Rules-based and generative AI systems
Medical-writing software can apply explicit expert rules, generate language from learned patterns, or combine both approaches. Understanding the distinction helps explain why tools differ in predictability, flexibility and the input they require.
System types
Artificial intelligence is the ability of a digital computer or computer-controlled robot to perform tasks commonly requiring human intelligence. Machine learning is a subset of AI that uses statistical methods to improve a model from experience. Deep learning is a subset of machine learning using large, multilayered neural networks.
Medical-writing software can use classic programming, symbolic AI or generative AI, alone or in combination. Classic programming provides functions such as buttons and menus and includes structured-content tools that store text blocks for reuse. Symbolic AI, also called an expert system or deterministic AI, applies rules and usually requires an extensive knowledge base developed for a particular task.
Expert systems resemble highly detailed templates: they incorporate extensive human knowledge and produce predictable output, but require structured inputs and adapt less readily to different tasks. Large language models resemble open-ended templates: their probabilistic output is more flexible but can vary and requires human oversight for accuracy and context. Generative AI predicts responses from patterns learned in large datasets, so exact reproduction can be difficult and errors are possible.
Natural-language generation creates text from structured data, while natural-language processing can turn text into structured data. Medical-writing automation can generate text from data or other sources, curate premade content from approved sources, or combine these functions. A tool can also combine technologies, for example using symbolic AI to generate text and generative AI to summarise or enhance it.
Model training
A language model uses training data to estimate the probability of the next word or sequence of words. Large language models are part of generative AI and are built through machine learning, specifically deep learning with neural networks. A transformer training approach described in 2022 combined initial learning from large amounts of unlabelled data with subsequent supervised learning from smaller labelled datasets.
Technical vocabulary
A token is the smallest unit of text a language model processes, often a word or part of a word. A 2025 example using the GPT-4o tokeniser split “Learning the tech lingo” into five tokens. The context window is the amount of text the model can process at once; a long regulatory document may exceed it.
Retrieval-augmented generation supplies relevant source material within a prompt to improve factual accuracy, selecting needed chunks from longer content. It retrieves relevant snippets from supplied or internet sources before generating a response that uses them. Multimodal models process images, video or audio as well as text and can help interpret figures and schemas. Hallucinations are incorrect content generated by a language model.
A bot is a program that executes commands or automates tasks. Assisted authoring generates human-language text from data inputs. Structured content authoring controls content through predefined rules, while digital data flow transfers information between systems or platforms without manual entry into documents.