article · 2026
Easier to Write, Harder to Read: AI-Assisted Authorship and Writing-Quality Decline in Social Work Conference Abstracts
SSRN Electronic Journal
School of Social Work bio: Brian E. Perron
Abstract
Objective. This study conceptually replicates Gartenberg et al. (2026), extending their methodology to a different scholarly community (social work), a different submission venue (conference abstracts), and a different family of AI-text detection models. Method. A corpus of 21,569 scientific-format abstracts submitted to the Society for Social Work and Research (SSWR) for conference years 2010–2026 was scored using three open-weight detectors: EditLens RoBERTa-large (primary), EditLens Llama-3.2-3B (within-family secondary), and academic-tuned desklib (cross-family). Classifications were anchored on the 95th percentile of each detector's 2010–2015 (no-LLM-exposure) baseline. A segmented regression tested for a step change at the 2025 conference cycle, the first cycle in which authors had full LLM availability across the entire writing window. Results. All three detectors identified a sharp step at the 2025 cycle. Under the primary detector (EditLens RoBERTa-large), the proportion of abstracts above threshold rose from a stable 5.0% across 2010–2015 to 19.6% in 2024, 33.3% in 2025, and 57.0% in 2026. The two secondary detectors yielded consistent patterns, providing converging evidence of a stylistic shift. Writing-quality metrics showed temporal and directional changes consistent with Gartenberg et al., including a statistically significant decline in Flesch Reading Ease. First authors with no prior SSWR submissions and non-Anglophone first authors exhibited the largest increases in AI scores. Discussion. Findings converge with Gartenberg et al. on direction, timing, magnitude, cohort patterns, and writing-quality decline. By the 2026 cycle, abstracts whose stylistic profile resembles AI-generated academic prose are no longer rare in social work scholarship; they constitute the modal pattern.