@inproceedings{raichura-2026-meridian,
title = "Meridian: A Bivariate Perplexity Architecture for Trustworthy and Equitable {AI} Text Detection",
author = "Raichura, Pranil",
editor = "Atwell, Katherine and
Borah, Angana and
Dementieva, Daryna and
Kreiss, Elisa and
Kotonya, Neema and
Liu, Jiarui and
Olson, Liz and
Wan, Ruyuan and
Zhao, Jieyu",
booktitle = "Proceedings of the Fifth Workshop on {NLP} for Positive Impact ({NLP}4{PI})",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.nlp4pi-1.27/",
pages = "312--325",
ISBN = "979-8-89176-495-8",
abstract = "AI text detectors are widely deployed in educational institutions to enforce academic integrity, yet they systematically misidentify authentic writing by non-native English speakers as AI-generated, with false positive rates reaching 61-72{\%} on English as a Second Language (ESL) student essays. This is a population-scale harm that falls disproportionately on the most linguistically vulnerable learners. Retraining detectors on learner data cannot close this gap: large language models can be prompted to mimic ESL writing patterns, and single-stream detectors are theoretically bounded to degrade toward random chance under such style-mimicry attacks. We present Meridian, a dual-stream architecture that maps each essay into a bivariate perplexity space with a Native-Expert (DistilGPT-2) axis measuring sub-word predictability and an ESL-Expert (character-level LSTM) axis capturing learner-specific morphological patterns. Meridian then classifies with an interpretable logistic boundary, deployable in under 2 seconds on CPU. Validated across three LLM families and two learner corpora (n=1,590), Meridian reduces wrongful ESL flags by 15.7x on held-out TOEFL11 with zero retraining, achieving 3.9{\%} overall false positive rate and 0.0{\%} for low-proficiency writers, while maintaining 98.3{\%} detection of AI-mimicked ESL text. These results demonstrate that equitable AI text detection for diverse, non-WEIRD populations is achievable through principled multi-stream feature design that encodes complementary linguistic signals rather than a single native-English reference norm."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="raichura-2026-meridian">
<titleInfo>
<title>Meridian: A Bivariate Perplexity Architecture for Trustworthy and Equitable AI Text Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pranil</namePart>
<namePart type="family">Raichura</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifth Workshop on NLP for Positive Impact (NLP4PI)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Katherine</namePart>
<namePart type="family">Atwell</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Angana</namePart>
<namePart type="family">Borah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daryna</namePart>
<namePart type="family">Dementieva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Elisa</namePart>
<namePart type="family">Kreiss</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Neema</namePart>
<namePart type="family">Kotonya</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiarui</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Liz</namePart>
<namePart type="family">Olson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruyuan</namePart>
<namePart type="family">Wan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jieyu</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-495-8</identifier>
</relatedItem>
<abstract>AI text detectors are widely deployed in educational institutions to enforce academic integrity, yet they systematically misidentify authentic writing by non-native English speakers as AI-generated, with false positive rates reaching 61-72% on English as a Second Language (ESL) student essays. This is a population-scale harm that falls disproportionately on the most linguistically vulnerable learners. Retraining detectors on learner data cannot close this gap: large language models can be prompted to mimic ESL writing patterns, and single-stream detectors are theoretically bounded to degrade toward random chance under such style-mimicry attacks. We present Meridian, a dual-stream architecture that maps each essay into a bivariate perplexity space with a Native-Expert (DistilGPT-2) axis measuring sub-word predictability and an ESL-Expert (character-level LSTM) axis capturing learner-specific morphological patterns. Meridian then classifies with an interpretable logistic boundary, deployable in under 2 seconds on CPU. Validated across three LLM families and two learner corpora (n=1,590), Meridian reduces wrongful ESL flags by 15.7x on held-out TOEFL11 with zero retraining, achieving 3.9% overall false positive rate and 0.0% for low-proficiency writers, while maintaining 98.3% detection of AI-mimicked ESL text. These results demonstrate that equitable AI text detection for diverse, non-WEIRD populations is achievable through principled multi-stream feature design that encodes complementary linguistic signals rather than a single native-English reference norm.</abstract>
<identifier type="citekey">raichura-2026-meridian</identifier>
<location>
<url>https://aclanthology.org/2026.nlp4pi-1.27/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>312</start>
<end>325</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Meridian: A Bivariate Perplexity Architecture for Trustworthy and Equitable AI Text Detection
%A Raichura, Pranil
%Y Atwell, Katherine
%Y Borah, Angana
%Y Dementieva, Daryna
%Y Kreiss, Elisa
%Y Kotonya, Neema
%Y Liu, Jiarui
%Y Olson, Liz
%Y Wan, Ruyuan
%Y Zhao, Jieyu
%S Proceedings of the Fifth Workshop on NLP for Positive Impact (NLP4PI)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-495-8
%F raichura-2026-meridian
%X AI text detectors are widely deployed in educational institutions to enforce academic integrity, yet they systematically misidentify authentic writing by non-native English speakers as AI-generated, with false positive rates reaching 61-72% on English as a Second Language (ESL) student essays. This is a population-scale harm that falls disproportionately on the most linguistically vulnerable learners. Retraining detectors on learner data cannot close this gap: large language models can be prompted to mimic ESL writing patterns, and single-stream detectors are theoretically bounded to degrade toward random chance under such style-mimicry attacks. We present Meridian, a dual-stream architecture that maps each essay into a bivariate perplexity space with a Native-Expert (DistilGPT-2) axis measuring sub-word predictability and an ESL-Expert (character-level LSTM) axis capturing learner-specific morphological patterns. Meridian then classifies with an interpretable logistic boundary, deployable in under 2 seconds on CPU. Validated across three LLM families and two learner corpora (n=1,590), Meridian reduces wrongful ESL flags by 15.7x on held-out TOEFL11 with zero retraining, achieving 3.9% overall false positive rate and 0.0% for low-proficiency writers, while maintaining 98.3% detection of AI-mimicked ESL text. These results demonstrate that equitable AI text detection for diverse, non-WEIRD populations is achievable through principled multi-stream feature design that encodes complementary linguistic signals rather than a single native-English reference norm.
%U https://aclanthology.org/2026.nlp4pi-1.27/
%P 312-325
Markdown (Informal)
[Meridian: A Bivariate Perplexity Architecture for Trustworthy and Equitable AI Text Detection](https://aclanthology.org/2026.nlp4pi-1.27/) (Raichura, NLP4PI 2026)
ACL