AI Slop is a human problem

Why is AI writing so bad?

It can one-shot an entire app, but if I ask it to help me write a 1000-word essay, it is riddled with non sequiturs and a lack of coherence. I fed it the outline, key ideas of this post, had a discussion about what I wanted to communicate and then asked Claude to write a post and what came out made me squirm. Every thread we touched got a paragraph, weighted equally. There was a certain blandness to it. So here I am writing this all by myself.

Is it an intelligence problem?

It is often said that being able to communicate your ideas clearly is a sign of intelligence and depth of understanding. My initial response to the sloppy writing was that AI doesn’t understand what it’s talking about, so it cannot talk about it well. Basically, it is not smart enough, and when it gets smart enough it will write well.

But really? I have used Claude effectively to plan out, review and build complex software. I have had interesting philosophical discussions. It has written snippets of poetry, “understood” entire manuscripts to find logical errors. AI agents have coordinated to solve unsolved math problems. In a lot of ways, it is much smarter than I am. In fact, when I asked it to explain a research paper to me, I found it much easier to explore and understand the nuances of the paper by talking to Claude. Even professors at the top of their field have told me about similar experiences.

This was not the case a few months ago, when I had to do a lot of reframing and “write like PG” or “use Google developer guidelines” prompts to make Claude talk to me in a way that was helpful.

So if you are precise with what you want, LLMs are getting increasingly good at responding well. But writing for an audience is a different type of skill. You are sharing a point of view. Is developing a point of view a function of intelligence?

To be clear, I am not asking it to produce art. We don’t need Shakespeare; all I am asking for is clarity - Feynman-style clarity about an idea translated to meet the reader where they are.

Is it a Context problem?

Since it is MY idea I want the AI to communicate, perhaps I haven’t shared it clearly enough. If I am clear about who I am writing for, what I want to focus on, examples, related concepts that inspired me, things I care about and the story I want to tell.

I will probably miss the point too if I have to communicate someone else’s idea to you. I would have to have deep conversations and make it my own idea first before I communicated it effectively.

So I refined the process further. I asked it to imagine the reader, gave it guidelines on how to tell a story, motivate questions then answer them, add some humour, figure out the most important points, spawn agents to verify each paragraph fed back into the core idea.

One idea that worked particularly well was to make it have a conversation with a reader and evolve the idea with them. The write-up got tighter and easier to read when I asked it to convert the conversation to an essay but it still lacked character.

You could argue that “lacked character” is bleeding into Shakespeare territory. I would say character and his unique choice of words in fact made Feynman so enjoyable and hence easy to understand.

Human-shaping the communication

Code is for computers to process. Writing is for humans to process. Humans have a limited attention span, we can’t hold more than 4 things in our brain at the same time, the content can’t be too hard or too easy or we get bored, we need jokes to not overload our brain.

There are 2 issues here:

1. Why did I have to do so much work to teach the AI how to present an idea in an essay, when it is otherwise pretty smart at understanding things?
2. Why does it still feel icky after all that work?

I think deep understanding leads to better writing only if the understander is human. Because they have to have human-chunked the problem to understand it, and now they are communicating it back to humans in human-chunks and it clicks. If you will allow me to anthropomorphise LLMs, the AI doesn’t need to human-ify the problem to process it. It has understood it in AI-terms and now it has to communicate it in terms the humans will understand. That’s why I had to specifically human-chunk it before it was legible. Think of a chimpanzee asking us to explain to another chimpanzee something about its banana preference and complaining that we didn’t quite get it right so we don’t understand bananas.

Even after we teach it how to dumb things down for us, it still speaks humanese with an AI-accent, which is where we find it not quite right.

I think we would be remiss to treat this inability as representative of its intelligence as we march towards ASI. It’s a different skill, which could serve as a bulwark against its ability to influence humans who don’t fully trust the being with the AI-accent. Until it eventually learns to speak with all the flaws of a native human.