Generative AI

You type 'an astronaut riding a horse' and get an image that looks deceptively real. Image generators are built into phones, chat apps and photo editors today. They are a creative tool and a forgery machine at the same time. To make sense of both, it helps to look at what an image even is to a computer.

Every image is a grid of numbers

A digital image consists of pixels, tiny dots of colour in a grid. To the computer each pixel is just a number, or a few numbers for the colour components. A photo with millions of pixels is therefore one huge table of numbers. And that is exactly why an AI can generate images: whatever can compute numbers can compute pixels.

From noise to image

Most image generators work on the diffusion principle. During training the model is shown millions of images that were covered step by step with random noise, and it learns to remove that noise again. Generation runs the process in reverse: the model starts with pure noise, like a TV screen with no signal, and computes it clearer in many small steps. Your text description steers every step, so that out of the chaos emerges exactly the scene you described.

Spotting deepfakes

The same technique also produces deepfakes: forged photos and videos of real people. Classic giveaways like odd hands, garbled text in the background or impossible shadows are getting rarer, though they have not vanished entirely. Do not rely on such glitches alone. Context questions matter more: who is spreading the image? Are reputable sources reporting the same thing? A reverse image search often reveals whether a picture is older or taken from a different context. And the more outrage an image is designed to trigger, the more scepticism is warranted.

Labels and provenance

Because the eye alone is no longer enough, technical and legal answers are emerging. With the AI Act the EU requires AI generated content to carry machine readable labels and deepfakes to be clearly disclosed, binding from August 2026. In parallel, camera makers and AI providers are building provenance credentials that travel with an image and record where it came from and how it was edited. None of these systems is perfect, but together they make forgery considerably harder.