Module 520 min read · Foundations

AI Ethics &
Real Risks

Most conversations about AI risk fall into one of two traps: either dismissing all concerns as science fiction, or catastrophizing about robots taking over. Neither is useful. This module is about the actual risks — the ones happening now, the ones coming soon, and what they mean for anyone who wants to use AI responsibly.

The risks worth worrying about

The most dangerous AI risk isn't a superintelligent system deciding to harm humanity. It's people using powerful but imperfect AI systems in ways that cause real harm — at a scale and speed that wasn't previously possible.

Here are the risks that deserve your attention:

Misinformation at scale
AI makes it dramatically cheaper and faster to generate convincing false content — fake news articles, synthetic images, deepfake videos, and coordinated disinformation campaigns. What used to require a team of people can now be done by one person in minutes. The 2024 election cycles globally showed what this looks like in practice: AI-generated audio clips of candidates saying things they never said, synthetic images of events that never happened, and automated social media accounts amplifying falsehoods at scale.
Algorithmic bias and automated discrimination
When AI systems are used to make decisions about people — loan approvals, hiring, medical diagnosis, parole recommendations, insurance pricing — the biases in their training data get encoded into those decisions and applied at scale. A biased human makes biased decisions one at a time. A biased algorithm makes them millions of times simultaneously, consistently, and often invisibly. Several real-world cases have shown AI hiring tools systematically downranking women's resumes, and risk assessment tools used in courts that assigned higher recidivism risk scores to Black defendants.
Power concentration
The most capable AI systems require billions of dollars to build and run. This means they are controlled by a small number of companies and governments. If AI continues to provide massive economic and strategic advantages, and access remains concentrated, the result is a dramatic amplification of existing power imbalances. This is one of the most serious structural risks — not because AI is malicious, but because it's a powerful tool in the hands of a small number of actors.
Job displacement without transition support
AI will automate significant portions of many jobs — not necessarily eliminating them entirely, but changing what skills matter. The risk isn't that all work disappears. It's that the transition happens faster than education systems, labor markets, and social safety nets can adapt — leaving large numbers of people economically stranded. The historical precedent from previous waves of automation is that new jobs do emerge, but the transition period is painful and unevenly distributed.
Erosion of human judgment and skill
When people outsource thinking to AI systems without understanding how they work or checking their outputs, two things happen. First, humans stop developing and maintaining skills that matter. Second, errors propagate without anyone catching them. A doctor who relies on AI diagnosis without understanding medicine, a lawyer who submits AI-written briefs without reading them, a student who uses AI for every assignment without engaging with the material — all of these create real risks, both individual and systemic.

Algorithmic bias: going deeper

Bias in AI deserves more than a bullet point because it's widely misunderstood. People often assume that if humans aren't intentionally encoding discrimination, the system will be fair. This is wrong for several reasons.

Bias in training data

If a model is trained on historical hiring data from a company that systematically underpromoted women, the model learns that pattern. It doesn't "know" it's discriminating — it's reproducing what was there. The data reflects the past, and the past was often unjust.

Feedback loops

If a biased model is deployed and its outputs are fed back into future training data, the bias compounds. A criminal justice risk model that overestimates recidivism for a certain group leads to longer sentences, which means more time in the system, which means more data confirming the original prediction. The model becomes more confident in a prediction that was wrong to begin with.

Proxy discrimination

Even if protected characteristics like race or gender are removed from the data, AI systems can infer them from proxy variables — zip codes, names, schools attended, browsing history. Removing the explicit variable often doesn't remove the discrimination.

The uncomfortable truth

There is no such thing as a neutral AI system. Every model reflects choices made by its builders — what data to train on, what to optimize for, whose feedback to use, which outputs to filter. These choices encode values, whether or not they're made consciously. "The algorithm decided" is never a complete explanation.

What responsible AI use looks like

Understanding these risks isn't just academic. It shapes how you should use AI tools in your own life and how you should think about AI systems that affect others.

Verify important outputs

AI is a starting point, not an endpoint. Anything that matters — facts, analysis, code, medical information, legal interpretation — should be verified through independent sources. Confidence in the output is not evidence of accuracy.

Maintain your own skills

Use AI to extend your capabilities, not replace them. If you use AI to write all your emails, you stop developing the ability to write clearly. If you use AI to do all your analysis, you stop developing judgment. The goal is augmentation, not dependence.

Be transparent about AI use

In professional and academic contexts, passing off AI-generated work as entirely your own raises serious ethical questions — and increasingly, legal ones. The norms around this are still developing, but the principle is straightforward: be honest about how you produced something.

Think about downstream effects

When you build systems that use AI to make decisions about people, you take on responsibility for those decisions. Understanding bias, testing for it, and building in human oversight aren't optional extras — they're baseline requirements for responsible development.

The bigger picture

AI is genuinely one of the most consequential technologies in human history. The people who understand it — really understand it, not just the hype — are in a position to shape how it develops and who it benefits. That's part of why this course exists, and part of why the governance and policy questions around AI are among the most important that this generation will have to answer.