When someone says a piece was written by a machine

Why detectors misfire on fiction

Third-party AI detectors do not read meaning. They measure surface statistics: how evenly sentence lengths are distributed, how predictable each next word is, how often familiar phrasings appear. Text that scores low on variation and surprise gets called machine made.

Genre fiction scores low on exactly those measures for reasons that have nothing to do with machines. Trope-driven vocabulary repeats because the genre expects it. Dialogue has a ritual rhythm, with short exchanges and repeated tags, that flattens sentence-length variance. Craft advice pushes writers toward plainer words and tighter sentences, which is to say toward the predictability the detector is scoring. Careful editing narrows the variation further. The better a writer follows the conventions of their form, the more machine-like they measure.

This is measured, not asserted. Peer-reviewed work has repeatedly found high false-positive rates on human text, with Liang et al. (2023) in Patterns finding detectors misclassified writing by non-native English speakers at strikingly high rates, and Weber-Wulff et al. (2023) in the International Journal for Educational Integrity finding the tools neither accurate nor reliable enough to act on. Detector vendors publish accuracy figures for their own tools; independent testing has not matched them.

What the record shows

• When it was first drafted, and when it was published

  The first timestamp is written once and never updated, so it cannot be moved later.

• Every saved revision since

  How many saves, across how many days, how many words were written and cut at each one, and which parts of the piece you kept returning to.

• Whether text was pasted in, and what became of it

  Pasting is ordinary. Writers paste their own earlier drafts, a line an editor sent back, a word a spell checker corrected. The record states the scale of what arrived that way and whether it is still in the piece, so a passage you pasted and later rewrote stops being described as pasted material.

• A fingerprint of the exact words

  A short code worked out from the text. Change one character and it changes completely, so a copy can be matched back to this dated record. Anyone can check it themselves with a standard tool.

• A timestamp nobody has to take our word for

  The fingerprint is committed to the Bitcoin blockchain through OpenTimestamps, so the age of the record can be verified without trusting Inkbreaker.

This record shows how the piece was written over time, including whether text was pasted in.

Questions people actually ask

Can an AI detector wrongly flag human writing as AI generated?

Yes, and often. Independent peer-reviewed testing has found high false-positive rates on human text, particularly for non-native English speakers and for writing that follows genre conventions closely. Detectors measure statistical surface features, not authorship.

How do I show I wrote something myself?

Show the work being made rather than arguing about the finished prose. A dated drafting history with revisions, writing time, and word-level changes is evidence a finished document cannot carry. Every piece written on Inkbreaker has one, and you can share it as a link or download it as a signed certificate.

What does "Composed in Inkbreaker" mean?

It is the mark a piece carries when the record shows it was built by typing here, over time, with no pasted text left in the finished piece. The provenance record is the underlying account; the mark is the short summary of it. Inkbreaker has no AI writing tools, so nothing here wrote any part of it.

Does Inkbreaker use AI to check my writing?

No. There is no AI anywhere on the platform. The Meridian scores prose with fixed, documented rules and arithmetic, and produces the same result for the same text every time. Moderation uses rule-based filters plus human review. Your writing is never sent to an external AI provider and is never used to train a model.

What if I drafted somewhere else and pasted it in?

The record says so, at the scale it happened, without treating it as wrongdoing. Writers draft on their phones and in notes apps, and bringing that in is normal. If you later rewrote the passage, the record reflects that too rather than describing text that is no longer there.

Does this cost anything?

No. Every published piece carries a public provenance record on any plan, free included. Pro adds the detailed evidence: the words written and cut, how the text arrived, and where revisions happened across the piece.