Experiments
How to make ChatGPT sound like you or any other celebrity
August 2023Teaching ChatGPT to talk like you
ChatGPT is busy right now. Literally. Log-in attempts are often met with “at capacity” messages because so many users are plugging in, asking it to write essays, generate ideas, write speeches, and all the other things people do with it.
Since the last time we wrote about AI copywriting, it has come a long way and has gotten much better at sounding human. At elespacio we’ve experimented with it for everything from keyword research to writing a video script in the style of Snoop Dogg, and it’s been a whole lot of fun.
“Will ChatGPT ever be able to knock up a convincing social post or wedding speech?”
But, clever as it is, you can still tell if something has been written by an AI bot rather than a person. You can spot that oddly impersonal tone a mile away. It will obligingly churn out text in the style of Snoop Dogg or Shakespeare if you ask it to, but only Snoop Dogg or Shakespeare can comment on how accurate the result is. And it has yet to really understand what humour is.
We wanted to know – is it possible to train AI to write like an actual living human, more specifically you, with all your peculiar writing habits? Will we get to the point where ChatGPT can knock up a convincing social post or a wedding speech, giving you all the credit but saving you the time and effort?
A little Googling brought up plenty of advice on how to train ChatGPT to write in a specific tone of voice. This is an abridged version, minus all the false starts, wrong turns and trials and errors, of how we tried to turn ChatGPT into an elespacio copywriter.
1. Tell it what to do
The first task is to ask ChatGPT to define the tone of voice and writing style, based on some text that you’ve previously written. But before that, it helps to tell it exactly what you want it to do and exactly how to do it.
We wrote this prompt (edited for brevity’s sake): “Hi! I want you to analyse the writing style, sentence structure, personality and tone of voice of several examples of text that I will give you. After each time I give you an example, say 'Thank you. Give me more". Do not tell me your analysis after each example. Wait until I say "All done" to let you know I've finished giving you examples. Then please tell me your analysis of all the texts.»
Thus prompted, ChatGPT should sit up and tell you it’s ready to go to work, with something like this: “Please go ahead and share the first text with me.”
2. Feed the beast
We then started giving it examples of text that we’d published, including short social media posts and longer articles and case studies. We tried to choose text that represented the elespacio tone of voice in all its quirky, pun-making glory, and we fed it as much as we could before we ran out of steam.
3. Get the analysis
Once we’d finished feeding the examples, we said “All done” and ChatGPT got to work analysing the style. Here’s a short version of what it came up with.
4. Give it a name
The analysis seemed fair enough, if a little vague. In retrospect, this should have been the point where we sent it back to the drawing board with more specific information and examples of our writing style. But you live and learn.
So we asked ChatGPT to save it. We said: “Thank you, that's great! Can we give this style of writing a name so that you will remember it in the future? Let's call it ETOV 1.”
5. Ask it to write something for you in that style
Next we asked ChatGPT to generate some content in the style it just described – i.e. in the ETOV 1 tone of voice.
We used the prompt: “Please write a short social media post for Instagram in the ETOV 1 tone of voice, summarizing the following text:” and we pasted the intro to an article.
6. Refine, refine, refine
What it came up with was…. nothing like the elespacio style. So we got to work pushing it as hard as we could, asking it to refine, adjust and rewrite.
We used prompts such as “Please write it in a more understated tone”, or “Please write it in under 50 words”, “Please don’t use exclamation marks”, and the slightly exasperated “Please be funnier”.
7. Success of a sort
After much back and forth, ChatGPT wrote a little social media post that could, just possibly, have been written by us:
"Let's unlock the potential of brand marketing and performance marketing and show them what happens when they ditch the rivalry and join forces, making other marketing strategies green with envy."
It was just one sentence and probably a fluke, but at this stage, we were ready to take it. So we gave ChatGPT a little pat on the back: “Thank you. That is great. That is a good example of the ETOV 1 tone of voice.”
The flush of success was shortlived, however, and the following attempts to get it to replicate this style weren’t so fruitful. If we had endless time on this earth, we would have repeated and tweaked the process again and again until we nailed it. But life is short.
No rest for the real human copywriter
So at the end of this experiment, would we trust ChatGPT to go off and create content in our tone of voice? Can the copywriter retreat to her beachside villa and spend the rest of her days prompting ChatGPT with one hand while holding a pina colada in the other? Sadly, no.
We managed to make ChatGPT come closer to sounding like us. But not close enough and certainly not consistently. Maybe after more tweaking and prompting and endless feeding of text, it might have come up with a passable impersonation. But it is quicker just to write the stuff yourself.
Being you is a complicated thing
“AI needs huge amounts of data to recognise language patterns, whereas humans are born to do it”
The whole exercise has been fascinating, however. The problems you come up against when trying to teach ChatGPT to write in your style are the same as the ones when teaching a human.
First, it is very difficult to describe words with words. One person’s idea of ‘funny’ is another’s idea of ‘cheesy’. What does ‘friendly’ sound like? It’s all subjective.
And second, it’s not enough just to identify the linguistic devices and figures of speech you use – most writers use metaphors, for example. What makes you unique is how, when, and how often you use them, and that’s more complicated. AI needs to be fed a huge amount of data for it to pick up those subtle language patterns, whereas a human is, in a very literal sense, born to do it.
Still friends with ChatGPT
To be fair to ChatGPT, it was never designed to be a content generator. It is, however, useful for all manner of writing task – repurposing content, writing ad variations, coming up with ideas, creating customer personas, suggesting synonyms, writing Snoop Dogg raps – the list is long.
So we will keep experimenting and having fun with ChatGTP. After all, given the rate that the technology is evolving, it could only be a matter of time before the copywriter can start packing her bags.
What we learned (and what we’d do differently):
1. It needs lots of help. ChatGPT is more likely to understand your writing style if you do. If you can pin down your writing habits, find the vocabulary to describe them – and go deep into the subtleties of how and when you use them, you might have more success than we did. Good luck.
2. It needs a lot of feeding. ChatGPT works by recognising patterns in data, so the more examples of writing you give it, the better.
3. It likes to know what to expect. Explain what you’re going to do and how and when you want it to respond. It saves you time and avoids confusion.
4. It likes it when you’re firm. When giving prompts, be precise, clear and blunt. ChatGPT might seem intelligent, but really it’s a simple creature.
5. But it’s good to be polite. That said, we found it was easier for us when we said ‘please’ and ‘thankyou’. Somehow it helps you keep track of what you’re trying to do. Weird, we know, but true.
6. Try to get it right the first time. It’s worth taking time to set up a clear process and give it as much information as possible at the start. The more adjusting and refining and to-ing and fro-ing you do, the more likely you are to go round in linguistic circles of hell. ChatGPT will start analysing the tone of voice analysis, or rewriting its analysis in the tone of voice – there are all kinds of ways to get lost in this maze.
7. It has a short memory. ChatGPT doesn’t save things from chat-to-chat, so you’ll have to keep going back into the initial chat whenever you want to use that tone of voice. And it’s a good idea to save a copy of the tone-of-voice analysis elsewhere, just in case.
Illustration: Midjourney
