Generative AI has rapidly become a technology that millions of people use every day. Applications such as ChatGPT, Copilot, and Claude answer questions, draft texts, and generate images. Yet most users have little understanding of what happens under the hood. That understanding is not optional: Article 4 of the EU AI Act (Regulation (EU) 2024/1689) requires organisations to ensure a sufficient level of AI literacy among everyone who uses or manages AI systems.
What is generative AI?
Generative AI is a class of artificial intelligence that can produce new content, such as text, images, audio, video, or code. This distinguishes generative AI from systems that only classify (e.g. detecting spam) or predict (e.g. calculating delivery costs). The EU AI Act describes the most capable generative models as "general-purpose AI models": systems trained on large amounts of data that can perform a wide range of tasks. Recital 99 of the regulation states explicitly: "Large generative AI models are a typical example of general-purpose AI models, as they can flexibly generate content such as text, audio, images or video, enabling them to perform a wide range of distinct tasks."
How does a large language model work?
A large language model (LLM) is a neural network trained on enormous quantities of text from the internet, books, and other sources. During training, the model learns statistically which words or word fragments (tokens) are likely to follow one another. The model does not store texts as memories; instead, it builds internal connections known as parameters. Modern models contain hundreds of billions of parameters.
When a user asks a question, the model uses those parameters to calculate which tokens are most likely to form a coherent and relevant answer. It generates text one token at a time, with each step conditioned on what has already been generated. This process is called autoregressive decoding. The result reads as fluent language, even though it is technically a sequence of probability calculations.
How does an image generator work?
Image generators operate on a different principle. The most common approach is diffusion: the model first learns what images look like by systematically adding noise to existing photographs and illustrations. It then learns to reverse this process, building a coherent image from random noise in a way that matches a text description. Models such as Stable Diffusion and the image generators embedded in Copilot and ChatGPT follow this principle. The text the user enters guides the generation process via a language model.
What is a prompt?
A prompt is the input a user provides to a generative AI system: a question, an instruction, context, or a combination of the three. The quality of the prompt largely determines the quality of the output. A vague prompt produces vague answers; a precise prompt with clear context, a specified tone, and defined boundaries produces more useful results. This is not a minor detail: the AI literacy obligations under the AI Act imply that users understand how to instruct AI systems effectively and responsibly, which requires direct knowledge of how to construct a prompt.
Why does AI hallucinate?
Hallucinations are outputs that are factually incorrect, invented, or misleading, presented with the same apparent confidence as accurate responses. They arise from the way language models are trained: models learn the most probable continuation of text, not the distinction between true and false. Facts that appear rarely in the training data are harder to reproduce reliably than language patterns that are consistently present.
Research by OpenAI (September 2025) shows that standard evaluation methods compound the problem: models are rewarded for providing an answer even when uncertain, which makes guessing strategically preferable to acknowledging that the answer is unknown. Hallucinations are therefore not a random error but a systematic consequence of how models are trained and assessed.
For responsible use in a professional setting, this means AI-generated output must always be verified against primary sources, particularly for factual claims, figures, and legal or medical information.
What does this mean for responsible use?
Understanding how generative AI works is a prerequisite for using it responsibly. Those who grasp that an LLM is not a search engine but a pattern-recognition system will scrutinise its output more critically. Those who understand how a prompt shapes output will communicate more effectively with AI tools. And those who understand why hallucinations occur will take the appropriate precautions when using AI in professional contexts.
Article 4 of the EU AI Act makes this a formal obligation, not a matter of personal preference. Providers and deployers are required to take measures ensuring that everyone who uses or manages AI systems possesses a sufficient level of AI literacy. A foundational understanding of how generative AI works is at the core of that obligation.