Foundations & Models
Generative AI
Earlier AI sorted, recognized, recommended — generative AI produces something itself: text, images, code, audio. This shifts the cost structure of entire activities, because the first draft is suddenly almost free. What remains valuable is what happens before and after: the right question, the selection, the check. Generating has become cheap; judging has not.
An example
The difference from earlier AI can be shown in a single sentence. Previous systems answered questions with a choice: Is this email spam, yes or no? Does this image belong to class A or B? Generative systems answer with something new — a text that did not exist before, an image nobody took. That makes them incomparably more versatile and at the same time harder to check: with a choice between two options, the error rate is measurable. With free text, you first have to define what wrong even means.
A closer look
Generative methods model the distribution of the data itself and draw samples from it. That explains two properties: the same input can yield different outputs, and the result is always a proposal drawn from what was learned, never a find.
Why it matters
This is exactly where most pilots fail. A classification model can be accepted on the basis of an accuracy figure; a generative system cannot. Anyone who tries anyway ends up with “looks good” as the acceptance criterion — and finds errors only if they happen to look. The work therefore lies in defining the yardstick before building: Which properties must a good output have, which errors are tolerable, which are not? Once that is settled, the check itself is usually cheap.
Commonly misunderstood
Because the draft is cheap, the effort shifts to checking. Teams that automate only the generation and leave the control unchanged often save nothing in the end.
Where this term does its work in the magazine.
2 articles use the term or take up its principle.
The Long View
Deskilling: What We Unlearn When AI Always Has an Answer
Anyone who never practices the fundamentals will later struggle to judge whether an AI result holds up.Read the articleFrom Science