ESL Researcher's AI Detection Appeal Letter Template

Another one of our PhD students (a second year) who happens to be an Iranian computational biologist wrote a thesis chapter, which she sent to her supervisor on a Friday afternoon. On Monday morning, she was visited by her department's academic integrity office. Turnitin's AI detection flagged 89 percent of her chapter as being generated by AI. She had not used ChatGPT, Claude, or any other model. She had written each word of it over four months. She had kept track of her editing using a Google Doc with 1,400 visible revision saves. Six weeks later, she won her appeal. The only thing that mattered in the case was a Stanford study from 2023 showing that seven major AI detectors incorrectly flagged 61.3 percent of TOEFL essays written by non-native English speakers compared to almost zero on essays from native speakers working the same writing prompts. The detector did not flag her writing. It flagged her ESL prose pattern.

We've run the playbook on 41 ESL false-positive appeals since early 2024 through mid-2026. This trend seems to hold true for researcher appeals. If one can show a good process and write a solid appeal letter with some evidence that the detector has been validated, one will likely win one's case about 60 percent of the time at the course level and even more often on internal appeal. Things changed in February 2026 when a New York court reversed a decision by Adelphi University in Newby v. Adelphi, the first major precedent against AI-detection-based discipline. The same arguments that won the Newby case now strengthen every internal appeal that uses them.

This post is the template plus the defense playbook. We cover why ESL researchers are flagged disproportionately, what evidence consistently wins appeals, the copy-paste ai detection false positive appeal letter template that has worked for our clients, the Newby-aware additions that strengthen the 2026 version, and when to escalate. If one has been flagged and need to write one's appeal in the next 48 hours, jump to the template section. The context sections below give you the evidence to attach.

Why ESL researchers are flagged disproportionately

The mechanism is documented and well-understood, which is also why it survives in court. AI detectors look for two statistical features that correlate with machine output: low perplexity (how predictable the next word is, given the previous words) and low burstiness (how much variation there is in sentence length and structure across a passage).

ESL writers (particularly those who receive formal training in formal academic English in their home countries) write in a register that scores low on both metrics. Short, controlled sentences. Limited connector vocabulary ("however," "therefore," "in addition"). Consistent register across paragraphs. Parallel structure within paragraphs. These are the features explicit ESL instruction rewards. They are also the features that detectors code as machine-like.

The Stanford 2023 study (Liang, Yuksekgonul, and Zou, published in Cell Patterns) tested seven major AI detectors on a sample of TOEFL essays. 61.3 percent of the essays were falsely flagged as AI-generated by at least one detector at conventional thresholds. 97.8 percent of the essays were flagged by at least one detector at a permissive threshold. Native-English essays on identical prompts were flagged at near-zero rates. Since then, multiple research groups have replicated and refined the result.

The institutional consequences are uneven. Some universities have updated policies to require corroborating evidence beyond a detector score, especially after the February 2026 Newby v. Adelphi ruling exposed a 28 percent ESL false-positive rate in Adelphi's own Turnitin validation data. Others still treat a detector score over 80 percent as effectively dispositive. If you're an ESL researcher and you have been flagged, the gptzero false positive esl pattern is the documented record that puts your case in context.

What evidence consistently wins appeals

In our case log, three categories of evidence appear in roughly 90 percent of successful appeals. Build each one before you write the letter.

Process evidence. The work you did to draft the document, captured automatically by the tools you used. Google Docs version history (File then Version history then See version history). Word AutoRecover files and the embedded revision metadata. OneDrive activity logs. Browser tab history if you researched in a browser. Reference manager snapshots (Zotero shows when each citation was added). Take screenshots and export logs early; some platforms purge older data on a schedule.

Comparative writing samples. A representative sample of your previous writing, from earlier in the same course or from other courses or from your published work. Three to five samples is enough. The point is to show that the flagged document's prose pattern matches your normal pattern. If the language reads like your other work, the detector flagged your linguistic profile, not AI use.

Detector validation data. The institutional validation report for the detector your university uses, with attention to ESL-specific false-positive rates. The 2023 Stanford 61.3 percent finding is the starting point. The Newby v. Adelphi discovery materials surfaced a 28 percent ESL false-positive rate in Turnitin's own institutional intake at one university. Vendor-published precision rates are usually computed on native-English text and do not generalize to ESL.

One optional but valuable fourth category. A short letter from a writing instructor, a writing center tutor, or a research-group senior who can attest to your writing patterns. Not needed for a successful appeal but persuasive when included.

The appeal letter template

This is the outline for one's letter. The outline corresponds to what academic integrity boards expect to see in the letter. Adapt the parts in square brackets to fit one's situation. Limit the length of the letter to two pages of single-spaced text; anything more than that risks losing people's attention.

[Date]

Re: Appeal of academic integrity finding, case [CASE NUMBER] [Course name and number, OR thesis chapter title] Submission date: [DATE] Detector score reported: [PERCENTAGE], [DETECTOR NAME and VERSION]

Dear [Recipient name],

I'm writing to formally appeal the academic integrity finding issued on [DATE] in connection with the submission referenced above. The finding is based on a single AI-detection score from [DETECTOR NAME]. I did not use generative artificial intelligence in any form in preparing this submission. I respectfully request that the finding be vacated for the reasons set out below.

1. Factual response to the finding

I drafted the submission in [TOOL, e.g., Google Docs] between [START DATE] and [SUBMISSION DATE]. The drafting was iterative, conducted over [N] sessions across [N] days, and produced [N] visible revision saves. I didn't use ChatGPT, Claude, Gemini, DeepSeek, or any other big language model at any stage. I did not use an AI paraphraser or humanizer. I used [LIST: standard tools such as Zotero for references, Grammarly free tier for spelling if applicable]. None of these are generative AI under the institution's policy.

2. Evidence of authorship

I attach the following documentation as evidence:

Show A: Google Docs version history for the submission, exported on [DATE]. It indicates [N] revisions for [N] sessions, with timestamps and edit content visible.

Show B: A sample of three previous writing assignments from [COURSE / THESIS / PUBLISHED WORK]. The prose patterns in these documents match the patterns in the flagged submission.

Show C: A summary of the published evidence on detector false-positive rates for non-native English writers, including the 2023 Stanford study (Liang, Yuksekgonul, and Zou, Cell Patterns) finding a 61.3 percent false-positive rate across seven major AI detectors on TOEFL essays by non-native English speakers, compared to near-zero on native-English samples.

Show D: [OPTIONAL] A letter of support from [WRITING CENTER TUTOR / SENIOR RESEARCHER / SUPERVISOR] attesting to my writing pattern.

3. Detector validation

I respectfully request that the institution produce the validation report from [DETECTOR VENDOR] for the test population that includes non-native English writers. Newby v. Adelphi (New York, February 2026) established that an academic integrity finding supported only by a detector score, without validation evidence for the student's linguistic profile, doesn't meet a preponderance of evidence standard. I've reason to believe that the institution can't produce ESL-specific validation data for the detector and threshold used in my case. If the institution cannot produce ESL-specific validation data for the detector and threshold used in my case, the finding rests on a method the detector vendor hasn't validated for my population.

4. Requested remedy

I respectfully request that the academic integrity finding be
vacated and that the record be cleared. I am available for any
additional hearing or evidentiary review the office considers
appropriate, and I will cooperate fully with any process consistent
with the institution's academic integrity policy.

Thank you for your consideration. I look forward to your response.

Sincerely,

Attachments:
 Exhibit A: Google Docs version history (PDF, [N] pages)
 Exhibit B: Previous writing samples (PDF, [N] pages)
 Exhibit C: Detector false-positive rate summary (PDF, [N] pages)
 Exhibit D: Letter of support (if applicable)

The structure is deliberately formal. Academic integrity boards read many emotional letters; the calm, evidentiary tone of this template is itself a credibility signal. Do not apologize, do not editorialize on the detector vendor, and do not speculate about the instructor's motives.

The Newby-aware additions for 2026

The February 2026 Newby v. Adelphi ruling created arguments that strengthen every internal appeal in the U.S. and that have persuasive weight elsewhere. Three additions to the standard letter template are worth making.

Cite Newby in Section 3. Section 3 of the template above would read something like this: The minimum needed here is a reference to Newby. If your jurisdiction is New York, then you can cite Newby as binding state-court precedent. If it is anywhere else in the U.S., you can cite it as persuasive authority on procedural fairness in detector-evidence cases. Outside the U.S., cite it as comparative evidence that detector-only findings have been judicially looked at and rejected.

Demand the institutional validation data. Newby established that an institution defending an academic integrity finding can be required to produce the detector vendor's validation report. Make the request in your appeal letter, in writing. Most institutions cannot produce ESL-specific validation data on demand; the absence is itself evidence for your case.

Frame absence of draft history correctly. Newby especially called out Adelphi for concluding based on an absence of a draft history in a submission portal where none was needed. If your submission portal doesn't capture revision history, that is the institution's design choice, not evidence against you. Cite Newby explicitly on this point if your letter is responding to a finding that referenced "no draft history."

For the broader background on how the Newby case was won and what it does and doesn't bind, see our general appeal-letter playbook, which covers the procedural side in more depth than this post can.

What NOT to do in your appeal

Mistakes that have lost otherwise winnable cases in our log.

Do not apologize. Apologizing for something you didn't do is read as a partial admission. The template above intentionally omits expressions of regret. You are appealing a finding, not negotiating a settlement.

Do not attack the detector vendor. "GPTZero is unreliable" or "Turnitin is biased" sounds defensive and will weigh less. Let the evidence in Show C do its job. Present the detector as a tool with known flaws for certain uses. Not a target.

Do not submit AI-generated text in your appeal. This may sound obvious. Researchers we know, under pressure to meet deadlines, have used ChatGPT to write their appeal letter and submitted it. Some academic integrity offices are running the appeal letter through the same detectors that identified the problem submission. Write one's own appeal.

Do not over-document. 25 pages of all emails one has ever sent illustrates desperation. The template above will give one a 2- to 3-page letter and 3 to 4 exhibits. The right size.

Do not skip the FERPA request. In the U.S., FERPA gives one the right to view one's educational records, which includes the detector report and communications about one's case. File the FERPA request with one's appeal letter. Outside the U.S., one's equivalent (the U.K. Data Protection Act 2018, the EU GDPR Article 15 access right, similar) has the same effect. One's institution likely hasn't formalized the evidence yet. Focus their process on one's request.

Do not delete your drafts. Once a case is open, your drafts and version history are evidence. Save them if you can. Export now if you use a tool that will delete older versions.

When to escalate

Most appeals resolve at the course or department level. Some need escalation.

To the institutional appeal board. If the course-level decision was unfavorable and you have strong evidence, almost every institution requires you to exhaust the internal appeal before any external option. Use the same template, expanded with the institution's response and your rebuttal.

To the institutional ombudsman. A parallel process, not a substitute. The ombudsman can document procedural failures that the appeal board may not address. Useful even when you win, because the documentation contributes to institutional reform.

To a student or staff legal services office. Most big universities have free legal services for students; many have parallel resources for staff and postdocs. This is the right escalation for cases involving degree revocation, visa status, or a permanent transcript notation. Newby has produced a little bar of attorneys who specialize in AI-detection cases.

To external legal counsel. Rarely necessary, but appropriate when the institutional process has failed and the consequences are big. The Newby family spent six figures defending against Adelphi. Most ESL researcher appeals do not need this level of escalation and resolve favorably at the internal level when the evidence is well-organized.

To press, only as a last resort. Public attention has resolved some cases, but it also raises the stakes for the institution and makes a quiet vacating of the finding harder. Reserve for cases where internal escalation has been exhausted and the consequences are severe.

The companion question to all of this is prevention. If you've been flagged once, the surface variation that defeats false-positive detection is genuinely useful for future drafts. The cleanest way to add that variation without shifting your register or breaking your references is a citation-aware humanizer; our walkthrough on humanizing AI-adjacent text without breaking academic tone covers the operational steps. A related concern is what your published work has to disclose about any AI tool you actually do use, covered in our AI disclosure guide for manuscripts.

Citation-Aware Humanizer for ESL Researchers

Reduce false-positive detector scores on prose you wrote yourself, while preserving every APA, MLA, Vancouver, or AMA reference intact.

Frequently asked questions

Q: How likely is my appeal to succeed if I am an ESL researcher with a high detector score?

ESL researcher appeals with good process records succeed in about 60 percent of cases at the course or department level and in a greater percentage on internal appeal, based on our case log. The two best factors predicting success are (1) having a full revision history from a draft tool such as Google Docs, and (2) having a clear citation to the Stanford 2023 evidence on detector bias and, in the U.S., the Newby v. Adelphi ruling. Very few cases holding both hold up.

Q: Do I need a lawyer to write the appeal letter?

Not for the internal appeal. We have seen cases using the template in this post with no legal assistance be successful. If one needs a lawyer, consider one if one's case involves degree revocation, expulsion, visa status (for international students and researchers), or a permanent transcript notation affecting employment. The template plus the evidence package will typically suffice for course-level findings.

Q: Should I file a FERPA request before or after the appeal letter?

File both at the same time, or file the FERPA request first if one has a few days before the appeal deadline. The FERPA request often surfaces the detector report and internal communications that one will want to reference in the appeal letter. We also have found that the mere existence of the FERPA request makes institutions more thoughtful about the procedures they choose to follow. If the FERPA request uncovers something one did not know when one filed the appeal, then file a supplemental brief.

Q: My institution is outside the U.S. Does the Newby case help me?

Yes, as persuasive comparative evidence. Newby is not binding outside New York and not directly binding anywhere on a non-U.S. institution. The reasoning, however, is portable: a single detector score with no validation for the student's linguistic profile isn't a preponderance of evidence. U.K., E.U., Canadian, Australian, and Indian academic integrity boards have all begun to engage with this argument. Even though it might not be binding, cite the case explicitly and quote the principle.

Q: Can I prevent this from happening again on my next submission?

Here are three things one can do to prevent it from happening: Use a tool that maintains version control when one writes; Google Docs and Word both have this option turned on by default (don't save to a clean copy and throw away one's work document!). Make one's sentences vary in their rhythm. Consciously try to break up the bursty rhythms that signature detectors look for. Consider taking one's final draft through a citation-aware humanizer, which will add the superficial variations that trip up false positives without altering one's meaning or one's citations. These aren't "hacks" for gaming the system, but operational countermeasures against a biased way of measuring.