Oryx AI — Auditable Research Report

Question: What does published evidence say about AI-generated citations in academic research, 2023-2026?

Status: APPROVED — Riya Nalwa approved this revised draft on 1 Nov 2026 at 14:35 IST (Gmail id 1a1965bb98a679e4); Amartya Kumar confirmed proceed at 14:37 IST. Human review DONE (Riya). Exported and published 1 Nov 2026. This is a local MVP test sample, NOT a client deliverable and NOT a pilot sale.

Retrieval note: Semantic Scholar: rate-limited (shared-IP budget).

Screening: 10 candidates screened, 4 included, 6 excluded — see 03-screening-log.md.

Framing

Sub-questions

1. What is the prevalence and nature of AI-generated hallucinations in academic citations?
   • Evidence required: Peer-reviewed studies, technical reports, or systematic analyses documenting instances where Large Language Models (LLMs) have fabricated citations (e.g., non-existent papers, incorrect authors, or mismatched titles).
2. How are academic publishers and institutions responding to the detection of AI-generated citations?
   • Evidence required: Official policy statements, editorial guidelines, and reports on the implementation of detection tools or screening processes by major academic journals and universities.
3. What are the ethical and integrity implications of AI-generated citations in scholarly communication?
   • Evidence required: Commentary, ethics committee reports, or editorial board discussions regarding the impact on trust, reproducibility, and the scholarly record.
4. What technical methods are currently being used to identify or mitigate AI-generated citations?
   • Evidence required: Technical papers or industry white papers detailing the efficacy of citation-checking software, cross-referencing databases (e.g., Crossref, DOI verification), and AI-detection algorithms.

Scope

• Timeframe: January 1, 2023, to the present date (as the 2026 endpoint is in the future, the scope is limited to available data up to the current moment).
• Domain: Academic and scholarly research, including preprints, peer-reviewed journals, and conference proceedings.
• Technology: Generative AI models (e.g., ChatGPT, Claude, Gemini) and their application in drafting or referencing academic literature.
• Geography: Global, with a focus on English-language academic publishing standards.

Exclusions

• Non-Academic Contexts: AI-generated citations in journalism, marketing, or general web content are excluded unless they directly inform academic practices.
• Future Predictions: Speculative essays or opinion pieces about what might happen in 2025-2026 that lack empirical grounding or current evidence.
• General AI Ethics: Broad discussions about AI ethics that do not specifically address the issue of citation integrity or hallucinated references.

False-premise check

CORRECTED in revision: the original draft wrongly stated the current year as 2024. Today is 1 Nov 2026 and the included sources include 2026 works, so the 2023–2026 timeframe is within available data. No date false premise remains.

Reframed Question: "What does published evidence say about AI-generated citations in academic research from 2023 to the present, and what are the emerging trends regarding their detection and mitigation?"

Findings

Executive answer

S2 and S3 discuss GenAI automating literature searches and manuscript composition, and the ethical and workflow pressures that follow [S2, S3]; they do not provide evidence of risk at scale. One published framework proposes competency standards for research leadership to benchmark institutional readiness on responsible AI use [S1], but it is a single framework, not evidence of a sector-wide shift. S2 argues the core ethical issue is the unethical use of AI rather than the technology itself [S2], but that is S2's argument, not a literature-wide consensus. Current literature highlights that GenAI is reshaping the traditional resources of editorial and reviewer attention [S3].

1. Prevalence and nature of AI-generated hallucinations

Confidence: Medium (The sources confirm the automation of literature searches but do not provide specific statistical rates of citation hallucinations.)

• Facts: GenAI is being used to automate parts of the research process, including literature searches [S3]. The use of AI tools in composing academic papers is described as "becoming rampant" [S2].
• Inference (speculative, not a source claim): If GenAI automates literature searches, fabricated or mismatched citations could follow; S3 does not define this as a structural risk inherent in current workflows [S3].
• Agent recommendation (not a finding, not a source claim): Researchers should verify all AI-generated literature references against primary databases to ensure accuracy.

2. Institutional and publisher responses

Confidence: High (The sources provide clear evidence of frameworks and editorial pressures.)

• Facts: A competency framework has been developed for university leadership to benchmark institutional readiness regarding responsible AI use [S1]. Academic journals are facing pressure as GenAI reduces the effort required to create manuscripts, thereby straining editorial attention and reviewer time [S3].
• Inference: S1's vocabulary could, if adopted, give institutions a shared standard for managing AI integration [S1], but S1 alone does not show a sector-wide movement from reactive to proactive policy.
• Agent recommendation (not S1's recommendation, not a finding): Universities could consider the competency framework provided by Zyphur (2026) to align research training with responsible AI standards [S1].

3. Ethical and integrity implications

Confidence: High (The sources explicitly discuss the ethical dimensions of AI in scholarship.)

• Facts: S2 identifies the core ethical issue as the "unethical use" of AI rather than the technology itself [S2]. S3 argues GenAI reshapes the resources editorial and reviewer attention rely on [S3]; it does not claim GenAI threatens the credibility of scientific signals themselves.
• Inference (hypothetical, not a finding): If automation of research workflows outpaces peer-review verification capacity, the scholarly record could be at risk; S3 discusses the editorial pressure, not this outcome [S3].
• Agent recommendation (not a finding, not a source claim): Scholarly communities should prioritize the development of ethical guidelines that distinguish between AI-assisted research and AI-generated fabrication.

4. Technical methods for identification and mitigation

Confidence: Low (The provided sources do not detail specific technical software or algorithms for citation verification.)

• Facts: Current literature on AI in academic writing is described as "scattered" [S4]. While bibliometric analyses are being used to track trends, the provided sources do not list specific technical tools or software for detecting AI-generated citations [S4].
• Inference: There is a gap between the identification of the problem (unethical use) and the availability of technical mitigation tools in the current literature [S2, S4].
• Agent recommendation (not a finding, not a source claim): Future research should focus on the efficacy of cross-referencing databases and AI-detection algorithms specifically tailored for citation integrity.

What the evidence does NOT show

• The evidence does not provide specific statistical data on the frequency or prevalence of AI-hallucinated citations in peer-reviewed journals.
• The evidence does not identify or evaluate specific technical software (e.g., Crossref, DOI verification tools) currently used to mitigate or detect AI-generated citations.
• The evidence does not offer empirical data on the success rates of current institutional policies in preventing AI-generated academic misconduct.

Evidence table (verified — UNVERIFIED rows are marked, never dropped)

Adversarial review flags (NOT fixes — reviewer must resolve each)

Here is the adversarial review of the research draft:

(a) Claims whose cited source does NOT support them 1. "Institutional responses are shifting toward competency frameworks for research leadership to manage these risks [S1]." (S1 is a single report/framework, not evidence of a broader "shift" in institutional responses.) 2. "The primary ethical concern is not the use of AI itself, but rather its unethical application in scholarly communication [S2]." (S2 states this is the author's argument/finding, not a consensus or established fact in the literature.) 3. "The use of AI tools in composing academic papers is described as 'rampant' [S2]." (S2 describes the use as "becoming rampant," which is a distinct difference in temporal scope and intensity.) 4. "The use of GenAI threatens the credibility of scientific signals, which rely on shared norms and limited human resources (editorial and reviewer time) [S3]." (S3 argues GenAI reshapes these resources; it does not explicitly claim it threatens the credibility of the signals themselves.)

(b) Material claims with NO citation 1. "Researchers should verify all AI-generated literature references against primary databases to ensure accuracy." 2. "Universities should adopt the competency framework provided by Zyphur (2026) to align research training with responsible AI standards." 3. "Scholarly communities should prioritize the development of ethical guidelines that distinguish between AI-assisted research and AI-generated fabrication." 4. "Future research should focus on the efficacy of cross-referencing databases and AI-detection algorithms specifically tailored for citation integrity."

(c) Overstatements, hedged claims presented as facts, or numbers/dates/quotations not traceable to a cited source 1. "Recent evidence indicates that Generative AI (GenAI) is increasingly used to automate literature searches and manuscript composition, creating risks for academic integrity [S2, S3]." (The sources discuss the potential for automation and ethical issues, but do not provide evidence that this is currently "creating risks" at scale.) 2. "Because GenAI automates literature searches, the potential for fabricated or mismatched citations is a structural risk inherent in current research workflows [S3]." (This is a speculative inference presented as a structural fact; S3 does not explicitly define this as a "structural risk.") 3. "Institutions are moving toward a standardized, evidence-anchored vocabulary to manage AI integration, suggesting a transition from reactive to proactive policy development [S1]." (This overstates the impact of S1, which is a single document, as a broad institutional movement.) 4. "The integrity of the scholarly record is at risk if the automation of research workflows outpaces the capacity of peer-review systems to verify the authenticity of citations [S3]." (This is a hypothetical scenario presented as a factual risk assessment.)

UNSOURCED claims (quarantined — never findings)

• Recent evidence indicates that Generative AI (GenAI) is increasingly used to automate literature searches and manuscript composition, creating risks for academic integrity [S2, S3].

• Confidence: Medium** (The sources confirm the automation of literature searches but do not provide specific statistical rates of citation hallucinations.)

• Recommendation:** Researchers should verify all AI-generated literature references against primary databases to ensure accuracy.

• Confidence: High** (The sources provide clear evidence of frameworks and editorial pressures.)

• Confidence: High** (The sources explicitly discuss the ethical dimensions of AI in scholarship.)

• Confidence: Low** (The provided sources do not detail specific technical software or algorithms for citation verification.)

• Inference:** There is a gap between the identification of the problem (unethical use) and the availability of technical mitigation tools in the current literature [S2, S4].

• Recommendation:** Future research should focus on the efficacy of cross-referencing databases and AI-detection algorithms specifically tailored for citation integrity.

• The evidence does not provide specific statistical data on the frequency or prevalence of AI-hallucinated citations in peer-reviewed journals.

• The evidence does not identify or evaluate specific technical software (e.g., Crossref, DOI verification tools) currently used to mitigate or detect AI-generated citations.

• The evidence does not offer empirical data on the success rates of current institutional policies in preventing AI-generated academic misconduct.

Sources (included — cited as [Sn])

1. Responsible AI in Academic Research: A Competency Framework for Research Training (2026). Michael James Zyphur. https://doi.org/10.61700/t31oy23grr

2. ETHICAL CHALLENGES OF USING AI IN ACADEMIC RESEARCH (2026). O.M. Lumanze. https://doi.org/10.52968/15068794

3. Academic Publishing in the Age of Generative AI (2026). Björn Hanneke, Frederik Hering, Jella Pfeiffer, Hajo Alexander Reijers, Oliver Hinz. https://doi.org/10.1007/s12599-026-01013-0

4. Artificial Intelligence for Academic Writing: A Bibliometric Analysis (2022–2026) (2026). Siti Hajar Abd. Hamid, Norah Md Noor, N H Abd Halim. https://doi.org/10.11113/itlj.v10.218

Excluded candidates are listed with reasons in 03-screening-log.md; they are never cited.

Metrics box (mechanical only — nothing estimated)

• Candidates screened: 10
• Sources included: 4
• Sources cited: 4
• Citation markers found: 16
• Invalid citation markers: none
• Evidence-table rows: 4
• Evidence-table rows verified: 4/4
• Claim sentences (original agent synthesis run): 27
• Cited claim sentences (original synthesis): 16
• Sourced-claim rate (original synthesis): 59% (mechanical sentence rule: a sentence counts as cited if it contains at least one valid [Sn] marker; headings excluded)
• REVISED-DRAFT recompute (Findings section only, same sentence rule, 1 Nov 2026): original draft 24 sentences / 16 cited = 67%; revised draft 24 sentences / 15 cited = 62%. The one-sentence drop is deliberate: the section-3 recommendation is now an agent recommendation with no citation, instead of appearing source-backed. No new evidence was added.
• Citation precision: NOT mechanically computable — see adversarial review flags above; human reviewer judges whether each cited source supports its claim.
• Overall confidence: not computed mechanically — see per-section confidence in Findings; reviewer confirms.
• Human review: DONE — Riya Nalwa, 1 Nov 2026.

Appendix: method note

Agent v0.4 staged run: framing, evidence acquisition, screening, synthesis, structured evidence table, mechanical table verification, mechanical claim-level verification, adversarial review, human gate. Sources: Semantic Scholar: rate-limited (shared-IP budget). Model: gemini-3.1-flash-lite. What would change the answer: additional primary sources, or resolution of the adversarial flags above.

Revision log (1 Nov 2026, agent revision — NOT a human approval)

• Resolved adversarial flags (a)1–4, (b)1–4, (c)1–4 by rewording to what S1–S4 actually support, and by labelling all four recommendations as agent recommendations, not findings and not source claims.
• Corrected the draft's false-premise note: it wrongly stated the current year as 2024; today is 1 Nov 2026 and 2026 sources are included.
• No new evidence added, no source added or removed, no metric recomputed — the original metrics box applies to the original draft. This revised draft has NOT been re-run through the agent's mechanical verification.
• Status remains REVISED DRAFT / needs_review. Human review is still required. Nothing in this file is approved, exported as final, published, or sent to any client.

Revision log, part 2 — metrics recompute (1 Nov 2026)

• Findings-section mechanical recompute using the agent's sentence rule: original draft 24 sentences / 16 cited (67%); revised draft 24 sentences / 15 cited (62%).
• The drop is not a loss of evidence: one sentence (the section-3 recommendation) is now correctly uncited as an agent recommendation. The original synthesis metrics (27 sentences / 16 cited / 59%) still stand as the job's run metrics and are not replaced by this Findings-section recompute.
• Status remains REVISED DRAFT / needs_review. Human review still required. Nothing approved, exported, published, or sent.

Oryx Centre for Strategic Futures · Contact: amartya@oryxcentre.dpdns.org · Home