# Generative AI **TL;DR:** AI that produces new artifacts (text, images, code, audio) in response to a prompt, rather than classifying or predicting from existing data. Generative AI is the category name for models that produce something new. You give them a prompt and they hand back an artifact: a paragraph, an image, a function, a voiceover, a spreadsheet formula. That sounds obvious now, but it's a genuine break from the previous decade of machine learning, which was overwhelmingly about *discrimination* rather than *generation*. Older systems answered questions like "is this email spam?" or "which of these five categories does this image belong to?" Generative models answer "write me the email" and "draw me the image." The mechanism is the same statistical trick applied at scale. An LLM predicts the next token given everything before it. An image diffusion model starts from noise and iteratively denoises toward something that matches your description. Neither has a plan, a concept, or an intent in the way a human author does. They are producing the most plausible continuation given their training data and your prompt. That's the source of both the magic and the failure modes: fluent output arrives whether or not the underlying facts are right, which is why hallucination is a property of the technology and not a bug someone forgot to fix. Generative AI is a broader umbrella than "LLM." Every LLM is generative AI; not all generative AI is an LLM. Image models, music models, video models, protein-structure models, and code models all live under the same tent. What unites them is the shape of the interaction: prompt in, artifact out, single turn. Nothing in that loop takes an action in the world, evaluates its own output, or decides what to do next. That last point is the distinction people get wrong most often. Generative AI produces an artifact and stops. Agentic AI wraps that same generative core in a loop with tools, memory, and a goal, so it can plan, act, check the result, and try again. A model that writes you a SQL query is generative. A system that writes the query, runs it, notices the syntax error, fixes it, reruns, and reports the number is agentic. The generative model is the engine; the agent is the vehicle. Vendors blur this constantly because "generative AI" was the 2023 label and "agentic" is the 2025 one, so plenty of products got renamed without gaining a loop. For teams making architecture decisions, the practical consequence is about failure containment. Generative systems fail visibly and cheaply: you read the output, decide it's wrong, and try a different prompt. A human is in the loop by construction. Agentic systems fail expensively, because the model acted before anyone reviewed anything. If your use case is "draft it and I'll check it," you want generative and you should not pay the complexity tax of an agent framework. If the value is in the actions, not the artifact, you need the loop, and you need the governance that comes with it. The other practical consequence is context. A generative model knows exactly what you put in the prompt and nothing else. Every generative product that feels smart about you (remembers your writing style, knows your codebase conventions, uses your company's terminology) is doing context engineering behind the scenes, assembling the right background into each request. The model is a stateless function; the intelligence you experience across sessions is a retrieval and memory system feeding it. ## Why it matters "Generative AI" is the budget line most enterprise AI spend still sits under, and it is the term that gets stretched to cover everything from a summarization endpoint to a fully autonomous agent fleet. Being precise about it changes what you buy and what you staff for. Generative use cases need prompt quality, review workflows, and cost-per-token discipline. Agentic use cases need permissions, audit logs, evaluation harnesses, and memory infrastructure. Teams that budget for the first and deploy the second are the ones that end up with an incident. ## Example A marketing team uses a generative AI tool to draft product descriptions. They paste in specs, the model returns three variants, a human picks one and edits it. That's generative AI: one prompt, one artifact, human approval built into the workflow. Six months later they want the system to pull live inventory data, check which SKUs lack descriptions, write them, push drafts to the CMS, and flag the ones it was unsure about. Same underlying model, completely different system: it now needs tools, a loop, memory of what it has already processed, and someone accountable for what it publishes. ## Related terms - [LLM (Large Language Model)](https://www.maximem.ai/glossary/llm) - [Agentic AI](https://www.maximem.ai/glossary/agentic-ai) - [AI Agent](https://www.maximem.ai/glossary/ai-agent) - [Hallucination](https://www.maximem.ai/glossary/hallucination) - [Prompt Engineering](https://www.maximem.ai/glossary/prompt-engineering) - [Context Engineering](https://www.maximem.ai/glossary/context-engineering) --- Source: [https://www.maximem.ai/glossary/generative-ai](https://www.maximem.ai/glossary/generative-ai)