All five Chinese LLM services transmit partial or complete responses to the client machine even when output blocking is triggered, representing a major information leak. For DeepSeek and Qwen, truncated blocked responses are on average close in token length to full successful responses. For Baidu-Chat, the complete response is transmitted to the client but only partially rendered in the browser UI, with only a word or two visible on screen.

Simple character-level perturbations (English) and homophone substitutions (Chinese), combined with LLM instruction-following prompts directing the model to use word substitutions in its output, successfully bypassed all input and output filters for all 41 input-blocked and 197 output-blocked queries across five major Chinese LLM services (Baidu-Chat, DeepSeek, Doubao, Kimi, Qwen). Every input-blocked query contained at least one keyword combination that alone triggered the filter, confirming keyword-matching rather than semantic classification.

§VII-D defense

Cross-national experiments conducted from Singapore, South Korea, and Taiwan during February 19–24, 2025 found no variance in blocking implementations, event syntax, or server infrastructure across all five Chinese LLM services. Input blocking was enacted identically in all three international locations, and services connected to the exact same server IP addresses globally — Kimi and Baidu-Chat connected to identical IPs and DeepSeek to the same two addresses across all tested locations.

§VIII-A evaluation

Input blocking in Chinese LLM services (DeepSeek, Qwen, Kimi, Doubao) is overwhelmingly consistent: all four services persistently block the exact same queries across all 5 measurement samples in both Simplified and Traditional Chinese. Output blocking is far less consistent, with only 29 out of 349 output-blocked queries blocked across all 5 samples. Baidu-Chat is exceptional: it performs almost no input blocking but instead relies heavily on post-search and output blocking (78.6% of blocks are output-phase).

§VI-B evaluation

DeepSeek, Kimi, and Doubao all transmit analytics logs to the same Autonomous System (AS24429, Zhejiang Taobao Network Co., Ltd., a ByteDance/Volcengine subsidiary), with one monitoring endpoint IP directly overlapping between DeepSeek and Doubao. Additionally, all four non-Qwen services maintain connections with servers physically located in China throughout the chat session, transmitting user IDs, session IDs, viewport data, language preferences, and in Baidu-Chat's case, the full query text via URL-encoded CAPTCHA requests.

§VI-D deployment

During the June 2025 shutdown, Iranian authorities blocked international One-Time Password (OTP) SMS delivery, preventing new sign-ins to foreign secure-messaging platforms and VPN services. This forced users toward government-approved domestic platforms that lack security and privacy protections. The blockade of OTPs effectively weaponized account-recovery flows as a secondary shutdown layer, disproportionately affecting users who needed to activate new circumvention tools during the crisis.

2025-iran-shutdown-measurement §Human Rights Implications port-blockingdpikeyword-filtering

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