In brief
Generative AI is the collective term for models capable of creating new content. For example, you provide a question, description, image or sound clip, and the model generates an output. This could be a draft text, but also an illustration, a spoken explanation or a piece of software. ‘Generative’ refers to the creation of output; it does not mean that this output is automatically correct, original or suitable for publication.
How can a model create something new?
During training, a model learns patterns from examples. This knowledge is captured in model parameters. During use – also known as inference – a query is processed using those learnt patterns. A language model typically constructs a sequence of tokens; an image model can generate an image through successive steps. You therefore do not usually retrieve a stored answer. The model calculates a new output within the limits of its training and the context provided.
Generating is different from recognising
A system that recognises a document as an invoice is performing a classification task. A system that writes an accompanying email for that invoice is generating text. These tasks can occur within the same workflow. A model can also interpret images as well as write text. It helps to specify exactly what you need: recognition, lookup, calculation or creating something new. Not every step requires a generative model.
A practical example
A team wants to make internal work instructions more accessible. The software first retrieves the current instruction. A language model then creates a brief explanation for a new employee. Another model may, if necessary, create a supporting image or spoken version. The original instruction remains the starting point. An employee checks the explanation before it is distributed. In this way, the system supports the generation of existing knowledge without taking over responsibility for the content.
What determines a useful result?
The brief, available information, choice of model and desired output format all work together. A compelling example says little about all other briefs. Therefore, also test for missing information, differing phrasing and complex cases. For text, accuracy of content is key; for images, composition, for example; and for code, functionality is also important. A usable result meets the agreed requirements and fits within the process in which it is assessed or used.
Where do you start as an organisation?
Choose one recurring task and describe what input is available, what the outcome should be and who checks it. Compare the time and cost of the entire task, including post-processing. Then determine whether local processing, an external API or a combination is suitable. Generative AI is particularly useful when the new content fits in with your day-to-day work at the right time and in the right form.
From concept to practice
How AITJE applies this
AITJE It starts with the task you wish to improve and the information available for that purpose. We determine where generative AI adds value and which steps are better carried out using standard software. We then compare suitable model routes, the necessary oversight and the costs of the entire process. In this way, a concept, text or other output becomes part of a practical workflow, with clear decisions regarding local processing and any external services.
View advice and analysisFrom explanation to application
Would you like to know what this means for your work?
