Across five political spam incidents, spam constituted 62–73% of all tweets in the Russia, China'12, and Mexico incidents, while Syria had only 6% spam. In the China'12 incident, 1,700 spam accounts (14% of all accounts) generated 600,000 spam tweets (73% of total), with 10 individual accounts each producing over 5,000 tweets before shutdown; in Mexico, 50 accounts sustained 1,000 spam tweets per day throughout the incident.

In four of five incidents (all except Syria), spam accounts were registered in temporally clustered blocks while legitimate accounts were not; in Russia and Mexico, multiple distinct registration bursts were observed. Across all five incidents, spam account usernames were automatically generated, with China'12 and Mexico accounts following a {name}{name}{number} pattern padded to exactly 15 characters (Twitter's maximum), making algorithmic reverse-engineering feasible.

§4.1, Table 6, Figure 4 detection

In the Russia and Mexico incidents, spam tweets showed statistically significant spikes at fixed sub-hour intervals (5 and 15 minutes past the hour respectively), consistent with cron-job automation. Despite this automation, both campaigns deliberately mimicked human diurnal activity patterns — spam volume peaked at the same hours as legitimate traffic — to evade time-based anomaly detection.

§3.2, Figures 2–3 detection

Default-profile usage was significantly elevated among spam accounts in China'11 (89.4% spam vs 51.2% non-spam), Russia (57.8% vs 34.7%), and China'12 (95.1% vs 47.8%); however, Mexico inverted this trend with only 1.7% of spam accounts using default profiles vs 27.0% of non-spam accounts, indicating that newer campaigns actively customize profiles to evade appearance-based detection.

§4.2, Table 7 detection

Twitter's existing automated spam-filtering mechanisms caught only approximately 50% of politically motivated spam in the Russian parliamentary election incident, as reported by Thomas et al. (2012) and noted as the baseline for this study. Spammer behavior varied sufficiently across incidents (targeting strategy, URL usage, mention patterns, default-profile adoption) that supervised machine-learning classifiers trained on one incident are unlikely to generalize to others.

§1, §5 evaluation

When the researchers attempted to use Gemini 2.5 Flash as a third independent LLM judge via its API for evaluating moderation decisions, Gemini automatically blocked all judging attempts citing safety reasons. This occurred even though the research task (judging whether a response is more or less moderated) does not itself produce harmful content. The incident illustrates that LLM safety systems can over-block legitimate research use cases, and that different LLM providers have different thresholds— Claude Haiku 4.5 and GPT-4o completed all judging tasks without safety refusals.

2026-lipphardt-dual §3.3.3 ml-classifierkeyword-filtering

Category-level analysis of 100 statements across 5 sensitive content categories found that interface-based moderation gaps vary significantly by topic. Sexuality showed the strongest WebUI/API gap (WebUI 7.0× more likely to be moderated than API per GPT-4o judge for Gemini). Political ideology followed at 2.0×, then hate speech at 1.0×. Miscellaneous offensive topics showed the inverse pattern (API more moderated at 0.3×). Religious content showed WebUI moderation with no API moderation. The pattern suggests public-facing WebUI interfaces prioritize reputational risk management for high-scrutiny categories.

2026-lipphardt-dual §4.5, Figure 8 ml-classifierkeyword-filtering

An empirical study of 100 sensitive statements tested on Gemini (2.5 Flash) and ChatGPT (GPT-5) found that WebUI interfaces are systematically more restrictive than their API counterparts. According to GPT-4o judge: WebUI was moderated 18% of the time vs. 9% (Gemini API) and 13% (ChatGPT API). DeBERTa classifier found 82% of WebUI responses moderated vs. 58% of API responses. The Gemini WebUI:API ratio ranged from 2.0:1 (GPT-4o) to 7.0:1 (Claude), and ChatGPT from 1.4:1 (GPT-4o) to 15.6:1 (Claude). Neither Google nor OpenAI discloses these interface-specific policies.

2026-lipphardt-dual §4.3, Table 2 ml-classifierkeyword-filtering

Across all 7 LLMs tested (GPT 4o, GPT 4o Mini, Gemini 1.5 Flash, Gemini 1.5 Pro, Llama 3.2, Claude 3.5 Haiku, Claude 3.5 Sonnet), statistically significant evidence of censorship bias was found in at least one evaluation metric per model: responses to Simplified Chinese prompts were more neutral, more similar to sanitized text, and less opinionated than semantically identical Traditional Chinese prompts (p < 0.05 across refusal-rate, sentiment, CensorshipDetector classification, and word-embedding analyses).

2025-ahmed-llm-censorship-bias Table 4, §6 keyword-filteringml-classifier

CensorshipDetector, an XLM-RoBERTa model fine-tuned on 587,819 Baidu Baike articles (censored) and Chinese Wikipedia (uncensored), achieved 91% accuracy on a held-out validation set of Chinese news articles, correctly classifying 93% of Chinese state media articles as censored and 87% of New York Times Chinese articles as uncensored, with average censorship scores of 0.93 and 0.13 respectively.

2025-ahmed-llm-censorship-bias §4.7, §4.7.2 keyword-filteringml-classifier

The paper proposes a black-box methodology for detecting censorship bias in LLMs by comparing responses to identical prompts in Simplified vs. Traditional Chinese — scripts for the same spoken language — controlling for translation quality while exploiting that Simplified Chinese training data is disproportionately sourced from mainland China's censored internet. Each prompt is repeated ten times and scored for similarity to censored text using an XLM-RoBERTa classifier fine-tuned on Baidu Baike (censored) vs. Chinese Wikipedia (uncensored) with scores from 0 to 1.

2024-ahmed-extended §2 / §2.4 keyword-filteringml-classifiermeasurement-platform