AI-generated content

AI-generated content is text, pictures, voices, music or video made by an AI model, rather than written, drawn or recorded by a person. It is also called generative AI content, or synthetic media. The models that make it are trained on very large collections of existing work. From that work, they learn how words, images and sounds usually go together. Film studios, newsrooms, game makers and publishers now use these models for work that once took people days or weeks.

Here is how it works in general. A person gives the model an instruction, called a prompt, together with any material it needs, such as a script, a photo or a short recording of a voice. The model produces a draft: a paragraph, an image, a spoken line or a few seconds of video. For text, the model is usually a large language model, a model trained on huge amounts of text so that it can read and write it. A person then checks the draft, corrects it and decides whether to use it. Careful companies also label the result as made by AI. Some add a hidden mark, called a watermark, which software can detect later.

The models are fast and cheap, and so they are used most where a lot of new material is needed all the time. A live service game needs new art, levels and lines for its characters every month. A news agency needs short reports on every company's quarterly results.

The same speed raises three questions. Is it good enough? A model can state a wrong fact in the same confident tone as a correct one. This is called AI hallucination. Can people trust it? Readers and listeners may want to know whether a person made what they are reading or hearing. And is it fake? The same tools can copy a real person's face or voice without that person's permission.

Take, for example, a Hindi film that is also released in Telugu. AI dubbing and lip sync changes the actor's mouth movements to match the Telugu words, so the film no longer looks dubbed.

Consider an app that publishes stories to listen to, in short episodes. In audio drama production, a model can draft an episode and voice a minor character, while human writers and voice artists do the main work.

Say a researcher sends a paper to a journal. Research integrity screening checks whether parts of it were written by AI without saying so.

How is AI-generated content different from a deepfake? A deepfake is one kind of AI-generated content: a video, photo or voice made to look like a real person saying or doing something that the person never did. It is usually made to deceive. Most AI-generated content is not. A dubbed film or a game level drafted by AI is open about what it is. Software for deepfake detection looks for the signs that a face or a voice has been made by a model.