PasteLint

Sources & Case Studies · Published investigation

Source signals about what happens between a draft and its destination.

Editors Optimize for Readers. Humanizers Optimize for Detectors.

Editorial preparation asks whether text communicates better. Detector optimization asks whether a classifier score changes. Those are different objectives.

By Guy Teichman 9 min read

The production question

A draft is not ready because its origin can be classified.

A printed page, a community argument about AI-assisted writing, and reports from a large publishing workflow appear to belong to the same debate. One concerns a prompt left in a book. Another compares an AI system with a sewing machine. The third concerns tools designed to change detector scores.

The useful connection is not whether AI was involved. It is what happened after a draft existed.

An editor asks whether the language is accurate, clear, proportionate, accessible, and ready for its destination. Detector optimization begins with a different question: whether an automated classifier will return a different score. A change can serve the second goal while damaging the first.

Working distinction

An editor asks, “Does this improve communication?” Detector optimization asks, “Will this produce a different classifier score?”

Case study one

The prompt reached the printed page.

Reports from the 2026 Cairo International Book Fair described printed novels containing what appeared to be chatbot instructions and replies. In one reported example, a response about continuing the narrative and changing its emphasis remained immediately before the story resumed. Images of the pages circulated publicly. [1] [2]

The image is useful because the defect is visible without asking a detector to identify anything. An instruction from the production process became part of the published artifact.

That does not tell us who drafted the surrounding prose, how much revision occurred elsewhere, or whether the reported books represent a broader publishing pattern. One report also records that the available attribution was incomplete and that a publishing representative could not verify every circulating claim. Those limits matter.

Case study two

The sewing machine does not decide which seams to keep.

In a WritingWithAI discussion, a writer compares AI-assisted writing with sewing by machine. The writer describes choosing the pattern, fabric, structure, emotional direction, and moments when generated material should be deleted or rewritten. The analogy argues that operating the tool is not the whole craft. [3]

The analogy does not settle whether AI-assisted authorship is legitimate, and this article does not need it to. Its durable point is narrower: the tool does not remove the need for selection. Someone still decides what belongs, what fails, what must be reworked, and what is ready to meet a reader.

That moves the useful inquiry away from the first draft. A draft can begin with dictation, notes, a collaborator, a model, a template, or a blank page. The editorial work remains visible in the decisions that follow: preserve, reject, reorganize, verify, clarify, and stop.

Editorial sequence

Origin does not replace preparation.

  1. DraftMaterial exists

    The source may be human, assisted, generated, copied, dictated, or assembled.

  2. JudgmentClaims and structure are examined

    Meaning, evidence, voice, names, boundaries, and destination requirements are reviewed.

  3. RevisionChanges remain accountable

    The writer or editor can compare, reject, repair, and verify them.

  4. PublicationThe artifact is ready for its audience

    Readiness depends on the editorial process, not a classifier’s theory of origin.

Case study three

Changing the score can make the text worse.

A long-running WritingWithAI discussion collects reports about tools intended to make AI-assisted text appear human-written to detectors. The thread mixes recommendations, promotions, anxiety, and criticism, so it should not be treated as a controlled comparison. It does, however, document the objective users bring to these tools: change how automated systems classify the text. [4]

Production reports collected during PasteLint research described a more concrete failure pattern after large-scale testing: wording became inflated, proper nouns were changed, Unicode punctuation was corrupted, and teams manually restored large amounts of original edited copy. In those accounts, the preferred final text was often the version that had already been edited for readers—not the version optimized for a detector.

These are practitioner reports, not universal measurements. They do not prove that every such tool damages every document. They do show why a classifier score is an incomplete acceptance criterion. If a transformation alters names, punctuation, tone, or meaning, a lower score does not repair the editorial loss.

Different acceptance tests

The destination determines what counts as better.

Editorial preparation

Optimize for use.

Can a reader follow it? Can a listener understand it? Are the claims supported? Are names and technical values intact? Will the text work in publishing, documentation, speech, or another downstream system?

Detector optimization

Optimize for scoring.

Did the classifier return a different label or percentage? That result says nothing by itself about accuracy, voice, accessibility, preservation, or publication readiness.

What PasteLint prepares

Text moves toward a destination, not away from a detector.

PasteLint prepares text for readers, speech systems, publishing, documentation, AI workflows, and other downstream destinations. That work can include removing paste damage, making bounded revisions reviewable, preserving uncertainty and structure, or adapting approved text for SSML.

Those transformations have acceptance criteria a person can inspect. Did the paragraph boundary survive? Did a proper noun remain intact? Did the statement keep the same degree of certainty? Will the abbreviation be spoken correctly? Does the revised sentence communicate more directly without claiming more?

PasteLint does not claim that these changes alter detector outcomes. A detector score is not the product’s destination and is not evidence that a document is ready.

The better question

What editorial process transformed the draft before publication?

“Was AI involved?” can describe provenance. It cannot tell a reader whether a claim was checked, a name was preserved, an instruction was removed, a paragraph was revised, or a proof was read.

Trust does not come from making the origin harder to classify. It comes from making the editorial process strong enough that the published work can withstand inspection.

Related reading

Continue with preservation, uncertainty, voice, and evidence.

Sources

Evidence inspected for this article.

  1. Egyptke: reporting and page image from the Cairo International Book Fair casesPrimary regional reporting; includes explicit verification limits and incomplete attribution.
  2. Gulf News: corroborating report on chatbot process text in printed novelsSecondary reporting based on circulating images and Egyptian media accounts.
  3. WritingWithAI: “The Sewing Machine Analogy”Practitioner argument and community discussion, not empirical authorship evidence.
  4. WritingWithAI: Humanizer Applications discussionHeterogeneous community reports, recommendations, promotions, and criticism.