Writing for Artificial Intelligence

It’s easy to feel both impressed and depressed by how good ChatGPT and the other LLMs have become at writing. A few years ago, I remember being shocked at the ability of even the more primitive GPTs to effectively imitate a writer’s style. Kafka. Didion. Faulkner. Maybe our sentences and paragraphs are really just a pattern. We revolve within our fractals of words and all it takes is a large enough sample size for an AI to mimic our signature moves.

But AI became bigger and stranger. It hallucinates and does funny stuff. It seems to need our writing, need our guidance—at least in its current form. I was fascinated by a prompt that was created to guide the new Apple Intelligence (which will be on our phones this Fall). “You are a helpful mail assistant which can help identify relevant questions from a given mail and a short reply snippet,” is how it starts and then goes on to give clear instructions. The AI is told to be “helpful.” Kind of weird.

The Apple prompt displays the art of “emotional engineering.” An AI will give you better answers if you tell it “to take a deep breath” before replying, or if you are polite and praise its responses, or if you inform the AI that it’s a “genius stock picker who knows how to find hidden advantages in the financial markets.” The scholar Ethan Mollick has described prompts as being “programs in prose.” He has a whole library of structured prompts that are fascinating to read through: “You are an AI instructional designer, helpful and friendly and an expert at tutoring. . . ”

The AI does better with a persona, with a character description. Wild.

P.S. This newsletter by Oliver Burkeman gives what I think is the definitive take on human creativity in the age of AI: “The point about a good novel produced by a human isn’t that only a human could ever have produced it. It’s that a human did in fact produce it.”