The proximate cause was watching a skilled copywriter lose a client because her work flagged on an AI detector. She had used AI. So did the detector. The client used AI to write the brief. Everyone in that chain was using AI. One person got penalized for it.

The deeper cause was recognizing that the real problem isn't detection — it's the writing itself. The text that gets flagged usually deserves to get flagged. Not because it was AI-generated, but because it's bad. Flat. Bloodless. Hedging every claim. Organized like a high school essay. Sounding like nobody wrote it because, effectively, nobody did.

DeepRewrite started as an internal editing pipeline — a way to take AI drafts and run them through something that actually resembles editorial work. Not synonym replacement. Not sentence shuffling. The real thing: concept-level restructuring, personality injection, the kind of pressure a good editor applies when they read something and think "this is technically correct and entirely unreadable."

The detection-proofing emerged naturally from doing that job properly. Which told us something important about what AI detectors actually measure: not whether a human wrote it, but whether an editor touched it. Those are different things. We're interested in the second one.

We are a small team. We run our own infrastructure. We write our own training data pipelines. We don't depend on OpenAI staying consistent, Anthropic not changing their output distribution, or any general-purpose model vendor keeping their token statistics stable quarter over quarter. Our models evolve with the problem. That matters when the problem is a moving target — and it is.

The Sarcastic Mode was built because one of us ran a corporate strategy document through the main engine as a test and changed the system parameters by accident. The result was sharper and funnier than anything we'd produced intentionally. We spent three days figuring out how to reproduce it on purpose. Then we shipped it as a feature.