Generative AI

Generative AI refers to a class of machine learning models that can produce new content rather than simply classifying or predicting patterns in existing data. Where a traditional classifier might label an image as 'cat' or 'dog', a generative model can produce an entirely new image of a cat on request, write a story featuring a dog, or compose music with a particular emotional tone.

The category covers a wide range of architectures and modalities. In text, large language models like GPT-4, Claude, and Llama are generative systems: they produce new text one token at a time based on a learned distribution over language. In images, diffusion models like DALL-E and Stable Diffusion generate new images by learning to denoise gradually from random noise toward coherent images. In audio, voice synthesis and music generation models produce new audio content from text descriptions or style prompts.

Generative AI became commercially significant in late 2022 with the release of ChatGPT and Stable Diffusion, which demonstrated that generation quality had crossed a threshold where outputs were useful for real-world applications. Since then, the capabilities of generative AI systems have expanded rapidly across all modalities.

The development of reliable and safe generative AI requires extensive human evaluation. Models can produce text that is fluent but factually wrong, coherent but harmful, or useful for most tasks but occasionally problematic. Human trainers provide the feedback that teaches these systems to be consistently helpful and to avoid generating content that is misleading or harmful.

Generative AI is the economic driver behind the AI evaluation economy: the large-scale deployment of generative systems creates ongoing demand for human quality evaluation.