# What is generative AI: how models make text, images and code

> Language models predict the next token and diffusion models turn noise into pictures. Here are the papers behind them, the products built on them and where they fail.

URL: https://claudegpt.si/news/what-is-generative-ai  
Published: 2026-11-18T14:59:20.730Z  
Author: ClaudeGPT Si Newsroom  
Section: Research  
Topics: Generative AI, Language models, Diffusion models, Transformers  
Publisher: ClaudeGPT Si

## Key takeaways

- NIST defines generative AI as models that learn the structure of their input data and generate new synthetic content, such as text, images, video and audio.
- Language models write by predicting the next token, while diffusion models make images by removing noise step by step.
- NIST calls hallucination a natural result of how generative models work, since they produce likely outputs rather than verified facts.

Generative AI is software that makes new content, such as text, images, code, video or audio, after learning the patterns in huge amounts of existing examples. The US standards body NIST, in its [generative AI profile](https://doi.org/10.6028/NIST.AI.600-1) published in July 2024, uses the federal definition: "the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content."

Put plainly, older AI mostly sorted things: spam or not spam, fraud or not fraud. Generative AI produces things. Ask ChatGPT for an email and it writes one. Ask an image model for a sailor on a stormy deck and it paints one. Under the hood, both are doing the same kind of thing: learning what typical data looks like, then producing a new sample that fits.

## How generative AI works

Every generative model is trained on examples and learns their statistical shape. What differs is how it produces something new. Two methods account for most of the products people use today.

### Text: predicting the next token

Chatbots run on [large language models](https://claudegpt.si/news/what-is-an-llm), or LLMs. An LLM reads text as [tokens](https://claudegpt.si/news/what-is-a-token-in-ai), small chunks of words, and is trained on one task: given the tokens so far, predict the next one. To answer a question, it predicts a token, adds it to the text, and repeats, one token at a time, until the reply is done.

The architecture behind nearly every LLM is the transformer, introduced in the 2017 Google paper ["Attention Is All You Need"](https://arxiv.org/abs/1706.03762) by Ashish Vaswani and colleagues. It dropped the step-by-step recurrent networks of the time for attention, a mechanism that lets every word in a passage look directly at every other word. The paper's translation model reached 28.4 BLEU on English to German, beating the previous best by more than 2 points, and its English to French model trained in 3.5 days on eight GPUs.

The next leap was scale. OpenAI's 2020 paper ["Language Models are Few-Shot Learners"](https://arxiv.org/abs/2005.14165) described GPT-3, a model with 175 billion parameters, about ten times more than any earlier non-sparse language model. Its finding: a big enough model can pick up a new task from a few examples in the prompt, with no extra training. It translated, answered questions and did 3-digit arithmetic, and the authors report that people often could not tell its news articles from human ones. Today's chatbots follow the same recipe: a large transformer trained to predict the next token.

### Images and video: turning noise into pictures

Most image and video generators use a different trick. A diffusion model, described in Jonathan Ho, Ajay Jain and Pieter Abbeel's 2020 paper ["Denoising Diffusion Probabilistic Models"](https://arxiv.org/abs/2006.11239), learns by watching images get buried in random noise, step by step, and practicing how to undo each step. To make a new image, it starts from pure noise and removes it a little at a time until a picture appears. On the CIFAR-10 image benchmark it scored an FID of 3.17, a measure of how close generated images are to real ones where lower is better, which the authors called state of the art.

Diffusion on raw pixels is slow and costly. The 2021 paper ["High-Resolution Image Synthesis with Latent Diffusion Models"](https://arxiv.org/abs/2112.10752) by Robin Rombach and colleagues ran the process on a compressed version of the image instead, which cut the compute needed, and added cross-attention layers so a text prompt can steer the result. That design is the basis of [Stable Diffusion](https://github.com/CompVis/stable-diffusion), which its makers describe as "a latent text-to-image diffusion model." The [diffusion model explainer](https://claudegpt.si/news/what-is-a-diffusion-model) covers it in more detail.

### Before diffusion: GANs

Before diffusion, the best-known image method was the generative adversarial network, or GAN, from Ian Goodfellow and colleagues' paper ["Generative Adversarial Networks"](https://arxiv.org/abs/1406.2661). A GAN trains two networks against each other. A generator makes fake samples, and a discriminator tries to tell them from real ones. Each pushes the other to get better, until the fakes are hard to spot.

## Examples of generative AI

Here is what generative AI makes today, with products described from their own pages as of November 18, 2026:

| What it makes | Example | What the maker says it does |
| --- | --- | --- |
| Text and chat | [ChatGPT](https://chatgpt.com/) (OpenAI) | Chat, web search, image creation, deep research and file uploads |
| Text and documents | [Claude](https://claude.com/product/overview) (Anthropic) | Drafts and edits writing, analyzes data, and creates documents, spreadsheets and slides |
| Code | [GitHub Copilot](https://github.com/features/copilot) | An AI coding agent that works from an issue through review to merging a pull request |
| Images | [Nano Banana](https://deepmind.google/models/gemini-image/) (Google DeepMind) | Image generation and editing built on Gemini, refined through conversation |
| Video | [Veo 3.1](https://deepmind.google/models/veo/) (Google DeepMind) | Video generation with native sound effects, ambient noise and dialogue, in 1080p and 4K |
| Music | [Suno](https://suno.com/about) | A browser-based studio and audio models for making songs |
| Speech | [ElevenLabs](https://elevenlabs.io/text-to-speech) | Text to speech in 90+ languages with its newest models |

The newest step is [agentic AI](https://claudegpt.si/news/what-is-agentic-ai): generative models that do not stop at producing an answer but take actions, such as running code or editing files, to finish a task.

## What generative AI gets wrong

The same design that makes these models useful makes them unreliable in specific ways. NIST's profile lists 12 risks. Three matter most for everyday use.

**Hallucination.** NIST calls it confabulation: "the production of confidently stated but erroneous or false content." It explains that this is "a natural result of the way generative models are designed." An LLM predicts likely next tokens, not true ones. Usually likely and true overlap. When they don't, the model states the wrong answer in the same fluent tone as the right one. NIST notes the problem is worst for open-ended, long answers and for topics that need expert knowledge, such as law. Check any fact that matters.

**Cost and energy.** Training and running these models takes large amounts of computing. NIST cites estimates that training one transformer LLM can emit as much carbon as 300 round-trip flights between San Francisco and New York, and that generative tasks such as summarizing use more energy than simpler ones such as sorting text into categories.

**Bias.** Models learn from data that carries society's biases, and they can amplify them. NIST also flags "performance disparities" between groups and languages, often from training data that does not represent everyone equally, and a homogenizing effect, where outputs drift toward the same styles and views.

## Generative AI vs other AI

Artificial intelligence is the broad field of making computers do tasks that seem to need intelligence. Generative AI is one part of it. Machine learning, where systems learn patterns from data instead of following hand-written rules, sits in between: most generative AI is machine learning, but most machine learning is not generative.

The useful line is between models that predict a label and models that produce content. NIST draws the same contrast when it calls text summarization a generative task and text classification a discriminative, or non-generative, one. A bank's fraud detector, a photo app that tags faces, or a recommender that picks your next video all choose from options or score inputs. They don't write anything new.

## Why it matters

The core ideas are older than the products: GANs date to 2014, transformers to 2017, modern diffusion to 2020. What changed was scale and the chat interface on top. The limits are just as built-in. A system that predicts the likely next word will sometimes predict a likely falsehood, so it works best as a fast first draft that a person checks.

## Frequently asked questions

### What is the difference between generative AI and AI?

AI is the whole field of making computers do tasks that seem to need intelligence. Generative AI is the part of it that creates new content, such as text or images, while most other AI classifies, scores or recommends from existing options.

### What are examples of generative AI?

Chatbots such as ChatGPT and Claude write text, GitHub Copilot writes code, Google DeepMind's Nano Banana makes images and Veo 3.1 makes video with sound. Suno makes music and ElevenLabs turns text into speech.

### Is ChatGPT generative AI?

Yes. ChatGPT runs on large language models that generate replies one token at a time by predicting what comes next, and it can also create images.

### Why does generative AI make things up?

Generative models produce outputs that match the statistical patterns of their training data, so a language model picks likely words, not checked facts. NIST calls the result confabulation, known widely as hallucination.

## Sources

- [Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)](https://doi.org/10.6028/NIST.AI.600-1), NIST
- [Attention Is All You Need](https://arxiv.org/abs/1706.03762), arXiv (Vaswani et al.)
- [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165), arXiv (Brown et al.)
- [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661), arXiv (Goodfellow et al.)
- [Denoising Diffusion Probabilistic Models](https://arxiv.org/abs/2006.11239), arXiv (Ho, Jain and Abbeel)
- [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752), arXiv (Rombach et al.)
- [Stable Diffusion repository](https://github.com/CompVis/stable-diffusion), CompVis
- [ChatGPT](https://chatgpt.com/), OpenAI
- [Claude overview](https://claude.com/product/overview), Anthropic
- [GitHub Copilot](https://github.com/features/copilot), GitHub
- [Nano Banana](https://deepmind.google/models/gemini-image/), Google DeepMind
- [Veo](https://deepmind.google/models/veo/), Google DeepMind
- [About Suno](https://suno.com/about), Suno
- [Text to Speech](https://elevenlabs.io/text-to-speech), ElevenLabs
