I care a great deal about words. As an editor, much of my work involves helping people express what they mean more clearly, and my studies in forensic linguistics gave me another reason to pay attention to the language people choose. A slight change in wording can alter what a reader understands, sometimes without the writer realising they’ve suggested something different.
This probably explains my interest in large language models, or LLMs, and in helping people use them. We’re giving instructions to an automated system through the same language we use with other human beings, often without considering how much we expect the recipient to fill in for us.
Take the instruction, ‘Make this sound exciting’. It seems straightforward enough, particularly when we’re typing quickly and want a better version of something we’ve written.
We know what exciting means, so we assume we’ve given the system a useful direction.
Yet my version of exciting could be spending an afternoon examining disputed evidence from a criminal trial. Somebody else might find that tedious and reserve the word for jumping out of an aeroplane, which would require considerably more persuasion on my part. Even when we’re describing the same event, the feeling we attach to it may be very different.
An LLM can draw on patterns in how people describe excitement and use the surrounding context to produce a plausible response. Its training involves learning relationships in language and predicting what comes next, with further training shaping how it responds to instructions. That gives it considerable ability to write about an experience, but it doesn’t give it our personal reference point for that experience.
So, when I ask it to make something exciting, I’m interested in whose version I’m going to receive. Which associations will shape the response, and how closely will they resemble what I had in mind? Unless I’ve explained myself, there’s a considerable amount of room for interpretation.
This is how a perfectly reasonable request can produce a paragraph announcing that I’m ‘absolutely thrilled’ about something I’d probably describe as interesting. The wording is fluent, the emotion is recognisable - but I wouldn’t say it. If I publish it unchanged, I’ve allowed the system to decide how enthusiastic I appear to be.
The same difficulty appears when we ask for an ‘unusual’ idea. Unusual compared with what? Something that feels unfamiliar in one workplace could have been standard practice elsewhere for years, and an idea that surprises a beginner might be thoroughly predictable to someone who knows the subject.
An LLM can recognise common patterns and suggest departures from them, so it would be unfair to say it has no basis for making that comparison. What it doesn’t automatically have is the comparison I’m making. It needs some indication of what I’ve already encountered and why the familiar approaches don’t meet my needs.
‘Unique’ asks even more of it. A system can generate an unexpected combination, but that doesn’t establish that nobody has thought of it before. Describing an idea as unique is easy; having enough evidence to justify that description is a different matter.
I think we sometimes overlook this because these words feel specific to us. They contain meaning drawn from our own experience, and we bring that meaning to the prompt without necessarily putting it into the prompt. Then we’re disappointed when the response doesn’t reflect something we never explained.
This happens between people too, of course. An editor receiving the instruction ‘make it warmer’ may need to establish whether the author wants a less formal tone or a more personal acknowledgement of the reader’s situation. The word gives us somewhere to start, but a conversation helps establish what the writer actually wants.
With AI, the speed of the answer can encourage us to skip that conversation. We receive a complete draft almost immediately, which makes it tempting to start judging the wording before checking whether our instruction was clear enough to produce what we intended.
‘Make it sound human’ is a particularly interesting example. Human beings write in all sorts of ways, including ways we find irritating. Some people are effusive, whilst others would feel uncomfortable publishing anything that sounded remotely like that. Asking for a human voice leaves a great deal undecided.
My own preferences are quite particular. Explaining them, and showing examples of writing I’ve edited myself, gives the system something more useful to work with.
That doesn’t mean every prompt needs to become an essay. Sometimes replacing an adjective with an explanation is enough. Rather than asking AI to make a workshop description exciting, I could explain what participants will be able to do and why I think they’ll become absorbed in it. That gives the response a reason for the enthusiasm, rather than an invitation to add enthusiastic language.
This matters to me as an editor because clarification should help an author become more recognisable in their writing. A polished sentence can still be a poor edit if it changes someone’s opinion or makes them sound like a person they’d never recognise.
AI can help us work towards that clarity, provided we remain involved in deciding what fits. It draws on patterns learnt from other people’s language; it isn’t simply lifting a stranger’s sentence and handing it over. Even so, those patterns can lead us towards familiar expressions that gradually replace our own way of explaining things. Without clarification, generative AI will come to recognise ‘that’s near enough’ as the real you.
Diane Hall is the founder of Evolve3 CIC and an editor with a background in forensic linguistics. She helps people use AI more thoughtfully, with particular attention to language and the human judgement needed to assess what comes back.