About
Why deslopify exists
An AI detector tells you who wrote a text. The Slop Score tells you how it reads. Different question, different answer. Here's why we built it that way, who it's for, and how it stays free.
The detector industry is selling the wrong thing
The published literature on AI text detection is consistent on three points. Untrained humans identify frontier-model output at chance. Detectors flag 61% of TOEFL essays by non-native English speakers as AI, halve their accuracy on paraphrased text, and run into a theoretical bound that no amount of better training can fix. Blind readers preferred AI-generated marketing copy to expert-human copy, so a low detector score isn't even correlated with quality.
Despite that, the AI-detection market keeps growing — sold to teachers, recruiters, and publishers as a binary verdict. The verdicts carry consequences. Cases are working through the courts now. We didn't want to ship into that.
What deslopify actually does
Three engines run against pasted text. Lexical (deterministic dictionary), structural (rhythm, parallelism, em-dash density), epistemic (a Sonnet 4.6 call that scores stance and specificity). Each returns a 0–100 sub-score and flags the spans driving its read. The composite is the Slop Score. Higher means sloppier — the brag is scoring 12.
The diagnostic surfaces six pinned reference texts at known bands (linked from /calibration) and shows you exactly which words and rhythms moved the number. The benchmarks pagedocuments what the score can and can't do, including the paraphrase-resilience deltas: −33 points under substantive rewrite vs −17 under surface paraphrase, which is roughly twice the resilience of register-blind detectors.
Who it's for
Writers who ship content where voice matters. Newsletter authors and ghostwriters watching their copy drift toward AI-shaped prose under deadline pressure. Marketing copywriters who need to know whether their landing-page revision actually sounds like the brand. Bloggers and founders writing their own essays. Students who use AI to draft and want to know whether their edit pass got the voice back.
Not for academic-integrity gatekeepers. We don't want this used to discipline anyone, and the researchsays it couldn't hold up there anyway. If you came here looking for an AI detector to flag student work, please read the limitations page first.
How it pays for itself
The diagnostic is free and will stay free. Scoring a paste costs us roughly half a US cent in Sonnet API calls, plus the deterministic engines (which are essentially free). Rate limits (10 scores per IP per day) and a 24-hour epistemic cache bound the worst case. Free scoring is sustainable because the per-scan cost is low enough to absorb.
The rewriter is where money happens. When you click Rewrite on a scored result, the same Sonnet model rewrites the text in a voice the engines won't flag. A rewrite costs us roughly two cents. We're launching paid access to the rewriter shortly — the waitlist is on the homepage. Pricing will be in the $10–15/month range for a generous monthly rewrite quota, plus likely a small one-time unlock for occasional users.
What we won't do
We won't market this as an AI detector. The research pagesexist to explain why that product category is structurally broken. We won't cross-sell to academic-integrity buyers. We won't add tracking pixels or analytics to the diagnostic — the privacy pagedocuments exactly what we send to Anthropic, store in Neon, and discard, and we'd like to keep that list short. We won't add a cookie banner the site doesn't need.
Where this is going
Next: the paid rewriter. Then, validation on the first 100 paid users — pricing, conversion, what the actual usage curve looks like. If that holds, French and Italian are the next locales (each requires an authored dictionary, structural patterns, a Sonnet prompt, and six pinned anchor texts validated against the live engine — see benchmarks for the bar). Maybe a small API for tools that want to embed the score. The public scoreboard the PRD originally proposed is deferred — sharing a score already works through the /score/<id> permalink, and a moderated public wall is a larger project.
Who builds this
Frank Goeltl, sole trader, operating from Kirchheim unter Teck, Germany. Operator details on the imprint; privacy commitments on the privacy page. The code, calibration anchors, and benchmark scripts are open at github.com/frankgoeltl/deslopify.me. Found something wrong? The waitlist email is the channel until there's a better one.