Artificial intelligence is being used by business and government for writing, and that is not going to change. Generative artificial intelligence, especially large language models, create content quickly and cheaply. This doesn’t mean these models are as useful for writing as you might hope. Here’s why.

Security. This is the obvious point. If your topic is sensitive, private or classified, you should not use publicly available tools like ChatGPT. These tools are focused on speed and convenience, not security. Even if your organisation has an in-house large language model, you still need to be careful about sharing sensitive information with it.

Currency. Large language models have been pre-trained on a vast number of parameters, but those parameters might not include everything relevant to your topic. For example, the pre-training won’t include the events from this morning that have triggered a ministerial brief. You can help by providing the model with detailed context and prompts, but this starts to erode the convenience of these models.

Efficiency. A document written by a large language model will only be as good as the prompts you give it. These prompts will determine the scope and structure of the document, the prioritisation of key messages and the level of detail in the document. Not all information is of equal value, and without good prompts, the model won’t know what messages to emphasise or leave out. You might need to write several prompts – leading to several drafts. By the time you write these prompts, it might have been quicker to write the document yourself.

We’re not against using artificial intelligence to help writing at work. It can help with low-risk, high volume writing tasks. It is also good for narrowly defined topics. However, writing for work requires important decisions about the purpose of your document, the amount of detail needed by your reader, and the actions required of your reader. This is most of the work when writing, and large language models can’t decide these things for you. Large language models are good at the easy part of writing – choosing which words go together. Humans are still needed for the hard stuff.