Data literacy is a policy question, not a training course

· Dhiraj Kumar Sah

Before I worked on AI, I worked on banking software for places with no bank. The system I helped build at a fintech company in Kathmandu let an agent with a point-of-sale device open accounts and move money in villages where the nearest branch was hours away. It was good work. It is also where I first noticed something that has followed me into every job since: the people whose lives the data described were, almost without exception, the people with the least ability to read it.

The account holder saw a receipt. The agent saw a screen. The bank saw a dashboard. The regulator saw a quarterly report. Each layer understood a little less about the individual and a little more about the aggregate, and each had more power over what happened next. Nobody designed it that way; it is just what happens when literacy about data is treated as a skill some people have, rather than a condition the system is responsible for creating.

The usual answer, and why it is not enough

When this comes up, the response is almost always a course. Teach people spreadsheets. Teach them what an average is. I have taught some of these courses, and they are worth doing. But they put the burden in the wrong place. A course asks the person on the receiving end of a decision to catch up to the system that made it. A policy asks the system to explain itself in terms the person can act on.

The difference is easiest to see in the small print. A loan declined "based on your profile" is a literacy problem for the applicant. A loan declined "because your income over the last six months was below X and here is how to dispute that figure" is a design decision by the lender, and it is the kind of decision that regulation can require.

Why this matters more now

AI systems are moving from describing people to deciding about them. The credit score becomes an agent that approves or declines. The dashboard becomes a model that flags a transaction. The quarterly report becomes a policy recommendation generated from a million records nobody has read. At each step the explanation gets harder to produce and easier to skip, because the system that produced the decision did not reason the way a person would, and the people deploying it may not fully know how it did reason.

That is why I have come to see the work I do now — on making AI agents accountable at the level of the runtime they execute in, rather than at the level of the instructions they are given — as a continuation of the problem I first saw on that point-of-sale device. A record of what the system did and why, written in a form someone outside the system can check, is not a technical nicety. It is the minimum condition for a person to be able to argue with a decision made about them.

Three things I would ask of policy

  • Require explanations that are actionable, not merely present. "The model determined" is a sentence; it is not an explanation. The test should be whether the affected person could do something different next time, or could dispute the decision with evidence.
  • Treat the audit trail as the product. If a system makes consequential decisions, the log of those decisions should be designed with as much care as the interface, and it should be legible to people who are not engineers.
  • Fund literacy as infrastructure, in the languages people actually speak. My own languages are Maithili, Nepali and English. The receipts in that village were in one of them. Most of the policy was in another. That gap is not an accident of translation; it is a choice about who the system is for.

None of this is specific to Nepal, or to Canada, where I live now. It is about the direction we are heading in everywhere: toward systems that know more and explain less. I would rather we build the explaining in from the start than add it later, after someone has been harmed and a regulator has asked why.