The Science of Why AI Still Can’t Write Like You — Marcus Moretti

Moretti, general manager of Spiral (Every’s AI co-writer), examines why AI writing remains stubbornly detectable despite models achieving PhD-level knowledge in other domains. The distinctive AI tells — “It’s not an idea. It’s a breakthrough,” “Delve,” lists of threes without “and” — persist because LLMs learn general speech patterns from massive training data, then post-training refines them into generic politeness.

New research on writing style reveals that the most distinctive parts of human prose are the unconscious ones — the structural patterns, rhythm choices, and word preferences that writers don’t deliberately control. These are precisely what LLMs fail to capture because they represent individual deviation from the statistical average, while LLMs are optimized to produce the statistical average.

Moretti connects this to Spiral’s product approach: rather than trying to make AI write like you from scratch, Spiral uses a chat-based process to understand intent, then produces multiple draft approaches. The goal is AI output that reflects the user’s voice through iterative collaboration rather than imitation.

RDCO Mapping

Directly relevant to our Sanity Check quality standards and the voice-match skill. The insight that distinctive voice lives in unconscious structural patterns — not vocabulary or topic choice — validates our approach of using the founder’s published work as voice training data rather than explicit style guides. Also connects to our draft-review skill’s detection of AI tells.