The rapid integration of generative artificial intelligence (GenAI) into applied linguistics has prompted growing interest in using GenAI outputs as research data. Although critical scholarship has examined issues such as bias, privacy, and invisible labor, little attention has been paid to whether GenAI-generated data are compatible with the epistemological commitments of anti-racist and decolonial research. This conceptual paper addresses the question: Can anti-racist linguistic research use GenAI as data? I argue that GenAI should be understood as a raciolinguistic infrastructure embedded in histories of colonialism, racial capitalism, and data colonialism rather than as a neutral producer of language. Drawing on anti-racist, decolonial, and Southern approaches to applied linguistics, I identify methodological risks associated with treating GenAI outputs as primary data: re-centering Whiteness and Global North norms, erasing communities by treating models as stand-ins for speakers, decontextualizing racism and ableism, and reinforcing ethically opaque and extractive infrastructures. At the same time, I outline conditions under which AI can be studied critically as evidence of institutionalized language ideologies and platform governance. The paper concludes by proposing a framework for linguistically just AI research that positions (Gen)AI as an object of critique while maintaining community-grounded knowledge as the epistemic center of interpretation. A r & aacute;pida integra & ccedil;& atilde;o da intelig & ecirc;ncia artificial generativa (GenAI) & agrave; lingu & iacute;stica aplicada tem despertado um interesse crescente no uso de conte & uacute;dos gerados por GenAI como dados de pesquisa. Embora a literatura cr & iacute;tica tenha examinado quest & otilde;es como vieses, privacidade e trabalho invisibilizado, pouca aten & ccedil;& atilde;o tem sido dedicada & agrave; quest & atilde;o de saber se os dados gerados por GenAI s & atilde;o compat & iacute;veis com os compromissos epistemol & oacute;gicos da pesquisa antirracista e decolonial. Este artigo conceitual aborda a seguinte quest & atilde;o: A pesquisa lingu & iacute;stica antirracista pode utilizar GenAI como dados? A pesquisa lingu & iacute;stica antirracista pode utilizar GenAI como dados? Argumento que a GenAI deve ser compreendida como uma infraestrutura raciolingu & iacute;stica inserida em hist & oacute;rias de colonialismo, capitalismo racial e colonialismo de dados, e n & atilde;o como uma produtora neutra de linguagem. Com base em abordagens antirracistas, decoloniais e do Sul para a lingu & iacute;stica aplicada, identifico riscos metodol & oacute;gicos associados ao tratamento de conte & uacute;dos gerados por GenAI como dados prim & aacute;rios: o recentramento da branquitude e de normas do Norte Global; o apagamento de comunidades ao tratar modelos como substitutos de falantes; a descontextualiza & ccedil;& atilde;o do racismo e do capacitismo; e o refor & ccedil;o de infraestruturas extrativistas e eticamente opacas. Ao mesmo tempo, delineio as condi & ccedil;& otilde;es sob as quais a IA pode ser estudada criticamente como evid & ecirc;ncia de ideologias lingu & iacute;sticas institucionalizadas e da governan & ccedil;a de plataformas. O artigo conclui propondo um quadro de refer & ecirc;ncia para uma pesquisa em IA linguisticamente justa, que posiciona a (Gen)AI como objeto de cr & iacute;tica, ao mesmo tempo que mant & eacute;m o conhecimento ancorado nas comunidades como centro epist & ecirc;mico da interpreta & ccedil;& atilde;o.
This study compared human-adapted and ChatGPT-simplified texts across three prompting strategies (general, example-based, and instructive) using 51 authentic English-language newspaper articles. Lexical complexity (density, diversity, and sophistication) and syntactic complexity (five length- and subordination-based indices) were analyzed using established computational tools and linear mixed-effects models. Although all simplified versions reduced syntactic complexity relative to the source texts, ChatGPT-simplified texts exhibited significantly higher lexical density and more aggressive syntactic reduction than human-adapted texts across all measures. Differences among the three prompting strategies were modest, though the instructive prompt consistently produced the syntactically simplest outputs. These findings suggest that ChatGPT and human simplification follow distinct linguistic profiles and that human oversight remains important when using AI-generated texts for instruction. (sic) (sic)(sic)(sic) 51(sic)(sic) (sic)(sic) (sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic), (sic)(sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic) ChatGPT(sic) (sic) (sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)((sic)(sic), (sic)(sic) (sic)(sic), (sic)(sic)(sic))(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic). (sic)(sic) (sic)(sic)(sic)((sic)(sic), (sic)(sic)(sic), (sic)(sic)(sic))(sic) (sic)(sic) (sic)(sic)(sic)((sic)(sic)& centerdot;T-unit & centerdot;(sic)(sic) (sic)(sic) (sic)(sic), T-unit(sic) (sic)(sic)(sic) (sic))(sic) (sic)(sic) (sic)(sic) (sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic)(sic)(sic)(sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic). (sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic) (sic)(sic) (sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic), ChatGPT (sic)(sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic) (sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic) (sic)(sic)(sic)(sic) (sic) (sic)(sic)(sic) (sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic). (sic)(sic)(sic)(sic) (sic)(sic) (sic) (sic)(sic)(sic) (sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic) (sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic)(sic)(sic). (sic)(sic)(sic) (sic)(sic)(sic) ChatGPT(sic) (sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic) (sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic), AI (sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic) (sic) (sic)(sic)(sic) (sic)(sic)(sic) (sic)(sic)(sic)(sic) (sic)(sic)(sic)(sic).
This study examines how finite clausal complexity relates to L2 writing performance by distinguishing clausal density (CD), the accumulation of finite subordinate clauses, from embedding depth (ED), their hierarchical nesting. Using 566 timed CET compositions from the Chinese Learner English Corpus, sentence-level CD and ED were manually annotated and analyzed across six CET-based writing-performance bands. Results show that both measures were weakly but significantly associated with variation across CET-based writing-performance bands and in different ways. CD displayed broader but modest variation across bands, whereas ED reflected a more constrained dimension of hierarchical organization. In joint models, CD retained a modest independent association, while ED overlapped empirically with CD and did not contribute independent information. Length-adjusted analyses further showed that the CD pattern was not reducible to composition length, although overall explanatory power remained limited. The findings indicate that finite clausal complexity should be treated as multidimensional, with clause accumulation and hierarchical embedding capturing related but non-identical aspects of finite clausal organization. However, neither CD nor ED should be interpreted as a strong diagnostic measure of L2 writing performance.
This study examines how epistemic commitment is temporally and syntactically staged across journal tiers in life-science research writing. Drawing on a 2.3-million-word corpus of 320 publications in biochemistry, genetics, and molecular biology, it compares higher-tier journals with a SCImago Journal Rank (SJR) value of 2.000 or above (approximately the top 14% of the subject category) with lower-tier journals below that threshold. It focuses on tense-marked epistemic stance verbs, selected hedging and boosting resources, and recurrent local framing sequences, where staging refers the temporal and syntactic packaging of epistemic commitment within a claim's local discourse environment. The lower-tier subcorpus showed a consistent reliance on past-oriented reporting frames and pointer-style environments, whereas the higher-tier subcorpus favored clause-integrated framing and a broader distribution of boosting resources, with present-oriented stance forms particularly characteristic of its review articles; several contrasts show moderate-to-strong log-ratio effects. The main contrast lies not in stronger versus weaker commitment alone, but in how commitment is packaged, integrated, and projected across publication strata. Given article-type asymmetry and whole-article rather than section-controlled comparison, the findings are interpreted as tier-associated tendencies rather than genre-independent effects. The study contributes to research on epistemic stance and supports corpus-informed ERPP pedagogy directed toward venue-sensitive claim framing. Denna studie unders & ouml;ker hur epistemiskt & aring;tagande tidsligt och syntaktiskt iscens & auml;tts i forskningsartiklar inom livsvetenskaperna p & aring; olika tidskriftsniv & aring;er, och diskuterar implikationer f & ouml;r st & ouml;d vid engelskspr & aring;kig publicering. Med utg & aring;ngspunkt i forskning om h & aring;llning (stance), metadiskurs, korpusbaserad lexikogrammatik och akademiska litteraciteter behandlar studien en lucka r & ouml;rande hur publiceringsniv & aring;er korrelerar med den spr & aring;kliga realiseringen av epistemiskt & aring;tagande inom ett n & auml;rliggande disciplin & auml;rt f & auml;lt. Med st & ouml;d i en korpus p & aring; 2,3 miljoner ord best & aring;ende av 320 publikationer i biokemi, genetik och molekyl & auml;rbiologi j & auml;mf & ouml;r analysen tidskrifter p & aring; h & ouml;gre niv & aring;, med ett SCImago Journal Rank (SJR)-v & auml;rde p & aring; 2000 eller h & ouml;gre-ungef & auml;r de & ouml;versta 14 % inom & auml;mneskategorin-med tidskrifter p & aring; l & auml;gre niv & aring; under denna tr & ouml;skel inom samma kategori. Studiens fokus ligger p & aring; tempusmarkerade epistemiska h & aring;llningsverb, utvalda hedgande och f & ouml;rst & auml;rkande resurser samt & aring;terkommande lokala inramningssekvenser kopplade till claim-integrering, d & auml;r iscens & auml;ttning f & ouml;rst & aring;s som den tidsliga och syntaktiska paketeringen av epistemiskt & aring;tagande i claimens lokala diskursmilj & ouml;. Delkorpusen fr & aring;n tidskrifter p & aring; l & auml;gre niv & aring; uppvisade en konsekvent tillit till d & aring;tidsorienterade rapporteringsramar och pekaraktiga (pointer-style) milj & ouml;er, medan delkorpusen fr & aring;n tidskrifter p & aring; h & ouml;gre niv & aring; gynnade satsintegrerad inramning och en bredare f & ouml;rdelning av f & ouml;rst & auml;rkande resurser, d & auml;r nutidsorienterade h & aring;llningsformer var s & auml;rskilt utm & auml;rkande f & ouml;r dess & ouml;versiktsartiklar; flera av dessa kontraster uppvisar m & aring;ttliga till starka log-ratio-effekter. Resultaten tyder p & aring; att den huvudsakliga kontrasten inte enbart handlar om starkare respektive svagare & aring;tagande, utan om hur & aring;tagandet paketeras, integreras och projiceras & ouml;ver olika publiceringsniv & aring;er. Med h & auml;nsyn till asymmetri mellan artikeltyper och j & auml;mf & ouml;relser p & aring; helartikelniv & aring; snarare & auml;n kontrollerade sektionsj & auml;mf & ouml;relser tolkas resultaten som niv & aring;relaterade tendenser snarare & auml;n genreneutrala effekter. Studien bidrar till forskningen om epistemisk h & aring;llning och st & ouml;djer korpusinformerad ERPP-pedagogik inriktad p & aring; venuesensitiv inramning av claims.
ABSTRACT Corrective recasts provide a target form, but successful overt repair may depend on processing resources. We tested whether concurrent verbal‐articulatory load reduces feedback‐triggered morphosyntactic repair in 120 B1–B2 learners of Modern Standard Arabic. A 2 × 2 between‐participants experiment crossed recasts with articulatory suppression across 1440 eligible initial‐error trials spanning case marking, agreement and derived verb patterns. Observed repair after recasts was 75% without suppression and 35% with suppression, versus 10% and 5% without recasts. A multilevel Bernoulli‐logit model of non‐repair showed a Feedback × Suppression interaction ( β = 1.10, 95% CrI [0.45, 1.73]). Supplementary right‐censored lognormal modelling also indicated attenuation of the recast benefit (interaction time ratio = 1.61 [1.25, 2.07]), although 68.75% of trials lacked observed repair and the event‐time interpretation is assumption‐dependent. Uncued post‐tests and manipulation checks provided secondary evidence. Recasts facilitated overt correction, but their advantage diminished under concurrent load. The findings support resource‐sensitive interference, not an isolated phonological‐loop mechanism; a matched nonverbal control is required for process‐specific attribution.
