Whether for brainstorming, drafting or proofreading, various generative AI platforms have quickly established a foothold in the writing process for many in the learning and teaching community. As a result, a large amount of the content we share and read is now being shaped by GenAI writing style and conventions.
What does this do for (and to) our writing? What changes when decisions are split between humans and the high-powered algorithms of platforms like Copilot, ChatGPT or Claude? To better understand, it’s worth taking a look at how these programs generate written content.
A (very brief) overview of how LLMs work
Large language models (LLMs – programs like Copilot, ChatGPT and Claude) are trained on massive amounts of data scraped from the internet, and operate on systems of pattern recognition based on this collected data. Your text input is converted into numeric tokens, with results determined by statistical probabilities found in sequences of tokens. You can use the OpenAI Tokenizer to see what this looks like in action. GenAI platforms do not all use the same tokenizers, and some suggest that improvements for future models might mean replacing tokenizers with new technology more adept at processing human language.
Gaining some insight into how an LLM works can help you to better understand how it interprets your writing, why it might give you certain responses, and how to evaluate whether those responses are what is most appropriate for your purpose.
Identifying AI writing conventions
The outputs of GenAI programs are changing rapidly with new updates, and there is still limited research available on the stylistic tendencies of AI. Additionally, AI detectors are unreliable and there is no way to tell for sure if text is AI-generated just by reading it. But with the rapid adoption of GenAI, some signs are beginning to emerge. Here’s a non-exhaustive list – with a disclaimer that these elements do not necessarily mean something is AI-generated, but do often show up in AI-generated text.
- Excessive usage of em dash – this one remains controversial, and in fact when I was searching for material on this point, Google’s AI Overview told me this was “a myth”. Still, the sudden proliferation of unspaced em dashes in digital content has become increasingly noticeable.
- Reliance on particular sentence structures – as an example, a comparative sentence structure often appears in the AI-generated writing that I have come across, which follows the formula of “it’s not just X, it’s Y”, or “you’re not just doing __, you’re doing __”. While I haven’t found peer-reviewed evidence to support this point (yet), a quick browse of Reddit and Open AI forums shows it’s an issue frustrating many users.
- Repetition – alongside commentary about GenAI’s apparent propensity for words like ‘delve’ and ‘meticulous’, a group of data science researchers from Northeastern University argue that repetition of “specific patterns of nouns, verbs and adjectives” shows up more often in AI-generated text than in that authored by humans.
- Vagueness – AI-generated writing seemingly lacks the precision and specificity that one might find in a well written piece from a human who’s knowledgeable in the topic they’re writing about. Bill Hu suggests that ChatGPT produces writing that is often more “generic and ambiguous” than human writing.
- Hallucinations – false and inaccurate information remains a problem for AI-generated content.
Putting your voice at the forefront
Different writing tasks call for different approaches. The processes for writing up a quick project overview or some meeting notes versus writing an extended editorial piece that draws on either your personal experience or field of knowledge, are not the same. So, if you’re using GenAI it makes sense to adjust that usage based on the outcome you’re hoping to achieve. Here are some things to keep in mind when you’re writing:
- Sincerity and authenticity – difficult to imitate, best when coming directly from a human being.
- The attention economy – there is so much content flooding into our lives every day. If your work reads as if it is AI-generated and doesn’t establish itself as distinct, people may not pay attention to it.
- What your unique voice can add – personalising a piece can make it much more engaging, so don’t be afraid to express your point of view or argument for compelling writing.
- Integrity – be confident that you can claim the work as your own ideas, consider your process and how it might impact your opinion if you were looking at someone else’s work created in the same way.
- Avoiding homogenisation – researchers at M.I.T., Cornell, and Santa Clara have been investigating the effects that GenAI can have on influencing the thought process of users, and suggest that using GenAI for creative activities can produce more homogenised, similar outputs.
- Honing your craft – sometimes you’re just trying to finish a task, and not aiming to improve your skills. But if you are trying to improve your writing, interventions and corrections from AI may hinder your development.
Really nice piece Rhiannon! Next time someone tries to tell me writing is dead b/c GPT, I can list off these reasons why not.
Lot’s of em dash easter eggs too 😉
Thank you David, writing will never die! I love an elegant em dash, but am also a little glad that so far LLMs don’t seem to go for for my usual preference of the spaced en dash ; )