AI evidence extraction

Marcas de inteligência artificial na escrita: vícios textuais, técnicas de autoria situada e protocolo de prompt para reconhecimento e revisão

Albert Bacelar · Educação & Inovação · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

3/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.64326/educao.v2i12.467

Methodology & findings

Study design

Qualitative corpus analysis of technical-documentary textual generation.

Primary method

None. This paper employs qualitative taxonomy development and hermeneutic analysis of textual patterns. No statistical hypothesis testing, null significance testing, or quantitative analysis is conducted.

Main result

The study found that "a escrita com aparência de IA nasce da combinação entre superfície polida, baixa materialidade, simetria sintática, falta de conflito, vocabulário de consenso e consequência fraca" (AI-generated writing emerges from the combination of polished surface, low materiality, syntactic symmetry, lack of conflict, consensus vocabulary and weak consequence). The authors propose that "o reconhecimento mais consistente exige leitura em camadas, com detector automático como instrumento auxiliar e insuficiente" (consistent recognition requires layered reading, with automatic detection as an auxiliary and insufficient instrument).

Reports effect sizes.

Research paradigm

Hermeneutic/interpretative; qualitative textual analysis

Author conclusions

The authors conclude that "o corpus analisado oferece contribuição prática: transforma impressões difusas sobre texto sintético em critérios observáveis. De um lado, descreve vícios: generalidade, fórmula, excesso, neutralidade, passividade, abstração, polidez vazia e fechamento previsível. De outro, organiza técnicas de recomposição: tese, frase matriz, cena, prova, causalidade, corte, ritmo, especificidade profissional e responsabilidade epistêmica" (the analyzed corpus offers practical contribution: transforms diffuse impressions about synthetic text into observable criteria. On one hand, it describes vices: generality, formula, excess, neutrality, passivity, abstraction, empty politeness and predictable closure. On the other, it organizes recomposition techniques: thesis, matrix phrase, scene, proof, causality, cutting, rhythm, professional specificity and epistemic responsibility). They further state that "a proposta final é escrever com autoria situada: decisão, cicatriz, método e consequência" (the final proposal is to write with situated authorship: decision, scar, method and consequence).

Risk of bias

No empirical validation: the taxonomy is developed through author interpretation without systematic coding or inter-rater reliability testing; Confirmation bias risk: the corpus selection and vice identification may reflect the authors' subjective expectations about AI writing patterns; Limited linguistic scope: framework is presented primarily in Portuguese context; generalizability to other languages and writing traditions unclear; No control comparison: no parallel analysis of human-written texts using the same framework to establish discriminative validity; Selection bias in corpus construction: no explicit description of how texts were selected or stratified; Subjective interpretation bias: the taxonomy relies on qualitative judgment rather than quantitative validation; Linguistic bias: the taxonomy may reflect Portuguese-language patterns and may not generalize to other languages; No inter-rater reliability testing reported; No comparison against established automatic detection tools with quantified performance metrics; Confirmation bias potential: authors developed taxonomy from texts they selected as AI-generated

Limitations

The paper explicitly states limitations regarding automatic detection tools: "Deep et al
(2025) descrevem falsos positivos, opacidade e viés linguístico em ferramentas de detecção" (false positives, opacity and linguistic bias in detection tools)
Additionally, the authors note that "tais marcas não constituem prova isolada de autoria automatizada, pois escritores humanos também podem reproduzir esses padrões" (such marks do not constitute isolated proof of automated authorship, as human writers can also reproduce these patterns)
The study lacks empirical testing of the proposed protocol on a validation dataset.

Open questions raised

The authors identify the need for domain-specific reinforcement in Medicine and Law due to text-mediated risks involving clinical care, professional confidentiality, institutional responsibility and decision-making. They note the insufficiency of automatic detection as a sole response and emphasize the need for layered human reading in conjunction with automated tools.
The paper identifies that recent literature (Deep et al. 2025, Kar et al. 2025, Fiedler & Döpke 2025) confirms "a fragilidade da detecção automática como resposta isolada" (the fragility of automatic detection as an isolated response) and that "a evidência converge para uma leitura por constelação: nenhum sinal isolado decide; o conjunto, o gênero e o risco orientam o julgamento" (evidence converges toward constellation reading: no isolated signal decides; the set, genre, and risk guide judgment). The paper proposes filling this gap with a layered recognition protocol combining automatic detection with human judgment across multiple textual dimensions.
The paper identifies gaps in automatic AI detection literature and proposes that future work should focus on: (1) implementation of the layered reading protocol in educational and professional contexts; (2) validation of the taxonomy across different text genres and languages; (3) testing the proposed prompt protocol in iterative writing cycles; (4) domain-specific application in Medicine and Law with empirical outcome measurement; (5) integration of automatic detection tools as auxiliary rather than primary mechanisms.

Data: The study states: "O estudo utilizou corpus documental textual analisado qualitativamente. A disponibilização do material deve respeitar a política editorial, os direitos autorais e as condições de anonimização aplicáveis" (The study used documentary textual corpus analyzed qualitatively. Material disclosure must respect editorial policy, copyright and applicable anonymization conditions). No specific dataset repository or URL provided.Extracted from: pdfAgreement 61%

Explore related topics

AI Manuscript Writing Academic Writing Content Analysis