What is generative AI?

The one-sentence version

Generative AI is a category of AI that produces new content, text, images, audio, code, structured data, by learning statistical patterns from existing content. Traditional AI classifies, predicts, or detects. Generative AI creates.

That distinction matters more than it sounds. A classification model gives you a label. A generative model gives you something that did not exist before the request. The downstream implications for what you can build, and for the risks you have to manage, are substantially different.

How it differs from traditional machine learning

Traditional ML systems learn to classify or predict. A spam filter classifies emails. A churn model predicts which customers will leave. A fraud detector identifies suspicious transactions. The output is a label, a score, or a decision.

Generative AI systems learn to produce. A language model produces text. An image model produces pixels. The output is content that did not exist before the request.

This changes the evaluation problem entirely. With a classifier you can measure accuracy directly: the model either got the label right or it did not. With a generative system, you have to decide what "right" means for open-ended output, which is harder and requires a different evaluation discipline. We covered that in what are LLM evals.

The main modalities

Text generation covers everything from answering questions to writing code to drafting contracts. It is the most mature and the most widely deployed category in enterprise settings.

Image generation produces photographs, illustrations, and diagrams from text descriptions. It appears most frequently in marketing, design, and training content workflows.

Audio generation includes voice synthesis, music, and sound effects. Voice synthesis in particular has become reliable enough for production use in customer support and internal tooling.

Video generation is maturing rapidly and already appears in enterprise workflows for marketing and training content, though quality and consistency requirements still limit the use cases where it ships without human review.

Code generation now powers a substantial fraction of software written by enterprises and is the category where productivity gains are most measurable. The impact on engineering team output is the most straightforward generative AI ROI story available right now.

Why 2023 was different

Generative AI is not new. Statistical language models have existed for decades. N-gram models, recurrent networks, and early transformers all generated text in some form. What changed around 2022 and 2023 was the intersection of three things at once.

Scale in training produced qualitative improvements in output quality that felt discontinuous rather than incremental. A model trained on enough data started to feel capable of general reasoning rather than narrow pattern matching.

Instruction tuning made models useful for general-purpose tasks. Earlier models were good at completing a given style of text. Instruction-tuned models could follow a directive, which made them useful for tasks that had not been anticipated at training time.

Broad API access let any organization deploy these systems without operating the underlying infrastructure. The combination made generative AI available to enterprise teams that previously had no path to it. You did not need a research team or a data center. You needed an API key and a production use case.

The gap between demo and production

The thing that surprises most enterprise teams is not that generative AI works. It is that making it work consistently is much harder than the initial demo suggests.

A model that produces excellent output 85 percent of the time is not production-ready if the other 15 percent produces output that damages your brand or gives a customer wrong information. The demo almost always hits the 85 percent case. The 15 percent case shows up under real traffic, with real users, asking things the demo never anticipated.

Building around that gap requires evaluation frameworks, guardrails, monitoring, and fallback logic that the demo never shows. This gap is the subject of most of what we publish here, and it is where most generative AI implementations break down. We covered it directly in why AI pilots fail.

Foundation models vs fine-tuned models

Foundation models are trained by AI labs on massive datasets and made available via API. Claude, GPT-4, Gemini, and Llama are foundation models. They are general-purpose: they can answer questions, write code, summarize documents, and do many other things without any customization.

Fine-tuned models start from a foundation and are further trained on domain-specific data to improve performance on a specific task. A model fine-tuned on medical notes will handle clinical language better than the base model. A model fine-tuned on legal documents will produce more accurate clause summaries.

Most enterprise teams build on foundation models directly, using retrieval and prompt engineering to handle domain specificity. Fine-tuning is appropriate when the foundation model consistently fails on a specific task and retrieval-based approaches have already been tried. We covered this decision in what is fine-tuning.

What enterprises actually build with it

Customer-facing chat and support automation is the most common starting point. Done well, it deflects a substantial volume of routine inquiries without the quality degradation that plagued earlier chatbot implementations.

Internal search and knowledge retrieval over company documents is often the highest-value application for knowledge-intensive organizations. Engineers, analysts, and support teams spend a large fraction of their time finding information. A well-built retrieval system cuts that significantly.

Automated drafting of contracts, reports, and communications reduces the time professionals spend on structured writing without removing human review from the loop.

Code assistance for engineering teams is measurable in ways most AI applications are not. Pull request volume, cycle time, and test coverage all move.

Data extraction from unstructured documents, PDFs, emails, forms, is an unglamorous category that produces some of the cleanest ROI numbers because the baseline is manual labor and the accuracy bar is concrete.

These applications share a common structure: a user or system submits a request, the AI retrieves relevant context, and the model generates a response that goes through some validation before reaching the end user. The details of that structure determine whether the system works in production. That structure is what we build.