Every: testing Deft, the model built to fix AI writing
What the article is about
Every staff writer Katie Parrott reviewed DFT v1, the first public model from Deft — a writing-focused AI research lab co-founded by Justin Murphy, a former political science professor and writer, and an AI researcher known as Rosmine. The article is an extended test across multiple writing tasks and delivery modes, asking whether Deft’s core technical approach actually produces better prose.
Context
Deft’s technical method is called distribution fine-tuning, abbreviated DFT. Rather than evaluating model outputs one at a time, the post-training process compares batches of model outputs against batches of human writing samples, and adjusts toward greater linguistic variety. The goal is to reduce the repetitive sentence structures, hedging phrases, and predictable rhetorical patterns that make AI writing identifiable to readers familiar with AI output. Parrott tested the model through a web console, through an API integration with a writing platform called Codex, and across both analytical essays and SEO article formats.
Key findings
DFT v1 does produce less predictable prose than mainstream AI writing tools. Syntactic variety is higher, and the writing avoids some of the more recognizable AI tics. Parrott compared it to GPT-5.6 Sol and found the output felt less formulaic.
The weaknesses are significant. Outputs were often dense and unclear, with awkward constructions and poor sequencing of information. Even in strict mode, the model fabricated unsourced details — a problem that undermines the core use case for analytical writing. The API integration was unreliable, sometimes returning partial drafts or instructions rather than content. Parrott’s conclusion is that variety alone cannot fix what ails AI writing. A model can produce sentences that don’t pattern-match to AI output while still failing at information hierarchy, reader engagement, and basic factual accuracy.
Who should read this
Writers and editors evaluating newer AI writing tools, particularly those who have found mainstream models too pattern-heavy. Also useful for anyone interested in the technical approaches researchers are taking to make AI prose less identifiable — the distribution fine-tuning method is worth understanding even if DFT v1’s current results are mixed.