Process Is the New Proof: Tracked Changes Beat Detectors

Newby v. Adelphi was decided by a New York court that vacated an academic integrity finding based on a single Turnitin AI score not being a preponderance of evidence in itself. Spring 2026 pilot at hundreds of institutions rolls out draft forensics that capture pasted text, writing time, and a full playback of a student's revision history. Clarity product, which Turnitin owns, is involved. April 10, 2026, Anthropic launches Claude for Word into public beta, enabling natively tracked changes, which means Claude's edits are added to Word's review pane as revisions. Every change is reviewable. All original lines remain intact. Curtin University turns off Turnitin's AI detection entirely because they've had more disputes where the detector was wrong than where it caught the students who cheated. Four signals in the first half of 2026 all push the same way.

Four signals are a pattern. The pattern is that the trusted evidence of authorship in academic writing has stopped being a detector score and started being a process record. Courts, journals, conferences, and the detector vendors themselves are converging on the same position: draft history beats probability scores.

This post is the brand argument we will make for the rest of 2026. We cover the four convergence signals in detail, the asymmetry between detector evidence and process evidence, what defensible process documentation actually looks like across the tools researchers already use, and how to build drafting habits that produce a record you can hand to an editor, a reviewer, or an academic integrity board without having to think about it. The reframe is operational, not philosophical: stop optimizing for the detector, start documenting the work.

The four convergence signals

The shift has not been theorized into existence. It has been pushed by four concrete 2025-2026 events, each from a different part of the academic ecosystem.

Newby v. Adelphi (February 2026, New York state court). The first major court decision about AI detection discipline reversed an Adelphi verdict against an ESL undergraduate. The result was narrow but the logic decisive: a detector score without validation for the student's linguistic profile does not make up a preponderance of evidence per the school's policy. Specifically, the court admonished Adelphi for using the lack of a draft history in the submission portal as evidence of guilt (although the portal did not require draft uploads). This adds process evidence (or the lack of it) to the legal record.

Turnitin Clarity (Spring 2026 pilot, broad rollout in progress). Turnitin's add-on to Feedback Studio offers complete playback of a student's writing process to instructors, including when the student pasted text, the length of the writing session, the construction time, and the draft history. With student permission, Clarity records keystroke rhythm and editing timeline. The product's stated frame is "assessment for integrity, not just detection." A major shift from the company that produced the AI detection score upon which academic institutions have leaned for three years. Now they're offering the process record as the more dependable signal. Claude from

Claude for Word (April 10, 2026, public beta). Anthropic ships as a native Microsoft Word add-in. When Claude makes an edit, it drops it in the Word review pane as a tracked change, preserving the original text as a deletion and adding the new text as an insertion. The author accepts or rejects each one in the native Word interface. The full edit history is available to the reviewer who opens the document. This isn't a feature add. It's an architectural choice that treats the tracked-changes record as the canonical interface between human author and AI assistant, exactly the record a journal or a thesis committee can use to verify the work. After

Curtin University (early 2026, Australia) and parallel institutions. Curtin concluded that the score was generating more disputes than it resolved, it shut off Turnitin's AI detection altogether. Vanderbilt and several Michigan State departments came to the same conclusion in 2025. None of these universities banned AI use. They took out the detection layer and pushed the integrity conversation to focus on disclosure, process, and tracked-changes evidence. At the same time, other universities (notably some Big Ten and Ivy League institutions in late 2025) doubled down on detection, making the field uneven but not changing the direction of the shift.

Four signals, four parts of the system, one direction. This is not an accident; this is the system rebalancing around easier-to-prove evidence than a custom probability score.

Why the shift makes sense

The enduring nature of the shift is because of the asymmetry between detector evidence and process evidence. Both types of evidence have known floors of accuracy. The detector floor is much lower.

Detector evidence is a probability score derived from a model that learns text features correlated with AI output. The Stanford 2023 study found seven major detectors that incorrectly identified 61.3 percent of TOEFL essays by non-native English speakers as AI-generated. The Adelphi institutional data surfaced in Newby discovery had a 28 percent ESL false-positive rate on the university's own Turnitin intake. Vendor-provided precision numbers are usually calculated on native-English text and don't generalize. The accuracy of the score depends on the writer's language background, the prompt distribution the detector was trained on, and the threshold the institution set. The author does not see any of these things at the time of submission.

Process evidence is the chain of events, timestamps, and revisions leading to the final document. Google Docs tracks all revisions automatically. Word AutoRecover and the built-in revision info track all revisions. OneDrive logs when the file was opened and saved. Browser tab history tells when the writer was doing research into the sources they cited. Reference manager software (Zotero, Mendeley, EndNote, Paperpile) timestamps each citation when it's added. All the evidence is created as part of the process. It can't be made after the fact without investing the same amount of time as the first writing.

This is what makes the shift load-bearing: The trust asymmetry. The detector score can be challenged on grounds of validation, threshold, and population bias; the institution defending the score needs to provide evidence they often do not have. The process record can be challenged on grounds of authenticity, which requires evidence the institution would need to provide against a writer who can point to platform-generated metadata. Almost always, the defender of the process record has the easier case.

This is not a humanizer-versus-detector argument. The humanizer changes the surface; the process record changes the evidence base. Even if every AI detector tomorrow improved to perfect accuracy on native English, the process record would still be the more useful signal because it speaks to authorship rather than to probability.

What defensible process documentation looks like

The good news is that almost every tool researchers already use generates process evidence as a default. The work is in capturing it and knowing where to find it when needed.

Google Docs. File then Version history then See version history. The full revision tree is preserved indefinitely, including who edited and when. Export as a PDF if a shareable record is needed; the export shows the timeline. The single most useful process tool available to academic writers in 2026, mostly because it's the default, is preserved indefinitely.

Microsoft Word. AutoRecover saves auto-saved versions in a hidden folder. Track Changes (Review tab then Track Changes) captures every edit you make once enabled. The workflow change worth making is to turn on Track Changes for your own drafting: start out with Track Changes turned on and it will capture all of your edits from that point forward. It's important to save versions explicitly with descriptive names ("draft-2026-07-12-section-3-rewrite") so the sequence is legible without opening each file. Most researchers do not turn Track Changes on for their own drafting, but if you want to track changes as you write, it is the best way to do so.

OneDrive, Google Drive, Dropbox. Cloud storage providers log file activity. Most academic IT services preserve this for at least 90 days; some longer. The log establishes when the file was modified, by whom, and from which device. A useful supplement when your primary editor's revision history is incomplete.

Reference managers. Zotero, Mendeley, EndNote, and Paperpile all record the date that a citation was added to your library and to which document. Process evidence on the research stage, it is evidence before and validation of the writing stage. Suppose a reviewer thinks one cited sources one didn't read. One can answer him or her with the timestamp when one read them.

Tracked changes export. This is the single most useful output to share one's process evidence with a reviewer, editor, or integrity board. The tracked changes are a Word or Google Docs document with all edits showing clearly (who did what, when). Tracked changes are enabled by default in our proofreader, but there's also an option to generate "raw text" if one need to see the clean version. We recommend this version for any submission where one might be asked about the writing process. Note that this can be generated natively in Claude for Word. Claude for Word produces the same artifact natively, which is the architectural point of the product.

Email and chat archives. Discussions with supervisors, co-authors, and writing center tutors. Save the threads. They establish the iterative thinking that produced the manuscript, which no detector can fake.

The drafting habits that produce defensible process

Five habits, none of which adds meaningful friction to your existing workflow, that compound into a defensible process record.

1. Draft in a tool that preserves revision history by default. Google Docs or Word with Track Changes on. Don't draft in plain text editors that do not preserve history; don't export to a clean copy and discard the working file. If your existing workflow involves drafting in a notebook and typing up later, the typed version is the draft; preserve it.

2. Save versions explicitly with dated, descriptive names. Once a week, save a snapshot of the current draft with a name like "manuscript-2026-07-18-after-methods-rewrite." The named versions make the sequence legible at a glance and survive cloud-storage retention purges.

3. Keep a one-line edit log alongside the manuscript. A plain text file in the same folder. Each entry is a date and a sentence: "2026-07-18: rewrote Section 3 methodology after meeting with supervisor." Five minutes per week, captures the thinking the revision history cannot show on its own.

4. Run AI-assisted edits through a tracked-changes pipeline. If you use Claude for Word, ChatGPT, or any other AI tool to help with editing, do the work through a pipeline that produces a tracked-changes record. Claude for Word does this natively. Our proofreader does this on every export. The record is what makes any future disclosure or integrity conversation tractable.

5. Archive the supporting research. Reference manager exports, downloaded PDFs in a dated folder, browser bookmarks of source pages. The research record validates the citation record, which validates the writing record. Each layer protects the layer above it.

Five habits, roughly 15 minutes per week of overhead, produce a process record that defends any future challenge from a detector, a reviewer, or an integrity board. For the broader question of how to document AI-tool use especially in the manuscript itself, our AI disclosure guide for manuscripts covers the disclosure side; for the appeal-side question of what to do if you've been flagged despite a clean process, the appeal-letter playbook walks through the response.

What this means for institutions, journals, and detector vendors

The shift has institutional consequences that the next 18 months will work out.

For universities, the operational question is whether to require draft history capture at submission, what consent framework supports it, and how to weigh the captured evidence against detector scores in disciplinary processes. The institutions that handle this well will publish a clear policy that names the trusted evidence and the threshold for action. The institutions that hedge will face Newby-style appeals on a regular basis.

For journals, the question is how to integrate process evidence into the peer-review workflow without adding friction the system can't absorb. Some journals will require tracked-changes uploads alongside the manuscript; some will accept process attestations in the disclosure statement. Some will outsource the verification to vendors building publisher-facing tooling. The journals that move first will set the de facto standard.

For detector vendors, the shift is existential. Turnitin's Clarity pivot is the case study: the company that built the detection score is now offering the process record as the more reliable signal. Vendors that refuse to make the same pivot will lose institutional contracts to vendors that do. The detection score won't disappear; it'll become one signal among several, weighted lower than the process record in any well-designed integrity process. Our note on Turnitin Clarity's process forensics covers what the product actually does and where institutions sit in the rollout.

The implication is the cleanest for researchers. The work to produce the record is little. The defensibility it provides is big. The reframe from "pass the detector" to "document the process" is the brand position we'll defend for the rest of the year.

Proofreader Built for the Process-Is-Proof Era

Tracked-changes export on every edit. Tool and version captured automatically. The record your editor, reviewer, or integrity board will trust.

Frequently asked questions

Q: Does "process is the new proof" mean AI detectors are going away?

Not entirely. Detectors will remain as one screening signal, specifically in undergraduate teaching, where the volume of submissions makes process-evidence review expensive. What's changing is the weight: a detector score on its own is no longer accepted as main evidence in disciplinary processes at well-designed institutions, and the Newby v. Adelphi ruling formalized that position in court. The score becomes the input to a closer review, not the conclusion.

Q: I draft on paper and type up later. Do I lose the process defense?

Not entirely, but the protection is weaker. The typed-up version is one's draft for process-evidence purposes; preserve the document from the moment one starts typing and let it accumulate revision history. Photographing or scanning one's handwritten notes and dating the images adds a second layer. The dual record (handwritten + typed revisions) is actually stronger than a typed-only record, because the handwritten layer is harder to make retroactively.

Q: What if my institution does not accept tracked changes as evidence?

File a FERPA request (in the U.S.; equivalent under U.K. Data Protection Act 2018, EU GDPR Article 15 elsewhere) to get one's full record. Submit tracked changes alongside any appeal regardless of whether the institution has formally accepted the format. The Newby ruling established that the absence of a clear process for considering this evidence is itself a procedural issue. Institutions are updating their policies in response, and providing the record proactively is part of the pressure that drives the update.

Q: Should I turn on Track Changes for my own drafting, even when I am the only author?

In our experience yes, and it's the workflow change most researchers benefit from. Track Changes on from the start of every document gives you a complete edit record without any more discipline. Hide the markup view for normal drafting (Review tab then Display for Review then No Markup), but leave the underlying record in place and export the tracked-changes version when needed. The friction is little and the protection is real.

Q: Does this apply outside the U.S., where Newby is not binding precedent?

Yes, for two reasons. First, the Turnitin Clarity rollout, Claude for Word, and journal-policy shifts are global, not jurisdiction-specific; the trusted-evidence shift is moving across academic publishing regardless of where one submits. Second, the Newby reasoning is portable: courts and integrity boards in the U.K., Australia, Canada, India, and the E.U. have engaged with the argument that detector scores alone do not make up a preponderance of evidence. Cite the case explicitly even where it's persuasive rather than binding; the reasoning travels well.