Co-opperative writing has been the first use I've found for these models that actually deeply appeals to me-- I'd seen the examples of people talking to LLMs as companions (Replika, Character.ai, etc), which were certainly interesting but didn't really hold much long term appeal. But using ChatGPT and GPT3 to speed up my creative writing has been amazing and I can't imagine I will ever get bored of it. I have an endlessly patient beta reader to nod its head and encourage me when I bounce ideas off it, add a paragraph to finish up the scene I was bored with, fill in the places where I don't have any particular ideas ("Okay, I need some kind of monster to jump out here-- give me five quick descriptions of an original fantasy monster that would be found in a desert")-- when before I probably would have just put a generic placeholder zombie or something.
With the advancements in AI technology, using language models like ChatGPT as writing assistants has become increasingly popular. However, it's crucial to understand the difference between 'AI-generated writing' and 'human-AI collaborative writing.'
In AI-generated writing, the writer provides a simple prompt to ChatGPT, such as 'write an essay about the mysteries of dark matter' or 'create a joke about why traffic lights always turn red just as you arrive.' The AI then produces the text, which the writer can copy and paste into a word processor. The focus here is solely on crafting a prompt that elicits comprehensive and accurate information from ChatGPT, regardless of the writer's knowledge on the subject. The final product is often passed off as the writer's own work.
Human-AI collaboration in writing involves the writer providing ideas and language, then using AI to improve it. That's how this blog post was made. I gave ChatGPT the prompt "polish for a LessWrong-style blog post," then added my own thoughts. ChatGPT transformed my rough ideas into a polished, well-structured piece, with good sentence and paragraph flow. The ideas are mine, but I was freed from worrying about the prose quality, so I could focus on idea flow and development.
Using ChatGPT to aid in writing improved the quality of my work, allowing me to write six pages of my master's thesis in just an hour with high-quality results. Though not perfect, any errors can easily be corrected by pointing them out to ChatGPT for revision. For example, in this blog post, ChatGPT initially mislabeled my work as a six-page thesis instead of six pages of my thesis, but with a simple correction request, the AI got it right on the second try.
This human-AI writing process starts with the writer's draft, which the AI improves for accuracy, readability and flow. Another way to use ChatGPT is by asking it to predict reader questions and concerns with a prompt.
For instance, let's try this approach right now.
I asked ChatGPT to write a biting criticism of this article, and it obliged:
Critics may argue against using AI like ChatGPT in writing, but those not in the "writing profession" can benefit. As a biomedical engineer, I aim to communicate information accurately, not create works of art. I use computers to analyze data, generate figures, and format documents. Why not use it to generate language as well? Writing in my field needs to be clear and serve its purpose, not have literary merit. ChatGPT helps me focus on the flow of my thoughts without interrupting to struggle with language. This human-AI approach lets me quickly get my ideas down, freeing me to focus on the content and not the language. It's a useful tool for anyone who writes for work or study, regardless of writing background.
Let's now use ChatGPT to anticipate reader confusion.
The impact of AI on writer's voice and style is uncertain. The text generated by ChatGPT has a tone and vocabulary similar to the writer, but with improved grammar and sentence structure. The AI may add flourishes, such as motivational sentences, and sometimes includes promotional material, like advertisements. This blog post serves as an advertisement for ChatGPT, and some promotional language had to be edited out. ChatGPT streamlines the writing process and improves the quality, but excessive promotional language can detract from the purpose of the writing. The source of the AI's promotional language is unclear.
I wanted to shorten the blog post, so I pasted a few paragraphs at a time into ChatGPT and asked it to rewrite them with all the important details. Sometimes, I had to make a few tries because ChatGPT would make them too short and some important transitions or details were lost. The original blog post was about three times longer.
In this process, I felt like I was switching between the roles of writer and editor. As an editor, I explained the flaws I saw in the AI's output and made suggestions on how to fix it. I kept making suggestions until ChatGPT produced something that met my expectations. This was much easier than trying to write everything myself. ChatGPT took care of the difficult part and I was able to focus on the creative side of writing.
Showcasing human-AI collaborative scientific writing
I used ChatGPT to describe a scientific process like DNA replication in an engaging way. The AI's language was almost sensuous, like a more mature version of the magic school bus. Here's a short example.
When I asked ChatGPT to describe DNA replication, the first output was a boring list of enzyme names and their roles, with a few creative descriptions tucked away at the end. I wanted to make the creative parts stand out, so I asked ChatGPT to move them to the beginning and mix them with the science. The result was a lively, easy-to-understand description that brings DNA replication to life for a lay reader
AI like ChatGPT can change technical and scientific writing. Scientists and engineers can focus on accuracy and ideas, while ChatGPT takes care of putting words together. This improves scientific communication and boosts productivity.
Many scientists struggle with writing, which can be boring and time-consuming. ChatGPT frees them from this burden and still gives a high-quality output for readers.
We put too much focus on the magic of writing and the writer. The real goal is a useful, timely product. ChatGPT helps us focus on that, following the advice of writing experts to publish often and not stress about the first draft.
As a writer, working with ChatGPT has been amazing. It lets me finally follow the advice to publish often and not get stuck on writing.
As AI tech becomes more common in writing, we must think about the impact of human-AI collaboration. The debate continues on whether it's plagiarism or tech adoption and the fear of misinformation and losing writing as a tool to gauge intelligence, creativity, and culture.
For writers trying human-AI writing, it's crucial to have terms to differentiate between copying AI-generated content and letting ChatGPT improve rough drafts. That's why I wrote this post on LessWrong. As we navigate this era, we must weigh the potential risks and benefits and understand the difference between copy-pasting and true human-AI collaboration. I list "human-AI collaborative writing" on my resume and hope for a world where it's a widely understood and appreciated skill.