Most writing about AI risk sits at one of two extremes. One says the machines will end us. The other says the worry is overblown and you should relax. Neither helps you decide whether to trust the summary an AI just wrote of your medical report.
This course is about the middle: the failures that already happen, to ordinary people, most days.
Three kinds of failure
It helps to sort problems by where they come from, because the fix is different in each case.
The system works as designed, and the design is the problem. A bank builds a model to predict who repays loans. It works. It also declines a whole district of a city, because historically few loans were made there, so few repayments were recorded there. Nothing is broken. The model learned what it was shown.
The system fails at the thing it claims to do. You ask for the citation for a legal case and get a case that does not exist, written in perfect legal style. The model is not lying — it has no concept of lying. It is producing plausible text, and a plausible citation is exactly what you asked for.
The system works, and someone points it at you. A voice clone of your daughter calls asking for money. The technology functioned beautifully. That is the problem.
Most of the news mixes these together. Keeping them apart is most of the skill.
Why "the AI decided" is almost always wrong
When a system denies you insurance, no AI decided anything. A company chose to buy or build a model, chose what to optimise for, chose what data to train on, chose the threshold at which a score becomes a rejection, and chose whether a human reviews it.
Every one of those is a human decision with a name attached. "The algorithm decided" is a sentence that moves responsibility from people to software. Watch for it. It usually appears in a press statement.
This matters for you practically: when something goes wrong, the question "which human made which choice?" gets you further than "why did the AI do that?"
The honest state of things
Some true statements that sit uncomfortably together:
AI systems are genuinely useful. A farmer in Maharashtra photographing a diseased leaf and getting a probable diagnosis in Marathi is a real gain. So is a deaf student getting live captions in a lecture hall.
They also fail in ways that are hard to see. A wrong answer arrives with the same fluent confidence as a right one. There is no tremor in the voice.
And the failures are not evenly spread. Speech recognition works worse on accented English. Face recognition has been measured to be less accurate on darker-skinned faces. The people who most need a system to work are often the ones it works worst for, and they are usually the least able to complain effectively.
What this course does
Seven more lessons, each on one failure mode: where bias comes from, hallucination and how to check, what you hand over when you paste, synthetic media, decisions about people made at scale, copyright, environmental cost, and using AI as a student without hollowing out your own learning.
No lesson will tell you to stop using these tools. That advice would be ignored, correctly. The aim is narrower and more useful: know what breaks, know how it breaks, and know what to do when it does.
Before you move on