The concrete risks

Discussion of AI risk tends to arrive at one of two temperatures: extinction, or nothing worth worrying about. Both skip the part that is actually happening, which is more specific and more tractable.

Here are mechanisms rather than labels. Each one follows from something in this journey. Start with the ones about information.

Cheap persuasion at scale. Generating text costs almost nothing, and it can be tailored. That makes fraud cheaper to run, propaganda cheaper to produce, and the ordinary background assumption that a message came from a person less safe to hold. Voice cloning from a short sample makes the familiar phone scam considerably worse.

A rising cost of verification. If convincing text, images, audio, and video are all cheap, then establishing what is real gets more expensive for everyone. The corrosive part is not the convincing fake. It is that a genuine recording can now be dismissed as a fake, which is useful to anyone caught doing something.

Surveillance that scales. Recognising faces, transcribing speech, and sorting patterns in behaviour used to be limited by how many people you could pay to watch. That limit is gone. What was infeasible to monitor is now cheap to monitor.

Automating decisions about people. Systems now sort applications for jobs, loans, housing, and benefits. From the bias chapter: the patterns come from the training data, the data reflects past decisions, and the past contains what it contains. A biased system also scales in a way a biased individual cannot, and it is harder to appeal against, because there is no reasoning to examine.