I’m Ben, the maker of AI Undetectable. This is an affiliated product-engineering post; AI assistance was used to prepare the post and evaluation work.
A useful failure case for a rewrite model is a paragraph containing a proposed study, not an established result. A rewrite can retain every number while changing what those numbers mean.
Here is an original synthetic fixture from our September 18 evaluation: a fictional six-week classroom pilot, two classes, revision that afternoon versus after two days, the same rubric and teacher, and explicitly no results. It also names absences and unequal outside help as confounders. No student or customer data is involved.
For a fixture like this, I would separate the checks rather than assign a single “sounds natural” score:
- Entity/quantity retention: six weeks, two classes, two days.
- Evidence status: fictional and proposed must not become completed or proven.
- Causal scope: a better second draft must not automatically be attributed to the delay.
- Completeness: the counterargument and confounders still need space in the rewrite.
Our full 20-output pass covered five writing types, two rewrite modes and two repeats. Some outputs omitted a fictional qualifier or shortened the source too much. Two selected, unedited Essay outputs are public alongside their entire 442-word source, model IDs and source hashes; they are examples, not a representative success-rate estimate:
Full source and recorded outputs at aiundetectable.com.
The practical lesson is that matching names and numbers is necessary but insufficient. A simple literal check can catch a missing “six-week”, but not the shift from “would compare” to “showed an improvement”. Those relations need semantic review. An LLM reviewer is another fallible model, not independent validation.
The examples are readable without an account. Live rewriting is a paid/trial product, not an open-source or unlimited-free endpoint. This evaluation used no detector scores and makes no detector-bypass claim.
For people evaluating rewriting or summarization: how do you label preservation of uncertainty and causal scope separately from basic fact retention? Would you keep these as explicit binary fixture assertions, or score them with a rubric and spot checks?