Generative AI: How Machines Learned to Create

Classification AI answers "what is this?" Generative AI answers "make me one of these." The shift from recognising content to producing it — text, images, code, audio, video, molecules — is why AI went from a back-office tool to something hundreds of millions of people use directly. Two model families did most of the work, and understanding them takes the magic out without taking the significance out.

Two architectures that changed everything

Transformers generate sequences one token at a time, each prediction conditioned on everything before it. Trained on enough text, a transformer learns grammar, facts, reasoning patterns and style as a side effect of predicting the next word. Large language models — GPT, Claude, Gemini, Llama and the rest — are transformers scaled up and then tuned to follow instructions. The same architecture, with tokens standing for image patches or audio frames, generates those too.

Diffusion models learn to reverse noise. Training adds random noise to images step by step and teaches the model to undo one step; generation starts from pure noise and denoises repeatedly, guided by a text prompt, until an image emerges. This is the engine behind Stable Diffusion, Midjourney and DALL·E, and — extended to sequences of frames — behind video generators like Sora, Veo and Runway. Diffusion has also produced state-of-the-art results in audio, protein design and molecule generation.

The two are converging: many current systems use transformers as the backbone of diffusion models, and language models increasingly generate images and audio natively.

Where it's being used

• Software. Code generation is the most economically significant application so far — see AI Coding Assistants.
• Knowledge work. Drafting, summarising, translating, extracting structure from documents, answering questions over a company's own data. Most enterprise deployments are this, and most are unglamorous and effective.
• Marketing and design. Copy variants, product imagery, localisation at scale. Cheaper and faster; the quality ceiling still belongs to humans, the floor has risen dramatically.
• Science. Generative models propose drug candidates, materials and protein structures for experimental validation. AI-designed molecules are in clinical trials.
• Media and entertainment. Concept art, previsualisation, voice synthesis, music. Also the site of the sharpest labour and copyright disputes.
• Customer interaction. Support agents that resolve rather than deflect, with escalation to humans — successful where the knowledge base is good, embarrassing where it isn't.

The problems that haven't been solved

Hallucination. Generative models produce plausible output, not verified output. Retrieval grounding and citation requirements reduce fabricated facts; they don't eliminate them. Any workflow that relies on a generated fact needs a check.

Deepfakes and provenance. Synthetic voice, image and video are now cheap and convincing enough to enable fraud — cloned executive voices authorising wire transfers, fabricated video in elections. The technical response is content provenance: the C2PA standard attaches cryptographically signed metadata recording how a file was created and edited, and major camera makers, software vendors and platforms have adopted it. Watermarking of AI output (Google's SynthID and similar) complements it. Neither is universal; detection tools remain unreliable. The EU AI Act now requires AI-generated content to be labelled.

Copyright. Whether training on copyrighted works is fair use, and who owns generated output, is being litigated in multiple jurisdictions with divergent early results. Licensing deals between AI companies and publishers, and opt-out mechanisms for creators, are the practical middle ground emerging while the law catches up.

Energy and concentration. Training frontier models consumes enormous compute, and inference at scale is now a significant driver of data-centre power demand. See Green Computing.

Using generative AI well

The deployments that work share traits: a specific task with a clear definition of good output; the model grounded in your own data rather than relying on memory; a human reviewing anything consequential; evaluation on real examples before launch and monitoring after; and honesty with users about what's generated. The ones that fail usually skipped one of those, most often the evaluation. Generative AI is a capability, not a product; the product is the workflow you build around it.