The One Thing Every AI Ethics Book Gets Wrong

The One Thing Every AI Ethics Book Gets Wrong

I have read more AI ethics books than any reasonable person should. It started as research and became something between a habit and a frustration. Over the past two years, while writing my own book on AI ethics, I have read the big ones, the small ones, the academic ones, the corporate ones, the ones that get quoted in TED talks and the ones that sit unread on desks in government offices.

And I have noticed something. They all make the same mistake.

They describe the problem brilliantly. And then they stop.

Let me be specific. Because this is not a vague complaint.

Pick up any well regarded book on AI ethics from the last three years. You will find an excellent explanation of bias. You will learn about the Amazon hiring algorithm, the COMPAS recidivism score, the facial recognition studies that showed error rates varying wildly across skin tones. You will read about the opacity of black box systems. You will encounter thoughtful discussions about privacy, about consent, about the philosophical tensions between different notions of fairness.

And if you are a product manager at a fintech company in Bengaluru, a developer working on a healthcare AI tool in Hyderabad, a policymaker drafting guidelines at a state government office, or a founder deciding whether your AI powered lending product is ready to ship, you will close the book and think: okay, I understand the problem. Now what do I actually do?

The book will not tell you.

It will tell you that fairness matters. It will not give you a checklist for auditing your model for bias. It will tell you that transparency is essential. It will not give you a template for documenting your AI system in a way that regulators, users, and your own team can understand. It will tell you that accountability is critical. It will not tell you how to set up an AI ethics review board, what authority it should have, or what happens when it says “do not ship” and the CEO says “ship.”

The books describe the weather. They do not give you an umbrella.

I say this with respect. Many of these authors are researchers I admire. The work they have done to map the landscape of AI harms, to document the case studies, to build the theoretical foundations of algorithmic fairness, that work is essential. Without it, we would not even have the vocabulary to discuss these issues.

But here is the uncomfortable truth. The gap between AI ethics research and AI ethics practice is not narrowing. It is widening.

Two hundred and thirty three reported AI incidents in 2024, according to the Stanford HAI AI Index Report. A fifty six percent increase from the previous year. Over two hundred AI ethics frameworks in existence globally, per a 2023 meta analysis by Correa and colleagues. And fewer than one in three organisations deploying AI have a formal policy for how that AI should behave, according to a 2025 ISACA survey of business and IT professionals.

The frameworks are not working. Not because the principles are wrong, but because principles without implementation are aspirations. And aspirations do not survive contact with a Tuesday morning shipping deadline.

I am going to make a claim that might sound arrogant but is actually just honest: the book I wrote is different.

Beyond Bias: The Four Way Test for Ethical AI is different not because the principles it espouses are new. Truth, fairness, privacy, accountability. These are the same principles that every ethics framework in existence has identified. The global consensus on what matters is remarkably strong. Eighty six percent of frameworks include transparency. Eighty one percent include fairness. Seventy one percent include accountability.

The book is different because of what it does with those principles.

Every chapter gives you something you can use on Monday morning.

Let me show you what I mean.

The chapter on bias does not just explain where bias comes from. It gives you a Bias Audit Checklist. Four phases: data audit, model audit, deployment audit, ongoing monitoring. Each phase has specific questions, specific tests, specific red flags. It tells you what to look for in your training data. It tells you how to test for proxy variables, those innocent looking data points like PIN codes and phone models that secretly stand in for caste and class. It tells you what to do when you find bias, not in the abstract, but step by step.

The chapter on transparency does not just argue that AI systems should be explainable. It gives you a Transparency Audit Checklist. Six parts: model documentation, decision traceability, data transparency, stakeholder access, regulatory compliance, ongoing monitoring. It tells you what to document, how to document it, and who needs access to that documentation.

The chapter on fairness does not just present the impossibility theorem, the mathematically proven fact that you cannot satisfy all fairness criteria simultaneously, and leave you in despair. It gives you a Fairness Audit Checklist that walks you through how to make defensible tradeoffs. Which fairness metric to prioritise and why. How to justify your choice to regulators, to users, to your own team.

The chapter on privacy gives you a Privacy Impact Assessment template. The chapter on agentic AI gives you an Autonomy Threshold Framework, a four tier system for deciding when a human must be in the loop. The chapter on regulation gives you a Regulatory Compliance Checklist tailored for three audiences: every AI deployer, companies operating across borders, and policymakers.

And the practice chapter, the one where everything comes together, gives you role specific implementation guides. Here is what a developer should do. Here is what a product manager should do. Here is what a business leader should do. Here is what a policymaker should do. Not theory. Action steps.

Why does this matter? Because the people who need AI ethics the most are not the people who have time to read philosophy.

The product manager with a shipping deadline does not need a nuanced discussion of Rawlsian justice. She needs to know: have I checked this model for bias? Is there a clear record of how it makes decisions? Can an affected person challenge the output? Am I comfortable explaining this system to a regulator?

The developer pushing code at midnight does not need a survey of two hundred global frameworks. He needs to know: what should I document? What should I test? What are the five things that, if I miss them, will cause real harm to real people?

The startup founder does not need an academic paper on the impossibility theorem. He needs to know: given that perfect fairness is mathematically impossible, what is the defensible position? What can I show a regulator, a journalist, or a courtroom that demonstrates I took this seriously?

The policymaker drafting India’s next AI guideline does not need another comparison of GDPR and DPDPA. She needs a framework for evaluating whether a proposed regulation meets the actual needs of the people it is supposed to protect.

Beyond Bias was written for these people. The people who are in the room when the decisions are made. The people whose Tuesday mornings determine whether AI systems serve the public or harm it.

I will tell you one more thing the book does that I have not seen in any other AI ethics text.

It starts from India.

The case studies are Indian. The regulatory analysis centres on the Digital Personal Data Protection Act, the India AI Governance Guidelines, the AI Ethics and Accountability Bill. The fairness discussion grapples with caste bias in datasets, linguistic bias in language models, the digital divide between urban and rural India. The privacy chapter examines what consent means when a Deloitte survey found that ninety one percent of people accept terms of service without reading them, when the enforcement agency has not yet brought a single case, and when the gap between what you are consenting to and what you can understand is wider than ever.

This is not an American framework with an India chapter stapled on at the end. This is a book written from the engineering floor in India, for the people building, deploying, and regulating AI in India, with global context that makes it relevant anywhere.

You should be able to use this book on Monday morning. That is the test. If you read it and think, “interesting,” I have failed. If you read it and think, “I know what I need to do tomorrow,” I have done my job.

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/