By Roman Aglaganov
In March 2023, a picture of Pope Francis in a huge white puffer jacket went viral. It looked like a fashion shoot; it was made with Midjourney. For many people it was the moment they realised AI images were no longer obviously fake.
How we got here
Modern image generators mostly use diffusion: the model learns to remove noise from images step by step, and then generates new images by starting from pure noise and “denoising” towards a text description. A research advance published in late 2021, latent diffusion, made this efficient enough to run on consumer hardware; it became Stable Diffusion, released openly in August 2022. Around the same time, OpenAI’s DALL·E 2 and Midjourney brought image generation to millions of people.
Then came video
In February 2024, OpenAI previewed Sora, which turned text prompts into strikingly realistic video clips up to a minute long, and released it publicly in December 2024. In May 2025, Google’s Veo 3 added something its predecessors lacked: synchronised sound, including dialogue. Silent AI movies had their talkie moment.
The real problem
The same tools that let a small studio storyboard a film let anyone fake a politician’s speech or a disaster that never happened. Detection tools struggle to keep up, which is why governments and companies are pushing for labels and content credentials that record how a file was made.
A useful habit: before sharing a too-perfect image or clip, ask where it came from. The pope never wore the jacket. It was still a very good jacket.
