← Back to website

What Is Accounting and Finance Research For in the AI Era?

When writing, coding and basic synthesis become cheaper, research does not lose all value. The basis of its value changes.
Dr Yuqian Zhang, Auckland University of Technology | October 2026

Executive Summary

AI does not make accounting and finance research worthless. It makes undifferentiated research output less scarce. The main question is therefore not whether a model can produce a plausible paper. The question is what a research contribution must now contain to deserve attention, editorial capacity and public trust.

The evidence in this brief points to a shift in scarcity. Crossref-indexed research article output in 11 selected journals rose 27.9% between 2015 and 2025. In the two finance journals with usable public submission records, however, submissions did not rise continuously after large language models became widely available. Editorial acceptance remained selective, and observed median time to an editorial decision ranged from 30 to 52 days across the reported years. Citation attention could not be used to compare journals reliably, while AI guidance is now common but uneven in scope, and stated guidance specifies few research criteria.

These findings do not support a simple story of research abundance leading to research collapse. They support a more specific claim: production is not the main bottleneck. Judgement, verifiable evidence, access to difficult settings, editorial evaluation and accountable use are more likely to determine value.

Table of Contents

  1. Executive Summary
  2. Introduction
  3. Background
  4. Data and Method
  5. Findings
  6. Discussion
  7. Research Opportunities
  8. Data Availability and Replication
  9. References
  10. Sources

Introduction

Generative AI can draft a literature review, write code, propose a regression and turn results into polished prose. The larger institutional effect is less clear. If these tasks become cheap, will research become more plentiful, less valuable or differently valuable?

This brief asks what makes research trustworthy and useful, not only what makes a paper easy to produce. A paper has little value if no one can establish what it contributes. A model or dataset has little value if no one can verify how it was built. A recommendation has little value if its authors cannot explain uncertainty or accept responsibility for error. Those conditions matter more when fluent output is abundant.

The report examines six dimensions: research production, citation attention, review and gatekeeping, visible AI guidance, evidence correction and stated editorial priorities. It then connects the evidence to a practical question: what should individual researchers, doctoral programmes, journals, departments and universities do next?

Central argument. The scarce academic asset is moving from the ability to produce a paper toward the ability to choose a worthwhile problem, obtain and understand difficult evidence, verify the chain from data to claim, and make a responsible judgement that others can audit.

Background

AI is not a single technology. In this brief, the term includes generative models that write and reason with language, code-generation tools, machine learning and automated data processing. The relevant concern is not that these tools can imitate an academic article. It is that they reduce the cost of many intermediate tasks while leaving the hardest social and evidentiary tasks intact.

Experiments show clear gains in some writing tasks. Noy and Zhang found that, among 453 college-educated professionals completing mid-level writing tasks, access to ChatGPT reduced average time by 40% and raised output quality by 18% [1]. Their experiment did not study accounting or finance scholarship. It does, however, show why drafting alone is a weak professional advantage.

Evidence about research ideas is more mixed. Si, Yang and Hashimoto recruited more than 100 NLP researchers to compare expert and model-generated ideas. The model-generated ideas were judged more novel, but slightly weaker on feasibility. The authors also reported limited diversity and failures of model self-evaluation [2]. Novelty in a blinded review is not the same as a defensible contribution.

A third experiment, on short-story writing, found higher individual creativity ratings for people given AI ideas, but greater similarity across AI-assisted stories [3]. The setting is not academic research. It nevertheless illustrates a possible collective problem: a tool can improve each person's output while narrowing the range of ideas the system produces.

Direct evidence from accounting itself is now emerging. Gardner, Krupa and Mammadov (2026) analysed 3,573 accounting papers across 10 journals and 4,166 abstracts across 22 journals, and found that AI-associated vocabulary rose by about 75% after 2023, with heavier use among junior authors and lower-ranked journals [12]. A separate study of more than 15 million biomedical abstracts estimated that at least 13.5% of 2024 abstracts showed signs of LLM processing [13]. These are vocabulary-based lower bounds, not verified tool use, and they say nothing about research quality. They do show that AI-assisted writing is already inside the published record.

These results suggest a distinction that matters throughout the rest of the report. AI may improve measured output quality on a defined task. It may also create common ideas, common errors and common forms of presentation. Neither result establishes the value of a longer-term research programme.

Data and Method

The report uses a frozen sample of 11 English-language accounting and finance journals: The Accounting Review, Journal of Accounting Research, Journal of Accounting and Economics, Contemporary Accounting Research, Review of Accounting Studies, Accounting, Organizations and Society, The Journal of Finance, Journal of Financial Economics, The Review of Financial Studies, Journal of Financial and Quantitative Analysis, and Review of Finance. The sample is purposeful and is not a global census.

Crossref [7] supplied article metadata. The queries used each journal ISSN, the journal-article type and a publication window from 2014 through 2026. Only issue years from 2015 through 2025 were retained. DOI duplicates were removed. A fixed, documented title rule excluded nonresearch records: editor and editorial matter, front and back matter, issue information, announcements and society business, prize notices, conference programmes, reviewer lists, placement advertisements, corrections, retractions, book reviews, discussion pieces and introductions. The final dataset contains 9,044 unique research article records from 11 journals. Two limits matter. First, Crossref typing and indexing are not identical to journal-specific definitions of a research paper, and some residual miscellany may remain. Second, the 2025 counts are the latest available at the access date: Journal of Accounting and Economics records show a large one-year fall that is more consistent with a deposit gap than with a publication decision. The 2025 endpoint is therefore provisional. At the 2024 endpoint the count is 887, a 23.5% increase on 2015.

Only two journals had public submission records that could be used under the evidence rules. The Journal of Finance publishes calendar-year totals, and The Review of Financial Studies publishes rolling May-to-May periods. Their definitions are not sufficiently similar to pool. Figure 1 therefore indexes each journal to its own first observed period and a companion table keeps the reported counts visible. The other nine journals publish only rolling, non-comparable metrics such as submission-to-first-decision timelines or current acceptance rates, so those values are not pooled here.

Reported submission totals and definitions
JournalReporting periodTotal submissionsDefinition
The Journal of Finance20151,227Calendar year; includes resubmissions
The Journal of Finance20201,333Calendar year; includes resubmissions
The Journal of Finance20231,307Calendar year; includes resubmissions
The Journal of Finance20241,066Calendar year; includes resubmissions
The Review of Financial StudiesMay 2022 to May 20231,534New and dual submissions; excludes resubmissions
The Review of Financial StudiesMay 2023 to May 20241,544New and dual submissions; excludes resubmissions
The Review of Financial StudiesMay 2024 to May 20251,474New and dual submissions; excludes resubmissions
The Review of Financial StudiesMay 2025 to May 20261,664New and dual submissions; excludes resubmissions

Citation attention was tested but not used as a comparative measure. For the 2020, 2021 and 2022 Crossref research articles, we matched DOIs to OpenAlex records [8] and summed citations in the three calendar years starting with the Crossref issue year. Of 2,655 cohort articles, 2,578 (97.1%) were matched by DOI; the unmatched records were mostly OpenAlex redirects to SSRN versions. Twenty-six matched articles (1.0%) had no recorded citation in the window. Journal shares ranged from 0.0% to 3.1%, with no more than five zero-citation articles in any journal, and the differences are not statistically distinguishable. The exercise is reported in the data appendix; it is not used to compare journals.

A citation cohort is the set of articles sharing a publication year. This measure does not count human readers. Automated access, citation lag, incomplete indexing and differences in citation cultures all affect the result. The evidence can show uneven visible attention. It cannot establish that human readers have disappeared.

AI guidance was collected from the latest usable official journal guidance or editorial-policy page for each journal. Nine pages were observed live on 22 October 2026. The Review of Financial Studies and Review of Finance use the latest usable archived pages (April 2025 and June 2024) because their live pages blocked automated access. Visible text was searched for generative AI, artificial intelligence, AI-assisted tools, large language models, LLM and ChatGPT. A page was coded as showing explicit AI guidance when an AI term appeared with a policy term such as disclosure, authorship, accountability, responsibility or prohibition. For Contemporary Accounting Research, the publisher page now directs authors to the CAAA author-information page, and the linked CAAA statement supplies the explicit terms. A missing mention means only that the guidance was not visible on the captured page. It does not prove that a linked publisher policy is absent.

Editorial priorities were recoded by hand from the substantive author-guidance and editorial-policy text of each journal's captured page, using a six-category codebook. A criterion counts only where the page states it as an author requirement or editorial expectation, with a quotable sentence. Navigation menus, JEL browse lists, copyright and cookie boilerplate, proof-stage notices and publisher-wide template paragraphs were excluded. This is a measure of stated criteria, not editorial behaviour.

Crossref update filters were also run for retractions, errata and corrections. The resulting record counts are not clean retraction rates because one event can appear in both the original record and a notice. No rate is inferred.

Terms used in this report have precise meanings. Crossref and OpenAlex are open bibliographic databases: Crossref records publication metadata such as DOIs and ISSNs, and OpenAlex adds citation links. A DOI is a persistent identifier for a publication, and an ISSN identifies a journal. Contribution is used in its ordinary scholarly sense: a defensible advance in knowledge or practice, not a count of outputs. An indexed record is one that a database has matched to a publication. A desk rejection is a rejection before external peer review. Identification concerns whether a relationship can be interpreted as causal rather than merely correlated. Replication means that another researcher can reproduce the analysis from the documented data and code, within legitimate access limits. Generative-AI guidance is an explicit statement about permitted use, disclosure or accountability.

Findings

Research production increased, but the submission story is uneven

Figure 1 compares the two journals with usable public submission records. Their reporting periods differ, so each series is indexed separately.

Figure 1: Submission trends differ at the two journals with public data

Index: first observed period for each journal equals 100. The Journal of Finance uses calendar years; The Review of Financial Studies uses rolling May-to-May periods. Sources: AFA [4] and SFS [5].

The Journal of Finance stands at 86.9 in 2024 on its 2015 base. The Review of Financial Studies stands at 108.5 in 2026 on its 2023 base, after dipping in 2025.

Submission counts tell a limited story. The Journal of Finance recorded 1,227 submissions in 2015, 1,333 in 2020, 1,307 in 2023 and 1,066 in 2024. The Review of Financial Studies reported 1,534 in its 2022-2023 period, 1,544 in 2023-2024, 1,474 in 2024-2025 and 1,664 in 2025-2026. These movements are inconsistent with a simple claim that AI caused a permanent increase in finance submissions.

Crossref-indexed research article output across the selected journals rose from 718 in 2015 to 918 in 2025, an increase of 27.9%. The path was not steady, and three journals published fewer articles in 2025 than in 2015. That endpoint is provisional: at 2024, two journals were below their 2015 level and the count was 23.5% above 2015. This is evidence of modest growth in indexed output, not an AI-driven explosion.

Figure 2: Crossref research article counts rose 27.9% from 2015 to 2025

Crossref-indexed research articles in 11 selected journals. Unit: article records. A fixed title rule excludes editorial matter, front and back matter, issue information, announcements, society business, prizes, conference programmes, reviewer lists, placement advertisements, corrections, retractions, book reviews, discussion pieces and introductions. Source: Crossref [7].

Accounting and finance counts fluctuate without a continuous rise. The two field series sum to 718 in 2015 and 918 in 2025, a 27.9% increase; the 2025 endpoint is provisional.

The data do not reveal one uniform trend because the journals use different definitions and because most journals in the sample did not provide submission counts that met the evidence rules. The absence of a clear post-2022 surge at these outlets does not imply that AI has no effect. It means the effect cannot be inferred from a short, selective series. Later research should test preprint volume, desk-rejection behaviour and submissions to lower-ranked and specialist journals.

Citation counts are not readership

We also tested whether citation attention had become more uneven. It had not, in any way the data can defend. After cleaning the article cohort and matching citations to the Crossref issue year, the pooled share of articles with no citation in their first three calendar years was 1.0%, and journal-level shares ranged from 0.0% to 3.1%. The zero counts are small, the intervals overlap and the measure is sensitive to when an article first appeared online. An earlier version of this brief reported large journal differences; that result came from a cohort that still contained front matter, announcements and articles with inconsistent publication years, and it is withdrawn.

This limitation matters for the original question. A fall in human readership remains plausible but unverified. Citation counts record visible scholarly use, not reading, and a zero count can mean no attention, delayed attention or an indexing gap. AI makes the distinction more important because text can be produced without being read.

Editorial gatekeeping remains selective, and decisions still take weeks

Acceptance and rejection rates add useful context. The Journal of Finance accepted 6.4% of 1,226 editorial decisions in 2023 and desk-rejected 43% of new submissions. Its median turnaround was 43 days. The Review of Financial Studies reported an 8.1% acceptance rate and a 33-day median turnaround in its 2024-2025 period, and 7.27% with a 30-day median in its 2025-2026 period [5].

Figure 3: Median time to an editorial decision ranged from 30 to 52 days

Median time to an editorial decision by reporting period. Gaps are shown without interpolation. The Journal of Finance series runs from 2015 to 2023; the 2024 and 2025 editor reports are published but could not be accessed for this brief. Sources: Journal of Finance editor reports [6] and SFS statistics [5].

The Journal of Finance reported medians between 40.5 and 52 days, including 48 days in 2021. The Review of Financial Studies reported 35, 33 and 30 days in 2023, 2025 and 2026.

The Journal of Finance reported a median of 48 days in 2021, 52 days in 2022 and 43 days in 2023. Its 2023 desk-rejection rate was also 43%. The Review of Financial Studies median fell to 30 days in its latest period. These figures show that fast triage is possible. They do not show that careful review has become easier or faster.

The evidence here cannot measure reviewer supply directly. But the pattern is consistent with a different bottleneck: the supply of reviewers who can test a claim, editors who can compare contributions, and researchers who can detect errors that a model will also miss. AI can help editors screen submissions. It can also increase the number of plausible but underdeveloped submissions requiring screening.

AI rules are visible on most author pages, but scope varies

Across the latest usable official pages, 8 of 11 journals showed explicit AI guidance in the visible text. The Accounting Review requires disclosure and states that AI cannot serve as an author [9]. Review of Accounting Studies states that large language models do not meet authorship criteria and that use should be documented [11]. Journal of Accounting Research, Journal of Accounting and Economics, Contemporary Accounting Research, Accounting, Organizations and Society, The Journal of Finance and Journal of Financial Economics also displayed explicit AI rules on their latest pages. The Journal of Finance, for example, publishes an AFA AI policy that makes authors responsible for fact-checking outputs, bars AI tools from co-authorship and asks for model versions and prompts when AI is used to create data or analysis [14]. The Review of Financial Studies, Journal of Financial and Quantitative Analysis and Review of Finance did not show explicit AI guidance on the pages captured.

Figure 4: 8 of 11 official journal pages showed explicit AI guidance

Latest usable official guidance or policy page per journal. Nine pages were observed live on 22 October 2026; The Review of Financial Studies and Review of Finance use the latest usable archived pages (April 2025 and June 2024). A missing page statement does not rule out a linked publisher policy. Source: official journal pages.

Eight journals show explicit AI guidance and three do not. The status reflects visible page text, not compliance or the complete policy environment of the publisher.

The variation should not be read as a journal ranking. Wiley, Elsevier, Springer Nature, Oxford University Press, Cambridge University Press and the American Accounting Association maintain different policy structures. A journal page may link to a publisher policy rather than reproduce it, and capture dates differ. The finding is narrower than it first appears: AI rules are now common, but most of them address writing, disclosure and authorship rather than data, code, model use or verification. A reader arriving at an author page will not always see the full policy environment, and three pages showed no explicit rule at all.

A publisher-level example makes the accountability issue explicit. Elsevier's journal policy permits AI support for manuscript preparation but states that AI tools must not substitute for human critical thinking, expertise and evaluation. Authors remain accountable and must verify sources, disclose substantive use and protect confidential material [10]. This makes clear where responsibility remains, but it does not show that the policy prevents error.

For accounting and finance, the policy question is broader than manuscript writing. AI may process earnings-call transcripts, propose empirical specifications, classify disclosures, generate code and produce recommendations. A journal rule that covers only grammar support would leave the research process largely unaddressed. Journals and universities need procedures for model choice, prompts, data retention, versioning and independent verification.

Corrections and retractions show why the evidence record needs maintenance

Crossref returned 20 retraction-update records, 21 erratum records and 28 correction records across the selected journals for 2015-2025. These are filtered record counts, not retraction rates. One event can appear more than once, and reporting practices differ.

The practical implication is not that the journals are unreliable. It is that a useful research record must include later correction, replication and data provenance. A publication is one point in a continuing evidence record, not its final state. When models can produce polished explanations of weak evidence, the ability to maintain and correct the record becomes more valuable.

Author guidance uses different ideas of valuable research

The recoded author-page sample states data transparency on 5 of 11 pages, integrity and accountability on 4, reproducibility on 3, contribution on 2 and research design on 1. No captured page states relevance to practitioners or policy. The counts are lower bounds: two journals' captured pages carried no substantive guidance, and three hold their policy in linked files.

Figure 5: Author guidance states few research criteria explicitly

Number of journals whose author-guidance or editorial-policy text explicitly states each criterion. Unit: journals. Source: manual coding of official author-page captures (2026 pages where available; archived pages for RFS and Review of Finance). Navigation, JEL browse lists, copyright notices and shared publisher template text excluded.

Data transparency is stated on 5 of 11 pages, integrity and accountability on 4, reproducibility on 3, contribution on 2 and research design on 1. Decision relevance is stated on none.

The gap is the finding. Only data transparency is stated by close to half the sample, and no journal states relevance to practitioners or policy as an evaluation criterion. Contribution to the literature is named by two journals and research design by one. Author guidance therefore specifies very little about what makes research valuable; most of that judgement stays with editors and referees, which is where AI-assisted submission volume will press hardest.

Discussion

The scarcity shifts

A better response to cheaper writing and coding is not to increase output. It is to shift effort toward assets that remain costly, difficult to copy or necessary for trust.

1. Problem selection

The ability to identify a question that matters to theory or a real decision becomes more valuable when articles are easy to generate. Judgment about importance cannot be automated by asking a model whether a topic is interesting.

2. Data and context

Proprietary audit data, loan records, internal controls, hand-collected governance information, field access and regulatory knowledge create a defensible research position. Public, standardised data remain useful, but their analysis is less distinctive when the analysis is easy to reproduce.

3. Verification

The supply of plausible claims may rise faster than the supply of people who can test them. Research infrastructure should make it easy to inspect data lineage, code, assumptions, failed tests and alternative explanations.

4. Accountability

Someone must decide which evidence is sufficient and accept responsibility when the recommendation is wrong. That decision requires domain knowledge, ethics and institutional judgment, not only model capability.

Basic coding and model use are no longer unusual advantages. They remain necessary. Basic writing is similar. A researcher can know Python, use a model well and still write a clear answer to an unimportant question. The advantage lies in combining technical ability with problem judgement, evidence access and accountable interpretation.

Critical skills need to work together

The practical response is not a longer list of software tools. Technical ability should support research judgement rather than replace it.

CapabilityWhat remains scarceEvidence of capability
Problem judgementRecognising a question that matters before the literature, a model or a trend labels it importantA coherent agenda connected to unresolved theory or consequential decisions
Research designChoosing a credible comparison, measure and identification strategyPre-specified claims, tests against rival explanations and explicit limits
Data and institutional knowledgeAccessing and understanding accounting rules, markets, firms, regulators and confidential settingsDocumented provenance, domain interpretation and lawful access to data that others cannot easily obtain
Computational auditKnowing when generated code or an automated classification is wrongIndependent replication, edge-case tests, sensitivity analysis and model evaluation
Evidence synthesisWeighing conflicting results without erasing disagreementA verified evidence map with stable identifiers and explicit uncertainty
Accountable judgementDeciding what the evidence means for a decision and accepting responsibilityClear recommendations, stated tradeoffs and a correction process

Programming and prompt fluency are still valuable because a researcher who can inspect data, build tools, evaluate a model and read its code has more control over the evidence. Model literacy is broader than prompt writing. It includes understanding training data, retrieval, context limits, uncertainty, evaluation, failure modes and the difference between fluency and reliability.

Productivity should mean more defensible work within a given time, not simply more output. The useful question is whether a workflow reduces the time and cost required to reach a defensible answer, or whether it merely increases the number of unfinished outputs. Workflow optimisation matters when it improves verification and reuse. It creates risk when it accelerates production faster than evaluation.

Workflow optimisation is necessary, but not sufficient

A useful AI workflow changes where human attention is spent. The objective is fewer errors, shorter iteration cycles and more auditable decisions. A workflow designed mainly to raise output volume can increase editorial burden without increasing knowledge.

  1. Frame a consequential question. State the decision or theoretical problem, the audience and what evidence would change an answer.
  2. Specify the claim before generating text. Identify the estimand, comparison, expected direction and limits.
  3. Map existing evidence. Use structured retrieval, maintain stable identifiers and verify every citation against the source.
  4. Secure data and rights. Document the source, permission, coverage and transformations. Keep confidential material out of tools that cannot protect it.
  5. Design before prompting. Let AI propose alternatives, but require human approval of the identification strategy, measures and falsification tests.
  6. Generate code with tests. Use the model to write code and tests, then rerun from a clean environment and compare outputs with an independent implementation.
  7. Red-team the result. Ask another model or colleague for the strongest competing explanation, then test the criticism against data.
  8. Write the argument, not just the prose. Use AI to critique structure and clarity, while a named human owns every factual claim and normative judgement.
  9. Preserve provenance. Retain data dictionaries, prompts where relevant, model versions, outputs, failed tests and decision notes.
  10. Maintain the contribution. Correct errors, update data, document replication and track use.

This workflow treats the model as a research instrument and a critic, not as a substitute for the researcher. The phrase “human in the loop” is insufficient if the human merely approves polished output. Oversight has to include the authority, time and competence to stop a result that does not meet the evidence threshold.

Publication remains a threshold, not a sufficient measure

The evidence reviewed here is not sufficient to show that universities have abandoned publication counts or replaced them with a new metric. A likely near-term outcome is hybrid evaluation: publication remains an eligibility threshold, while appointments, promotion and funding use a broader evidence portfolio. Publication is likely to remain visible because it provides peer evaluation, a durable record and a common language for comparison. Its weakness is that a count cannot distinguish a carefully verified contribution from a technically compliant paper.

A credible next step is a research contribution portfolio. It would report contribution across several dimensions rather than hide them inside one undisclosed number.

DimensionWhat the evidence should showMain gaming risk
Problem importanceA consequential unresolved question and a credible claim about who would be affected by the answerTopic labels and novelty claims without consequence
Evidence strengthResearch design, measurement validity and survival under credible alternativesComplex robustness tables around a weak question
Provenance and reproducibilityData and code access where lawful, plus a traceable chain from source to resultOpenness as a badge with unusable files
Decision useDocumented use by researchers, firms, regulators or standard settersCitation padding and stakeholder name-dropping
Maintained valueCorrections, updates, replication and reuse over timeOne-off publication without maintenance
Scholarly stewardshipReview, data curation, mentorship and infrastructure that improve shared research capacityCounting service volume without quality

The portfolio is a design proposal, not a validated evaluation system. It has costs. Collecting evidence is burdensome. Weights can be gamed. Qualitative evidence can be captured by powerful institutions. Researchers without proprietary data may be disadvantaged. A portfolio should therefore expose its components and permit contextual weighting rather than claim that one number is neutral.

A defensible implementation would separate three layers: contribution quality, process integrity and external use. Contribution quality concerns problem selection and evidence. Process integrity concerns data, code, disclosure and correction. External use concerns citations, policy or practice. The three layers answer different questions and should not be collapsed without explanation.

Three scenarios for 2030

The following scenarios are analytical possibilities. They are not forecasts. Their purpose is to make assumptions explicit and to identify indicators that could distinguish them.

ScenarioConditions and indicatorsConsequence for research valueMain risk
More output, weaker attentionSubmission volume rises while human reading and reviewer supply do not. Editors desk-reject more and visible attention stays concentrated.Problem selection, verification and trusted synthesis become more valuable.Evaluation becomes a competition in volume and visibility.
Quality checks become standardJournals and institutions adopt evidence labels, replication checks, data-access registration or contribution portfolios.Identification, data provenance and maintained evidence become core assets.New metrics become expensive and gameable.
Embedded researchFirms, regulators and standard setters commission and use research more directly.Decision relevance, implementation and accountability gain status.Funding relationships weaken independence and public credibility.

How to prepare

Individuals. Build one or two deep problem agendas rather than a large set of disconnected papers. Develop a defensible data or field position. Learn to specify causal designs, test code independently and use models against a written evidence standard. Keep a portfolio record that shows what changed in understanding or practice, not only where work was published.

Doctoral programmes. Teach research design, measurement, data engineering, model evaluation and research integrity as one connected sequence. Require students to replicate a published result, document a data pipeline and defend the assumptions of an AI-assisted analysis. Writing instruction should focus on argument, uncertainty and evidence, not only prose generation.

Journals. Publish clear rules for AI use across writing, data work, modelling and review. Require enough provenance for readers to understand the chain from data to claim. Invest in reviewer support and experiments with verification, registered reports and evidence audits. Publish the limitations of operational statistics when comparable data are unavailable.

Departments and universities. Stop using paper counts as the only evidence of capability. Use contribution portfolios with transparent dimensions and contextual weights. Fund long-lived data infrastructure and research software. Recognise replication, correction and data curation as research work. Audit any new metric for exclusion, gaming and administrative cost.

The preparation plan can begin before institutions agree on a new metric. Over 90 days, map the data and fields that provide a genuine research advantage, establish a source and code ledger, and run one AI-assisted workflow with independent verification. Over 12 months, aim for one contribution that can be defended, one reusable data or method asset, one output aimed at a decision maker, and one act of replication or stewardship. The objective is not to use more tools. It is to build a record that remains credible when tools are cheap.

Research Opportunities

The evidence suggests several research questions that accounting and finance can answer better than general commentary on AI.

These projects require access that will itself become part of the research contribution. The most valuable studies may use confidential review records, firm data, regulator access or longitudinal fieldwork. Public aggregation is useful for describing the environment. It is less useful for identifying the mechanisms that determine which research survives scrutiny.

Data Availability and Replication

All compiled datasets, the methodology note, the codebook and the Python replication script are available below. The script reads the saved CSV files, prints the statistics used in this report and summary, and regenerates all six PNG figures.

FileDescriptionDownload
README_methodology.txtMethodology notes, sources, compilation methods and known limitationsTXT
replicate.pyPython replication script that regenerates every chart and statistic in the reportPY
journal_scope.csvJournal sample and inclusion decisionsCSV
submissions_outputs.csvReported submission counts for the two journals with public recordsCSV
submission_index.csvWithin-journal index for the submission chartCSV
crossref_articles.csvArticle-level Crossref recordsCSV
crossref_article_counts.csvJournal-year Crossref article countsCSV
field_article_counts.csvField-year Crossref article countsCSV
openalex_attention_articles.csvArticle-level OpenAlex cohorts and citation windowsCSV
attention_cohorts.csvJournal-year attention cohortsCSV
attention_journal_summary.csvPooled journal attention measuresCSV
openalex_coverage.csvCrossref-to-OpenAlex coverage checksCSV
review_workload.csvReported decision times and ratesCSV
ai_policy_status.csvDated journal-page policy observationsCSV
ai_policy_latest.csvLatest usable policy observation per journalCSV
editorial_priorities.csvCoded journal-category observationsCSV
editorial_priority_summary.csvAggregate codebook countsCSV
corrections_retractions.csvCrossref update-record countsCSV
capability_model.csvAnalytical task and responsibility modelCSV
contribution_scorecard.csvProposed contribution portfolioCSV
scenario_assumptions.csvLabelled 2030 scenariosCSV
analysis_statistics.csvKey derived statistics used in the reportCSV

The public files reproduce the reported measures. They do not include paywalled editor reports or copyrighted journal text. Source locators and permitted local verification records are documented in the methodology.

References

[1] Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192. https://doi.org/10.1126/science.adh2586
[2] Si, C., Yang, D., & Hashimoto, T. (2024). Can LLMs generate novel research ideas? A large-scale human study with 100+ NLP researchers. arXiv:2409.04109. Preprint.
[3] Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290
[4] American Finance Association. Journal Statistics. https://afajof.org/journal-statistics/. Accessed 22 October 2026.
[5] Society for Financial Studies. RFS Statistics. https://sfs.org/rfs-statistics/. Accessed 22 October 2026.
[6] Schoar, A. (2024). Report of the Editor of The Journal of Finance for the Year 2023. The Journal of Finance, 79(2), 1697-1706. https://doi.org/10.1111/jofi.13329
[7] Crossref. REST API. https://api.crossref.org/. Accessed 22 October 2026.
[8] OpenAlex. Works API. https://api.openalex.org/works. Accessed 22 October 2026.
[9] American Accounting Association. Guide for Authors: The Accounting Review. https://aaahq.org/Research/Journals/The-Accounting-Review/Guide-for-Authors. Accessed 22 October 2026.
[10] Elsevier. Generative AI policies for journals. https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals. Accessed 22 October 2026.
[11] Springer Nature. Review of Accounting Studies submission guidelines. https://link.springer.com/journal/11142/submission-guidelines. Accessed 22 October 2026.
[12] Gardner, J., Krupa, N. R., & Mammadov, B. (2026). Generative artificial intelligence in the writing process of accounting research. Accounting Open, 2, 100010. https://doi.org/10.1016/j.accop.2026.100010
[13] Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11. https://doi.org/10.1126/sciadv.adt3813
[14] American Finance Association. Editorial Policies, including Artificial Intelligence Disclosure Policies. https://afajof.org/editorial-policies/. Accessed 22 October 2026.

Sources

SourceUseAccess or observation date
Crossref REST APIArticle output and update records22 October 2026
OpenAlex Works APICitation cohorts and coverage22 October 2026
American Finance Association journal statisticsJournal of Finance submissions, decisions and review time22 October 2026
Journal of Finance editor reportsAnnual decisions, desk rejection and turnaround2015-2021 and 2023 reports; 2022 comparison from the 2023 report
Society for Financial Studies statisticsReview of Financial Studies submissions, acceptance and turnaround22 October 2026
Official journal guidance pagesVisible AI guidance and coded editorial criteriaNine live pages on 22 October 2026; two archived pages (2024-2025)
Noy and Zhang (2023)Writing-task productivity evidencePublished 2023
Si, Yang and Hashimoto (2024)Research-idea generation evidencePreprint, 2024
Doshi and Hauser (2024)Individual creativity and collective similarityPublished 2024
Elsevier generative-AI policyPublisher accountability and disclosure rules22 October 2026

The complete source and claim ledgers contain exact locators, local snapshot records and limitations. Those files remain in the project evidence folder for audit and replication.