One thing that anyone working with ChatGPT will quickly notice is that it has a particular voice. It’s as Vauhini Vara observed, ChatGPT’s responses throughout its versions has grown increasingly sterile in many ways. However, ChatGPT is a Language Learning Model, meaning that it can potentially be trained, even with a smaller sample of writing. How does it adapt to writing in different styles when given a selection of writing to emulate? That’s for us to find out now.
In this quick experiment with the publicly available ChatGPT 3.5, I will use a sample of my own writing to see how ChatGPT deals with.
Giving ChatGPT Writing Samples
It is important to note however that a lot of creatives are supremely uncomfortable with AI using their works in any capacity. Imagine having something you’ve spent years on be slotted into a database for use without your prior consent. Even if legally it is still up in the air, I view it as a matter of decency that we respect their wishes when it comes to employing it for our own personal use. That is why I will be providing a small sample of my own writing that I feel comfortable enough letting the OpenAI database gobble up.
Another general rule is that applies to data in general is that the larger corpus, or body of writing you give ChatGPT, the better the outcomes usually are. I will be offering a relatively small sample size of my writing, which will likely affect the quality of the experiment.
ChatGPT writing a story without a reference
Here is what ChatGPT will output given a simple premise and no reference to emulate. If anyone is interested in the full text, do let me know, but we can already gleam the writing style from this. It is very fairytale-esque, with a conversational tone as if it is a story told by a campfire.
ChatGPT with a reference
I provided ChatGPT with a 1,000 word sample of a eastern-inspired fantasy draft to emulate. Here it is apparent that the writing style does shift to include more isolated and grounded descriptions of the scenery. ChatGPT effectively transferred over the flow of the narration, at least on a syntax level.
Below is a section of my writing sample that ChatGPT clearly draws heavily from in its opening paragraph.
Takeaways
The reading experience does certainly change and the version that draws from a reference reads as less generic, but the effect still feels surface level if the metric for success is emulating a style. You can tweak the language of ChatGPT to be as granular as you would like, but that still demands the kind of skillful editorial engagement that you would need to edit your own works. As far as emulating style is concerned, at least in an ethical capacity, ChatGPT is still not there yet.
It would be interesting to see if ChatGPT can be prompted to write to a theme, but prior experience with line-by-line prompting tells me that it may be more trouble than it’s worth. I do not see a future where ChatGPT can convincingly emulate authorial style in the context of a creative writing piece. There are many voices out there, but when we put the work in ourselves we still have our own for now it seems.