Generative AI: The Technology That Learned to Create

In our last post we established that AI is not one thing. It is a spectrum of technologies, from the algorithm that filters your spam to systems that can analyse medical scans and make decisions that affect people's lives. This piece focuses on a specific and significant point on that spectrum: generative AI, the technology behind Large Language Models (LLMs) like ChatGPT, Claude, Gemini, and a growing number of applications that are reshaping how organisations work, how content is produced, and how decisions get made.

Generative AI is, at its core, a system that creates. It generates text, images, audio, video, and code in response to instructions. What makes it different from earlier AI is not just what it produces, but how it learned to produce it. These systems were trained on vast quantities of human-generated content: books, articles, websites, social media posts, images, recorded speech, scientific papers, legal documents, and more. The scale is genuinely difficult to comprehend. We are talking about a significant proportion of everything humanity has ever written, published, or posted online being absorbed into a system designed to learn the patterns within it.

The analogy that comes closest is this: imagine the most widely read person who has ever existed, someone who has absorbed the entire output of human knowledge and expression, but who has never actually lived a single day of human experience. They can construct a technically perfect sentence about grief without ever having felt it. They can describe the texture of injustice without ever having experienced it. That is both the extraordinary capability and the fundamental limitation of generative AI. They are not tools which are able to ‘think’. They are large language systems which have an incredible ability to predict the next likely set of words based on the probability, weighting and the datasets that they have been trained on.

The mechanism is a little like predictive text on your phone, but scaled to a degree that makes the comparison almost absurd. When you type a message and your phone suggests the next word, it is drawing on patterns from your previous messages. Generative AI draws on patterns from the entirety of recorded human language and does so with a sophistication that produces outputs which can be genuinely indistinguishable from human writing. Anyone who has encountered an autocorrect fail, the phone confidently inserting entirely the wrong word at entirely the wrong moment, has a small window into what happens when pattern-matching goes wrong without context or understanding. We will come back to that in the next piece, because it matters enormously when these systems are being used to make real decisions about real people.

What generative AI has changed, practically speaking, is the barrier to creation. Writing, coding, designing, analysing, translating: tasks that previously required specialist skills, significant time, or both, can now be initiated by anyone with access to a tool and an internet connection. That democratisation of capability is genuinely significant and, approached thoughtfully, genuinely useful. 

However, it also comes with questions that neither the law nor the industry has yet answered satisfactorily. The creative industries have raised urgent concerns about the use of their work to train these systems, often without consent or compensation. In February 2025, over a thousand artists released a silent album in protest at UK government proposals they described as the legalisation of music theft. The legal position in the UK remains unresolved. Under current UK law, it is not entirely clear whether wholly AI-generated content benefits from copyright protection at all, and Parliament is actively considering removing that protection entirely. The question of who owns what, when a human and an AI produce something together, remains open, and they sit at the heart of what kind of creative economy and knowledge ecosystem we want to build.

There is also a consent question that extends well beyond the creative industries. Generative AI is already being used across sectors you engage with every day, whether you have chosen to interact with it directly or not. Financial institutions use it to assess creditworthiness. Employers use it to screen job applications. Healthcare systems are beginning to use it to support clinical decisions. In England, Palantir, a US data analytics firm, holds contracts with at least ten government departments including the NHS, processing data through systems that most patients neither know about nor have a meaningful mechanism to decline. The question of whether you should be told, clearly and in advance, when an AI system has played a role in a decision that affects you is a governance question that we have not yet resolved. Burying that information in terms and conditions is not transparency or treating customers fairly. It is an attempt to mitigate legal liability, which is a different thing entirely.

These are real concerns that deserve real consideration. But there are also some fantastic examples of how generative AI can really support industries and drive innovation. The University of Sussex for instance is testing AI-generated classes to support trainee teachers. New teachers often face their first real-world classroom without the ability to rewind or practice high-stakes interactions (like managing disruptive behavior or addressing complex student needs). By creating a sandbox in which they can role play and practice it enables student teachers to build their confidence and pedagogical intuition before they ever step in front of real children.  

UNICEF has pioneered a workflow that addresses one of the most persistent bottlenecks in international development: the ‘knowledge silo’ problem. Humanitarian agencies produce vast amounts of internal reporting, but these documents are often too voluminous for human analysts to synthesise at scale. Researchers use generative AI to collate these documents and produce high-quality, structured executive summaries and thematic extractions, which can be used to inform policy. 

This is the reality of generative AI in 2026. It is not coming. It is here, operating in the background of systems most people use without thinking about them. The tools themselves are neither good nor bad. What matters is who builds them, on whose data, with whose consent, and with what accountability when they get things wrong.

Next up: What happens when AI gets it wrong and who is accountable for the harms caused?