AI evidence extraction

Leveraging Artificial Intelligence in Scholarly Publishing

Fatima Zahra Ouariach, Meziani Khouloud EL, Ouariach Soufiane · 2025

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

10/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.44

FWCI

Top 10%

Impact

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.71426/jassh.v1.i1.pp1-8

Methodology & findings

Study design

Secondary data analysis synthesizing verified evidence from 2021-2025.

Main result

The study found that "the integration of Artificial Intelligence into scholarly publishing is a double-edged sword that is currently cutting deeply into the fabric of academic tradition." Key quantitative findings include: active AI usage surged from 37% to 58% globally between 2024-2025; grant writing preparation time reduced by approximately 90% (from 30-50 days to 3-5 days) with AI assistance achieving a 50% success rate compared to 10-20% baseline; retractions increased 40% from approximately 10,000 in 2023 to over 14,000 in 2024; DOI hallucination rates reached nearly 30% in sciences and over 60% in humanities; and detection tool bias against non-native English speakers was documented as "High" for GPTZero.

Research paradigm

Critical interpretivism

Author conclusions

The authors conclude: "This study concludes that the academic ecosystem has irreversibly crossed the event horizon of AI integration. The efficiency gains are too immense to roll back; the grant writing and literature synthesis capabilities of current models offer a solution to the unsustainable workload of the modern researcher. However, this efficiency comes at the cost of transparency and trust." They further argue that "Universities must abandon the punitive approach to AI (bans and detectors) in favor of critical AI literacy" and that "Publishers must accept that the 'version of record' can no longer be guaranteed by pre-publication review alone."

Risk of bias

Publication bias: study relies on industry reports and peer-reviewed publications, potentially missing gray literature or unpublished negative findings; Selection bias: industry reports from Stanford HAI, Elsevier, and Clarivate may reflect corporate perspectives on AI adoption; Geographic bias: data sources appear predominantly Western-centric despite claims to address Global North/South disparities; Language bias: analysis acknowledges models trained on English-language corpora but does not systematically disaggregate by language in all data presented; Temporal bias: rapid evolution of AI between 2021-2025 may render early data points obsolete by publication; Funding source bias: reliance on industry reports from major publishers (Elsevier, Wiley, Springer) which have vested interests in normalization of AI tools; Selection bias in curated industry reports and peer-reviewed articles; Potential publication bias favoring high-profile studies; Temporal lag in secondary data aggregation; Geographic bias in training data of AI systems discussed; Language bias (English-centric) in data sources; Funding source bias from industry reports (Elsevier, Clarivate); Bias in AI detection tools against non-native English speakers (documented as 'High' for GPTZero); Training data bias in LLMs (English-language, Western-centric corpora); Selection bias in secondary data sources (dependence on published/available reports); Potential underreporting of undeclared AI use due to incentive structures; Geographic representation bias (data primarily from Global North institutions)

Limitations

The authors note that "traditional primary data collection methods such as longitudinal surveys would likely yield obsolete results by the time of publication" and that their secondary data analysis approach, while appropriate for capturing rapid technological change, relies on aggregated data from divergent sources
The study acknowledges that "the reliance on these tools by universities to adjudicate academic misconduct cases is ethically suspect" and that "the 'don't ask, don't tell' environment where smart AI use is rewarded, but clumsy AI use is punished" suggests enforcement mechanisms are inadequate
No explicit discussion of limitations in representativeness across non-English publishing systems or developing academic ecosystems is provided.

Open questions raised

Longitudinal studies needed to track long-term impact of AI assistance on cognitive development of early-career researchers
Research into watermarking and provenance technologies to authenticate human origin of critical data without relying on biased textual analysis
Geopolitical dimension: how divergent AI adoption rates in China vs. West will reshape balance of scientific power
Deeper investigation into how AI shapes thought processes rather than binary 'is it AI?' question
Longitudinal studies tracking long-term impact of AI assistance on cognitive development of early-career researchers
Research into watermarking and provenance technologies for authenticating human origin of data

Data: Stanford HAI AI Index 2025 [21]; Elsevier Insights 2024 [1]; Clarivate Pulse of the Library 2024 report [8]; Stanford HAI AI Index 2025 (referenced as [21]); Elsevier Insights 2024 (referenced as [1]); Clarivate Pulse of the Library 2024 report (referenced as [8]); Additional peer-reviewed articles from information science, higher education, and ethics journals (2021-2025); Stanford HAI AI Index 2025; Elsevier Insights 2024; Clarivate Pulse of the Library 2024 reportCode: Not mentionedExtracted from: pdfAgreement 67%

Explore related topics

Hallucination Control Research Integrity Literature Review