Will LLMs replace technical writers?

Put simply, the great leap that artificial intelligence has made with large language models (LLM) is that it can now write any text. "Writing texts" is also the common perception of what technical writers do. This leads to the conclusion that all technical writers could be replaced by ChatGPT. Can these models really produce the texts we expect from technical writers?

Challenges and limitations of LLMs in technical communication

Factual accuracy plays a decisive role in product texts that explain the operation, functionality and maintenance of products. Two things come together here that are difficult for Large Language Models. LLMs do not have the device whose operation is being explained in front of them, for example. It does not even know of its existence. It is merely led into a context of text fragments by our questions and tasks ("Describe the connections of the PowerPump-5000").

If it is a well-known device about which a lot has been published on the Internet, there will be a lot of text fragments about it. In this case, the performance of the LLMs is impressive. If the device is new or less common, the best you will find in this context are text fragments about similar devices (other pumps and their connections). Generative AI then makes use of these fragments. On the one hand, this is an impressive transmission performance that suggests real intelligence, but it is usually completely wrong in factual terms - how is the AI supposed to know any better?

So you could say that LLMs are best at writing texts that already exist a thousand times. That's why they are often used for SEO texts. This ability is less useful in technical communication, unless it is a pure copycat product that offers the same functionalities as a thousand others.

This is not the end of the story either: serious AI players are already combining different AI techniques and integrating additional knowledge into the text generation of LLMs via knowledge graphs, for example. This is how we get the GenAI to write texts that match the device that customers have in front of them. And they provide answers that a service technician, for example, can rely on. At Empolis, we speak of Trustworthy AI.

The role of the LLMs is reduced in this set-up. We have to design the knowledge contribution ourselves, regardless of whether we contribute product knowledge and features from a knowledge graph or determine similar cases using case-based reasoning. Nevertheless, the combination of knowledge-based AI and LLMs is exciting.

Opportunities for technical writing

Especially if we expand the use case - from writing texts to conducting a dialog in which we understand what users are saying, provide a context for the conversation, present content precisely against this background, answer queries, etc. These are all skills that LLMs bring to the table when they "feel at home with the content".

Even if this combination brings its own challenges: It is worth investing energy here, because here we are at one of the great visions of technical communication: providing exactly the information that a user needs in a specific situation.

What does this mean for the work of technical editors? It could actually mean that less needs to be written. But many of the activities that have always made up technical communication will of course remain the same - compiling and checking information, consulting experts from the development department, etc. - after all, new information has to come from somewhere.

In the future, technical writing will be particularly challenged when it comes to structuring and organizing information - e.g. in the form of knowledge graphs, which are used to feed and control the LLM dialogues.