Qualitative Analysis of Censorship: AI Workflow

Authoritarian information environments move fast: in August 2025 reporting, China Digital Times documented how podcasts, social media, publishers, and academics navigate shifting red lines. This post shows researchers and analysts how to run a repeatable, defensible qualitative analysis of censorship across modalities (podcasts, Weibo threads, book publishing logs and academic outputs) and how to operationalize the pipeline in www.evidano.com. You’ll get a concise 7-step workflow (ingest → code → compare → visualize), coding priorities tuned to the tactics described in the reporting (e.g., ‘battery squad’ algorithm hacks, platform moderation signals), and practical notes on multilingual transcription, metadata capture, and secure handling. Target audience: UX researchers, policy analysts, and academic teams studying Chinese media and digital censorship. Ethics note: treat collected content as research data, secure consent where required, and avoid diagnostic or legal interpretations: this guide is analytic, not legal advice.

Fast take: why this matters for researchers

China Digital Times’ August 21, 2025 roundup traces how creators and audiences shift conversations into liminal spaces (podcasts, coded comments, overseas journals) as censorship opens and closes. Full source: www.chinadigitaltimes.net/2025/08/navigating-chinese-censorship-in-podcasts-publishing-scholarship-and-social-media/.

Payoff: map where content moves, how audiences signal solidarity, and which moderation signals reveal censorship events.
Use-case: rapid thematic coding of podcast transcripts + cross-segment comparisons (platform, country, author affiliation) to show migration and self-censorship patterns.

Findings snapshot

Date	Item	Why it matters	Source
2025-08-21	CDT roundup of censorship across domains	Frames cross-domain tactics and recent examples researchers can code for	www.chinadigitaltimes.net/2025/08/navigating-chinese-censorship-in-podcasts-publishing-scholarship-and-social-media/
2025-08-22	Podcast 'Eight and a Half Minutes' removed from Chinese platforms	Example of platform enforcement after sensitive topic (Jimmy Lai) discussed	www.chinadigitaltimes.net/2025/08/navigating-chinese-censorship-in-podcasts-publishing-scholarship-and-social-media/
2021	Ximalaya disclosure of AI screening (ASR) use	Platform-level automated speech recognition used for pre-release screening	www.ft.com/content/5a8bb423-62a3-4c42-bd54-566f05a9ea47
2025-07	The China Journal: 'Disappearing Research' (Leng & Plantan)	Empirical typology of self-censorship with migration of topics to English journals	www.journals.uchicago.edu/doi/10.1086/735160

What happened (quick summary)

Recent reporting shows censorship in China is multi-layered and uneven: platform moderation, publisher-level review, and top-down political cycles all shape which topics survive. Creators use migration (publishing overseas), alternate modalities (podcasts), and coded engagement (the ‘battery squad, ’ food comments, cat photos) to keep sensitive conversations visible while evading automated or manual takedowns.

Academic work documents how rising political pressures since Xi Jinping narrowed published topics and pushed China-based scholars to shift sensitive findings into English-language or overseas journals. Meanwhile platforms have increasingly adopted AI tools (automatic speech recognition and keyword filters) to screen audio and text before release.

Qualitative analysis of censorship: Coding priorities

Modality & metadata

Capture modality (audio, social post, book, journal article), platform, upload timestamp, and any downstream takedown or removal events.

For podcasts: store episode audio, platform approval metadata, and ASR transcript confidence scores where available.

Tactics and signals to code for

Coded engagement tactics (battery-level comments, cat photos, keyboard mash) as markers of collective algorithmic evasion.

Self-censorship indicators (topic omission, euphemism usage, hedging language), publisher-level prepress edits, and migration signals (English-language reposts).

Actors & relationships

Tag actors (host, platform moderator, publisher, university committee) and relational signals (formal takedown, shadowban, account suspension).

Capture follower counts and prior moderation history when available to compare enforcement thresholds.

Temporal patterns

Record dates of publication, removal, and re-posting to identify opening/closing windows of permissiveness.

Look for correlated enforcement across platforms or post types following high-sensitivity events.

How Evidano helps: map the reporting to a repeatable pipeline

Ingest & transcription

Problem: multi-format audio and Chinese-language posts with inconsistent metadata.

Evidano solution: upload podcast audio, apply custom-dictionary ASR, and redact PII automatically while preserving timestamps for quote-level evidence.

Multilingual normalization & translation

Problem: content migrates between Chinese and English, complicating theme detection.

Evidano solution: built-in translation with a custom dictionary ensures consistent keyword mapping across languages for accurate thematic grouping.

Coding, theme extraction & cross-segment analysis

Problem: manual coding is slow and inconsistent across modalities and studies.

Evidano solution: import codebooks, run AI-assisted coding, generate thematic and frequency analyses, and compare segments (platform, author location, time-window) in a few clicks.

Visual reports & evidence for stakeholders

Problem: stakeholders want concise proof of pattern (e.g., a takedown wave).

Evidano solution: export co-occurrence networks, hierarchical code→subcode visualizations, and clickable quotes to show exact traces of migration and algorithmic evasion.

Security & reproducibility

Problem: sensitive datasets risk exposure and model training by third parties.

Evidano solution: data encryption, no third-party model training on customer data, reproducible pipelines and audit logs for compliance and peer review.

7-step workflow: from raw posts to stakeholder memo

Run this as a two-week pilot to surface patterns described in the reporting:

1) Ingest: collect podcast audio, social threads, publisher metadata, and scraped comments into Evidano; preserve timestamps and original URLs.
2) Transcribe & translate: run ASR with a custom dictionary for names, organizations, and euphemisms; auto-redact PII if needed.
3) Initial codebook: seed codes from the CDT reporting (modality, tactic, actor, enforcement) and import into Evidano.
4) AI-assisted coding: apply model-assisted coding, review edge cases, and lock a validated codebook.
5) Cross-segment analysis: run frequency and co-occurrence analyses by platform, author location (China vs. overseas), and time windows around removal events.
6) Visualize & annotate: produce co-occurrence maps and timeline charts; attach clickable evidence quotes for each claim.
7) Stakeholder memo: export a short executive memo with visuals and appendix of raw evidence links for verification.

Conclusion & next steps

Censorship in China is dynamic; researchers need a reproducible approach to track where content moves and how communities adapt. The reporting summarized here (see original: www.chinadigitaltimes.net/2025/08/navigating-chinese-censorship-in-podcasts-publishing-scholarship-and-social-media/) provides concrete tactics to code for, now map them at scale with tools that preserve evidence and privacy.

Ready to try this pipeline on your corpus? Start a pilot or request a demo at www.evidano.com to import transcripts, run thematic + cross-segment analyses, and produce shareable, auditable reports.

Keep reading

Browse all articles