Do rewrites that hide authorship also fool AI detectors?
A paper claims heavily rewritten AI text becomes both unattributable to humans and undetectable as AI-assisted, but tests only the attribution half. Does rewriting actually evade detection systems?
The paper claims a second erasure stacked on the first. Its abstract says heavily rewritten messages "may also evade AI-text detectors, making them difficult both to attribute to their human authors and to identify as AI-assisted," and it calls this double erasure. The contribution list repeats the point and cites two detector references, [16, 19]. The excerpt contains no detector experiment: it names no detector, reports no detection rate, and describes no sample of rewritten text run through one. The double erasure is an implication the paper asserts beside its attribution result, not something it measures.
The excerpt places the detector claim directly after its stylistic-convergence analyses, but it gives no mechanism that links the two. The attribution results show that rewriting leaves a model less able to tell authors apart. Nothing in the excerpt shows that an AI-text detector responds to the same signals, so a reader has to supply that link as an inference, and the paper does not make it.
The question sits against Can humans detect AI text if machines can measure it?, which holds that AI text can differ measurably from human text while human judges cannot perceive the difference. Double erasure asks the reverse question for rewritten text: whether rewriting narrows the measurable gap that detection depends on. The sibling insight, How much does AI rewriting erase distinctive author voice?, supplies the first half, the attribution loss, and that is the half the excerpt measures. Together the two claims would describe text that is neither attributable to its author nor flagged as AI-assisted, a case the attribution result alone cannot reach.
What the excerpt does not establish is which detectors were involved, on what texts, at what thresholds, and whether anything holds for the three assistants and three registers used in the attribution study. The detector half should be treated as a hypothesis for a direct test. The paper says it releases its protocol and prompts, so such a test could start from the same rewriting setup. If the claim held, a detector-based check on workplace email would miss rewritten messages that attribution also cannot trace back to their authors, but that conditional is as far as this excerpt can carry it.
Inquiring lines that read this note 48
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How reliably can humans and AI detectors identify machine-generated text?- How do lay readers differ from classifiers in detecting AI text?
- Are detector errors on AI text systematic or random by design?
- Can AI detectors reliably distinguish human from machine-generated text?
- Can a classifier distinguish machine-written text from poor human writing?
- What signals do AI text detectors actually measure in their classification?
- Does the same rewriting that erases authorship also narrow measurable AI text markers?
- Can detection systems identify AI text rewritten to match human author style?
- Would detectors trained on unaltered AI text catch heavily rewritten versions?
- Can AI-rewritten text still be detected as machine-modified?
- Can text detection methods distinguish between AI collaboration and delegation?
- How accurate is Originality.ai's detector at identifying AI-written content?
- What false-positive rate would indicate the classifier harms legitimate human writers?
- Can language models detect AI-generated text in blind evaluation tasks?
- Can AI detectors confuse distinctive writing style for machine authorship?
- Can AI text detection improve enough to help evaluators make better decisions?
- How accurate are automated AI text detectors compared to human judgment?
- Can machine-readable crawler intent declarations solve attribution problems?
- Do writers edit AI assistance enough to fool content filters?
- Can explanation-based AI safeguards work in real-time writing interfaces?
- Does AI writing assistance make different authors sound more alike?
- Does asking AI to preserve voice recover lost authorship signals?
- How does hiding AI use from readers differ from showing it to collaborators?
- What tolerance limits exist for AI visibility in shared writing?
- Do writers claim authorship without feeling they wrote the words?
- When AI becomes invisible in writing tools, do writers stop disclosing it?
- Why do people withhold AI credit even when using personalized text generation?
- How do attribution norms for human ghostwriters compare to AI usage patterns?
- Do human readers still recognize authors after heavy AI rewriting?
- How much of AI-assisted comments remain the writer's own words?
- Does the 'feel of AI' in unedited posts trigger audience backlash and detection?
- What makes readers suspect AI involvement in academic writing they evaluate?
- Can readers detect AI involvement in writing when not explicitly told?
- Does directly adopted AI text require disclosure even if methodology is unchanged?
- Can commercial AI detectors accurately identify AI-written application essays?
- Do admissions penalties follow actual AI detection or suspected authorship?
- What heuristics do readers use to detect or fail to detect LLM writing?
- Did reviewers successfully circumvent ICML's hidden-instruction watermark detection method?
Related concepts in this collection 2
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
How much does AI rewriting erase distinctive author voice?
Does heavy AI rewriting weaken the computational signals that identify individual authors? The question matters because it bears on whether AI assistance erases stylistic distinctiveness—a possible cost of polish and consistency.
the attribution loss this question extends; the detector half is untested here
-
Can humans detect AI text if machines can measure it?
AI-generated text shows measurable differences from human writing across multiple linguistic dimensions, yet human judges consistently fail to identify it. Why does the gap between what is measurable and what is perceptible exist?
the measurable-gap claim that double erasure would narrow from the rewriting side
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- The Assistant Erased You: Measuring Loss of Authorship Signals in AI-Mediated Communication
- What Influences Readers' and Writers' Perceived Necessity of AI Disclosure?
- Understanding Reader Perception Shifts upon Disclosure of AI Authorship
- "That's AI Slop, You Bot!" Studying Accusations, Evidence, and Credibility in Online Discourse Towards LLM-Generated Comments
- Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews
- Linguistic markers of inherently false AI communication and intentionally false human communication: Evidence from hotel reviews
- Hidden Prompts in Manuscripts Exploit AI-Assisted Peer Review
- The AI Ghostwriter Effect: When Users Do Not Perceive Ownership of AI-Generated Text But Self-Declare as Authors
Original note title
Heavily rewritten messages may also evade AI-text detectors — a double erasure the paper names but does not test