Please computer, help me writing!
I have complained a bit about the capabilities of AI on this blog. Time to say something positive about AI too: I really hope AI will help me write before I get too old to write. Because I badly need it.
I only recently learned about the concept. Late to the party as always, I guess. The idea is that a human writer and an AI should cooperate, as described in for example this article. The human has an idea and starts writing. Based on statistical methods, the AI suggests text sequences that seem appropriate. The human then decides which text sequences to accept, reject or modify.
I very much want to be that human. Because I actually need some help. I struggle with writing and have always done so. Very often I have a thought in my head that seems to make sense. Occasionally, I take the leap and try to transform such a thought into a text. Sometimes I score the jackpot and the text gets rather good with little effort. More often I write and rewrite, copy and paste and delete and rewrite again. After several days or weeks of on-and-off writing, the text finally becomes readable. Almost as often I write a text, read it and discard it, because I see that this particular thought sounded better inside my head than outside it.
The new and the familiar
Why is it like that? Why do so many thoughts look good as thoughts, but bad as texts? There are at least two possibilities:
The thought wasn't that good after all. "The obscurely said is the obscurely thought", is a saying in Sweden, apparently after Bishop Esaias Tegnér said so in a speech in 1820. If a thought can't be expressed clearly, that is often because it actually is not clear at all. Writing it down can help me realize that. Then it is back to the drawing table for that thought.
The thought actually is really clear and logical, but I struggle to fuse it with what normally is on people's minds. Here lies the center of the art of writing, I think: To fuse something readers don't know with what they already know.
An interesting text must contain both things the readers don't know and things they know. If there is nothing new to learn from a text, reading it is just a waste of time. But in order to present something new, it is crucial to use a well-known framework. Spelling and grammar are the basics. Current AIs have already solved that part decently. What really distinguishes good writers from bad writers these days is the "fluff". Just like spelling and grammar must be familiar to the reader, many whole sentences need to consist of information that is also already familiar to the reader. Those sentences act as a bridge between the known and the new: In order to explain something unknown, one needs to point to something known.
I struggle with that part. Year after year, I read book after book and incorporate them into my own thoughts. In the light of my own library of references, many thoughts really make sense to me. But when I try to write them down, I realize this makes sense to me only because I have read a certain set of books. A very large AI database should be able to tell me what decently educated people actually know and agree on. It could tell me "this isn't generally accepted knowledge, it needs to be further explained" when I write things that are only established knowledge inside my mind.
A database of past writing could also give me clues about where to start from. What are other people's minds dwelling on, really? Where do their thoughts start? The answer should lie in what they commonly write about. And that is what databases of human writing are experts of. They can't think for themselves. But they know more than any human about the aggregate of past human thought. When I write, I struggle to make educated guesses about what other people think. The AI's guesses would be immensely more educated than mine.
Overall, I know that writing is a lot about predictability. Only one, single thing should surprise the reader: the intended message of the text. All other parts should be as predictable as possible: Spelling, grammar, structure, text length, paragraph length, the use of under-titles, the progression of the message, the use of anecdotes, the share of text used for anecdotes… I try very hard to write texts that are structured just the way people expect. An AI has immensely more data than I do over what people expect. Please, computer, give it to me!
Could an AI tell me anecdotes?
One area I struggle especially hard with is anecdotes. Anecdotes are the stuff that separates good writers from mediocre writers. The ability to casually drop a few eye-catching examples without losing the narrative of the text is more or less a prerequisite for good writing. The Economist, which is close enough to a journalistic gold standard to have their own style guide, almost always begins their articles with an anecdote about something entirely different, but still related to the real subject of the article. Only very good, and very well-informed, journalists are skilled enough to do that.
For the rest of us, holding on to a narrative while searching for anecdotes is just too difficult. So I wonder: Could an AI do that for us? Could an AI provide us with useful anecdotes that support our narratives and help bring forth our message? I'm thinking of something like a specialized search engine for anecdotes: I give the AI my text and point to the specific part where I want to fit in an anecdote. The AI could then trawl its infinite database of anecdotes and give me one or two or ten suitable anecdotes to choose from.
A good anecdote machine would not make all good writers redundant. But it would help a lot of mediocre writers like myself close the gap to the good writers. With good enough writing software, everyone who knows what a readable text looks like could also produce one. That gives two opportunities:
In a best-case scenario, people with lots of knowledge and great ideas will be able to compose texts that the majority wants to read. This is in contrast to today when knowledge and ideas are often outcompeted by raw writing skill.
In a worst-case scenario, the internet will be overwhelmed with machine-produced bullshit. If everyone is able to produce great writing then everyone will be writing. And most people, frankly, have little more to say than "in-group good, out-group bad". In this case the software that produces great text will have to be countered with software trying to sift the vast amounts of great text for something with actual content.
There is a clear risk for the worst-case scenario but the upside in the best-case scenario is great enough for me to still wish for it. Today’s technology has removed all practical barriers for communication between people. What remains is our personal inability to share our thoughts. If AI-enhanced software could overcome that inability there would truly be no barriers for the free exchange of information any more. Whether this will be heaven or hell remains to be seen.
I love your blog with Anders! Don't overthink writing. Just understand that
1. Quality will naturally vary, and
2. Some ideas take multiple drafts to hash out.
In exactly the same way, not all apple pies are delectable. One has only a little time on Earth to pick apples, cut them, and bake them into pies. A good cook loves all his creations, but still admits that some end up burned, and thrown into the compost pile to be eaten by homeless people and rabid raccoons.
> An interesting text must contain both things the readers don't know and things they know.
When I read this sentence I thought about a podcast "99% invisible" which does this job perfectly. It's usually about things that seem to be "basic" and explains their background. The last episode I listened to was about potted plants. I'm not a big fun of plants myself (I regard them as a kind of a waste of time) but I loved to listen how potted plants were used in Victorian era, how they were brought to Europe and why people like plants. The connection here would be plants - something really common and the historical and economical background for them.