive choice. The Nepali state has 123 spoken languages on its books. Every AI deployment that touches citizens needs to be honest about which it serves and which it does not.
Sovereignty: where the model runs is where the policy lives
A health diagnosis system run on servers in Singapore is not really under Nepali control. The model’s weights are owned by the vendor. The data sent to it is processed under a foreign jurisdiction. If the vendor turns the system off, the hospital using it has no recourse. If the vendor changes the model’s behaviour, the hospital finds out by surprise.
Case: hospital diagnostic AI. A teaching hospital in Kathmandu adopts an AI system that helps radiologists read X-rays. The system is excellent. It also runs entirely on infrastructure owned by a US company. Two years in, a US sanctions change makes it harder for the vendor to operate in Nepal. The hospital’s diagnostic workflow now depends on a foreign policy decision in Washington. The patients in Kathmandu do not know this.
The sovereignty question is not nationalism. It is operational continuity: who can keep the system running, who can audit it, who is liable when it fails, and what happens if any of those parties decides to walk away. For a country that wants to use AI in critical infrastructure — health, energy, finance, justice — these are first-order questions, not afterthoughts.
The realistic answer is rarely “build everything ourselves.” It is closer to: prefer open-weights models that can run on local hardware; build local capacity to evaluate and operate models; insist on contractual terms that survive vendor change. The point is not autarky. It is not building dependencies you cannot afford to lose.
The Nepali AI literacy demand
Put the four questions together and you have a working definition of what AI literacy demands of Nepal.
- Read articles about AI knowing that the most interesting questions are not “is the model big?” but “for whom does it work, for whom does it fail, and who is accountable when it does?”
- When AI is procured, ask about the training data, the language coverage, the failure modes, and the operational sovereignty before signing.
- Build local capacity to audit, label, and refine systems for the Nepali population. This is not a research project; it is infrastructure.
- Teach the next generation to be editors, judges, and curators of AI output — not just users or producers.
If a generation of Nepalis can do those four things, the country will navigate the AI transition reasonably well. If it cannot, it will become a market for foreign systems with poor local fit. The cost of the second path is much greater than the cost of the first.
Check your understanding
Quick check
—A facial recognition system 'achieves 97% accuracy overall' but performs much worse on travellers from specific ethnic communities. The most accurate explanation is:
What comes next
We close the chapter, and the course, with a short summary section that gives you the one paragraph you should be able to say in your own words about AI. If you can hold that paragraph steady, this course has done its job.