# Sebastian Mallaby says AI still really sucks at writing anything longer than about 800 words

The author’s AI rule is simple: let the machine brief you, but don’t let it become the part of your brain that writes.

From [The Tim Ferriss Show](https://listenerreader.com/shows/tim-ferriss-show)  
By ListenerReader Culture Desk, published 2026-11-12  
Canonical: https://listenerreader.com/story/sebastian-mallaby-says-ai-still-really-sucks-at-writing-anything

Sebastian Mallaby told The Tim Ferriss Show that AI is useful for research, but he would never use it to write because it is "terrible for anything longer than about 800 words." Mallaby’s larger claim is that writing is not a task to be automated away, it is the process by which he figures out what he believes.

That is a pleasingly cranky line to draw in 2026, when every productivity wizard with a ring light is trying to convince you that your next book, startup memo, wedding toast, and apology text should be co-authored by a chatbot that sounds like a LinkedIn post wearing a blazer. [Sebastian Mallaby](https://listenerreader.com/guests/sebastian-mallaby), to his credit, is not doing the fake monk routine. He does use [artificial intelligence](https://listenerreader.com/topics/artificial-intelligence). He just doesn’t want it touching the part of the job where the soul, or at least the argument, lives.

> What I would never do is get the AI to write because frankly, it's not very good at long form. In fact, it really sucks.
>
> Sebastian Mallaby, on the episode ([6:17](https://www.youtube.com/watch?v=EBgWAnX1sNs&t=377s))

The funny thing is that Mallaby’s most quotable AI take is also the most vulnerable one. The "800 words" threshold has the vibe of a man measuring a flood with a yardstick. Tim Ferriss immediately pushes him with the obvious objection: AI keeps getting better, and there are already tests where professional readers can’t reliably tell a writer’s real prose from machine imitation. Mallaby doesn’t really dispute that. He retreats to a stronger place, not "the machine can’t," but "I won’t."

## His real rule is not anti-AI. It is anti-outsourcing-the-self.

Mallaby’s distinction is sharper than the usual AI panic. He is happy to let a model do the scavenger work. If he is interviewing a scientist, he’ll use a large language model to digest the person’s papers, compare them with related work, and give him a faster path into the material. That is not laziness, in his telling. That is a better warm-up.

> It's cutting to the chase and nourishing me intellectually.
>
> Sebastian Mallaby, on the episode ([6:04](https://www.youtube.com/watch?v=EBgWAnX1sNs&t=364s))

This is where Mallaby sounds less like a technophobe and more like a person who has actually used the thing. His Google Maps distinction is good: offloading a narrow, annoying task is fine. Offloading the entire act of thinking is not. Nobody becomes more human by memorizing the route to the airport. But if you stop forming sentences, stop making judgments, stop discovering that your first idea was mush and your second idea was worse, then yes, you’ve handed over something expensive.

> It's not fine to offload all thinking, [laughter] right?
>
> Sebastian Mallaby, on the episode ([5:03](https://www.youtube.com/watch?v=EBgWAnX1sNs&t=303s))

The verdict on the claim: the 800-word part is probably too tidy, and it may age badly by lunch. AI can already produce long passages that are competent enough to fool tired readers and depress working writers. But Mallaby is right about the higher-order danger. The point of [writing](https://listenerreader.com/topics/writing) is not merely to produce a document-shaped object. It is to make contact with your own confusion and then beat it into a shape someone else can follow.

## The prepared mind, now with fewer shortcuts

Mallaby’s billboard phrase, borrowed from Louis Pasteur, is "prepare your mind." He connects it to venture capitalists anticipating a market, to Ilya Sutskever recognizing the importance of transformer architecture at [OpenAI](https://listenerreader.com/topics/openai), and even to a New England Patriots interception against the Seattle Seahawks. The point is not that smart people get lucky. It is that they have already built the mental hooks for luck to grab.

> The risk with large language models is that we just get lazy and whenever we need to know something, we just get it to tell us what to think.
>
> Sebastian Mallaby, on the episode ([3:30](https://www.youtube.com/watch?v=EBgWAnX1sNs&t=210s))

Ferriss adds a useful anxiety: people have become weirdly helpless without GPS, and knowledge workers could suffer the same muscle loss with AI. Mallaby resists the analogy in its simplest form, but the gym metaphor hangs over the whole exchange. If you let the machine lift every weight, do not be shocked when your mind starts looking like it skips leg day.

Mallaby’s best answer is not a policy, it is a boundary. Use the model to get to the thinking faster. Don’t let it do the thinking you most need to do. The next version may write 2,000 decent words, then 10,000. Fine. The question gets more brutal, not less: if the machine can imitate the output, what part of the work are you still unwilling to give away?

## Source episode

How To Prepare Your Mind for the Age of AI — Sebastian Mallaby

- YouTube: https://www.youtube.com/watch?v=EBgWAnX1sNs
