AI for Academic Surveys: Ethics & Detection

• AI for Academic Survey is the exploration of methods, systems, and ethical frameworks used in integrating AI into manuscript preparation.
• The study distinguishes between mere grammar correction and substantive rewriting, highlighting varying disclosure requirements and ethical perceptions.
• Technical AI-detection methods are applied to assess manuscript revisions, ensuring alignment between emerging policies and academic integrity.

AI for academic survey refers to the set of methods, systems, perceptions, and ethical frameworks surrounding the use and detection of AI—particularly LLMs such as ChatGPT and Bard—in the preparation, revision, and assessment of scholarly manuscripts for academic journals. This domain concerns both voluntary and mandated disclosure practices, the ability of technical systems to detect AI-mediated writing, and the evolving landscape of academic integrity, institutional policy, and scholarly communication.

1. Perceptions of AI Use in Manuscript Preparation

Empirical evidence demonstrates that academic researchers differentiate sharply between types of AI assistance. A survey of academic economists found that only 22% of respondents believed that using a tool like ChatGPT for grammar correction necessitates disclosure, whereas 52% viewed AI-mediated “rewriting” as warranting explicit reporting. The underlying belief is that grammar corrections are a low-level, mechanical intervention, while text rewriting constitutes a substantive modification of author contribution.

Subgroup differences are substantial: native English speakers are generally more likely than non-natives to support disclosure, especially around rewriting, and early-career researchers (students or postdocs) show greater support for disclosure than professors. These differences are closely linked to ethical stances; individuals regarding AI assistance as “unethical” are about three times more likely to endorse reporting its use (regression coefficient β₃ ≈ 0.557 for an "Unethical" dummy variable, highly significant in OLS analysis).

This divergence in perception underscores the centrality of ethical frameworks and academic norms in shaping attitudes toward AI deployment in the scholarly review pipeline (Chemaya et al., 2023).

2. Disclosure Practices and Policy Developments

Patterns of voluntary disclosure closely follow perceptions. While only 22% favored disclosure for grammar correction, 52% supported acknowledgment for AI rewriting. Comparative analysis with research assistant (RA) help finds near parity between ChatGPT and RA assistance in reporting preferences. Traditional tools (Grammarly, Microsoft Word) receive less scrutiny for disclosure—rates as low as 5–14% for such tools—implying that familiarity or “conventionality” shapes the perceived threshold for acknowledgment.

Publishers and conferences are actively responding. For instance, Elsevier requires authors to disclose AI use for improving the language and readability of manuscripts, enforced via a dedicated disclosure section. Appendix B in (Chemaya et al., 2023) lists a mix of mandatory and voluntary institutional policies, revealing an ongoing transition toward codification. The field is thus witnessing the emergence of both voluntary norms and formalized mandates, with major journals and conferences moving toward increased transparency requirements.

3. Technical Detection of AI-generated Content

The use of LLMs in academic writing has catalyzed the development of AI-detection services. For quantitative assessment, tools such as Originality.ai assign a content-level “AI score” from 0 to 100%. Controlled experiments show that abstracts rewritten or grammar-fixed by GPT-3.5 consistently yield higher AI scores than unaltered originals, with rewriting generating the largest rightward shift in score distributions (as seen in the paper’s histograms and Figure 1). However, there is overlap at the margins: 24.2% of cases showed grammar correction leading to higher AI scores than full rewriting, pointing to the stochasticity and context-sensitivity of detection.

The efficacy of these services remains holistically robust—uncorrected abstracts have low scores, while AI-revised content is correctly flagged. Yet, minor changes in prompt design (such as the granularity of “one paragraph” instructions) can nontrivially shift AI scores, indicating sensitivity to often-invisible instruction engineering. Robustness was supported via statistical checks using 75th/90th percentiles and test–retest reliability.

Despite high accuracy in the disambiguation of unedited versus AI-revised text, current detectors sometimes fail to reflect academic consensus on what constitutes a meaningful level of AI intervention. As such, technical identification and scholarly norms are partially but not fully aligned (Chemaya et al., 2023).

4. Implications for Academic Integrity

The integration of AI tools into manuscript preparation introduces both positive and negative implications for academic integrity. On the benefits side, AI can streamline writing for non-native speakers, enhance clarity, and allow greater focus on research innovation. On the risk side, several major threats are identified:

• Introduction of material errors, such as incorrect mathematics or hallucinated references, which may pass unnoticed if overrelied upon.
• Increased risk of plagiarism, as AIs generate text based on a vast corpus of existing works, sometimes inadvertently mimicking or copying prior content.
• Black-box opacity, where the user cannot fully audit the text creation process, potentially undermining scholarly accountability and transparency.
• Variable perceptions of ethicality: Only 6% of respondents regarded grammar correction as unethical, but 39% held using ChatGPT for rewriting to this standard, reflecting deep divides within the academic community.

This landscape calls for an ongoing dialogue on academic standards, with statistical models showing ethical perceptions as the single strongest predictor of support for disclosure (see regression model: Report = β₀ + β₁ Native + β₂ Professor + β₃ Unethical + ε, with β₃ highly significant).

5. Challenges, Limitations, and the Evolution of Policy and Detection

Several institutional, technical, and epistemological challenges are identified:

• Policy is often lagging behind practice; less than half of surveyed academics were aware of formal institutional rules regarding AI use.
• There is a “cat-and-mouse” dynamic between AI developers and detectors: prompt engineering, model upgrades, and use of adjacent paraphrasing tools can compromise detection reliability. Thus, the detection landscape is inherently unstable and subject to rapid innovation.
• Current studies are heavily concentrated in economics and related disciplines; the generalizability of attitudes and practices to other scientific domains remains uncertain and is an important avenue for future empirical work.

Looking forward, the expected trajectory includes:

• Institution of more granular disclosure policies that differentiate between minor (e.g., grammar) and major (e.g., rewriting or novel content generation) uses of AI.
• Increasing demands for transparency of both AI use and even the specific prompts employed, to facilitate peer review scrutiny.
• Continued improvement and adaptation of detection services, as models like GPT-4 and successors further blur the boundaries between human and machine authorship.
• Calls for fresh research—especially qualitative, vignette-based studies—to elucidate the underlying sources of ethical concern regarding different categories of AI assistance.

6. Synthesis and Outlook

AI for academic survey and manuscript preparation presents a landscape of rapid cultural, technical, and ethical transformation. Differential attitudes are primarily driven by the perceived substance and ethicality of AI contributions, with grammar fixes largely accepted and rewriting triggering substantial concern and demands for disclosure. While detection services reliably flag AI-altered texts in a statistical sense, they are imperfect proxies for academic consensus on what constitutes problematic AI involvement.

Both detection and policy are poised for further evolution. More nuanced, context-aware guidelines and adaptive detector thresholds are needed to align technical capacity with normative expectations. The long-term credibility of scientific literature will require not just effective detection technology, but also institutionalized standards, explicit reporting requirements, and a robust ecosystem of ethical deliberation and community dialogue (Chemaya et al., 2023).