Pen pals

Advice for writing with large language models

This is a free post from Our Long Walk, my blog about the economic past, present, and future of South Africa. If you enjoy it and want to support more of my writing, please consider a paid subscription to the twice-weekly posts that include my columns, guest essays, interviews and summaries of the latest relevant research.

---

I gave my second-year economic history students a difficult assignment this previous semester: write an essay.

While students have been writing essays for centuries, today’s students face an extra challenge: they now have a digital co-author capable of doing the work for them. Welcome to the world of large language models like ChatGPT.

My advice to them was to use all available tools to write the best essay. The only constraints were the 1000-word limit, and that the topic had to be something related to economic history.

I’ve been grading student essays for fifteen years. This year was the worst. By far.

So let’s dissect what went wrong.

A surprisingly large number of students thought that ChatGPT (or similarly capable LLMs) could simply write the entire essay. They would, presumably, enter a topic – or perhaps they might even ask ChatGPT for a topic – and then ask it to produce 1000 words about said topic.

But how do I know that it was a machine rather than a man that wrote these essays? There are several tell-tale signs. LLMs are more likely to use certain words: ‘captivate’, ‘tapestry’, ‘resonate’, ‘testament’, and ‘delve’, for example. Ask it to write an essay and it is very likely to use the phrase ‘a complex and multifaceted narrative’. And the last paragraph will usually start with ‘In conclusion’. In bulleted lists (a no-go for an essay), LLMs tend to highlight the title of each paragraph.

The most obvious giveaway of a ChatGPT-created essay is the reference list. Many students did not even check to see whether the references they purportedly cited actually existed. (One student cited me as the author of a paper that I never wrote!)

Fortunately, most students did not succumb to the temptation of having a LLM write their entire essay. But it was clear that many used it for introductory or concluding paragraphs, or as a way to ‘fill out’ the essay when they ran out of ideas. These paragraphs were often easy to identify: the tone would suddenly change, the frequency of LLM words would suddenly increase, or there would be a sudden absence of grammatical and language errors.

What struck me most, though, was that students who used LLMs to write (a large part of) their essays failed to ask interesting questions. This year’s essays resembled Wikipedia entries more than scientific papers with a research question, literature review, hypothesis, method of analysis, results, and conclusions. And they were bland, boring and, to be totally honest, bereft of any original thought.

The reason, of course, is that large language models are designed to generate text that sounds plausible based on patterns in the data they were trained on, rather than to produce genuinely insightful analysis. [Hold on, let me paste the last two paragraphs into ChatGPT and ask it to complete the paragraph. Here goes.] They excel at mimicking the structure and style of human writing but lack the capacity for true critical thinking, creativity, and the development of new ideas. As a result, the essays they produce often lack depth and originality, leading to work that is technically sound but intellectually shallow.

The above is an example of how ChatGPT can be used to improve writing. The idea was mine, but it helped articulate my thoughts more clearly and efficiently. It provided a coherent continuation of my argument, allowing me to focus on refining my ideas rather than getting bogged down in the mechanics of writing.

While AI tools can be valuable aids in the writing process, they should complement, not replace, the creative efforts that are essential to producing meaningful academic work.

Indeed, academic researchers like Wharton School professor Ethan Mollick have emphasised the complementary nature of these tools. In a Wired article last year, Mollick noted how LLMs can 1) help you find the right word, like a thesaurus, 2) find inspiration, 3) do research, 4) offer creative options, and 5) review your work. None of this is to substitute the job of the author; instead, it is to make many of the mundane tasks of the author easier by providing additional resources and insights that enhance creativity and efficiency.

[Let’s use this as an experiment. I did not like how ChatGPT completed that final sentence. The use of two nouns or adjectives is another tell-tale sign of LLMs. To avoid this, simply ask it to minimise its use. Or ask it, as I did, to produce a sentence orthogonal to this one.]

None of this is to substitute the job of the author; instead, it is to make the job of the author easier by enhancing the overall writing process through efficient, accessible assistance.

[Yes, that’s much better.]

The good (and perhaps somewhat scary) news is that the LLMs we use today are the worst of their kind. Put differently, they will only improve. That means that in the not-too-distant future it will be impossible to tell whether a sentence is LLM-created or not.

This is wonderful news for those of us who write often. Soon, we can all be Stephen Kings, Barbara Cartlands or Wilbur Smiths, authors known for their prolific writing. I’m most excited about automating the increased reporting load now required of university faculty – true for many other professions as well. And it is not just writing: LLMs and their AI collaborators can already perform pretty advanced data analysis, create wonderful artwork (something I use frequently on this blog), read images and videos, and even produce original songs and films. [I pasted this column into ChatGPT, asking it for three verses and a chorus. I also asked for music styles. I then pasted the info into Suno.ai. Just. Listen. To. This.]

But all of this improvement creates an obvious problem for those of us in the business of assessing student performance. If students can indeed write their essays with the aid of LLMs, what is the point of assigning essays in the first place?

One option, as some History departments at South African universities have now implemented, is to have students write essays in person, with pen and paper. Of course, that eliminates the ability to use ChatGPT and other tools. What you assess is really the student’s own understanding, critical thinking, and ability to articulate ideas without external aid. [The Oxford comma is another ChatGPT signature.]

But this method also has shortcomings. It limits the scope of assignments to what can be accomplished in a short, supervised period, and may not reflect the student's true capabilities in a more flexible, real-world context. [The straight apostrophe is also classic ChatGPT.]

The point is that as proficiency with ChatGPT becomes increasingly valuable, students must learn to use these technologies sensibly and, dare I say it, ethically to avoid falling behind their peers. For example, Google has just announced Illuminate, which turns academic papers into radio shows. And there will soon be a gazillion other applications of LLMs, ranging from real-time translation to personalised tutoring, from data visualisation to virtual research assistants. It would be irresponsible for us as teachers to outsource this crucial aspect of their education.

So what is the alternative? As always, when a new technology disrupts the status quo, the best response is to embrace it with open eyes and patience. Have students play around with it. Play with them. Show them what you’ve learned and learn from them. Instead of just handing out essay topics, perhaps discuss the potential of these tools – and their limitations. This collaborative approach applies to everyone: managers and their teams, teachers and their pupils, parents and their children.

We are all students, all the time. Technology will not stop its advance. Better to exchange frequent letters with our newest pen pal, learning and growing together.

An edited version of this article was published on News24. Support more such writing by signing up for a paid subscription. The image was created with Midjourney v6.