What is generative AI?

Definition

What is generative AI?

Generative AI is a subfield of artificial intelligence that creates new content on its own. Unlike analytical AI, which classifies data or makes predictions, generative AI produces text, images, audio, video and code.

Generative vs. analytical

Analytical AI classifies – it detects spam, assesses risks or sorts images. Generative AI creates something that did not exist before. It writes email drafts, generates product images or produces program code.

The abbreviation GenAI has become established as the short form. Both approaches are based on machine learning but differ in their goal.

Breakthrough through ChatGPT

OpenAI released ChatGPT at the end of 2022 and turned generative AI into a mass phenomenon. Within two months the service reached 100 million users.

Since then, Google, Anthropic, Meta and dozens of other companies have released their own generative models. The market has been growing at triple-digit rates ever since.

Technology

How does generative AI work?

Generative AI learns statistical patterns from large datasets and uses them to create new content. The basis is machine learning – algorithms that learn from data instead of being hard-coded.

Training and training data

Generative models are trained on billions of data points – text, images, code repositories. In the process the model learns to predict the next element in a sequence: the next word, the next pixel, the next line of code.

This process requires thousands of GPUs and takes weeks. The training costs for a large foundation model are in the tens of millions.

Transformer architecture

Most generative models are based on the transformer architecture – a neural network that Google researchers published in 2017. Transformers process data in parallel rather than sequentially and thereby capture long-range relationships.

The central mechanism is called attention. It determines which parts of an input are relevant for the output. GPT, Claude, Gemini and Llama all use this architecture.

Foundation models and fine-tuning

Foundation models are pre-trained base models that serve as the basis for various applications. A single model can translate texts, answer questions and write code.

Through fine-tuning, companies adapt these models to their own data and tasks. Alternatively, prompts – textual instructions – control the behaviour of the model without additional training.

Distinction

What distinguishes generative AI from classic AI?

Classic AI analyses and classifies existing data – generative AI creates new content. Both belong to artificial intelligence but pursue different goals.

Discriminative vs. generative

Discriminative models distinguish between categories. They detect spam, assess credit risks or sort images. The result is a decision: yes or no, category A or B.

Generative models produce outputs. They write texts, create images or produce audio. The result is a new artefact.

Data processing

Classic AI needs structured, labelled training data. A spam filter needs thousands of emails manually marked as “spam” or “not spam”.

Generative AI learns from unstructured raw data. A language model reads billions of web pages without anyone categorising the content individually.

Interaction in practice

Both approaches complement each other. A customer service system uses classic AI to categorise incoming enquiries and generative AI to formulate the answers.

The line blurs with modern models. GPT-5.4 and Gemini 3.1 can analyse and generate – depending on what the prompt requires.

Language models

What role do large language models play?

Large language models are the most important category of generative AI for text tasks. They process and generate natural language at a level that comes close to human communication.

Transformers as the foundation

All leading LLMs are based on the transformer architecture. They were trained with hundreds of billions of parameters and can summarise texts, translate, answer questions and write code.

Model size alone does not determine quality. Architecture, training data and fine-tuning are at least as decisive.

The leading LLMs in 2026

GPT-5.4 from OpenAI routes requests internally between a fast and a deeper reasoning model. Claude Opus 4.6 from Anthropic relies on a context window of one million tokens and precise instruction following.

Gemini 3.1 Pro from Google uses a mixture-of-experts architecture and processes text, image, audio and video natively. Llama 4 Maverick from Meta is the strongest open-source model with 17 billion active parameters and 128 experts.

Practice

What examples of generative AI are there?

Generative AI includes specialised models for text, image, code, audio and video. Each category has its own architectures and market leaders.

Text generation

ChatGPT (OpenAI) is the most used generative AI application worldwide. It answers questions, writes texts and analyses documents based on GPT-5.4.

Claude (Anthropic) processes contexts of more than one million tokens and is suitable for long documents and complex analyses. Gemini (Google) is integrated directly into Google Workspace.

Image generation

GPT Image 1.5 from OpenAI leads the current benchmarks and generates photorealistic images with precise text rendering. Midjourney V8 generates in native 2K resolution and is five times faster than the previous version.

Flux 2 from Black Forest Labs delivers the most realistic skin textures. Stable Diffusion 3.5 remains the strongest open-source model and runs on your own hardware.

Code assistants

GitHub Copilot suggests code completions in real time and generates entire functions from comments. Developers report productivity gains of between 20 and 50 percent.

Claude Code from Anthropic works directly in the terminal and takes on complex programming tasks including refactoring and debugging. Cursor combines a code editor with AI-supported code generation.

Audio and video

Suno creates complete pieces of music including vocals from text descriptions. ElevenLabs generates natural-sounding speech and can clone voices.

Runway Gen-4.5 is currently the leading model for AI video production and is used in the professional film and advertising industry for special effects and scene generation.

Fields of application

What can generative AI do for companies?

Generative AI automates knowledge-intensive tasks and gives employees time for strategic work. Its use ranges from content creation and customer service to software development.

Marketing and content

Generative AI creates drafts for blog articles, social media posts, product descriptions and email campaigns. Editorial review stays with humans – the model does the groundwork.

Image generation delivers campaign visuals and visualisations without photo shoots. Companies use it to test concepts before releasing budget for final production.

Customer service and communication

AI-supported chatbots answer standard enquiries around the clock in natural language. That reduces processing time and relieves service teams of routine questions.

Generative AI summarises long email threads, creates meeting minutes and drafts replies. The human checks and sends.

Software development

Code assistants suggest functions, write tests and document existing code. They speed up the entire development cycle from prototype to deployment.

In data analysis, generative AI creates SQL queries from natural-language questions and produces reports from raw data.

Limits

What risks does generative AI bring?

Generative AI produces convincing content that can be wrong, biased or legally problematic. Responsible use requires clear guardrails.

Hallucinations

Hallucinations arise when a model invents facts that sound plausible but are false. The model generates statistically likely text – it has no concept of truth.

In sensitive areas such as law, medicine or finance, humans must check the outputs. Retrieval-augmented generation (RAG) reduces hallucinations by giving the model access to verified sources.

Bias in training data

Bias arises when training data over-represents certain perspectives. Models trained mainly on English-language texts represent other cultures less well.

Companies should check AI outputs for distortions before they are published. Diverse training data and regular audits reduce the problem.

Data protection and the GDPR

Anyone who enters personal data into cloud-based AI tools must comply with the GDPR. Many models process data on servers outside the EU.

Alternatives are on-premise deployments with open-source models such as Llama 4, or API providers with European data centres and data processing agreements.

Copyright and deepfakes

Generative models were trained on copyrighted works. Whether their outputs enjoy protection or infringe rights is currently being decided by courts worldwide.

Deepfakes – AI-generated images, videos or audio that imitate real people – are used for disinformation and fraud. Companies need internal guidelines for AI use and labelling obligations for generated content.

FAQ

Frequently asked questions about generative AI