Living with AI — a personal ethics · AI, Ethics, and Society

The previous section ended on a sober note about how slowly policy moves. It is worth being honest about how slowly: a Data Protection Bill that has been in drafting for the better part of a decade will, optimistically, be in force in eighteen months, and meaningfully enforced in three to five years. A procurement standard takes a year from drafting to circular. A sectoral rule, even when the regulator wants to issue it, takes a consultation cycle. Meanwhile, the technology arrives every Tuesday. Between today and the moment the policy machinery catches up, the question of what is acceptable — what gets forwarded, what gets disclosed, what gets refused — will be answered, in practice, by the cumulative small decisions of millions of individuals. The most important regulation of AI in Nepal, for the next several years, will not be the one passed in Singha Durbar. It will be the one each thoughtful person quietly imposes on themselves.

Why personal practice is the load-bearing layer

There is a familiar move in technology debates that treats individual behaviour as too small to matter. The structural problems are structural, the argument goes; demanding that ordinary people behave well in the absence of rules is at best naive, at worst a way of letting institutions off the hook. There is real force in that argument, and this course has spent five chapters honouring it — Nepal needs better rules, real enforcement, sectoral regulation, a working data-protection regime.

But the argument has a quieter side that gets less attention. Norms come before law. The reason it is illegal to drink and drive in Nepal is not that the legislature woke up one morning with the idea; it is that, over decades, enough individuals refused enough drinks to make the prohibition socially possible to enforce. The reason it is socially unacceptable in most newsrooms to publish a private photograph without consent is not the Personal Privacy Act 2018; it is the cumulative editorial decisions of two generations of journalists who decided, individually, that they would not. By the time the law arrived, the practice had largely formed. The law ratified a norm; it did not create one.

AI is now in the same place. What it is acceptable to forward, generate, disclose, deploy, and refuse will be settled, in Nepal, long before the law has anything specific to say. The individuals making the settling are everyone reading this. Five commitments, none of them heroic, would meaningfully shift the trajectory.

One — verify before forwarding

The most common harmful AI interaction in Nepal today is not a deepfake; it is a forwarded WhatsApp message with a synthetic image attached, propagated by people who did not mean any harm and would have stopped had they paused. Treat any AI-generated or AI-suspect content as a draft, not a verdict. Before forwarding a striking photo, a damning voice clip, an alarming statistic, ask the small set of questions a working journalist would ask: where did this originate, who first published it, does a reverse image search show the same image dated earlier or in a different context, is the voice consistent with the public record, is the claim independently reported by an outlet that would have something to lose if it were wrong. Most of the time, ninety seconds of looking will resolve it. The remaining cases — the genuinely ambiguous ones — are the cases where forwarding is least defensible, because you are passing on something you yourself do not know is true.

Two — disclose when AI materially shaped what you publish

The norm forming around AI use in writing, design, code, and analysis is the norm that will decide whether trust survives the next decade. A useful, low-friction rule: if AI materially shaped what you publish or submit — drafted the argument, produced the image, wrote the code, generated the data summary — say so, briefly, where the reader can see. Not as an apology. As a fact about how the work was made. “Drafted with AI assistance and then edited and fact-checked by me” is a one-line disclosure that costs nothing and tells the reader exactly what they need to know. The version of this norm that fails is the version in which disclosure is optional and inconvenient enough that the people who skip it gain a quiet advantage over the people who do not. The version that succeeds is the version in which disclosure becomes a small badge of seriousness, the way citing your sources is a badge of seriousness in a paper. We are, collectively, choosing which version this becomes. Each individual choice contributes.

Three — push your institution to audit

Most readers of this course are, or will be, embedded inside an institution — a bank, a hospital, a school, a ministry, an NGO, a tech firm. Most of those institutions are quietly deploying or considering AI systems that touch people’s lives. Almost none of them are auditing those systems with the seriousness the deployment warrants. You do not need to be the chief technology officer to change this. A well-prepared internal note from a mid-level employee — “we are about to deploy X, here are three questions I have not seen answered, here is the cost of getting it wrong” — is, in most Nepali institutions, the most influential intervention on a deployment decision that has happened. Institutions respond to internal pressure that comes with specifics. The specifics are within reach of anyone who has finished this course.

Four — protect those who cannot protect themselves

The harms surveyed in this course are not evenly distributed. They fall hardest on those who cannot push back: elderly relatives who do not know that a familiar voice on the phone can be synthesised; students whose schools are deploying AI grading without explanation or appeal; workers whose performance is monitored by systems they have not been shown; women whose images are being used to train models they did not consent to; daily-wage labourers whose creditworthiness is now being scored by a model that has never seen anyone like them perform well. Personal ethics that stops at the boundary of personal benefit is a thin ethics. The richer version asks, regularly, who in my immediate radius is exposed to this, and what can I do for them today. Sometimes it is a five-minute conversation with a parent about a scam pattern. Sometimes it is filing a grievance on behalf of a colleague. Sometimes it is teaching a student how to ask whether an AI system has the right to make the decision it just made about them. The cumulative weight of these small acts is, in practice, a substantial portion of the protection most vulnerable Nepalis will actually receive.

Five — refuse the most extractive products

There is a class of AI product whose business model is, fundamentally, the extraction of attention, dignity, or data from people who do not understand the trade they are making. Surveillance-by-default consumer apps. Deepfake-as-a-service tools whose primary documented use is the production of non-consensual intimate imagery. “Companion” products marketed to lonely teenagers. Engagement-maximising recommender systems that have measurably harmed adolescent mental health in every jurisdiction that has bothered to study them. Convenience is real. So is the cost. A personal ethics worth having declines to be a paying customer for the most extractive products on the menu, even when they are convenient, and especially when the cost is being paid by people who are not in the room.

The course ends, the choices continue

This is the last section of the course. Six chapters ago, we began with the claim that ethics is mostly present-tense and bureaucratic — not a debate about runaway superintelligence, but a series of specific, ordinary decisions being made right now about who gets credit, who gets watched, who gets believed, and who gets to write the rules. We have walked through the bias that hides inside headline accuracy, the privacy regime that defaults to extraction unless someone insists otherwise, the labour market that is being quietly reshuffled while the conversation focuses elsewhere, the trust crisis that is arriving in Nepali public life faster than any of our institutions are prepared for, and the governance window that is closing while we look at it.

What we have not done — what no course can do — is decide which side of any of this you will stand on when it touches you. That part is not in the syllabus. It is in the inbox, the WhatsApp forward, the procurement meeting, the family conversation, the moment in which a tool offers you a small convenience in exchange for somebody else’s dignity. Those moments will continue arriving long after the course closes. The hope of this course is not that you remember every claim it made. The hope is that, the next time one of those moments arrives, you notice it as a moment, and choose deliberately.

What comes next

This is the end of the chapter readings. The final exam for the course, AI, Ethics, and Society, is now available at /courses/ai-ethics-society/exam. It draws on the full six chapters — why ethics now, bias and fairness, privacy and data sovereignty, work and the economy, truth and media, and governance — through ten multiple-choice questions. The pass threshold is eighty percent, and a course certificate is issued on a passing attempt. Take it when you are ready; the material does not expire, and neither do the choices it has tried to put in front of you.