"If the right to privacy means anything, it is the right of the individual, married or single, to be free from unwarranted governmental intrusion." - William J. Brennan, Jr.
The full “trigger word” or keyword lists used by Big Tech aren’t publicly published – that is what makes them secret. But we do know quite a bit from leaks, academic studies, content moderation policy documents, and transparency reports.
Here’s a clear breakdown 👇
🧱 1. The reality: Lists exist, but they’re proprietary
Each major platform (Meta, X/Twitter, YouTube, TikTok, LinkedIn, etc.) maintains internal keyword or phrase lists used by:
These lists are enormous, multi-language, and constantly evolving.
They are never fully published because doing so would:
Enable evasion and “gaming the system”
Create PR controversies over politically or culturally sensitive inclusions
Expose moderation bias or state-requested influence
So, officially, the companies only describe categories of content moderated, not the actual trigger terms.
🏢 2. Partial transparency sources (what is known)
The following are legitimate published or leaked clues:
Meta (Facebook, Instagram):
“Dangerous Individuals and Organizations List” (leaked in 2021 by The Intercept): included thousands of groups or individuals tied to hate, terrorism, militias, etc. This list functions like a keyword/entity trigger.
Community Standards page details banned categories (violence, hate speech, etc.), though not the exact words.
Internal moderation manuals (Project Daisy, leaked to Motherboard) show patterns, not terms — e.g., slurs + target class = violation.
Twitter/X (before 2023):
Transparency reports referenced “Hateful Conduct” and “Misinformation” classifiers that use internal term banks.
Independent researchers like the Stanford Internet Observatory have identified sample keyword clusters via the API (before rate limits changed).
YouTube / Google:
Has published research on “classifier fairness” and word embeddings used in moderation (e.g., Google Research’s Perspective API) — some training lexicons are public Perspective API toxic terms dataset on GitHub.
“Hate and harassment” categories use lists of slurs + context detection models like BERT fine-tuned on Jigsaw Toxic Comment Dataset.
TikTok:
Leaked internal moderation rules in 2020 (The Guardian, Intercept): “Do not recommend” list included political phrases, LGBTQ terms (later revised), Tiananmen references, and other “sensitive” content.
Keyword-level suppression lists were country-specific.
Microsoft / LinkedIn:
Uses blocklists for “profanity, hate terms, and unsafe domains.” Internal terms not public; however, researchers found that flagging strongly correlates with the Hatebase lexicon and Google’s “badword” open list.
Governments:
In some regions (China, Russia, Turkey, EU DSA enforcement desks), governments provide official “blocklists” or terms to monitor.
The Global Internet Forum to Counter Terrorism (GIFCT) maintains shared hash databases of terrorist content and associated keywords, which member companies use collaboratively.
🧩 3. Open-source and academic mirrors
Academics and civil society groups maintain approximate keyword sets used to model Big Tech moderation:
These aren’t the same internal lists the platforms use, but they give you a reasonable map of the categories that attract algorithmic scrutiny.
⚖️ 4. What’s “flaggy” in current systems
Across platforms, the types of words or phrases that are highly weighted in classifiers include:
Explicit sexual content (esp. minors, body parts, sexual offers)
Violent or self-harm instructions
Slurs or ethnic / religious insults
Political crisis nouns (“election rigged,” “bioweapon,” “crisis actor,” etc.)
Medical claims (especially COVID, vaccines, reproductive health)
Conspiracy signifiers (“plandemic,” “new world order,” etc.)
Weaponized language (call-to-violence verbs, “go attack,” etc.)
Again — it’s context + keyword + account history that decide if it’s flagged.
🧭 5. How to map “trigger profiles” safely
For activists or alternative-media projects (like your Digital Freedom Awareness Week), a responsible approach is:
Use context-sensitive testing — small posts, slightly altered phrasing, monitor engagement and label behaviors.
Compare safe vs. flagged variants to map edge boundaries.
Archive key phrases and timestamps privately (don’t publish full flag words).
Cross‑post through federated or open platforms (Mastodon, Odysee, Matrix) to gauge normalized reach.
Reference sheet:
Open-source and partially leaked “trigger” keyword resources by category
Note: Platforms do not publish their internal trigger lists. The sources below are open datasets, research corpora, and notable leaks that approximate what gets flagged. Use these for awareness and testing, not as definitive platform lists.
Hate speech and harassment
Hatebase Lexicon (multilingual)
What it is: Crowdsourced/global lexicon of hate terms with context metadata (language, region, target).
Use: Build awareness of slur categories and regional variants.
Mostly image/video classifier-based; few public term lists. Research uses “NSFW/violence” taxonomies (e.g., OpenNSFW variants). No authoritative keyword list.
Spam, fraud, coordinated inauthentic behavior
Spam/SEO keyword lists (open NLP)
What it is: Heuristics for marketing spam, phishing terms, crypto scam phrases.
