1. Ballotta et al. (2023) is described inconsistently. In 3.2.2 it is SHAP decomposition of ESG ratings (governance 47%, environmental 28%); in 4.2 it is attention visualization of 10,000 sustainability reports (38%/27%/21% attention weights; 73% regulatory scenarios). These cannot both be correct. Which method and findings are actually in the cited paper?
2. Kimbrough et al. (2024) appears misrepresented. 3.1.3 cites it as a difference-in-differences design around mandatory ESG disclosure regulations showing AI-measured readability reduces rating disagreement; the reference title concerns voluntary ESG reporting and rating disagreement. Please clarify the design, treatment, and whether AI/readability was measured.
3. Citation attribution errors in 3.2.2. BERT/FinBERT are attributed to Liu et al. (2020), but that reference is RoBERTa; FinBERT is not cited. LinkBERT is attributed to Bingler et al. (2022), but the cited paper is about ClimateBERT, not LinkBERT. Please correct or provide exact sources.
4. Table 1 overstates evidence. The readability row lists “correlates with ESG ratings” as a key benefit without qualification, while the text acknowledges correlational and mixed evidence. This should be qualified or removed.
5. Internal inconsistency in search reporting. “Initial searches yielded approximately 450 unique records. After removing duplicates…” is contradictory: records cannot be both unique and then de-duplicated. Please report the actual flow numbers.
