Frozen text
The exact submitted article is preserved at intake so later edits do not change the object of study.
Recruitment phase
What can an AI detector know about authorship?
This independent study compares detector labels with documented writing workflows—not to ask whether prose looks machine-like, but whether the label readers see accurately describes how the work was made.
Led by Michael Kloth · The Art of Photography AI
01 / The question
A score can describe textual patterns. Can it describe provenance?
AI detectors evaluate features of finished text. They cannot observe who formed the argument, where a sentence came from, what an editor changed, or why the writer kept it.
This study places those two kinds of evidence side by side: the detector’s output and a contributor’s documented, post-specific workflow.
02 / What is compared
Four records. One article.
Workflow attestation
The writer records what happened on that specific article, including whether AI flagged problems or supplied language.
Detector result
Pangram’s label, score, flagged passages, scan date, and interface version are recorded without overwriting earlier runs.
Review
The label is evaluated against documented provenance, with limitations and exclusion decisions retained.
03 / Commitments
Designed to be inspectable.
- Transparent position. The researcher uses AI openly and developed the study after seeing detector labels applied to his own work.
- Post-specific evidence. General habits are not substituted for what happened on the article being studied.
- Granular permission. Participation, attribution, quotation, dataset inclusion, and future contact are separate choices.
- Contrary results count. A null result—or a finding favorable to the detector—will be reported plainly.
04 / Initial cohort
Substack writers with a candid record of how they work.
The first phase is intentionally invitation-led. It begins with known contacts and writers who have already discussed their use—or non-use—of AI publicly.
Check eligibility“The study is not asking a detector to read a writer’s mind. It is asking whether the label shown to readers survives contact with documented provenance.”
05 / Study status
Recruitment comes before infrastructure.
Current phase: interest and cohort validation
The full intake and analysis workflow will proceed only if a sufficiently trustworthy cohort participates. This recruitment gate is part of the protocol, not an afterthought.
Contribute to the record
Your process is part of the evidence.
Expressing interest takes a few minutes. It is not consent to participate and does not require you to submit an article.
Learn about participation