Source-linked AI summary
Who Acts, Who Knows, Who Answers? A Corpus-Assisted Discourse Analysis of Agency, Epistemic Responsibility, and Accountability in Generative AI Higher Education Research
Biranchi Poudyal
TL;DR
This study asks how GenAI scholarship distributes action, knowledge, and answerability, a question shaped by the labels used for AI in higher education. It analyzes titles and abstracts through corpus-assisted discourse methods. AI is commonly represented as acting, while learners more often know, judge, and verify, and human or institutional accountability remains distinct from AI’s functional agency.
Problem
GenAI labels in higher education shape who is represented as acting, knowing, and answerable, yet prior reviews rarely analyze the grammar of the literature as evidence.
Source: §1 · pp. 1–2Method
The study applies corpus-assisted discourse analysis to titles and abstracts, examining how GenAI scholarship formulates action, epistemic work, and responsibility.
Source: §1 · pp. 2–3, §3.4 · p. 7Results
AI is represented with substantial functional agency, while students more often know, judge, and verify, and educators and institutions bear most obligations amid passive or nominal language.
Source: §8 · pp. 14–15Takeaways & Limitations
Keeping performance, authority, and accountability distinct helps clarify who acted, warranted a knowledge claim, could dispute a decision, and answers for consequences.
Source: §8 · p. 15Because the corpus covers titles and abstracts, its claims concern public-facing argumentation rather than entire articles.
Source: §7 · p. 14
Abstract
from arXiv · showhide
Generative artificial intelligence (GenAI) is increasingly described in higher education as a tool, collaborator, evaluator, proxy, and infrastructure. These labels are not neutral: they shape who is seen as acting, knowing, and, crucially, being answerable when someone uses AI. This study examines how scholars distribute such responsibility across 366 English-language titles and abstracts published between November 2022 and 14 August 2026. A corpus-assisted discourse analysis combined frequency counts, five-word collocations, actor-predicate associations, obligation-clause coding, and heuristic analysis of passives, nominalisations, and reporting metonymies. The final corpus contained 91,405 tokens, drawn from 8,734 unique records screened after deduplication. AI was the most frequently named actor, appearing 2,050 times and entering 447 predicate associations, most involving action. Students, by contrast, were more often associated with knowing, judging, and verifying, yet no sentence ever made AI explicitly responsible. Responsibility fell largely on educators, institutions, and policy, or disappeared through passive and nominalised constructions. Even frequent phrases such as "responsible AI" and "responsible use" rarely identify who is accountable for what. The findings reveal a persistent gap between giving AI functional agency and assigning normative accountability.
1. Introduction
GenAI scholarship has diversified beyond questions of use, but the language used to describe agency and responsibility remains underexamined. This study examines how titles and abstracts formulate action, epistemic work, and accountability.
1. Introduction
- Different labels for GenAI imply different distributions of action, judgment, knowledge, and fault, so the analysis asks how those formulations relate to explicit accountability.The study distinguishes tool, interlocutor, proxy, and infrastructural representations and asks whether recurring formulations preserve the distinction between operant performance and accountable judgment.Source: §1 · pp. 1–3
- As research has broadened from integrity and appropriate use to learner capabilities, policy, and institutional governance, questions have shifted toward what human capability remains observable when GenAI is used.Related work spans learner agency, perceived agency, writing, cognitive offloading, institutional responses, and pedagogical proposals.Source: §1 · pp. 1–2
- The study treats recurring language about GenAI agency and responsibility as evidence, a dimension few prior reviews have analyzed.It focuses on titles and abstracts, among the most consistently visible representations of an author’s argument.Source: §1 · pp. 2–3
2. Agency, Epistemic Work, and Accountability as Discourse
The paper distinguishes grammatical and functional agency from epistemic participation and answerability. It uses discourse frameworks to examine how language represents actors, processes, obligations, and GenAI’s operational roles.
2.1 From Grammatical Activation to Answerability
- Agency can be grammatical, function-based, or epistemic; participation in knowledge work does not confer equal authority or accountability.Epistemic responsibility entails answerability, including a bearer, decision or object, anticipated action, standard, audience, and possible consequence or redress.Source: §2.1 · pp. 3–4
2.2 Transitivity, Modality, and Social Actors
- Transitivity and social-actor analysis track how clauses activate, passivate, include, exclude, or impersonalize participants, while modality expresses obligations.Corpus-assisted discourse analysis identifies recurring patterns, while concordance reading examines their meanings in context; frequency alone does not establish motivation.Source: §2.2 · p. 4
2.3 Tool, Interlocutor, Proxy, and Infrastructure
- Tool, interlocutor, proxy, and infrastructure are descriptive framings for comparing how GenAI descriptions attach action, responsibility, and obligation.These framings do not prescribe accountability: interlocutor language can mask delegated judgment, while tool language can support it; infrastructure framing can clarify or obscure responsibility.Source: §2.3 · pp. 4–5
3. Method
The study analyzes public-facing titles and abstracts in peer-reviewed English-language GenAI higher education research. It screens records, constructs a corpus, and applies corpus-assisted discourse methods with stated safeguards and limitations.
3.1 Design and Scope
- The study uses corpus-assisted discourse analysis of public documentary materials, with titles and abstracts as its common textual unit.Eligibility covered English-language journal articles, reviews, and peer-reviewed conference papers focused on GenAI in higher education from 1 November 2022 to 14 August 2026.Source: §3.1 · p. 5
3.2 Search and Duplicates
- Searches across Scopus, Web of Science, IEEE Xplore, and SpringerLink yielded 11,175 exported rows; deduplication left 8,734 unique records.Springer results were a non-random, non-exhaustive sample because exports used relevance-ranked results for 2024–2026.Source: §3.2 · pp. 5–6
3.3 Screening and Corpus Freeze
- Screening produced a final corpus of 366 documents: 8 from 2023, 35 from 2024, 97 from 2025, and 226 from 2026.The screening path excluded 8,368 of the 8,734 unique records; 48 Springer-only titles lacked a corresponding abstract.Source: §3.3 · p. 6
3.4 Corpus Preparation and Analysis
- Actor–predicate associations required an actor to precede a verbal predicate within six intervening words in the same sentence; collocations used symmetric five-word windows.The study also coded obligation expressions by bearer and used heuristic counts for passives, nominalisations, and reporting metonymies; these grammar counts were diagnostic indicators, not perfect annotations.Source: §3.4 · pp. 6–7
3.5 Rigor, Reflexivity, and Research Integrity
- The researchers used safeguards against treating AI-active language as moral agency or automatic displacement, but follow-up verification was not always possible.Productive collaboration and automated verification were retained as negative cases, and counts were manually rechecked after the corpus freeze.Source: §3.5 · p. 7
4. Results
Across the corpus, AI is most often represented as acting, while learners more often occupy epistemic roles and institutions or educators bear obligations. Passive and nominalised language can leave the bearer of responsibility unclear.
4.1 Corpus Composition and Lexical Priorities
- “Responsible AI/GenAI” appeared 83 times and “responsible use” 66 times, but the corpus rarely named an accountable relationship.“AI responsibility” and “student responsibility” never appeared as exact phrases, and only one sentence linked AI and accountability, stating AI is not accountable.Source: §4.1 · pp. 7–8
4.2 Who Acts, Knows, and Verifies?
- AI was the most frequent actor label (2,050) and had 447 strict actor–predicate associations, while students received predicates across knowledge, judgment, and verification more often than AI.Action accounted for 78.7% of AI references; knowledge, judgment, and verification together appeared in 29 AI associations versus 84 student associations.Source: §4.2 · p. 8
4.3 Collocation and Operational Representations
- AI collocations combined instrumental and normative vocabularies, while student collocations more often concerned agency, learning, thinking, engagement, and accountable identity.AI-action predicates were highest in tool records, at 119.2 per hundred records; verification-judgement combinations were highest in proxy discourse, at 26.1 per hundred records.Source: §4.3 · pp. 8–9
4.4 Who Answers?
- Of 166 obligation or answerability expressions, 37 named institutions or policy actors, 20 educators, 6 students, and 33 left the bearer unspecified or used impersonal clauses.Twelve AI/system clauses described how systems ought to be used, framed, or restricted, but none assigned them moral or institutional consequences.Source: §4.4 · pp. 9–10
4.5 Passives, Nominalisation, and Metonymic Disappearance
- Heuristic counts found 800 passive constructions, 780 without an agent reference, 1,152 responsibility-related nominalisations, and 241 reporting metonyms.Nominalisations can name norms without specifying who acts; institutional names appeared, but their actual decision-making roles often remained unclear.Source: §4.5 · pp. 10–11
5. Discussion
The discussion distinguishes GenAI’s functional agency from epistemic authority and normative accountability, while finding responsibility language often leaves the accountable relationships incomplete. These patterns vary across representations, motivating a four-part distinction for analyzing agency and accountability.
5.1 Functional Agency Without Normative Equivalence
- AI performs consequential functions but is not treated as an epistemically authoritative or accountable agent; no “strictly responsible” predicate is assigned to it.The authors caution against treating grammatical activation as personhood while noting that human accountability can exceed available control.Source: §5.1 · p. 11
5.2 Responsibility Recognized but Incompletely Relational
- Responsibility terms can name multiple possible duties, while obligations often appear impersonal or through policy, curriculum, and assessment metonyms.Providers are the least visible actor group, narrowing ecological responsibility toward an abstract system and an individual user.Source: §5.2 · p. 12
5.3 Variation by Representation
- Tool discourse has the most AI-action predicates, proxy discourse emphasizes student verification, and infrastructure discourse uses the largest proportion of nominalizations.Interlocutor discourse has the fewest AI-action predicates, though dialogue can facilitate reflective judgment.Source: §5.3 · p. 12
- Responsibility cannot be inferred from representation labels alone; analysis should examine operations, standards, contestability, and who bears consequences.A partner may remain constrained by learner judgment, while a tool may carry out much of the epistemic work.Source: §5.3 · p. 12
5.4 Contributions to Research on Epistemic Agency
- The paper distinguishes grammatical activation, functional agency, epistemic authority, and normative accountability as related but non-equivalent concepts.The literature has ethical vocabulary but lacks a clearly articulated conventional language for identifying who assumes practical accountability.Source: §5.4 · pp. 12–13
6. Implications for Research, Pedagogy, and Governance
The implications call for making responsibility relationships auditable, teaching verification as a skill, and assigning accountability in proportion to control. Researchers are also urged to describe AI operations and human decisions clearly.
6. Implications for Research, Pedagogy, and Governance
- Research claims should identify the responsible actor, decision or knowledge object, required action, standard, audience for justification, and consequence or redress.The suggested formulation replaces “verify AI outputs” with specified student checks, educator evidence criteria, and institutional dispute procedures.Source: §6 · p. 13
- Verification should be taught as a skill involving source assessment, domain knowledge, uncertainty, and judgment, with assessment drawing on revision evidence and student defense.Disclosure alone is not presented as sufficient evidence of students’ reasoning.Source: §6 · p. 13
- Accountability should track control: providers, institutions, educators, and students have distinct responsibilities within their respective authority and capacity.The passage also says detection scores should prompt investigation rather than automatic acceptability decisions, and urges active, concise descriptions of AI’s actions and limits.Source: §6 · pp. 13–14
7. Limitations and Future Research
The study is limited by its title-and-abstract corpus, incomplete coverage, and approximations in dates and coding. The authors recommend full-text and coding-validation studies before firm confidence in the conclusions.
7. Limitations and Future Research
- The corpus uses titles and abstracts, which carry less information about methods and relationships within articles, and is probably not exhaustive.Available database filters were used, but Springer records without abstracts were excluded and IEEE coverage was limited to journal articles.Source: §7 · p. 14
- The corpus skews toward a partially complete, rapidly evolving 2026, while date fields and rule-based actor-predicate coding introduce approximations.The coding also misses longer-range dependencies such as pronoun reference or negation, and representation and publication-type coding add abstraction.Source: §7 · p. 14
- The authors call for full-text corpus work, validation of actor-predicate rules through parsing and double coding, and checks of abstract-to-article language correspondence.These studies are proposed as steps that should precede firm confidence in the conclusions.Source: §7 · p. 14
8. Conclusion
The conclusion finds that GenAI has functional agency without normative equivalence, while students more often perform epistemic work and educators and institutions bear obligations. It argues that scholarship should distinguish action, authority, and accountability.
8. Conclusion
- GenAI carries functional agency, while students are more associated with knowing, judging, and verifying, and educators and institutions bear most obligations.Providers remain peripheral, and passive or nominal language often obscures responsibility relationships.Source: §8 · pp. 14–15
- Scholarship should distinguish performing an action, holding authority over it, and being accountable for it.The authors argue that these distinctions help clarify who acted, warranted a knowledge claim, could dispute a decision, and answers for consequences.Source: §8 · p. 15