What Is Generative AI?

Something unusual is happening. A technology that was, until recently, a specialist research interest has become a general-purpose tool used by hundreds of millions of people. It writes, reasons, explains, translates, codes, and converses. It arrived not as a gradual improvement but as a step change — and the pace has not slowed.

For most people in professional life, the honest position is somewhere between curiosity and confusion. The tools are clearly useful. The underlying technology is obscure. The claims made about it range from sober to hysterical. Cutting through that noise to understand what this technology actually is — how it works, what it can reliably do, and where it falls short — is the foundation for using it well.

That is what this paper is for. No jargon that isn't explained. No hype in either direction. Just a clear account of what generative AI is, grounded in how the technology actually functions.

These three terms are used interchangeably in conversation, but they describe different things. Understanding the distinction helps enormously in understanding what today's tools can and cannot do.

Generative AI is notable because it produces original output, not just a classification or a prediction. It does not retrieve a pre-written answer from a database. It constructs a response, word by word, from patterns it absorbed during training. That is what makes it feel qualitatively different from earlier AI systems — and what makes it genuinely useful for open-ended creative and analytical work.

The current generation of AI tools did not appear from nowhere. They are the result of a seventy-year research arc, with a small number of pivotal moments that each changed the trajectory significantly.

A large language model is a neural network trained on a massive quantity of text to predict — with high accuracy — what comes next in a sequence of words. "Large" refers to the number of parameters (the learned numerical weights) inside the model. Modern LLMs contain hundreds of billions of parameters, each adjusted during training to improve prediction accuracy across trillions of words of text.

The scale is difficult to convey. To give it some shape:

The result of this process is a model that has absorbed the statistical structure of language at a scale no human could approach — and, in doing so, has internalised a remarkable amount about how the world works, how arguments are constructed, how code functions, and how facts relate to each other. This is not because anyone programmed it with these things. It is because they are present, implicitly, in the text it was trained on.

The analogy that holds best is this: if you read everything ever written, you would know a very great deal — not because anyone taught you, but because knowledge is embedded in language. An LLM has done the reading. What it has not done is understand it in the way humans do — a distinction with important practical implications, covered in section 6.

When you type a message to an AI and it responds, a specific sequence of operations occurs. Understanding this sequence dissolves a lot of the "magic" — and explains both the model's impressive capabilities and its characteristic failure modes.

Step one: tokenisation. Your text is not read as words. It is broken into tokens — chunks of characters that may be a whole word, part of a word, or a punctuation mark. This is how the model processes language internally.

Step two: attention. The transformer architecture processes all tokens simultaneously and calculates relationships between them. "Paris" relates to "agreement." "2015" relates to "signed." These relationships — encoded as attention weights — allow the model to understand context, not just isolated words.

Step three: prediction. The model produces a probability distribution across its entire vocabulary — tens of thousands of tokens — for what should come next. It selects the next token based on this distribution, then repeats the process, generating one token at a time until the response is complete.

This mechanism — prediction based on learned statistical patterns — is simultaneously the source of the model's remarkable capability and its most important limitation. It is very good at producing text that looks like the correct answer. It is not, in a meaningful sense, reasoning its way to truth. This is why LLMs can be fluently, confidently wrong — a phenomenon known as hallucination — and why verification of factual claims always matters.

The clearest mistake people make when starting with AI is treating it as either a search engine (factual lookup) or a human expert (reliable judgement). It is neither. Understanding what it genuinely does well, and where it reliably fails, determines whether you deploy it effectively or frustratingly.

The limitations are not reasons to avoid LLMs — they are design constraints. A system designed around these constraints, with appropriate verification steps and human oversight where it matters, is highly effective. A system that ignores them and treats the model as an infallible oracle will produce unreliable results and erode trust quickly.

Several distinct families of large language model are now widely deployed. They differ in size, architecture, capabilities, and the organisations that built them — but they share the same fundamental mechanism described above. Understanding the landscape helps in choosing the right tool for a given context.

The market is moving fast. Benchmarks that distinguish models today will be superseded by new releases within months. The practical guidance is to choose a model based on the specific task — some excel at code, others at long documents, others at instruction-following — rather than assuming one model is universally best. Most enterprise AI platforms give you the choice.

Several persistent misconceptions shape how people approach — and often misuse — these tools. Addressing them directly is more useful than leaving them to be discovered through frustration.

• It is searching the internet for answers. It is not. A standard LLM has no live internet connection. It generates responses from patterns learned during training. When it gives you a statistic or a citation, it has predicted what that information looks like — which is why specific facts should always be verified. (Models with web search tools are a different matter, covered in a later paper.)
• It understands what it writes. This is philosophically contested and practically important. The model predicts plausible continuations of text with extraordinary accuracy. Whether that constitutes understanding is an open question — but behaviourally, it does not understand in the way a human expert does. It has no beliefs, no intentions, and no model of the world. It has learned patterns.
• It is conscious or has feelings. It does not. LLMs produce text that describes emotions because they were trained on text written by humans describing emotions. The language is mimicked, not experienced.
• Bigger models are always better. Not for every task. Smaller models, properly prompted and given relevant context, often outperform larger ones on specific tasks — while being faster and cheaper to run. The right model is the one appropriate to the task, not the largest available.
• Once an AI says something is correct, it is. The model has no access to ground truth. It can be confidently wrong, and it cannot tell the difference from the inside. Confidence in the output does not imply accuracy. Verification always matters for consequential outputs.
• It will replace human judgement. For well-defined, repetitive language tasks — drafting, summarising, classifying — it reduces effort significantly. For tasks requiring contextual judgement, ethical reasoning, or accountability, it is a tool in human hands, not a replacement for those hands.

Generative AI has been described — with varying degrees of credibility — as the most significant technological shift since the internet, the industrial revolution, and the invention of the printing press. These comparisons are difficult to evaluate in the moment. What is observable is narrower and more useful: a specific class of work has become dramatically faster and cheaper to do.

Writing first drafts. Summarising long documents. Classifying and extracting information from text. Translating between languages. Generating functional code. Answering questions from a body of documents. Each of these tasks, done at professional quality, previously required skilled human time. Each of them can now be accomplished in seconds by a well-configured AI system at negligible marginal cost.

The organisations that treat this as a tactical time-saver — using AI to do one task slightly faster — will extract some value. The organisations that treat it as a structural opportunity — redesigning workflows around what AI can reliably do, reserving human effort for what it cannot — will extract a great deal more. The technology does not determine which category you fall into. The design choices do.