Every: 13 beliefs about AI writing
Published on Every in September 2026, Mike Taylor’s piece describes his shift from dismissing AI-written work to publishing it under his name. Rather than a how-to guide, the article is a set of working principles — things he has found to be true about how the quality of AI writing depends on the inputs the writer provides, not primarily on the model chosen.
What the article is about
Taylor argues from sustained experience with multiple models and writing contexts. His central position is that by 2026, the best language models can produce publishable-quality drafts — but only when the prompts they receive contain material that could not have been generated by AI alone: personal conviction, lived detail, and a clearly held point of view.
The 13 beliefs
Taylor lists 13 principles. Some address the mechanics of prompting: the prompt should be larger than the output, embedding substantial personal material so the model has something specific to realize. Write original content not replicated in training data: distinctive premises rather than well-covered topics.
Several address the content itself: write about topics where you hold a genuine position, include data or visualization that grounds abstract claims, preserve all drafts because earlier versions often contain usable material that later revision cleans away.
Others address the writer’s judgment: author social media content personally rather than delegating it, because social voice is difficult to replicate through prompting and valuable precisely because it reflects a specific person. Interesting lived experiences are the strongest material foundation — not because AI cannot invent, but because invented material lacks the specificity that makes writing worth reading.
The piece addresses volume too: effort does not guarantee success; what scales is not labor but strong ideas paired with consistent publication. Taylor’s test for whether a collaboration is working is whether the output is something you are willing to publish under your name — not whether it was produced faster.
Key method
The framing Taylor uses throughout is that AI realizes the writer’s inputs rather than adding to them. This places responsibility for quality on the material that goes in — the specificity of the premise, the clarity of the conviction, the distinctiveness of the personal detail. Models that receive underdifferentiated prompts produce underdifferentiated prose, regardless of capability.
He is direct about one of the beliefs: modern language models can produce work you are willing to publish. This is stated as a current observation about where the capability threshold sits, not a universal recommendation.
Who it is useful for
Writers, journalists, and content professionals who have tried AI assistance and found results inconsistent, and want to understand what determines quality before attributing failures to the model. The principles apply regardless of platform. The article is most useful to people who already have a working method they want to examine critically — and to those wondering why AI-assisted drafts keep requiring the same amount of rewriting despite better models.