Index of everything on this site: https://ai.hiina.space/index.md
On community rules, off-the-shelf LLMs under-remove. Across 95 subreddits, GPT-3.5 reached a median accuracy of 63.7%, precision of 83% and recall of 39.8%, and 86.9% of errors were false negatives. In a real-time setting precision fell to 6.2%. Median F1 dropped from 0.68 to 0.53 between March and June 2023 snapshots of the same model. Adding the parent comment fixed 35% of errors (Kumar et al., ICWSM 2024). A subreddit-aware fine-tuned BERT beat ChatGPT, and giving ChatGPT the written rules did not help (F1 0.81 versus 0.68 with rules, Decoding the Rule Book, EMNLP 2025). On contested cases models do worse. Across about 4.3M actions in 24 subreddits, roughly 1 in 7 cases were disputed between moderators, and "almost half of all gray area cases involved automated moderation decisions". Llama 70B scored macro-F1 0.61 on undisputed human cases and 0.51 on gray ones, and other models ranged down to 0.39 (Gray Area, arXiv; AAAI 2026 inferred from a copyright line). On PluRule (2,885 rules across 1,989 communities), the best model scored 57.7% against a 50% trivial baseline, and "low-effort" rules scored about 43% (aggregator page only, paper not read, benchmarklist). Self-promotion and off-topic rules are the hard case, and they are exactly where agent spam would sit.
The bigger risk for an outsourcing moderator is deference. In a CIKM '26 paper, a prior score placed in metadata shifted ratings in 7 of 8 models (|d| up to 0.71) and cut acceptance by 22.3 points for Claude 4.5 Sonnet. On a compliance task, anchored metadata blocked 48% of error corrections. Chain-of-thought did not help, and a "disregard metadata" warning helped only on the compliance task (arXiv; accepted, conference 7 to 11 Nov 2026). Provenance cues work the same way. A human-versus-LLM authorship cue shifted pairwise verdicts by 6 to 45 points depending on model and dataset, with acknowledgement at 0% for most models and 2% at most (The Judge Who Never Admits, preprint; quality judging, not moderation). A separate study reports "label dominance", where judges attend to the authorship label more than to the content (abstract only, arXiv). Human moderators show a related effect: AI-presented citations were treated as verified in a small Wizard-of-Oz study (n=50, preprint, Power Echoes). LLM-Mod also found GPT-3.5 reversed its moderation decisions when questioned, with no new justification (abstract-level, CHI EA 2024).
AI-text detection, the nearest existing "this came from a machine" signal, has a poor record. Seven detectors falsely flagged non-native English essays at an average rate of 61.22% (Liang et al., preprint). Stack Overflow in 2023 told moderators to stop using detectors over an "alarmingly high rate of false positives" tied to suspensions of low-contribution users, which triggered a moderator strike (DevClass). Reddit moderators interviewed said "no tool is reliable enough to fully automate enforcement" and relied on account and style signals instead (arXiv).
(inference) Never hand an LLM moderator a raw provenance flag or classifier score and ask for a verdict. Compute the combination outside the model, use the result to pick the action band, and give the LLM only the content, the rules and the thread context, which is where it adds value. The LLM moderator also reads untrusted posts, so it is a prompt-injection target. As a companion review of prompt-injection research noted, twelve published injection defenses fell to adaptive attacks at over 90% success (Willison). Its tools should be limited to labeling.
Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space