Research
The evidence behind the company.
Writence started from one finding: AI-writing detectors falsely flag non-native English writers far more often than they flag native ones. Four layers of evidence — mathematical, empirical, real-world, and human — point the same direction. Here is the case, with sources, and what we build in response.
The thesis
The writers who work hardest at English are the ones most often accused of not writing their own words. That isn’t a discipline problem. It’s a measurement problem — and the measurement is biased.
We don’t claim detectors are useless or that anyone is acting in bad faith. We claim something narrower and better-supported: on second-language English, today’s detectors are wrong often enough that no one should treat their output as proof.
FOUR LAYERS OF EVIDENCE
Why we’re confident the bias is real
No single study carries the argument. The case is strong because four independent kinds of evidence agree.
Mathematical
Detectors can’t stop flagging ESL text
A detector that separates “AI” from “human” by surface statistics is measuring the same things that make second-language English distinctive: lower word-level variety, more frequent words, simpler sentence shapes. Garland (2026) frames this as a structural limit — past a point, lowering false positives on native writing raises them on non-native writing. It is a trade-off baked into the method, not a bug a vendor can patch away.
Empirical
The studies keep finding the same gap
Stanford (Liang et al., 2023) found leading detectors flagged a majority of TOEFL essays by non-native writers as AI-generated, while near-perfectly clearing essays by native US students. Later work (Maryland, 2025; and others) reproduces the direction of the effect across detectors and datasets. The bias is consistent, not a one-paper artifact.
Real-world FPR
False-positive rates far above the claim
Vendors typically advertise false-positive rates below 1%. Liang et al. (2023, Stanford) found leading detectors misclassified 61.22% of TOEFL essays by non-native writers as AI-generated. On a class of a few hundred international students, even a fraction of that rate means real people wrongly accused.
Emotional
The human cost: “flagxiety”
Surveys put the share of US students who feel anxious about being falsely flagged at roughly three-quarters — and report it running about twice as high among international students, who already carry the heaviest burden of proof. Self-censoring your own clearest sentence to avoid a detector is a tax paid before a single accusation is ever made.
Figures are drawn from published studies and independent testing; see the full article for sources and exact methodology. Reported ranges vary by detector and writing sample.
OUR RESPONSE
Don’t argue with the detector. Document the writing.
You can’t prove a negative to a black box. So we built the opposite of a detector: a cryptographic Authorship Certificate. As you write, your editor records an append-only chain of edits, signed with ed25519 keys, that anyone can verify independently.
The certificate documents the process by which a piece of writing came to exist. It is evidence of how the work was made — not a legal verdict, and not a guarantee against any detector. We don’t promise to beat detection. We give writers a fair, checkable record.
How the Authorship Certificate works →- Append-only edit chain — never mutated, never deleted within retention.
- ed25519 signatures, so the record can’t be forged after the fact.
- Publicly verifiable: anyone with the link can check it, with no account.
- Consent-first: the chain only records when a writer opts in.
- Documents process; does not claim to prove innocence or pass detectors.
WRITENCE PAPERS
What we publish, and why
Short papers setting out our position, the mechanism behind it, and what we are not claiming. Written to be checked, adopted or argued with by people who do not work for us.
Paper 01
Provenance, not detection
The question a detector asks has a moving answer. The question “how was this text produced” does not. The full mechanism — signing, key publication, third-party verification — with nothing withheld that a verifier needs.
Read the paper →Paper 08
A policy that does not misfire
For the person who writes the rule, not the student accused under it. Five things a policy needs, three patterns to avoid, and a model clause you may copy without attribution.
Read the paper → Read the evidence.
Decide for yourself.
The full article walks through every study, the math, the real-world numbers, and the human cost — with sources.
Evaluating Perplexity Bias in Academic Detection
Quantitative assessments demonstrate that standardized models consistently mistake concise, rule-based second-language phrasing for synthetic generation.
The protocol generates a cryptographic signature confirming the record’s integrity, without exposing unpublished findings.