AI evidence extraction

Prompt injection in manuscripts: exploiting loopholes or crossing ethical lines?

Shuchen Tang, Zilong Li · Research Integrity and Peer Review · 2026

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

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.1186/s41073-025-00187-7

Methodology & findings

Study design

Mixed methods approach combining: (1) Qualitative content analysis of academic integrity codes, institutional ethics guidelines, and policy documents from prominent research institutions and publishers using deductive coding; (2) Quantitative questionnaire survey with stratified sampling of authors, peer reviewers, and journal editors (n=194 after excluding neutral responses) using Likert-scale questions and open-ended items, with randomized design comparing ethical judgments when AI involvement was disclosed versus undisclosed..

Sample

N = 194, 12 groups

Primary method

Descriptive statistics for Likert-scale responses to identify trends in the data; thematic analysis of open-ended responses to explore motivations and rationales behind ethical judgments. Randomized design comparing ethical judgments across disclosure conditions.

Main result

The study found that transparency plays a critical role in ethical judgments regarding prompt injection. When AI involvement was explicitly disclosed, "72% of respondents classified Prompt Injection in Manuscripts as a form of misconduct," whereas "only 9% of participants viewed Prompt Injection in Manuscripts as misconduct" when AI involvement was not disclosed. Additionally, "80% of respondents expressed strong support for greater transparency in AI use," highlighting widespread recognition of transparency's importance for maintaining academic integrity.

Reports effect sizes.

Research paradigm

Mixed methods (qualitative content analysis and quantitative survey)

Author conclusions

"The integration of AI technologies into academic publishing, particularly within the peer review process, offers substantial benefits in efficiency and rigor. However, it also introduces significant ethical challenges, with prompt injection emerging as a major concern... A critical aspect of addressing AI-related misconduct is transparency. Our study found that when AI involvement in the review process is explicitly disclosed, prompt injection is widely viewed as unethical. However, the lack of transparency-when AI's role is not clearly stated-creates ambiguity, making it harder to judge the ethical implications of AI-assisted peer review. This underscores the urgent need for clear, standardized policies that define AI-related misconduct and mandate transparency in AI use throughout the academic process."

Risk of bias

Self-selection bias in survey respondents (acknowledged by authors); Potential non-response bias from academics less concerned with AI ethics; Survey respondents may have been more aware of AI issues than general population; Scope limited to text-based manipulations, excluding multimodal vulnerabilities; Exclusion of neutral (3-point) Likert responses from analysis may introduce bias; Self-selection bias: respondents aware of or concerned about AI issues more likely to participate; Sample limited to text-based prompt injections only; Potential response bias in self-reported ethical judgments; Stratified sampling may not fully represent broader academic community; Self-selection bias in survey respondents; Potential bias toward participants more aware of or concerned about AI issues; Sample may not represent broader academic community; Limited scope to text-based prompt injections only; Survey design with randomized disclosure condition may introduce social desirability bias

Limitations

The study acknowledges several limitations: "First, the self-selected sample may have introduced bias, as respondents who were more aware of or concerned about AI-related issues were more likely to participate
As such, the findings may not fully represent the broader academic community
Additionally, the study focused on text-based prompt injections within the context of peer review, leaving multimodal AI vulnerabilities-such as those in image or data-based submissions-untested."

Open questions raised

Near-total absence of systematic empirical research on prompt injection in manuscripts (acknowledged as central research gap)
Lack of understanding of multimodal AI vulnerabilities beyond text-based injections
Need for further empirical research to develop and test enforceable safeguards balancing AI efficiency with ethical integrity
Need for experimental research on AI transparency and auditing mechanisms
Gap in understanding ethical justifications for prompt injection in non-disclosed AI scenarios
Insufficient exploration of cultural and disciplinary differences in ethical judgments

Extracted from: pdfAgreement 69%

Explore related topics

AI Peer Review Peer Review Mixed Methods