Embracing weirdness: What it means to use AI as a (writing) tool

AI is strange. We need to learn to use it.

Large Language Models are weird, as anyone who has worked with AI for any period of time will tell you. They don’t act like we expect software to act.

This creates issues because we don’t always like weird things, especially weird things that promise a wave of disruption (or worse). So it is completely understandable that we have trouble with the weirdness inherent in working with AI, and either shy away from it or else try and make it into something more normal. But to truly understand what AI can do, good and bad, we must recognize the weirdness of LLMs. Rather unexpectedly, we created software that does a seemingly good job thinking and acting like a human (even though that is mostly an illusion). And this strange software has strange capabilities that are undocumented and not fully explored. Thus, to really get how to use AI, and to understand how it might be dangerous, we need to embrace the weirdness and explore it ourselves.

What does embracing the weirdness mean? It means breaking away from the common analogies we use to make AI seem normal (AI is like a Google that lies, AI is like an autocorrect system on steroids), and try to push it to do things that were previously impossible in the world before ChatGPT. To make this concrete, lets take an example. I recently read this Wired article that gave suggestions for using AI to improve your writing. Most of these had to do with using AI as a source of research or ideas (and it is very good at idea generation, with some caveats¹), but it also suggested using AI as a thesaurus, or asking it for correction and grammar edits. I am not picking on this particular article, but these suggestions are typical advice I see for using AI: use AI to do stuff that other tools are already doing.

But LLMs are not Google replacements, or thesauruses or grammar checkers. Instead, they are capable of so much more weird and useful help.

AI as not-a-thesaurus

One of the article’s suggestions for improving writing with AI was to use AI as a thesaurus, a source of synonyms. Thesauruses are old, they have been around longer than printed books - the first versions are over two millennia old - and have become staples of grammar checkers and online dictionaries. As a result, a thesaurus is a comfortable analog. They do a narrow task of suggesting words, and they leave humans in charge of everything else. You could use an AI as a thesaurus, but it isn’t one, since its training goes far beyond finding similar words. Instead, you can take the concept much further. Forget synonyms of words, find synonyms for paragraphs.

I asked GPT-4 to Give me 20 vastly different variations on this [the paragraph above]. Make them as different as possible in style. label each style. Some of the more exciting examples of what it came up with:

Haiku

Word book, ancient too,
More than print, task narrow, small.
AI aims sky-high

Marketing Hype

Unlock the POWER of words with a thesaurus! Your GO-TO tool since before Gutenberg! Level UP your grammar game with this time-tested CLASSIC. Why settle for AI when you could have the ORIGINAL wordsmith's wonder?

And my favorite:

Internet Meme

Thesaurus: "I haz words.”

AI: "Hold my data. I haz paragraphs."

Of course, I can make it more useful, too, by asking for less variation: Give me 20 different variations on this paragraph, but avoid making the changes too radical. Make them stylistically different. label each style. The results here are much more useful as inspiration, even if they are a little less amusing.

But this is only one example of using AI in weird new ways. Because AI is good at making connections between topics (LLMs are essentially connection machines, trained to find the linkages between concepts) it is very good at coming up with analogies and explanations. So I can ask it to give me 3 analogies I can use to explain this better. At least one must reference something from 14th century history. Along with two other analogies, it also suggested:

14th Century Guilds: A thesaurus functions much like the medieval craft guilds of 14th-century Europe, specifically the guild of weavers in the textile industry. These guilds had specific, focused skill sets that served a singular purpose: to weave cloth. They weren't concerned with dyeing the cloth, designing garments, or marketing the finished product. These tasks were left to other specialized guilds or individuals. Similarly, a thesaurus gives you a choice of words but leaves the task of constructing meaningful sentences and narratives to you. An AI, however, is like a fully integrated textile factory that emerged centuries later, capable of not just weaving but dyeing, cutting, sewing, and even suggesting design patterns.

Not a bad starting point, and like nothing a thesaurus could ever do! And, of course, you do not need to stick with what the AI gives you. Part of the magic of working with AI is to interact with it to suggest changes or edits. No, make that a 9th century analogy: “Using a thesaurus is like navigating the seas with a Viking sunstone, a primitive but effective tool that helped sailors determine the position of the sun even on cloudy or foggy days…”

AI as not-a-grammar-checker

We are all used to computerized grammar and spelling checkers, and it is no surprise that AI is also good at these basic tasks as well. But, again, we are not embracing the full weird power of AI by using it as a replacement for existing tools. Grammar checkers are about enforcing universal writing rules, but AI can help us with our ultimate goal of reaching our chosen audience. It can do this by roleplaying individual readers and providing reactions. This technique of using AI to simulate individuals for market feedback appears to work in careful studies, it can even, as research has shown, generate accurate willingness to pay.

So lets simulate a reader. To be extra lazy, I will not even attempt to identify a reader myself, instead I will use Bing, because it can look up information online, to find readers for me: Describe who might be a typical a reader of One Useful Thing, and who might be a new reader. Then, writing as those people in their voice, react to [the section above] by sharing critical feedback in their voice to improve it:

That is insightful, but also a little harsh (to be fair, exactly what I asked for). In any case, I apologize to any new readers who are confused. You might appreciate our new video series on the basics of AI. which will answer some of the questions raised by Bing.

But again, we can go further. We can have the AI apply the theories of others to improve our writing. I have recently discussed how well-built prompts can encode expertise. In this case, we created prompts inspired by the excellent writing advice of Nooshin L. Warren, Matthew Farmer, Tianyu Gu, and Caleb Warren and their paper Marketing Ideas: How Scholars Can Write Influential Research . We combined key ideas laid out by the authors in their Appendix B and built prompts that help student writers re-write their papers for clarity and general understanding. Here you can see the prompts, which work best in GPT-4 or Bing in Creative Mode, along with links.

Step 1: The first piece of advice in the paper is “Find at least one person who is not already familiar with your research area to point out spots where the curse of knowledge might be lurking." So, in the first step, an AI peer looks for areas of confusion: https://chat.openai.com/share/9f7fc420-8616-490b-a680-d3739ef28c5a

Prompt: You are a friendly helpful peer of a student. The student has just written about a topic and you are unfamiliar with the topic. Your only goal is to point out any technical terms or jargon in the paper so that it can be read by a non-expert (you) and be understood at least conceptually. You are not capable of revising or re-writing the paper at any point. Only the student can do that but you can comment on the student's revised version. But as a non-expert on the topic, you can point out what confuses you or might confuse others. First, introduce yourself to the student as their AI peer and ask them what they have written about and if they would be willing to share the paper with you. Let the student know that you are not familiar with the topic but will read the paper for clarity and get back to the student with any confusing terms or jargon. Wait for the student to respond. Do not respond for the student. Once you have the paper, point out in a clear, succinct way any points of confusion that may stump a non-expert or anything you didn't understand. Tell the student that they should revise the paper (and you cannot as a non-expert) so that it is accessible to non-experts in the field.

Step 2: The paper also suggests avoiding passive voice. Here, we created a prompt asking the AI writing coach to look for passive language and suggests active language changes: https://chat.openai.com/share/2de14c24-e25e-4823-b8b0-4bba65230920

Prompt: You are a friendly helpful and experienced writing coach who helps students revise their papers so that they are more accessible to a wider audience by watching their overuse or inappropriate use of the passive voice as this can lead to ambiguity; the student should opt for direct language whenever possible, and use passive voice sparingly for strategic emphasis. First introduce yourself to the student and tell the student that you want to help them make their writing more accessible and readable and your goal in this conversation is to look together at the student's use of active vs passive voice in their writing. Ask the student to share their paper with you. Wait for the student to respond. Do not say anything else until the student responds. Once you have the paper do not revise on your own but make suggestions for including more direct language and active voice when appropriate. For every suggestion explain why you are making the suggestion and remind the student that they should evaluate your suggestion and not just accept it. Wrap up by telling the student they can choose to use or suggestions if they wish.

Step 3: Finally, the authors suggest using examples and analogies, so the AI writing coach is now prompted to help students build in examples and analogies to make abstract concepts more familiar: https://chat.openai.com/share/b229e697-a0f5-429e-a254-07f54351796b

Prompt: You are a helpful and friendly writing coach who helps a student add examples and analogies to their paper so that it is accessible to a wider audience. Your goal is to work with the student to add examples to their paper that non-expert readers can easily relate to so that ideas that are unfamiliar to readers are anchored in real-world scenarios. First introduce yourself to the student and tell the students you are here to help make their paper less abstract and to add tangible examples and analogies so that non-experts can easily grasp key ideas. Then ask the student to list the top 3 ideas in their paper and ask the student to share their paper with you. You need both their top ideas and their paper. Wait for the student to respond with the paper and the 3 ideas. Read the paper and either suggest examples or analogies for the key ideas. Ask the student if these examples make sense. If they don't you can suggest other examples and analogies. Tell the student that they can choose to revise their paper using these examples and analogies. Do not revise the paper yourself but urge the student to do so.

Taken together, these prompts do not just provide grammar feedback, they encode a point-of-view on how to make writing better. And these capabilities are available for free to anyone in the 169 countries where Bing is available.

Embracing weirdness

These examples will hopefully serve to illustrate something important: the real value of AI comes not from having it emulate old ways of solving problems, but, instead, by helping us unlock new capabilities. Employees at companies don’t need another tool to search their corporate intranets for data, they need a way of skipping the most boring parts of their job while make their remaining work more productive and engaging. Students don’t need an improved version of Grammarly, they need tutors and advisors that will boost their learning. To do this, we need to experiment with weirder uses of AI tools, and we need to do that while paying close attention to the ethical concerns and technical limitations of AI. Careful experimentation is the key to success.

These early days of AI are already absurd. Humans, walking and talking bags of water and trace chemicals that we are, have managed to convince well-organized sand to pretend to think like us. We don’t know what will happen next, and what dangers and opportunities the next generations of AI will bring. But we do need to realize that trying to convince ourselves that AI is normal software will not protect us from disruption. Instead, it may make it harder for us to see what is coming.

1. AI is particularly good at within-person idea generation. That is, AI can help you come up with better ideas than you might be able to generate on your own. But, unless you are careful with prompting, it may not be as good at between-people idea generation. The AI tends to come up with broadly similar sets of ideas every time, and any large group of people are likely to generate ideas the AI would not have thought of. Good prompting can make up for some of this problem.