How Does Generative AI Work?
Generative AI is just a part of the larger field of machine learning, which in turn is part of the overarching field of artificial intelligence. Tools based on generative AI create new content in the form of text, images, videos, etc.
How Do AI Tools Work?
The various tools generate content based on the instructions they receive from users, known as the prompt. The quality of the instructions is crucial for the quality of the result. For example, one can specify the scope of the answer, the style for a text or an image, specify that certain things should be excluded, or that certain perspectives should be compared, etc. Depending on the complexity, the answer is usually delivered within a few seconds (for shorter texts) or minutes (for images or other larger tasks). If you want to adjust, refine, or develop the result, you can give follow-up instructions.
Questions that involve and can be answered with more strictly rule-based, formulaic text – program code, mathematical formulas – have good chances of getting good answers, but translations between different languages are also constantly improving, and overall, the result is often – though not always! – impressive.
Probability, Not Truth!
The basis for all this is that the so-called language models that have been developed have been trained to recognize patterns in unimaginably large amounts of material – text, images, etc. When a tool using one of these language models receives a prompt, the question is interpreted based on patterns: a question containing these words in this sequence is likely to be answered with these words in this sequence (or, for example, with an image with this appearance; images can also be part of the training data). The generated answer thus constitutes an assumption based on statistical probability – not on any actual understanding of the content. Each feedback on the result in the form of new prompts refines these assumptions.
However, the process can sometimes deliver answers that contain completely false statements, constructed to fulfill what is requested in the instructions. For example, if you ask a question and request an answer that includes references to relevant sources, the result may be a list where the names of the journals, year, and volume number are correct, but the specified articles are missing when you check the references. They are entirely made up and cannot be found anywhere else on the internet, while other references may turn out to be correct.
More and more tools, however, also function as search engines, and do not only search material in a closed language model. The interpretation of the prompt also makes them look for information from the open internet to answer the questions and prompts they receive. But exactly how an answer came about can be difficult or impossible to know, and the quality of the actual references given is not guaranteed to be good. Therefore, the basic principle for users remains to always see the result they get as raw material, rather than an answer: it must be critically reviewed, evaluated, and often adjusted before it can be used (or sometimes discarded!).
Future Development
Generative AI continues to develop. Already now, enhanced AI support is available in programs such as Word, PowerPoint, Excel, and similar. There are also already tools for automatic subtitling of videos, as well as what is said in online meetings, in real-time, etc. As more material is fed into the models, and more people use the tools, write more detailed prompts, and provide feedback on the results, their answers get better and better.
At the same time, one can also see how the large, general tools, such as Microsoft’s Copilot or Google’s Gemini, are getting competition from smaller, but more specialized tools for specific purposes. Language models can receive special training (so-called fine-tuning) for more specific purposes, such as providing feedback on academic texts, grading exams, or writing code. Tools can also be limited so that no training data or user data leaves the organization.