AI That Creates: Images, Audio & Video

Everything so far has been about text. The same underlying idea — learn patterns from millions of examples, then generate something new — also produces images, voices, music and video. The mechanism differs, and the difference is worth two minutes because it tells you how to get good results.

Image models mostly work by denoising. Take a real photo, add a little visual static, then more, then more, until it's pure noise. Now train a model to run that backwards: given a noisy image and a description, predict what the slightly-less-noisy version looks like. Do that repeatedly, starting from pure static, and an image emerges that matches the description. The model isn't collaging bits of photos it has seen. It's walking from noise toward something that fits your words.

That's why prompting for images is a different craft from prompting for text. A text model wants instructions. An image model wants a description. You get much better results by naming four things: the subject, the style, the composition, and the lighting. Compare "a logo for a coffee shop" with "a minimal logo for a coffee shop, single-line drawing of a coffee cherry, black on cream, lots of negative space, flat vector". Same tool, entirely different output — and the second one is specific enough that you can iterate on one variable at a time.

The tells are real, though they've moved. Hands and teeth are largely solved; text inside images is much better but still unreliable at small sizes. What gives generated images away now is subtler — jewellery and buckles that don't quite connect, background objects that dissolve under inspection, patterned fabric that doesn't line up across a seam, and a certain glossy over-lit sameness. If it matters, zoom in on the edges and the background.

Voice is further along than most people realise. Synthetic narration is genuinely good, and cloning a voice from a short sample is a product feature, not a research demo. The remaining tell is prosody: AI voices flatten out on long sentences and don't quite land emphasis. Music generation will produce a complete, coherent track from a text description; it's excellent for backing and placeholder work, less so where a specific artistic intent matters. Video is the fastest-moving of the three, and the honest summary is that it's astonishing for short clips and still difficult for anything needing continuity between shots.

Two practical cautions, both of which catch people out. First, rights. Who owns AI-generated output, and whether it can be used commercially, depends on the tool's terms and on where you are — several jurisdictions have held that purely AI-generated work can't be copyrighted. If it's going on something commercial, read the licence for that specific tool rather than assuming.

Second, disclosure. Generated media of real people is trivially easy now and causes real harm. Don't make a real person appear to say or do something they didn't. Where a synthetic voice or face could reasonably be mistaken for genuine, label it. Most platforms are moving toward requiring this anyway, and it costs you nothing to be ahead of it.

The working posture for all of it: treat generated media as a fast, extraordinarily cheap first draft. Brilliant for exploring ideas, mocking something up, or showing a designer what you mean. Not a source of truth, and not automatically yours to sell.