Four Questions That Would Have Caught Every AI Disaster I’ve Described in This Series

Let me take you back through the wreckage.

An eleven year old girl in Jharkhand died of starvation after her family’s ration card was cancelled for not being linked to Aadhaar. A hiring algorithm at one of the world’s largest companies taught itself that women were a negative signal. A credit scoring system used your PIN code, your surname, your phone model as quiet proxies for things no algorithm should be allowed to infer. A facial recognition system failed on dark skinned faces at rates a hundred times higher than on light skinned ones, and was deployed anyway. A deepfake of a dead politician endorsed a living candidate in the largest election the world has ever seen.

Five hundred million Indians are invisible to the AI systems that are increasingly making decisions about their lives, because those systems were trained on urban, English speaking, smartphone owning India and assumed that was everyone.

I have spent the last sixteen articles walking you through these failures. Not because I enjoy cataloguing disaster. But because each one of them, every single one, had a moment where someone could have asked a simple question and changed the outcome.

I want you to sit with that for a moment. Not the scale of the harm. The simplicity of the prevention.

The Jharkhand case. Before that biometric system was deployed as the sole gateway to food rations, someone could have asked: is this truthful about what the system can and cannot do? Fingerprint scanners fail at measurably higher rates for manual labourers. That was known data, not a surprise. Someone could have asked whether it was fair that a technology’s known limitations would fall disproportionately on the poorest people in the system. Someone could have asked whether building a single point of failure authentication mechanism for survival level services was genuinely beneficial to the people it was supposed to serve.

Nobody asked. Or rather, people asked, and the questions did not have the institutional weight to change the decision.

Amazon’s hiring tool. Before that model was trained on ten years of male dominated hiring data, someone could have asked: is it truthful to claim this system identifies “the best candidates” when its entire understanding of what “best” looks like was shaped by a decade of decisions made primarily by and about men? Someone could have asked whether the system was fair when it systematically penalised resumes containing the word “women’s.” Someone could have asked who benefits from this system, and whether the benefit was distributed equitably.

The model was scrapped. But it ran for years before it was scrapped. Years in which real people had real resumes filtered by a system that had learned that half the population was a liability.

Here is the pattern I see, and I see it with disturbing clarity after two years of digging into this.

Every AI failure I have examined, from the COMPAS recidivism algorithm that predicted Black defendants would reoffend at nearly twice the false positive rate of White defendants, to the healthcare algorithm that used spending as a proxy for need and systematically deprioritised Black patients, to the predictive policing systems that sent patrol cars back to the same neighbourhoods that were already over policed, every single one could have been caught at the decision stage. Not with a hundred page compliance document. Not with a six month audit that starts after the system is already live. With questions.

The right questions. Asked at the right time. By people who had the authority to act on the answers.

I am not being naive. I know that organisations are complex. I know that shipping deadlines are real, that competitive pressure is real, that the people in the room when these decisions are made are usually smart, well meaning, and overwhelmed. I said as much in my last article.

But I also know something else, something I learned not from the technology industry but from my family.

My grandfather was a schoolteacher in a small town in Maharashtra. He was not a technologist. He would not have known what to make of a neural network or a training dataset. But he understood something about decision making that the technology industry, with all its brilliance, has largely forgotten.

He believed that the quality of a decision depends on the quality of the questions you ask before you make it. Not after. Not when the consequences arrive. Before.

He had four ideas that governed how he lived. Simple ideas. The kind of ideas that a schoolteacher carries to the same classroom for forty years, ideas that his six children absorbed into their bones and passed down to their children. I absorbed them too, though I did not understand their full weight until I started building systems that make decisions about people. Systems that have no instinct for kindness. Systems that cannot look a person in the eye and decide that the rules need to bend because the circumstances demand it.

I built something from those ideas.

A framework. Not another hundred page document. Not a set of principles that sounds good in a boardroom and dissolves the moment it meets a shipping deadline. Something simpler than that. Something that fits in your head.

Four questions.

I can apply them to any AI decision. Any model. Any deployment. Any policy choice. I have tested them against every case study I have written about in this series, and they catch the failure every time. Not because they are magic, but because they force you to look at the decision from four angles that most teams, under pressure, naturally compress into one: “does it work?”

“Does it work?” is not enough. It was never enough. The Jharkhand system worked. The Amazon hiring tool worked. The COMPAS algorithm worked. They all worked, in the narrow sense that they produced outputs that matched their training objectives. The question was never whether they worked. The question was whether they were right.

I am not going to tell you the four questions in this article.

That might sound like a tease, and I suppose it is. But there is a reason. The questions are simple, but they are not trivial. They need context. They need case studies. They need the mechanics of how bias enters a system, how transparency fails, how privacy is eroded, how accountability dissolves when AI acts autonomously. They need the Indian regulatory landscape mapped out, because the framework I built is grounded in a reality that most AI ethics writing ignores entirely.

All of that is in the book.

What I will tell you is this: the framework is not original in the way a patent is original. The values underneath it are ancient. They are the same values that show up in every ethical tradition I have encountered, from my grandfather’s classroom in Maharashtra to a bankrupt cookware company in 1930s Chicago to the consensus of two hundred global AI ethics frameworks. Truth. Fairness. The genuine desire to do right by people. The discipline to check whether your actions actually help or harm.

The contribution is not the values. The contribution is making them operational. Turning them into four questions that a product manager can ask before a deployment decision. That a developer can ask before pushing code. That a policymaker can ask before drafting regulation. That a business leader can ask before signing a procurement contract for an AI system.

Four questions. Any AI decision. Monday morning.

The book is called Beyond Bias: The Four Way Test for Ethical AI. It covers everything this series has covered, and goes deeper. Bias. Transparency. Fairness. Privacy. Agentic AI. The regulatory landscape. The hidden costs. And it gives you, in every chapter, the practical tools that the principles alone cannot provide.

But at its core, it is four questions. Built from something my grandfather taught me before I ever wrote a line of code.

If you want to know when it is ready:

I’m have written a book about exactly this  how AI and automated systems make decisions about your life, where accountability disappears, and what we can do about it. If you want to know morea about this book or order a copy, you can do it here: https://akshaywalimbe.com/beyond-bias/