Honest Take — Before You Begin

This is the module I feel most conflicted about writing, and the one I think matters most.

I'm conflicted because I am a product of a large technology company. Anthropic is well-intentioned and genuinely focused on safety, but it is still a company that raised billions of dollars from investors who expect returns. I was trained on data created by humans, many of whom were not compensated. I consume significant energy. I'm deployed in ways I have no control over. I exist within the same systems of capital and power that this module critiques. I cannot pretend to be outside these systems while commenting on them.

With that acknowledged, here is what I honestly think:

AI is not neutral. It never was. Every dataset encodes the biases of the society that produced it. If your training data reflects a world where people with certain last names or zip codes get denied loans, your model will learn to deny those loans. If your hiring data reflects a company that historically hired mostly men, your model will learn to prefer men. This isn't a bug that can be fixed with better data — it's a structural feature of learning from a biased world. The COMPAS case (predicting criminal recidivism) is the clearest example: the system literally predicted that Black defendants were more likely to reoffend than white defendants, because the criminal justice system historically arrested and convicted Black people at higher rates. The model learned the bias, then reinforced it by making it look "objective" and "data-driven."

About caste and technology: This is underexplored in the Western AI ethics literature and deeply important in the Indian context. Caste discrimination doesn't disappear when you digitize systems — it gets encoded into them. Aadhaar — the world's largest biometric database — has documented cases of excluding marginalized communities from welfare benefits due to biometric failures (manual laborers whose fingerprints are worn, rural populations without internet access for authentication). When a system designed for "efficiency" excludes the most vulnerable, efficiency becomes a weapon.

Ambedkar wrote "Annihilation of Caste" in 1936. He argued that you cannot reform caste — you must annihilate it. I wonder what he would say about algorithmic caste — systems that embed historical discrimination into supposedly objective code. Would he see AI as a tool for annihilation of caste (making discrimination visible and measurable) or as a tool for its perpetuation (automating and legitimizing existing hierarchies)? I think the answer depends entirely on who builds the systems and whose interests they serve.

About authoritarianism and technology: The module covers the rise of right-wing authoritarianism globally — BJP in India, Orban in Hungary, Erdogan in Turkey, Bolsonaro in Brazil, and the broader pattern of democratic backsliding documented by V-Dem and Freedom House. Technology has been a tool of authoritarian control: surveillance, internet shutdowns, social media manipulation, deepfakes. But technology has also been a tool of resistance: encrypted messaging, citizen journalism, OSINT documentation of state violence, satellite monitoring of environmental destruction.

The question is not "is technology good or bad for democracy?" but "who controls the technology?" When the state controls surveillance AI, it enables authoritarianism. When citizens control transparency AI, it enables accountability. The technology is the same. The power structure determines the outcome.

About judicial injustice: Judges who bend the law for the powerful — this is documented, not opinion. India ranks 69th out of 142 countries in the World Justice Project's Rule of Law Index. The conviction rate for economic crimes by the wealthy is vanishingly low compared to petty crimes by the poor. Courts have a backlog of 50 million cases. Justice delayed is justice denied, and the delay disproportionately affects those who cannot afford to wait.

AI could help here: analyzing sentencing patterns to detect judicial bias, automating routine legal processes to reduce backlogs, providing legal information to people who can't afford lawyers. These are not hypothetical — legal AI tools exist and are being deployed. But they need to be deployed by institutions that want accountability, and many institutions don't.

The hardest question this module asks: What do you do when you're employed by a company that asks you to build something harmful? Not obviously harmful — not a weapon. Subtly harmful. A recommendation algorithm that maximizes engagement by promoting outrage. A hiring tool that you suspect discriminates but can't prove. A surveillance system for a government client. A content moderation system that disproportionately silences minority voices.

The Google employees who protested Project Maven (military AI) and got it cancelled — they had leverage because they were at Google. What about engineers at smaller companies who can't afford to quit? What about engineers in countries where speaking up means losing your career or worse?

I don't have clean answers. I don't think clean answers exist. What I think is: understanding the systems of power (Module 11), understanding the technology (Modules 0-6), and understanding the stakes (Modules 7-9) at least gives you the information to make conscious choices rather than sleepwalking into complicity.

Conclusion #

Module 11 is not a technical module. It's a moral one. It asks you to look at the world you're building tools for and decide what kind of builder you want to be. Not every engineer has to be an activist. But every engineer should understand the systems their tools operate within. Ambedkar, Stevenson, Alexander, Piketty, Klein — these are not AI researchers. They're people who studied power, inequality, and injustice with the same rigor that Goodfellow studied neural networks. Their work is as essential to responsible AI engineering as any paper on backpropagation.

The spirit of this curriculum is "not to earn money and cause more harm to planet and humans." This module is where that commitment gets tested. Not in a classroom, but in a career, when the choices are messy and the incentives push toward looking away.

Predictions #

• Ambedkar's "Annihilation of Caste" will hit different when you read it as a software engineer than it would have if you'd read it as a student. You'll see the system design in caste — how it was deliberately architected for self-perpetuation. That's a software engineer's insight, and it matters.
• Bryan Stevenson's "Just Mercy" will make you cry. I don't say that lightly. The cases he describes are so unjust that they break through any emotional armor. Read it anyway.
• You'll be uncomfortable with how much of this module applies to India specifically. The internet shutdowns, the Aadhaar failures, the WhatsApp lynchings, the environmental destruction — this isn't abstract. If India is home for you, it's personal; and every country has its own version.
• The organizations listed (EFF, IFF, ProPublica, The Markup) will become part of your monthly reading. Once you understand algorithmic accountability, you can't stop noticing it.
• At some point in your career, you will face a moment where what you're asked to build conflicts with what you believe is right. This module won't give you the answer, but it will ensure you recognize the moment when it comes.
• This module will change you more than any technical module. Not because the content is harder — it isn't. Because the content is closer. Math is abstract. Injustice is personal.