PREPRINT  ·  SEPTEMBER 24, 2026

Synthetic scientific thought

Authorial voice, AI-mediated translation, and the circulation of knowledge

Maria Christina da Silva Firmino-Cervera
Universidade Federal do Sul e Sudeste do Pará (UNIFESSPA) · Electronic Textual Cultures Lab, University of Victoria

Abstract

This essay proposes “synthetic scientific thought” as a heuristic category for investigating what may happen to the discursive inscription of knowledge when academic texts circulate through translation and other forms of mediation by artificial intelligence. The proposal begins from a distinction: preserving propositions and achieving fluency do not ensure the preservation of the gestures through which researchers delimit claims, inscribe voices, take positions, and affiliate themselves with disciplinary traditions. Grounded in Socio-Discursive Interactionism and studies of authorial voice and academic translation, the article defines the concept, describes the conditions of its application, and examines exploratory examples previously made available by the CF-C Program. It engages with recent research on stance, subjective distortion, terminological variation, and the convergence of scientific English, acknowledging findings that point both to risks and to possibilities for preservation and broader participation. The concept does not designate thought produced by a system, nor does it permit an inference of absent human responsibility from a text. It names a hypothesis about the circulation of utterances whose propositional form may remain recognizable while differences in positioning become less legible. The Omega Test is presented as a future, delimited investigation of this hypothesis, without anticipating its results or treating local observations as proof of a global historical transformation.

Keywords: synthetic scientific thought; informational gestures; authorial voice; academic translation; artificial intelligence; epistemic justice.

1. A question about what remains legible

A translated article may continue to report the aim of the research, the procedures employed, and the result reached. It may also come to state with greater certainty what the author had presented as a possibility, erase the attribution of an idea to another voice, or replace a disciplinary term with a generic expression. In either case, a quick reading may recognize “the same content.” A situated reading finds differences in how that content is proposed, delimited, and related to a research community. It is in this distance between propositional equivalence and enunciative trajectory that this essay situates its question.

The CF-C Program, developed to analyze informational gestures in AI-mediated academic translation, calls reductions or reconfigurations that make operations inscribed in the source text less discernible discursive-informational flattening. The exploratory study made available with six Applied Linguistics abstracts showed that a situated taxonomy can distinguish terminological, axiological, enunciative, and argumentative transformations, although that corpus and its individual coding do not support generalizations about academic translation as a whole (Firmino-Cervera, 2026). The new Omega Test protocol expands the textual unit to Introduction+Conclusion pairs and establishes a delimited comparison; its data collection and evaluation do not yet constitute results of this preprint.

My theoretical hypothesis is that, if mediations of this kind become recurrent across extended chains of production, translation, revision, and circulation, scientific texts may exhibit increasing formal regularity while some traces of their situated elaboration become less visible. I provisionally call this possibility synthetic scientific thought. The expression requires a qualification from the outset: the objects available for analysis are utterances, practices, and conditions of circulation; an author's thought is not directly observable in the textual product. Nor does this proposal provide a global measure of the historical prevalence of the phenomenon.

2. Interlocutions and point of departure

The central contribution of this essay is to propose a conceptual category for naming a specific form of opacity in AI-mediated scientific circulation: not the simple loss of content, but the reduced legibility of the gestures through which a study situates itself, sets its limits, engages with disciplinary traditions, and assumes enunciative responsibility. The argument thus differs from general diagnoses of machine translation quality or stylistic homogenization because it shifts attention to the relation between preserved propositional form and transformed enunciative trajectory.

In Socio-Discursive Interactionism, textual analysis articulates the general infrastructure of the text, textualization mechanisms, and enunciative mechanisms (Bronckart, 1999, 2008). For the purposes of CF-C, this distinction makes it possible to locate modifications at Level 1, in global organization; at Level 2, in connection, cohesion, and other textualization relations; and at Level 3, in voices, modalizations, and positions taken. Authorial voice is not reducible to the presence of “I” or “we”: it may reside in the caution of a hypothesis, the choice of a concept, the way a claim is attributed, or the decision to restrict its scope. Studies of metadiscourse show how interactional resources participate in the relationship among researcher, text, and reader (Hyland, 2005).

In this framework, the notion of informational gestures, proposed within the CF-C Program, enters into dialogue by analogy with the tradition of professional or didactic gestures studied by Bucheton and Soulé (2009) and is articulated here with Socio-Discursive Interactionism. This approximation does not equate teaching action with operations inscribed in the scientific text. In the CF-C Program, informational gestures designate linguistic-discursive operations that lend orientation, value, and legibility to meanings in scientific texts: modalizations, attributions of voice, terminological choices, delimitations of scope, argumentative sequences, and markers of theoretical affiliation.

Academic translation had already brought the stability of voice into question before large language models became widespread. Huang and Li (2023) identified modality shifts in Chinese scientific abstracts translated into English. Huang, Jia, and Li (2025), examining stance in Applied Linguistics abstracts, describe a negotiation between author and translator positions; they also report broad preservation of authorial epistemic stance in result-presenting moves. The literature therefore offers both a question and a counterpoint: transformation is possible, but it is neither inevitable nor uniform across resources, passages, and situations.

More recent research broadens the field without coinciding with the object of this essay. Shafiabadi and Yvon (2026) investigate changes in subjective positioning in the machine translation of texts of a different kind. Liu and Zhu (2026), in a preprint on the revision of second-language writing, report a reduction in dialogic engagement resources and variation between models in epistemic markers; they do not study the translation of the Brazilian articles proposed here. On another plane, Filimonovic and colleagues (2026), also in a preprint, estimate convergence of AI-assisted scientific writing toward U.S. scientific English, leaving open the question of what this means for inclusion. Ugwuanyi and colleagues (2026) discuss both opportunities for participation and risks to the diversity of varieties of academic English. None of these works, on its own, demonstrates the erasure of an individual authorship in scientific circulation.

The terminological problem adds a requirement. Dahan and colleagues (2026) show that glossaries can improve the accuracy and consistency of scientific translation while suppressing legitimate variation present in the source. Protecting an important term is not equivalent to imposing a single form in every context. An analysis of gestures must ask which difference serves a function in the text before classifying it as a loss or demanding its stabilization.

3. Delimiting “synthetic scientific thought”

I propose using synthetic scientific thought as a name for one possible configuration of scientific text circulation: a configuration in which algorithmic interventions preserve recognizable propositions and valued forms of academic communication but reduce, displace, or make less perceptible the informational, enunciative, and disciplinary gestures through which research positions are made distinct. By gestures, I mean situated textual operations — such as modalizing a claim, attributing it to another voice, preserving a technical term, delimiting scope, or marking theoretical affiliation — that guide the reader as to how a proposition should be understood. “Synthetic,” in this definition, qualifies the mediated configuration of the utterance and its circulation; it does not assert that a machine thinks, produces knowledge autonomously, or assumes the responsibility borne by the people and institutions involved.

The concept requires three distinctions. First, the absence of an explicit first-person marker is not, in itself, erasure: impersonal styles may be authorial decisions situated within a discipline. Second, a lexical or formal alteration is not automatically a loss: the function the feature served in context and the legitimate possibilities of the target language must be examined. Third, a modification observed in a text does not directly reveal how the author experienced writing, what she thought, or how readers actually interpreted the translated version. The analyzable gesture is a textual indication, not transparent access to consciousness or proof of reception.

These distinctions make it possible to formulate a condition of application: there are grounds to speak of flattening when a contextualized comparison between source and version shows a relevant operation in the original, a transformation in the version, and a recoverable justification for the loss of its discursive function. There are further grounds to investigate synthetic scientific thought when such transformations cease to be only local cases and enter an explicitly delimited hypothesis of recurrent circulation under documented mediation practices. Even so, moving from a case to a collective trend requires other corpora, observation over time, and studies of publication and reading practices.

Table 1. Proposed scales of inference for the concept. The final row describes a future research agenda, not a result of this essay.

4. Two illustrations and what they do not demonstrate

In the exploratory CF-C study, the expression “valores da colonialidade,” present in the abstract by Pinto, Dalben, and Audi (2024), appeared in one translation output as “colonial values” (Firmino-Cervera, 2026). The English version retains the lexical nucleus “colonial” but may shift the reference from the concept of coloniality to an evaluation of values associated with the colonial. Reading requires recovering the article's full argument: the relationship among expression, theoretical affiliation, and context makes it possible to assess whether the output empties the critical position of its force. The occurrence does not justify saying that every translation of “colonialidade” will have the same effect.

In another abstract, the expression “supervisoras de ES,” whose abbreviation referred to Estágio Supervisionado, received an English output that interpreted it as supervision of special education programs (Sól, 2024; Firmino-Cervera, 2026). This is a shift in institutional referent that can be documented by comparing the source, the expansion of the abbreviation, and the translation. The occurrence demonstrates the usefulness of checking conditions of authorship and disciplinary context; by itself, it does not prove a thesis about every configuration of authorial voice or about the range of available models.

The two cases belong to a preliminary investigation of six abstracts, with its own categories and conditions, already made available as a preprint (Firmino-Cervera, 2026). They serve here as illustrations of analytical legibility, not as sufficient evidence of prevalence. The first brings together a disciplinary term and an axiological position; the second shows how an apparent ambiguity can create a different object of research. Legitimate language uses may vary between fields, and a later revision of the text may translate these passages more appropriately. The CF-C Program also records questions of structural condensation in a separate investigation in the English–Portuguese direction; that material is not presented here as additional confirmation of the hypothesis of historical circulation.

5. From mediated text to scientific circulation: a hypothesis in stages

The broader proposition of this article depends on a chain of inference. The first stage is to observe, in a localized comparison, an identifiable textual function in the original. The second is to describe its trajectory in versions obtained under specified conditions and, when there is variation, to record its distribution. The third is to ask how such versions are selected, revised, and published. The fourth is to investigate whether similar alterations, accumulated across journals, fields, and languages, come to restrict the diversity of positions that the reader can recognize. CF-C's exploratory studies support the first stage and motivate the second; they do not establish the third or fourth.

The adjective “scientific” marks a particular feature of this circulation: knowledge claims depend on attribution, method, scope, interpretive dispute, and the possibility of contestation. A translation that converts a hypothesis into a certainty or erases the difference between a technical term and its everyday neighbor changes the conditions under which a claim can be examined by others. This does not automatically turn a true proposition into a false one. It potentially changes how someone can be answerable for it, situate it, and reproduce the steps that sustain it. The visibility of these relationships is a relevant property of scientific text.

The relationship with epistemic justice must remain precise. Fricker (2007) describes injustices that affect someone in their capacity as a subject of knowledge. A difference between two versions of a text is textual evidence of mediation; without data on practices, positions, and reception, it is not enough to demonstrate that a specific epistemic injustice occurred. The normative hypothesis is that recurrent loss of situated differences may contribute to unequal relations of recognition, especially if some researchers must pass through more mediations to participate in international circulation. The same mediation may reduce linguistic barriers and broaden participation. These consequences require investigation in their own right and do not exclude one another in advance.

6. The Omega Test as a delimited investigation

The current Omega Test protocol, dated September 2026, provides for ten Brazilian articles originally published in Portuguese between 2015 and 2020: five in Medicine and five in Applied Linguistics. Each article provides an Introduction+Conclusion pair. Five additional, independent pairs will be used to calibrate the researcher's single codebook, prompts, and glossaries; no calibration pair will enter the evaluation. The time window preceding the widespread use of generative models reduces one source of uncertainty about the preparation of the originals, without proving that they were written without any technological mediation.

The original text constitutes C1. Three PT-BR→EN translation conditions will be applied separately to each section: C2, Google Translate; C3, ChatGPT with a simple instruction; and C3.1, the same ChatGPT with a CF-C instruction and frozen glossaries. The priority C3–C3.1 contrast will examine the effect of the set of instructions within the same documented model. C2 provides a baseline from another system, so differences relative to C3 and C3.1 may also reflect differences between systems. The plan provides for five runs per section and condition, up to 300 translation outputs, while retaining the article as the unit of inference.

The analysis will preserve the three levels of the text, with interpretive priority given to authorial voice at Level 3. Each occurrence identified in the source will be aligned with the translated versions; changes introduced without a corresponding source gesture will be recorded separately. The protocol provides for recoverable examples and justifications, variation between runs, negative cases, and limits of inference. The CF-C Taxonomy v2.0 supplies fixed theoretical categories; calibration will adjust the operational rules of the codebook before the evaluation corpus is coded. The design allows trajectories to be described in these articles and conditions; on its own, it does not measure historical trends in scientific production.

Testable proposition 1. Some versions may retain relevant propositional content and alter enunciative functions identifiable in the source; the analysis will also need to record versions that preserve them.

Testable proposition 2. The C3–C3.1 contrast may show a reduction, maintenance, or introduction of alterations, depending on the article, gesture, and run; no uniform benefit from the CF-C instruction is presupposed.

Testable proposition 3. The distribution of alterations across the three levels may make Level 3 interpretively relevant without assigning to it terminological or architectural losses that belong to other layers.

If the relevant gestures are preserved under the conditions examined, if the divergences arise from justifiable linguistic choices, or if the CF-C instruction introduces new losses, these findings will restrict or require a reformulation of the initial hypothesis. An instrument that can record preservation and contrary cases is more useful for research than a vocabulary that recognizes only what it expected to find.

7. Contribution and research program

The contribution proposed by synthetic scientific thought is to bring together, under a provisional name, a question that crosses scales: how local differences in enunciation may matter for the diversity of what circulates as science. The literature on modality in abstracts, stance in translation, and the convergence of scientific English provides close interlocutors; CF-C adds an analytical pathway from gestures inscribed in Brazilian source texts to their trajectories in translations under specified conditions. This formulation does not claim terminological exclusivity or priority over every work in the field. Its value will depend on the clarity of its criteria and its capacity to accommodate results that run counter to the initial expectation.

The concept also suggests a practice of reading and revision. Authors, translators, and editors may ask which claims became more categorical, to whom an idea came to be attributed, which terms maintained their disciplinary affiliation, and which delimitations of scope remained visible. Recording the model, version, instruction, date, and interventions allows the mediation to be reconstructed more precisely. The adoption of glossaries must coexist with the possibility of functional variation, as demonstrated by the findings of Dahan and colleagues (2026). These are implications argued from the analytical framework; their effectiveness outside this project still requires study.

Later studies could compare versions of articles actually published over time, observe editorial decisions and reading practices in multilingual circulation, or test other language pairs and genres. Only then would it be possible to move from identifying altered gestures to estimates of frequency, social distribution, and possible collective effects. Open science increases the availability of versions, records, and instruments for this work, but availability in itself does not ensure diversity of voices or equitable recognition.

8. Conclusion

Synthetic scientific thought names, in this essay, a hypothesis about the legibility of the situated elaboration of knowledge in AI-mediated texts. The propositional form of a study may cross languages and systems while some of the enunciative choices that qualified it change function; it may also cross them with its voice preserved. The analytical task is to distinguish these pathways, document the conditions under which they occur, and prevent the fluency of the final version from prematurely closing the question.

The exploratory examples made available by CF-C make this question concrete. The Omega Test will give it a more delimited comparative examination, preserving the difference among source, translation conditions, coding, and interpretation. The historical and social reach of the concept remains an open agenda. Making the human trace legible means, here, creating conditions in which mediation can be described, discussed, and revised without attributing to the text more than can be concluded from it.

References

BRONCKART, Jean-Paul. Atividade de linguagem, textos e discursos: por um interacionismo sociodiscursivo. São Paulo: EDUC, 1999.

BRONCKART, Jean-Paul. O agir nos discursos: das concepções teóricas às concepções dos trabalhadores. Campinas: Mercado de Letras, 2008.

BUCHETON, Dominique; SOULÉ, Yves. Les gestes professionnels et le jeu des postures de l’enseignant dans la classe: un multi-agenda de préoccupations enchâssées. Éducation et didactique, v. 3, n. 3, p. 29–48, 2009. https://doi.org/10.4000/educationdidactique.543.

DAHAN, Nicolas; PENG, Ziqian; YVON, François; BAWDEN, Rachel. Improving Term Evaluation in Machine Translation: Variation Matters. In: Proceedings of the 17th Conference of the Association for Machine Translation in the Americas, 2026. p. 101–134. https://aclanthology.org/2026.amta-research.7/.

FIRMINO-CERVERA, Maria Christina da Silva. Discursive-Informational Flattening in Scientific Abstracts: A Pilot Study with the Taxonomy of Informational Gestures (CF-C). Zenodo, 2026. https://doi.org/10.5281/zenodo.22019234. Preprint, English and Portuguese versions.

FILIMONOVIC, Dragan; RUTZER, Christian; MACHER, Jeffrey; WEDER, Rolf. Does Scientific Writing Converge to U.S. English? Evidence from Generative AI-Assisted Publications. arXiv, arXiv:2511.11687v2, 2026. https://arxiv.org/abs/2511.11687. Preprint.

FRICKER, Miranda. Epistemic Injustice: Power and the Ethics of Knowing. Oxford: Oxford University Press, 2007. https://doi.org/10.1093/acprof:oso/9780198237907.001.0001.

HUANG, Yueyue; JIA, Hui; LI, Dechao. Transvocal stance in academic translation: A rhetorical analysis of grammatical stance in translated applied linguistics English research article abstracts. Journal of English for Academic Purposes, v. 74, 101472, 2025. https://doi.org/10.1016/j.jeap.2025.101472.

HUANG, Yueyue; LI, Dechao. Translatorial voice through modal stance: A corpus-based study of modality shifts in Chinese-to-English translation of research article abstracts. Lingua, v. 295, 103610, 2023. https://doi.org/10.1016/j.lingua.2023.103610.

HYLAND, Ken. Metadiscourse: Exploring Interaction in Writing. London: Continuum, 2005.

LIU, Ao; ZHU, Shanhua. The Cost of Perfect English: Pragmatic Flattening and the Erasure of Authorial Voice in L2 Writing Supported by GenAI. arXiv:2605.13055, 2026. https://arxiv.org/abs/2605.13055. Preprint.

PINTO, Lori de Santana; DALBEN, Tatiany Pertel Sabaini; AUDI, Luciana Cristina da Costa. A difícil integração do inglês como língua franca e o livro didático: análise de atividades de compreensão oral em livro didático de língua inglesa. Trabalhos em Linguística Aplicada, v. 63, n. 3, p. 564–578, 2024. https://doi.org/10.1590/01031813v63320248674693.

SÓL, Vanderlice dos Santos Andrade. Tríade rizomática de formação docente de Língua Inglesa: uma proposta de problematização da relação universidade-escola. Trabalhos em Linguística Aplicada, v. 63, n. 1, p. 225–239, 2024. https://periodicos.sbu.unicamp.br/ojs/index.php/tla/article/view/8669111.

SHAFIABADI, Nazanin; YVON, François. Biases in Translation: Assessing Opinion Distortion in Machine Translated Texts. In: Proceedings of the Fifteenth Language Resources and Evaluation Conference, 2026. p. 8596–8614. https://doi.org/10.63317/2pjio9ho8rxg.

UGWUANYI, Kingsley; MAIR, Christian; DOVCHIN, Sender; ERDOCIA, Iker; KUTEEVA, Maria; AIREMIONKHALE, Esther. Generative AI and linguistic diversity in academic writing and publishing: Perspectives from World Englishes. Journal of English for Research Publication Purposes, published online in 2026. https://doi.org/10.1075/jerpp.00035.ugw.

Declaration of AI use

Artificial intelligence tools were used to support preliminary drafting and bibliographic searching. The conceptual formulations and interpretations are the author's, who assumes responsibility for the content of the manuscript.

• Maria Christina da Silva Firmino Cervera
•  

• Synthetic scientific thought: Authorial voice, AI-mediated translation, and the circulation of knowledge v.2.0 - published 23 Sep 2026 in Preprint