Generative Artificial Intelligence Policies under the Microscope

Since the rise of ChatGPT, generative artificial intelligence (GenAI) technologies gained widespread popularity, impacting academic research and everyday communication.^5,10 While GenAI offers benefits in task automation,⁹ it can also be misused and abused in nefarious applications,⁷ with significant risks to long-tail populations.⁶ Professionals in fields such as journalism and law still remain cautious due to concerns related to hallucinations and ethical issues, but scholars in computer science (CS), the field where GenAI originated, appear to be cautiously, yet actively exploring its use. For instance, Liang, W. et al.³ report the increased use of large language models (LLMs) in the CS scholarly articles (up to 17.5%), compared to mathematics articles (up to 6.3%), and Liang, W. et al.² report that, between 6.5% and 16.9% of peer reviews at ICLR 2024, NeurIPS 2023, CoRL 2023, and EMNLP 2023 may have been altered by LLMs beyond minor revisions.

Considering researchers’ increasing adoption of GenAI, it is crucial to establish usage policies to promote fair and ethical practices in scholarly writing and peer reviews. Some conferences, such as ICML 2023 accurately highlighted the confusion and questions surrounding GenAI use, with concerns regarding the novelty and ownership of generated content. Previous research examined GenAI policies by major publishers such as Elsevier and Springer,⁵ but there is still a lack of clear understanding of how CS conferences are adapting to this paradigm shift. Hence, this Opinion column provides a summarized overview of the “scholarly writing” policies of CS conferences and major computing societies: ACM, IEEE, and AAAI (2024 and 2025 if available; otherwise 2023 and 2024) and offers recommendations. Conferences studied in this analysis were selected based on CSRankings,^a including leading conferences in each CS subfield. Our analysis shows that many CS conferences have not established GenAI policies and those that have, vary in leniency, disclosure, and sanctions. Policies for authors are more prevalent compared to reviewers, and some address code writing and documentation. These policies are evolving, as demonstrated by conferences such as ICML 2023, which initially prohibited LLM-generated text but later clarified its allowance for editing author-written content. However, by and large, adoption remains inconsistent across conferences, creating uncertainty in their application.

Table 1 lists the conferences considered, alongside their leniency ratings on a 5-point Likert scale. Three authors independently annotated the policies after agreeing on rating criteria. Given the ordinal nature of the ratings, we used Krippendorff’s alpha (α = 0.832, satisfactory) to assess interrater reliability. Final ratings were assigned through majority voting. Nevertheless, we acknowledge that variations in ratings may arise from the subjective nature of the task, largely attributable to the ambiguity in policy language.

Most conferences with GenAI policies were somewhat lenient for authors (ratings of ‘3’ or ‘4’). The 2025 author policy for WWW was assigned a rating of ‘3’ as it allowed rephrasing and ICRA was assigned ‘4’ as it allowed different types of usage with disclosure. Interdisciplinary conferences such as UIST and VR had lower leniency ratings of ‘1’ (‘no AI-generated content allowed) and ‘2’ (for example, allowing grammar/spelling edits), respectively. In contrast, several AI conferences, including CVPR, were highly permissive of GenAI use with rating ‘5’ (for example, no restrictions), likely due to a greater familiarity of the AI field with LLMs and their perceived benefits. In comparison, conferences adopted more restrictive policies for reviewers. Particularly, ICRA, VIS, and IMC were rated ‘1’, ‘2’, and ‘3’, respectively, likely stemming from concerns over content leakage and limitations of LLMs in handling complex tasks.¹

Area-Level Trends

We compared the trend of GenAI policy adoption across different CS areas, following the classification from CSRankings: AI, Interdisciplinary, Systems, and Theory. For AI, 61.5% of conferences had GenAI policies for authors in Year 1, which increased to 92.3% in Year 2 (Figure 1a), indicating that conferences in AI field are the most active in adopting GenAI policies for authors. Conferences in the Interdisciplinary area showed 26.7% in Year 1 and 60% in Year 2, making them the second most proactive in introducing GenAI policies for authors. While these two areas are leading in GenAI policy adoption for authors, the Systems area lagged behind the overall trend in Year 1 (20.7%) but saw a 13.8% increase in Year 2 (34.5%), implying that this area is gradually aligning with the overall trend. A similar pattern was observed for reviewers, with AI (Year 1: 7.7%, Year 2: 30.8%) and Interdisciplinary (Year 1: 6.7%, Year 2: 40%) leading GenAI policy adoption, while the Systems area (Year 1: 3.4%, Year 2: 3.4%) is gradually converging with the broader trend. In contrast, no conferences in the Theory area had GenAI policies for authors or reviewers. This may be due to the conservative nature of the area or lack of active GenAI usage in writing theory articles.

In terms of author policies, AI area shows higher leniency ratings compared to the overall average (see Figure 2a). Additionally, the Systems area shows very high leniency ratings (Year 1: 4.00, Year 2: 3.7) compared to overall average (Year 1: 3.5, Year 2: 3.61). In contrast, the leniency ratings in the Interdisciplinary area are relatively lower (Year 1: 2.50, Year 2: 3.33). This may imply that, despite being more proactive in adopting policies than Systems, Interdisciplinary area tends to be more cautious, imposing stricter guidelines on the use of GenAI.

While conferences with GenAI policies tend to be more lenient toward authors, with an average leniency rating of 3.50 in Year 1 and 3.61 in Year 2—both above 3—the policies for reviewers are notably more restrictive (see Figure 2b). The average leniency rating for reviewers dropped from 3.00 in Year 1 to 2.18 in Year 2, falling below 3, and no conferences granted reviewers a leniency rating above 3 (see Table 1). This likely reflects concerns about the potential risks of using AI tools in the peer review process, such as the possibility of leaking sensitive or unpublished information and inadequate capability of LLMs for performing tasks of a high level of expertise and nuanced judgment.¹ Since reviewers have access to confidential work, conferences may be more cautious to ensure data security, privacy, and the protection of intellectual property.

Temporal Trends

We examined the GenAI policies for two consecutive years, 2024 and 2025, and when unavailable, 2023 and 2024. Overall, conferences are moving toward introducing new GenAI policies. In Year 1, only 18 (28.1%) out of 64 conferences had GenAI policies, either for authors or reviewers, which increased to 32 (50%) in Year 2. Specifically, for authors (see Figure 1a), only 18 (28.1%) had GenAI policies in Year 1, increasing to 31 (48.4%) in Year 2, a 20.3 percentage point increase. In contrast, for reviewers (see Figure 1b), only 3 (4.7%) conferences had GenAI policies in Year 1, rising to 11 (17.2%) in Year 2. While many conferences are increasingly adopting and publicly sharing GenAI policies for authors, they are slower in providing clear guidelines for reviewers. Conferences may be unaware of reviewers’ needs or could be implementing GenAI policies for reviewers internally without making them publicly available. In either case, there is a noticeable disparity in how GenAI policies are adopted and communicated for authors versus reviewers.

Several conferences adopted GenAI policies during the latter year, including SIGMOD and EMNLP for authors, and SIGCSE and VIS for authors and reviewers. Once a policy was established, similar degree of leniency was maintained for the latter year, with a few exceptions (for example, SC shifted to a more restrictive policy in 2024). These findings suggest a cautious, evolving approach to GenAI policies, with some conferences adopting clearer guidelines with increasing GenAI use.

Society Versus Conference-Level Trends

We reviewed the GenAI policies of ACM, IEEE, and AAAI, as many CS conferences are affiliated with these societies. While conference-specific policies differ in disclosure requirements and flexibility, society-level policies permit authors’ GenAI use with disclosures. ACM and AAAI also permit reviewers to enhance their reviews using GenAI, as long as the submissions remain unexposed to these systems. In contrast, IEEE has no policies for conference reviewers. None of these policies mention sanctions for authors or reviewers. As Table 1 summarizes, some conferences do not have conference-specific GenAI policies for authors and refer to the society-level policies—for example, SIGCSE, VIS, IMC, while others have established their own GenAI policies for authors, including CVPR (IEEE) and UIST (ACM). Still, many conferences appear unaware of society-level GenAI policies, as the conference websites do not mention them. For reviewers, few conferences adhere to society-level policies (for example, IMC), while many follow their own (for example, CVPR).

Recommendations

Among the 64 conferences analyzed, 32 (50%) adopted GenAI policies for authors or reviewers in two years, but only 11 (17.18%) addressed reviewers. Theory and Systems fields lagged behind, with no Theory and 34.5% Systems conferences adopting policies for authors by Year 2, compared to 92.3% in AI and 60% in Interdisciplinary fields. Notably, 14 of 32 policies emerged in Year 2, and no conference allowed GenAI authorship. Many ACM and IEEE conferences appear unaware of society-level policies, showing inconsistent adoption. Policies also lack sanctions for non-compliance, which is essential to enforce rules and prevent misuse.

A notable gap exists in reviewers’ GenAI policies, perhaps stemming from concerns about exposing sensitive information. Yet, this risk can be mitigated by configuring LLM settings to ensure models neither retain nor learn from user-provided data.⁴ LLMs, for example, OpenAI’s ChatGPT and Google’s Gemini, allow users to opt out of model training, providing added safeguards.⁴ Moreover, LLMs provide enterprise versions with data security (for example, ChatGPT^b and Anthropic’s Claude 3^c) or accommodate privacy requests.^d With the rise of GenAI, conferences without policies should establish guidelines for authors and reviewers, including evaluating AI use in submissions. In addition, GenAI policies at the area or society level would ensure consistency, benefiting both authors and reviewers, especially for resubmissions. Moreover, implementing professional development initiatives, training programs, and balanced enforcement strategies can promote responsible AI use.⁴

GenAI is a transformative technology that enhances efficiency across disciplines, and fields that fail to adapt, risk falling behind. We recommend lenient use of GenAI in conferences, especially to support non-native English speakers. GenAI should enhance, not alter, the author’s work,⁴ while reviewers’ judgments must be independent. Both authors and reviewers should disclose GenAI use transparently and take full responsibility. Ultimately, scholars must verify content for accuracy and ethical compliance, as GenAI tools are imperfect and prone to hallucinations.⁸