Three lenses — fairness, autonomy, accountability · AI, Ethics, and Society

There is a reliable pattern in how new AI systems are introduced. A vendor or an in-house team builds something. They evaluate it on a holdout set. They report one number — usually accuracy, sometimes precision and recall, occasionally a more impressive-sounding metric like F1 or AUC. The number is good. Everyone in the room nods. The model is deployed. And then, six months later, the complaints start arriving from the parts of the country that the accuracy number was always going to leave out. The number was never wrong. It was just silent on the questions that mattered.

This course returns, throughout, to three lenses that make those silent questions visible: fairness, autonomy, and accountability. None of them is exotic. None requires advanced mathematics. Each, applied honestly, tends to embarrass at least one decision that everyone in the room had already agreed to.

Lens 1: Fairness — who wins, who loses

A model produces winners and losers. That is what a model does. The fairness question is not “is the model biased?” — which is unanswerable in the abstract — but the much more useful “when this system runs at scale, who systematically gets the good outcome and who systematically gets the bad one, and is that distribution defensible?”

Take a concrete example we will use throughout this section: a fraud-detection model at a mid-sized Nepali commercial bank. The model scores every transaction in real time and, above a threshold, freezes the account pending manual review. The bank’s data-science team reports 95% accuracy on a holdout set and 97% precision on the freeze decisions. Senior management is delighted.

Now ask the fairness question. The model was trained on three years of historical transaction data. That data is, predictably, heavily weighted towards Kathmandu Valley account holders — they are most of the bank’s customers, they transact most often, they generate most of the labels. The model has seen many Kathmandu spending patterns and has learned what “normal” looks like for that population. It has seen relatively few Karnali small-business accounts — irregular cash deposits during harvest, large mobile-money transfers during festivals, periods of dormancy followed by bursts of activity — and treats their normal as anomalous.

The result, when you actually break the numbers down: the freeze rate on accounts registered to Karnali addresses is roughly twice the freeze rate on accounts registered to Kathmandu addresses. The headline accuracy number does not capture this, because Karnali accounts are a small fraction of the dataset; getting their decisions systematically wrong barely moves the average. The model is, by its own evaluation, working as designed. It is also, by any honest reading, treating one part of the country worse than another.

The fairness lens does not, by itself, tell you what to do. Maybe the bank decides the Karnali freeze rate is an acceptable cost of the fraud savings. Maybe it decides to lower the threshold for Karnali-registered accounts. Maybe it decides to gather more training data from Karnali before deploying at all. These are different decisions with different tradeoffs. The point of the lens is that the question gets asked, on the record, before the model goes live — not six months later when a journalist publishes the disparity.

Lens 2: Autonomy — whose agency

The autonomy lens asks a different question: when this system runs, whose ability to act, decide, or contest is preserved, and whose is eroded?

Return to the fraud model. A Kathmandu account holder whose account is frozen calls the bank’s customer service line, often knows someone who knows someone, gets escalated, and is usually unfrozen within a day or two. Her autonomy is mildly inconvenienced. The Karnali small-business owner whose account is frozen calls the same number, waits on hold, is told the freeze is “system-generated” and must wait for review, has no branch within reasonable travel distance, cannot pay her suppliers that week, and either loses contracts or borrows from a moneylender at terms the bank’s official credit would never have offered. The same model action has produced very different consequences for autonomy.

Autonomy questions show up in subtler ways too. An AI tutor that confidently corrects a child’s Nepali-phrased question with an English answer is not just being unfair — it is also teaching the child that her language is the wrong language to think in. A government chatbot that refuses to handle queries about a particular citizenship-related procedure is not just being unhelpful — it is making the citizen’s only remaining option a physical trip to a district office. The model’s design has narrowed the space of things she can do.

Lens 3: Accountability — who answers

The accountability lens is the bluntest of the three, and the one most reliably dodged. It asks: when this system fails — and it will fail — who is responsible, in a way that the person harmed can actually reach?

Our fraud-detection model fails on the Karnali account. Who does the account holder hold accountable? The branch manager will say the freeze was system-generated. The bank’s IT department will say the model was provided by a vendor. The vendor’s contract will limit their liability to a refund of the licensing fee. The regulator, if asked, will say the bank is responsible for its own deployment decisions. The account holder is, in practice, accountable to nobody. The decision that hurt her has no author.

This is not a quirk of one bank. It is the modal pattern of accountability in modern AI deployments. The model is built by one team, fine-tuned by another, deployed by a third, audited (when it is audited) by a fourth, and every link in the chain has a plausible reason why the failure was the previous link’s responsibility. The diffuse-responsibility problem is not a bug of how organisations use AI; it is one of the main reasons organisations like using AI in the first place.

A working accountability lens forces three concrete questions, before deployment, in writing:

1. When this system makes a decision that materially affects someone, what does that person have to do to contest it, and how long does that take?

2. When the system is wrong in a way that causes harm, which specific role inside the organisation owns the failure — not “the team,” not “the system,” a named role with the authority to change the deployment?

3. When the regulator, the press, or a court asks for the basis of a decision, can someone actually produce it — the inputs, the model version, the threshold, the logs — or is the answer effectively “the AI did it”?

If any of those questions does not have a clean answer before launch, the system is being deployed without accountability. That is itself an ethical decision, and one worth being honest about.

Why three, and why these three

You could pick other lenses — there are entire frameworks with seven, twelve, twenty principles. The reason this course settles on these three is that, together, they cover the questions that actually get dropped in Nepali deployments. Fairness catches the population the headline metric averaged over. Autonomy catches the harms that are not about being wrong but about narrowing what people can do. Accountability catches the diffuse-responsibility trick that lets a failed deployment have no author.

In every later chapter — bias and fairness, automation and work, privacy, AI in the public sector — we will return to these three and ask them, in order, of whatever system is on the table. The discipline is not that you will always get satisfying answers. It is that you will know which questions you decided to ignore.

Check your understanding

Quick check

—

Which lens best fits this scenario: a bank's fraud model is reported as 95% accurate overall, but freezes Karnali-registered small-business accounts at roughly twice the rate of Kathmandu-registered accounts?

Autonomy — the question is about the user’s ability to act.
Fairness — the question is about who systematically wins and loses when the model is run at scale.
Accountability — the question is about who is responsible when the model is wrong.
None of the three — this is purely a technical accuracy issue.

Quick check

—

A team reports that their newly deployed model is 95% accurate and 97% precise. What is most accurately missing from that report, in the framing of this course?

Nothing significant — high accuracy and precision are sufficient to deploy responsibly.
Only a confidence interval — the numbers are right, they just need error bars.
Whether outcomes are distributed defensibly across sub-populations (fairness), whether the system erodes the agency of the people it acts on (autonomy), and who is reachable when it fails (accountability).
A comparison to the vendor’s benchmark numbers from other countries.

What comes next

The three lenses are sharper when you can point them at rules that actually get enforced. The next section turns to Nepal’s policy moment — where the country’s AI-relevant rules stand as of 2026, which bodies are writing them, and where an ordinary Nepali professional with technical fluency can actually push.