The word that carries the meaning is "sampling"
Most explanations of generative AI stop at "it creates new content". The useful part is how. During training, the model fits a distribution over its data: which pixel arrangements, sound waveforms, or token sequences are plausible. Generating is then drawing one point out of that distribution, guided by your prompt.
Three consequences follow, and all three shape how you work:
- Variance is the design, not a bug. Two runs draw two samples. Prompt engineering narrows the region being sampled; it never collapses it to a point. Only a fixed seed does that.
- Plausible beats correct. The model optimizes for "looks like the training data", so a confidently wrong hand with six fingers is a perfectly reasonable sample. This is the mechanism behind AI hallucination.
- There is no retrieval step. Nothing is being pasted from a library. That is why you cannot ask for a specific real object and expect fidelity, and why supplying a reference image works so much better than describing one.
The families you will actually touch
You do not need the research map, only the four kinds of engine that sit behind the tools you use daily.
- Diffusion models. Almost every image and video model in production. Start from noise, remove a bit of it per step, guided by the prompt. Their knobs (steps, guidance scale, denoising strength, seed) are the knobs you see in the interface.
- Autoregressive transformers. Text, and increasingly image and audio. Predict the next token given all previous ones. Chat models and the LLM that rewrites or captions your prompt sit here.
- GANs. Largely displaced for open-ended synthesis, still the default for upscaling, face restoration, and other narrow one-shot transforms where speed matters.
- Flow and consistency models. The current push toward the same quality in one to four steps instead of thirty, which is what makes near-real-time preview possible.
What changes in a real workflow
Working with gen AI is closer to directing than to operating software. Software returns the same output for the same input; a sampler returns a family of outputs. The practical adjustments:
Budget for batches, not takes. A single run is a coin flip. Generate four to eight, then select. Cheap models you can run eight times often beat one run of an expensive one.
Move control out of the prompt when the prompt stops working. Words are a weak steering wheel. If the same problem survives three prompt rewrites, the fix is structural: a reference image, a first frame, a mask, a control input, or a fine-tune.
Version what you can reproduce. Model name, model version, seed, and the exact prompt string. Without those four you cannot get back to a shot you liked, and model versions do get retired.
Treat cost as part of the creative decision. Every generation is metered compute. Iteration strategy (how many samples, at what resolution, before you commit to an upscale) is a real part of the craft.
Where it stops
Generative AI is weakest exactly where production is strictest: exact text on a sign, consistent identity across a sequence, hands and joints, physical causality, and anything that needs to be true rather than plausible. None of those are fixed by a longer prompt. They are fixed by constraining the sample, or by doing that part in a compositing tool afterward and letting the model do what it is genuinely good at, which is producing a plausible world quickly.