AI Writing Lab Deft Launches Beta, Claims 86% of Outputs Read as Human to Pangram

On August 19, AI writing lab Deft, co-founded by John Morrow (@rosmine) and @jmrphy, launched its eponymous public-beta writing model with a core pitch of making AI writing indistinguishable from human writing: official figures claim 86% of user queries are judged fully human by Pangram. The model is a small beta that may still make errors, and its launch announcement was written by the model itself. Its Distribution Fine-Tuning (DFT) technique has sparked debate over both detector efficacy and writing quality.

Confirmed

• Deft was co-founded by @rosmine (John Morrow) and @jmrphy; the beta launched August 19 with a model-written announcement.
• Officials claim 86% of user queries are judged fully human-written by Pangram in current tests.
• The model performs better on analysis, essays, and creative writing, but weaker on marketing copy and news; users are advised to add prompt detail and adjust style options.
• Per @HamelHusain, Deft's research paper argues standard SFT makes model output formulaic (slop signs), overusing certain tokens or phrases; Deft proposes Distribution Fine-Tuning (DFT) as a post-training remedy.
• The product includes customization: bespoke models trained to match a company's voice, turning rough ideas or drafts into natural prose, premised on a future of customized human-AI knowledge-production workflows.

Unconfirmed

• The 86% Pangram pass rate is a single-source official figure without independent verification; one repost headline framed it as outputs still detected as AI, contradicting the official framing.
• Ethan Mollick argued that tools built to fool AI detectors typically produce bad writing; how Deft's actual text quality holds up awaits beta testing.

Why it matters

• Reposters such as @TuhinChakr, @soumitrashukla9 and @koltregaskes focused on the model's anti-detection capability; some argue that once model writing reaches or exceeds skilled-human level, detectors must fail because no detectable features remain. If the 86% figure holds, it directly challenges the AI-text-detection and content-moderation ecosystem.
• Mollick's critique represents the opposing view: fooling detectors shouldn't come at the cost of quality, and a model's worth rests on readability, not pass rates; Deft's DFT research on formulaic writing is a technical response to that concern.
• Deft's customized workflows and distribution fine-tuning mark an attempt to move writing AI from generic generation toward humanlike output plus enterprise style customization.

2026-08-19 ~ 2026-08-21 · 15 related posts

• New AI writing lab Deft launches public beta model — TheZachMueller · 2026-08-19
• [source] Deft AI Launches Human-Style Writing Tool with Custom Company Model Training — soumitrashukla9 · 2026-08-19
• AI detectors in ruins? New Deft model claims 86% of queries judged as human-written — soumitrashukla9 · 2026-08-19
• Deft Launches AI Writing Lab Focused on Human-Like Output Quality — soumitrashukla9 · 2026-08-19
• Ethan Mollick: AI detector evasion tools produce bad writing — eldonredwards · 2026-08-19
• New AI lab Deft launches with a model that wrote its own announcement — zacharynado · 2026-08-19
• Deft Model Claims to Evade AI Detection with 86% Human-like Score — churchkey · 2026-08-19
• [source] Deft model fixes LLM formulaic writing with Distribution Fine-Tuning — HamelHusain · 2026-08-20
• Deft AI writing lab launches with 86% human-written content rate — nwilliams030 · 2026-08-21

6 near-duplicate retellings: soumitrashukla9 · TuhinChakr · cephaloform · TuhinChakr · koltregaskes · rohanpaul_ai