“An Endless Stream of AI Slop”: How Developers Discuss the Burden of AI-Assisted Software Development

We analyzed 1,154 posts across 15 documents, developing a codebook of 15 codes organized into three thematic clusters: Review Friction, Quality Degradation, and Forces and Consequences.

Every number in this summary sentence describes the study: corpus size (1,154 posts, 15 documents) and codebook size (15 codes, three clusters). It repeats the abstract sentence E1 and adds the number of documents.

Our analysis of developer discourse about AI slop identified 15 codes in three thematic clusters.

Every number in this summary sentence describes the study: the size of the codebook (15 codes, three clusters). It repeats the codebook size that E1 states and gives no quantitative result.

We qualitatively analyzed how developers discuss AI slop in 1,154 Reddit and Hacker News posts, developing a codebook of 15 codes organized into three thematic clusters: Review Friction (how AI slop burdens reviewers, erodes trust, and prompts countermeasures), Quality Degradation (damage to codebases, knowledge resources, and developer competence), and Forces and Consequences (systemic incentives, mandated adoption, craft erosion, and workforce disruption).

Every number in this summary sentence describes the study: the size of the corpus (1,154 posts) and the size of the codebook (15 codes, three clusters). The sentence states no quantitative result.

The clusters and their content are a qualitative main result; the paper states no quantitative main result, so the record has no main result for it.

This yielded 16 documents: 13 Reddit threads (R01, R02, R03, R04, R05, R06, [...] R07, R08, R09, R10, R11, R12, R13) and three Hacker News result pages (H01, H02, H03).

Gives the size of the collected material, which describes the study. Recorded from the method because it splits the documents into kinds (13 Reddit threads, three Hacker News result pages), which no summary sentence gives.

The footnote lines 11 and 12 run between two lines of this sentence in text.txt, marked with [...]. The next two sentences on the same page say that R09 was excluded and that 15 documents (1,154 posts) remained, a size that E1 and E2 already state.

The final codebook consists of 15 codes: one rhetorical code (sarcastic-skepticism) and 14 topical codes.

The size of the codebook describes the study. Recorded from the method because it splits the 15 codes into kinds (one rhetorical, 14 topical), which no summary sentence gives.

Of all 1,154 posts, 978 (84.7%) received at least one code, yielding 1,603 codings.

Says how much of the corpus the study coded, which describes the study. The paper does not compare the share across groups or over time or relate it to anything else.

The caption of Fig. 2 on page 4 gives the same number of coded posts: “Code frequency distribution across 978 coded posts.”

On average, these posts received 1.6 codes, showing that developers often address multiple themes in one post.

Every main result of the paper is qualitative, so no main result rests on this count from the coding.

The sentence draws a finding from the coding density of the annotated corpus. The abstract, the introduction, and the conclusion state only qualitative main results, and none of them mentions how many codes a post received.

The three most frequent topical codes are structural-drivers (256, or 26.2% of coded posts), ai-limitations (227), and slop-mitigations (226), which together account for 44.2% of all 1,603 codings.

Every main result of the paper is qualitative, so no main result rests on these counts from the coding.

Fig. 2 on page 4 prints the same counts and adds the shares the sentence leaves out: structural-drivers 256 (26.2%), ai-limitations 227 (23.2%), slop-mitigations 226 (23.1%). Its caption gives 978 coded posts as the base of the shares, while the sentence gives its combined share as a share of the 1,603 codings.

The rhetorical code sarcastic-skepticism (155) ranks fourth, indicating that ironic framing was common in this discourse.

Every main result of the paper is qualitative, so no main result rests on this count from the coding.

Fig. 2 on page 4 prints sarcastic-skepticism as 155 (15.8%), below slop-mitigations at 226 (23.1%). The extraction broke the code name across two lines at its hyphen, and the quote writes it as the one word sarcastic-skepticism.

Independent evidence is consistent with these concerns: A large-scale comparison found that AI-generated code carries more high-risk security vulnerabilities than human-written code [15], and a survey of AI programming assistants found that developers value them for speed but struggle to get output that is correct and controllable [7].

Reports the results of cited work [15] and [7], not a result of this study. Recorded because the discussion sets them beside this paper's own findings.