Self-Admitted GenAI Usage in Open-Source Software

Finally, we examine the longitudinal impact of GenAI adoption on code churn in 151 repositories with self-admitted GenAI usage and find no general increase, contradicting popular narratives on the impact of GenAI on software development.

Repeats the main result that B4 states, and its 151 is the number of repositories analyzed.

We assess the longitudinal impact of GenAI adoption on software evolution using a repository-level analysis of code churn in 151 projects, showing that GenAI adoption does not lead to a general increase in churn and that effects are stronger for generation tasks.

Repeats the two main results that B4 and B5 state, and its 151 is the number of projects analyzed.

Our analysis revealed patterns of corrective actions following code generation.

Repeats the main result that B6 states, without a quantitative part of its own.

We cannot conclude that there is a general trend toward increasing code churn.

Repeats the main result that B4 states, without a number of its own.

Therefore, with our data and methodology, we cannot confirm this claim.

Repeats the main result that B4 states, as a verdict on the GitClear claim.

In summary, this task-level analysis reveals that GenAI-assisted generation tasks appear more rework-prone than other forms of assistance.

Repeats the main result that B5 states, without a number of its own.

Moreover, a majority of our survey participants agreed that the regulation of GenAI usage is necessary in open-source projects.

Repeats the count that R36 gives, in words, and no broad statement states the main result that it would serve.

Our analysis revealed that developers primarily use GenAI tools for code generation, natural language translation, and code refactoring.

Repeats the main result that B1 states, without the counts.

Contrary to claims [...] in the GitClear report, which was extensively discussed in the software development community [36, 37], we did not find an increasing code churn after GenAI adoption.

Repeats the main result that B4 states, without a number of its own.

A footnote is printed between the two halves of the sentence in text.txt at the page break.

The RQ1 results show that developers most often use GenAI for code generation, natural-language translation, and refactoring—with explicit acknowledgment as the dominant purpose.

Repeats the two main results that B1 and B3 state, without the counts.

Together with our RQ3 task-level analysis, which shows that generation tasks are more rework-prone than optimization or maintenance, this suggests that GenAI output has a provisional nature.

Repeats the main result that B5 states, as the ground for an implication.

The RQ3 results show no systematic increase in code churn after GenAI adoption, contradicting prominent industry claims.

Repeats the main result that B4 states, without a number of its own.

Another important purpose was regulation (see RQ2).

Adds a second purpose category to the purposes finding that B3 states, so it divides that result rather than stating one of its own.

Table 6 (page 7) gives the codes of the category Guidance and Best Practices as Set example 25, Exclusion of usage within the project 18, and Regulation of usage within the project 10; the sentence gives no number.

We did find 15 repositories with an overall upward trend in line-based code churn after the first GenAI mention. However, for most of them, the slope was negative.

Names the exception to the main result that B4 states, so it divides that result rather than stating one of its own.

The second sentence starts with "However" and refers to "them", so it is quoted together with the sentence before it. Table 9 (page 11) prints the parts of the sentence's 15 and no total: for line-based churn Upward with #Positive slope 5 (16.7%) and #Negative slope 10 (33.3%). The arithmetic is left to a later step.

In addition, we also found 15 repositories with a downward trend.

Names a subgroup of the repositories behind the main result that B4 states, so it divides that result rather than stating one of its own.

Table 9 (page 11) prints the parts of the sentence's 15 and no total: for line-based churn Downward with #Positive slope 3 (10.0%) and #Negative slope 12 (40.0%). The arithmetic is left to a later step.

Translation followed with 50 instances, while optimization and maintenance tasks accounted for 34 and 26 instances, respectively.

B1 would still stand, because it names code generation, natural language translation, and code refactoring, and maintenance is not one of the categories it names.

Maintenance tasks accounted for 26 instances.

split S1

Table 3 (page 5) prints the parts of the sentence's 26 and no total: Label revision 8, README revision 7, Document revision 4, Changelog revision 2, Prompt refinement 1, Color suggestion 1, Dependency upgrade 1, Version update 1, and Comment revision 1. The arithmetic is left to a later step. The sentence pairs its items with its numbers by "respectively", so the first number is optimization's and the second is maintenance's, which is how C2 and R1 divide it.

In other cases, changes were reverted (1), AI-generated comments were deleted (2), or the generated code was commented out (1).

B6 would still stand, because it names bug fixes, refactorings, reversions, and deletions, and commenting code out is not one of the actions it names.

In other cases, the generated code was commented out (1).

split S2

Among them, 13 referred to code that was initially generated using GenAI tools and then changed.

B6 would still stand, because this is the count for a subgroup of the 20 follow-up actions that C6 gives.

In addition to the 13 human actions that followed AI code generation that we discussed above, we found seven human actions following the generation of configuration and validation files or an unclear role of the GenAI tool.

B6 would still stand, because this is the count for the remaining subgroup of the 20 follow-up actions that C6 gives.

This manifested itself in several ways, such as offering guidance (53 mentions), flagging areas needing attention (23 mentions), and addressing GenAI limitations (4 mentions).

B3 would still stand, because it states that acknowledgment was a common purpose and these are the counts of the three other purpose categories.

Table 6 (page 7) prints only the per-code counts and no category totals: for Guidance and Best Practices 25, 18, and 10; for Quality Assurance 11, 10, and 2; for GenAI Limitations 3 and 1. The arithmetic is left to a later step.

Of the 151 repositories with self-admitted GenAI usage, 119 had a significant difference in file-based churn, and 120 had a significant difference in line-based churn (p < 0.05).

B4 would still stand, because its first sentence summarises the RDD trend analysis, whose counts C10 gives from Table 9, while this sentence reports the Wilcoxon signed-rank comparison of Table 8.

Table 8 (page 11) gives the same numbers in its Sum sig. column, 119 for file-based and 120 for line-based churn, against 32 and 31 not significant. The two analyses count significant repositories differently: Table 8 (page 11) gives 119 of 151 and 120 of 151 as significant under the Wilcoxon signed-rank test, and Table 9 (page 11) gives 26 of 149 and 30 of 149 as significant under the RDD analysis. The sentences following B4's first one, in the Summary RQ3 box on page 12, give 15 repositories with an upward trend and 15 with a downward trend, the numbers the RDD paragraph on page 12 reports.

Eleven repositories had an increasing file-based churn with a medium-to-large effect size, and 15 had an increasing line-based churn with a medium-to-large effect size.

B4 would still stand, because these are the counts for one effect-size subgroup of the repositories in Table 8.

Table 8 (page 11) prints the parts of both numbers and no combined medium-to-large row: for file-based churn the increasing counts are medium 9 and large 2, and for line-based churn medium 7 and large 8. The arithmetic is left to a later step.

A decreasing churn was more common: 49 repositories had a decreasing file-based churn with a medium-to-large effect size, and 49 repositories had a decreasing line-based churn with a medium-to-large effect size.

B4 would still stand, because these are the counts for one effect-size subgroup of the repositories in Table 8.

Table 8 (page 11) prints the parts of both numbers and no combined medium-to-large row: for file-based churn the decreasing counts are medium 23 and large 26, and for line-based churn medium 24 and large 25. The arithmetic is left to a later step.

For file-based churn, an overall upward trend with a negative slope after the cutoff date was most common (12 repositories).

B4 would still stand, because this is the count for one of the four RDD patterns rather than the result as a whole.

Table 9 (page 11) gives the same number for file-based churn, Upward trend with #Negative slope 12 (46.2%).

For line-based churn, an overall downward trend with a negative slope was most common (12 repositories). However, there were almost as many repositories (10) with an overall downward trend, but a positive slope after the cutoff date.

B4 would still stand, because these are the counts for two of the four RDD patterns rather than the result as a whole.

The text and Table 9 put the 10 repositories in different patterns. The sentence gives them as an overall downward trend with a positive slope after the cutoff date, while Table 9 (page 11) prints 10 (33.3%) for line-based churn under Upward trend with #Negative slope and prints 3 (10.0%) under Downward trend with #Positive slope. Table 9 gives the sentence's 12 as Downward trend with #Negative slope 12 (40.0%). The second sentence starts with "However" and is quoted with the sentence before it.

We found 15 repositories with a significant upward trend in file-based churn and 15 with a significant upward trend in line-based churn. However, most of them had a negative slope.

B4 would still stand, because these are the counts for the repositories that are the exception to it.

Table 9 (page 11) prints the parts of both numbers and no trend totals: for file-based churn Upward with #Positive 3 (11.5%) and #Negative 12 (46.2%), and for line-based churn Upward with #Positive 5 (16.7%) and #Negative 10 (33.3%). The arithmetic is left to a later step. The second sentence refers to "them" and is quoted with the sentence before it.

In addition, 11 projects had a significant downward trend in file-based churn and 15 had a significant downward trend in line-based churn.

B4 would still stand, because these are the counts for one of the trend directions rather than the result as a whole.

Table 9 (page 11) prints the parts of both numbers and no trend totals: for file-based churn Downward with #Positive 4 (15.4%) and #Negative 7 (26.9%), and for line-based churn Downward with #Positive 3 (10.0%) and #Negative 12 (40.0%). The arithmetic is left to a later step.

In addition to code generation, we found that internationalization and natural language translation are common use cases for LLMs in open-source software projects (49 instances, see Table 3).

B1 would still stand, because this gives the count for the natural-language part of the translation category whose count C2 gives.

The text gives the translation count differently in two places: this sentence gives 49 instances (page 13) and the results sentence of C2 gives 50 instances (page 5). Table 3 (page 5) prints Translation as Text 49 and Code 1.

However, our RDD analysis revealed that three repositories (3/26, 11.5%) had a significant upward file-based code churn trend with a positive slope (p < 0.05).

B4 would still stand, because this is the count for one of the four RDD patterns and the paper presents it as the exception to that result.

Table 9 (page 11) gives the same number for file-based churn, Upward trend with #Positive slope 3 (11.5%).

To contextualize this finding, we note that a larger proportion of repositories (12/26, 46.2%) showed an upward trend with a negative slope, i.e., churn increased, but the increase decelerated over time.

B4 would still stand, because this is the count for one of the four RDD patterns rather than the result as a whole.

Table 9 (page 11) gives the same number for file-based churn, Upward trend with #Negative slope 12 (46.2%).

In addition, 26.9% (7/26) showed downward trends.

B4 would still stand, because this is the count for one of the four RDD patterns rather than the result as a whole.

The text and Table 9 give the share of downward trends differently. The sentence gives 26.9% (7/26) for downward trends, while Table 9 (page 11) prints file-based Downward as #Positive slope 4 (15.4%) and #Negative slope 7 (26.9%), so the sentence's figure matches the negative-slope row alone.

Using this concept as a lens to study how GenAI tools are integrated into OSS projects, we analyze a curated sample of more than 200,000 GitHub repositories, identifying 1,292 such self-admissions across 156 repositories in commit messages, code comments, and project documentation.

Every number in this sentence describes the study: the size of the repository sample, the number of mentions identified, and the number of repositories they come from.

Using a mixed methods approach, we derive a taxonomy of 32 tasks, 10 content types, and 11 purposes associated with GenAI usage based on 1,292 qualitatively coded mentions.

Every number in this sentence describes the study: the sizes of the code book and the number of coded mentions. The taxonomy itself is a qualitative finding.

We then analyze 13 documents with policies and usage guidelines for GenAI tools and conduct a developer survey to uncover the ethical, legal, and practical concerns behind them.

Its only number, the 13 documents analyzed, describes the material of the study, and the concerns it reports are a qualitative finding.

We introduce self-admitted GenAI usage as an empirical lens for studying GenAI adoption in open-source software and curate a dataset of 1,292 self-admitted GenAI usages across 156 GitHub repositories.

Every number in this sentence describes the study: the size of the curated dataset and the number of repositories it covers.

Using a mixed-methods approach, we derive a taxonomy of GenAI usage comprising 32 development tasks, 10 content types, and 11 purposes, grounded in a qualitative analysis of the identified usages.

Every number in this sentence is a size of the code book, which describes the study, and the taxonomy itself is a qualitative finding.

We empirically analyze how open-source projects govern GenAI usage by examining 13 policies and usage guidelines, contextualized through a developer survey.

Its only number, the 13 policies examined, describes the material of the study, and the sentence otherwise only announces an analysis.

Our initial sample of GitHub projects contained 207,062 repositories distributed across Python (77,542), JavaScript (48,500), TypeScript (37,424), Java (25,160), and C# (18,436).

Gives the size of the initial repository sample, split by language, which describes the study.

Figure 1 (page 3) reports the same size as 207,062 repositories in step 1 of the data collection process.

Our final sample of GitHub repositories, obtained in February 2024, contained 14,785 GitHub repositories distributed across Java (5,060), C# (3,544), TypeScript (2,464), Python (1,875), and JavaScript (1,842).

Gives the size of the repository sample the results rest on, split by language, which describes the study.

Table 1 (page 3) gives the same size in its caption as n = 14,785, and Figure 1 (page 3) reports 14,785 repositories in step 2.

Running the above retrieval process on all repositories yielded 3,004 mentions of GenAI tools: 1,572 in commit messages, 397 in source code comments, and 1,035 in documentation files.

Gives the number of mentions retrieved, split by artifact kind, which describes the material of the study.

Figure 1 (page 3) reports the same total as 3,004 mentions in step 3 and gives no split by artifact kind.

We calculated a sample size of 341 mentions (of 3,004) to achieve estimates with a 95% confidence level and a 5% confidence interval.

Gives the size of the sample that the two authors labeled, which describes the study.

The inspection resulted in disagreement between the two authors for only 14 cases (4% of the sample).

Counts the disagreements in the coding procedure, which describes the study rather than a result.

The resulting 95% confidence interval was estimated to be (0.87, 0.95), indicating an “almost perfect” agreement.

Reports an inter-rater agreement score that supports the coding procedure, which describes the study.

In total, we identified 1,292 true-positive mentions of GenAI tools that were aligned with our inclusion criteria. We found true-positive mentions in 156 repositories (11 Python, 12 JavaScript, 37 TypeScript, 47 C#, and 49 Java repositories).

Gives the size of the coded material and the number of repositories it comes from, split by language, which describes the study.

The two sentences are quoted together because the split by language belongs to the size that the first sentence states. Figure 1 (page 3) reports the same total as 1,292 usages in step 4.

A preliminary analysis identified 1,008 mentions related to Copilot for Pull Requests.

Describes the composition of the coded material and which mentions were deferred during coding, not a result.

The quote leaves out the footnote marker attached to the last word.

To build the code book, two authors independently analyzed 284 GenAI mentions.

Gives the size of the material used to build the code book, which describes the study.

The caption of Table 3 (page 5) gives the same number in n = 284 + 1,000 + 8 = 1,292.

The analysis yielded “substantial” to “almost perfect” agreement levels on task (k = 0.81 − 0.89), content (k = 0.95 − 0.99), and purpose (k = 0.79 − 0.92), according to standard guidelines for interpreting k [20].

Reports inter-rater agreement scores that support the coding procedure, which describes the study.

Our analysis identified 32 distinct task categories in which developers use GenAI tools in their workflows.

Gives the number of codes in the task code book, which describes the study.

Our analysis identified three main categories of AI-generated content in open-source software projects, organizing ten dis[...]tinct codes (see Table 5).

Gives the number of categories and codes in the content code book, which describes the study.

The page break falls inside the word "distinct" and Tables 5 and 6 are printed between its two halves in text.txt.

Of the 1,009 instances of code PR description in Table 3, 1,000 originated from a single repository and are discussed separately above.

Describes how the coded material is distributed over repositories, which describes the study.

Table 3 (page 5) gives PR description as 1,009, the same number as the sentence.

Our analysis identified 11 different purposes for GenAI mentions in software projects, grouped into four main categories (see Table 6).

Gives the number of codes and categories in the purpose code book, which describes the study.

Using our sample of GenAI mentions, we found 28 mentions related to policies and usage guidelines around GenAI tool usage.

Gives the number of mentions that the RQ2 analysis rests on, which describes the study.

Table 7 presents detailed examples drawn from 13 documentation files and commit messages in 12 GitHub repositories, where the last column indicates the number of mentions identified in the software artifact.

Gives the number of policy documents and repositories that the RQ2 analysis rests on, which describes the study.

Table 7 (page 9) lists thirteen policies, P1 to P13, and their last column gives the per-artifact mention counts 1, 1, 1, 9, 1, 1, 1, 1, 4, 1, 1, 1, and 5.

We were able to determine the email addresses of 30 contributors. In total, we received eight survey responses, which we analyzed using a combination of open coding and card sorting.

Gives the number of contributors contacted and the number of survey responses analyzed, which describes the study.

The two sentences are quoted together because the response count is read against the number contacted.

Five developers highlighted the necessity of regulating the usage of GenAI tools in software projects.

No broad statement states the main result that it would serve: the survey findings of RQ2 are qualitative and the record has no broad statement for them.

The discussion builds on this count without a number (page 13, "Moreover, a majority of our survey participants agreed that the regulation of GenAI usage is necessary in open-source projects."), and the boxed answer for RQ2 (page 10) states only the kinds of policy found.

We found 13 policies and guidelines on GenAI usage in open-source software projects, including strict policies prohibiting GenAI usage, policies requiring attribution, but also guidelines encouraging contributors to use GenAI, for example, for translating natural language text.

Its only number, the 13 policies analyzed, describes the material of the study, and the kinds of policy it lists are a qualitative finding.

We included 151 repositories with true positive GenAI mentions that did not prohibit the use of GenAI tools.

Gives the number of repositories that the RQ3 analysis rests on, which describes the study.

The caption of Table 8 (page 11) gives the same size as n = 151, and Figure 1 (page 3) reports 151 repositories for the RDD analysis in step 4.

For the 149 repositories included, we identified four patterns, that we describe in the following:

Gives the number of repositories that the RDD analysis rests on, which describes the study.

The caption of Table 9 (page 11) gives the same size as n = 149.

They found that approximately 54% of the generated code lines were integrated and only 2.5% of the generated snippets were later modified.

Reports the results of cited work, which the paper sets beside its own churn result, not a result of this study.

The 2025 follow-up report [52], based on an expanded dataset of 211 million lines through 2024, reported a rise in code churn from 4.5% in 2023 to 5.7% in 2024, a 39.9% drop in refactoring, and a 17.1% increase in copy-pasted code.

Reports the results of cited work, which the paper sets beside its own churn result, not a result of this study.

In our mixed-methods study design, we first mined more than 200,000 GitHub repositories, isolating 1,292 true-positive GenAI mentions across 156 projects.

Every number in this sentence describes the study: the size of the repository sample, the number of mentions, and the number of projects.

Qualitative open coding of these instances and subsequent card sorting yielded taxonomies of 32 assisted tasks, 10 content types, and 11 usage purposes.

Every number in this sentence is a size of the code book, which describes the study, and the taxonomies themselves are a qualitative finding.

We complemented this content analysis with a survey of project contributors and a systematic review of 13 project-level policies and guidelines.

Its only number, the 13 policies reviewed, describes the material of the study.

In addition, we performed a regression-discontinuity (RDD) analysis of code churn in the 149 repositories that contained sufficient data to study the impact of GenAI adoption on open source software projects.

Its only number, the 149 repositories, is the size of the material the RDD analysis rests on, which describes the study.

While acknowledgment of GenAI usage is common (RQ1), contextual metadata such as prompts, model versions, or scope of generation is rarely recorded.

Cannot tell what the frequency word "rarely" rests on, so it is recorded here for the checker.

The paper gives no count for how often the generation context is documented. The discussion states the same in words (page 13, "A more standardized approach for documenting GenAI contributions is required, since most self-admitted GenAI usages did not document the generation context beyond brief summaries."), and the codes of Table 6 (page 7) count purposes, not whether a mention records prompts, model versions, or the scope of generation.