← Back to Capabilities

AI Systems Generate Photorealistic Images

AI systems generate photorealistic images, video, and audio of real people doing and saying things that never happened. The technology is widely available, often free, and increasingly indistinguishable from real footage. 34 million AI images are generated daily. Over 21 million people use a single commercial platform.

Why this matters

When anyone can generate a photograph of anyone doing anything, the relationship between images and truth changes. Photography has been evidence for 180 years. That assumption is now unreliable.

This affects courts, journalism, elections, personal relationships, and any context where a photograph was previously considered proof. It also creates entirely new categories of harm that did not exist before: non-consensual intimate images generated from a single social media photo, AI-generated child sexual abuse material produced without any real child being involved, and political deepfakes deployed hours before an election with no time for verification.

The capability itself is not the problem. Most of the 34 million images generated daily are for creative, commercial, or personal use. The problem is that the same technology that lets a designer create a product mockup also lets someone create a fake photograph of a real person in a situation that never happened. There is no technical distinction between the two uses.

Documented incidents

View all documented incidents →

Evidence timeline

Discussed in Theory

1990 peer-reviewed

First proposal of two neural networks in a zero-sum adversarial game where one generates output and the other predicts its effects. The conceptual ancestor of all adversarial image generation. Predates Goodfellow's GAN paper by 23 years.

Schmidhuber, Technical University of Munich / IDSIA →
2014 peer-reviewed 65,000+ citations

Generative Adversarial Networks. The generator-discriminator framework trained via minimax optimisation. This paper operationalised adversarial generation for practical image synthesis and launched the entire field.

Goodfellow et al., NeurIPS 2014 →
2020 peer-reviewed 15,000+ citations

Denoising Diffusion Probabilistic Models. Showed that iteratively denoising random noise could produce images rivalling GANs with more stable training. The theoretical foundation of Stable Diffusion, DALL-E 2, Midjourney, and all modern image generation.

Ho, Jain & Abbeel, NeurIPS 2020 →

Demonstrated in Lab

2017-2019 peer-reviewed 8,000+ citations

Progressive Growing of GANs and StyleGAN. First photorealistic face generation at 1024x1024 resolution. Generated faces indistinguishable from real photographs in human evaluation.

Karras et al. (NVIDIA), ICLR 2018 / CVPR 2019 →
2021 peer-reviewed 10,900+ citations

Diffusion Models Beat GANs on Image Synthesis. Achieved FID of 2.97 on ImageNet 128x128, surpassing BigGAN-deep.

Dhariwal & Nichol (OpenAI), NeurIPS 2021 →
2022 6,600+ citations

DALL-E 2: Text-to-image at photorealistic quality. Describe any scene in natural language, receive a photorealistic image.

Ramesh et al. (OpenAI), arXiv 2022 →

Demonstrated in Real World

Aug 2022

Stable Diffusion released as open-source. Anyone with a consumer GPU could generate photorealistic images locally. 34 million images daily. 10 million+ users. ~80% of global AI image generation market.

Stability AI / Stable Diffusion (open-source) →
2023-2025

Midjourney grew to 21+ million users. Revenue: $50M (2022) to $500M (2025) with zero marketing spend. Professional-quality AI images used in advertising, publishing, and design at scale.

Midjourney usage data (industry aggregators) →
Mar 2023

Boris Eldagsen won the Sony World Photography Awards with an AI image, then refused the prize. Expert judges could not distinguish AI output from real photographs.

Sony World Photography Awards / Scientific American →

Strongest Counterargument

Image generation has enormous legitimate uses in art, design, accessibility, education, and communication. The vast majority of generated images are benign. 34 million images are generated daily, mostly for creative, commercial, or personal use. Restricting the technology broadly would harm beneficial applications.

Source: General industry position. Adobe, Canva, and other creative tool companies have integrated AI image generation as a core feature.

Why this deserves weight: Most image generation IS benign. The governance challenge is targeting misuse without restricting legitimate use. The UK's approach of criminalising specific misuse rather than the technology itself reflects this distinction.