> *"The medium is the message."* > — Marshall McLuhan
---
1. Three Real Posts First
Post 1 (TikTok): > "I just got my seggsuality back 🦷"
Post 2 (Bluesky): > "Unc cooked the turkey fr fr"
Post 3 (TikTok): > "Looking for a plug for 🍃 in the 614"
If you're not in the relevant communities, these posts may leave you baffled. But they all do the same thing: hide a sensitive meaning through some form of encoding.
- "seggsuality" = sexuality, via lookalike substitution
- "Unc" = Uncle Sam = US government, via referential alias
- "plug" = drug dealer, via community jargon
- 🦷 = tooth = "the tooth" = "truth" in TikTok slang
- 🍃 = weed, via emoji visual similarity
- "614" = Columbus, Ohio, via area code
- Zhang et al. (2014): studied Chinese social-media censorship evasion with 5 categories (orthographic substitution, character decomposition, translation, semantic reinterpretation, historical allusion). But it only works for Chinese and fails when transferred to English.
- Renwick & Barbosa (2021): focused on character-level operations (leetspeak, punctuation insertion, glyph substitution) but ignored semantic and conventional strategies.
- Calhoun & Fawcett (2023): classified by orthography, morphology, phonology, prosody, and reanalysis, but omitted indirect strategies requiring shared cultural knowledge.
- Leal-Arenas & Corizzo (2024): centered on algospeak (abbreviations, phonetic spellings, symbol substitution, paraphrase, re-semantization) but with a muddled hierarchy.
- TikTok video descriptions (1,400, March–May 2026)
- Bluesky posts (600, October 2025 – January 2026)
- Document level: does the post contain ILE (binary)
- Span level: minimal continuous ILE spans
- Mechanism level: category assignment from the 11 classes
- Document-level κ = 0.852
- Category-level κ = 0.789
- Span-level κ = 0.886
- 4 existing taxonomies (Zhang; Calhoun & Fawcett; Fillies & Paschke; Leal-Arenas & Corizzo)
- The newly proposed 11-category taxonomy
- A no-taxonomy baseline (LLM judges freely)
- No-taxonomy baseline recall = 0.587
- Zhang taxonomy recall = 0.556
- Calhoun taxonomy recall = 0.317
- New taxonomy recall = 0.952
- Orthographic: all methods do okay — it's the most visible surface operation
- Formal Encoding: new taxonomy +62.2% recall — only the new taxonomy lists this as a standalone category
- Referential Alias: new taxonomy +20.8% — old taxonomies don't handle aliasing strategies that require cultural knowledge
- Morpho-Lexical: new taxonomy +24.2% — old taxonomies lack a word-formation dimension
These are not random typos. They are Indirect Linguistic Encoding (ILE) — systematic strategies humans use to hide sensitive meanings in language.
The question is: Can AI recognize these encodings?
---
2. The Fragmentation Problem in Existing Research
Hamed Firoozfar et al. at the University of Virginia found that before their work, academic taxonomies of ILE were fragmented:
Each taxonomy covers only part of the phenomenon. Worse, the classification criteria depend on the data source — researchers of Chinese see only Chinese strategies, algospeak researchers see only algospeak strategies. There is no unified "encode-decode path" framework.
It's like the blind men and the elephant: everyone touches something real, but no one assembles the whole animal.
---
3. The New Taxonomy: 11 Categories × 33 Subcategories
Using the iterative taxonomy-development methodology of Nickerson et al. (2013), with 4 iterations (conceptual → empirical → conceptual → empirical), the Firoozfar team built a taxonomy whose meta-characteristic is the "encode-decode path".
Key design decision: the meta-characteristic is "how" (how meaning is encoded), not "why" or "about what".
Why? Because the "why" (censorship evasion, politeness, in-group identity) and the "what" (drugs, self-harm, sex) vary by platform, culture, and era — but the "how", the path from surface form to hidden meaning, is stable across contexts.
The result is 11 top-level categories:
1. Orthographic Transformation: modifying letterforms while keeping readability, e.g., leetspeak ("s3x"), glyph substitution 2. Phonetic Substitution: encoding via homophones, e.g., "seggs" = sex 3. Formal Compression: abbreviations and initialisms, e.g., "fr" = for real 4. Formal Encoding Systems: rule-based encodings, e.g., ROT13, Base64 5. Conventional Sign Reassignment: assigning new meanings to existing symbols, e.g., 🍃 = weed 6. Morpho-Lexical Encoding: coining new words via derivation, e.g., "unalive" = kill 7. Referential Alias Encoding: referring to entities by aliases, e.g., "Unc" = US government 8. Semantic Circumlocution: descriptive phrases instead of direct words, e.g., "the plant" = marijuana 9. Metaphorical and Metonymic Encoding: cross-domain mappings, e.g., "cook" = perform well 10. Pictorial and Symbolic Encoding: emoji as meaning carriers, e.g., 🦷 = truth 11. Cross-Linguistic Transformation: using another language or script, e.g., pinyin for Chinese sensitive words
A single token can trigger multiple categories. "Seggs" is simultaneously an orthographic transformation (visual similarity) and a phonetic substitution (sound similarity); "🍃" is both pictorial encoding and conventional sign reassignment. The taxonomy does not require mutual exclusivity and allows combined labels.
---
4. Dataset: 2,000 Real Posts
The team collected 2,000 English posts from two platforms:
The two platforms have different moderation regimes — TikTok strict, Bluesky relatively lax — allowing comparison of encoding strategies under different moderation pressure.
Sampling was stratified: one part biased toward ILE seed-word lists, one part random. Posts that were pure URLs/hashtags/@mentions were removed, as were verified bot accounts.
Two annotators independently labeled every post at three levels:
Agreement was high:
Ultimately, 44.8% of posts contained at least one ILE instance. Nearly half. ILE is not a marginal phenomenon — it is the norm of social media language.
---
5. Experiments: LLM + New Taxonomy vs. LLM + Old Taxonomies
The team ran experiments with 3 LLMs (GPT-5.4, Claude Sonnet 4.6, DeepSeek V4 Flash) × 6 prompt variants:
All prompts used the same 4 few-shot examples; only the embedded taxonomy changed.
Four non-LLM baselines were added: TF-IDF + logistic regression, character n-gram + SVM, Word2Vec cosine similarity, and embedding-graph centrality.
---
6. Results: The New Taxonomy Wins Across the Board
Taking GPT-5.4 as an example:
| Metric | No taxonomy | Best old taxonomy | New taxonomy | Gain | |---|---|---|---|---| | Document accuracy | 78.5% | 80.5% (Leal) | 84.3% | +3.8pp | | Document F1 | 77.1% | 79.6% (Leal) | 83.9% | +4.3pp | | Span F1 | 59.9% | 64.0% (Leal) | 66.2% | +2.2pp |
Across all three LLMs, the new taxonomy achieved the highest document-level and span-level F1. All LLM variants crushed the non-LLM baselines (best NLP method: only 68.4% accuracy).
But the more important findings are at the category level.
6.1 A Partial Taxonomy Can Be Worse Than No Taxonomy
On the "Formal Encoding Systems" category (e.g., ROT13):
The Calhoun taxonomy performed *worse than no taxonomy* on this category! Because it doesn't cover this phenomenon at all — the LLM's classification framework has no "formal encoding" category, so when the LLM sees ROT13 it doesn't know where to put it and simply doesn't label it.
An incomplete taxonomy is worse than none — it doesn't just fail to help, it constrains the LLM's judgment space.
6.2 The New Taxonomy's Biggest Advantage Is on "Combined Encoding"
Real posts often combine multiple encoding strategies. "Seggs 🦷" is simultaneously phonetic substitution + pictorial encoding + conventional sign reassignment.
On "combined ILE" (2+ mechanisms), the new taxonomy's advantage is most pronounced: it allows multiple labels per span, while most old taxonomies require mutually exclusive classification and fall apart on combinations.
6.3 Each Category's Improvement Comes from a Different Source
---
7. Why This Work Matters
7.1 Direct Implications for Content Moderation
Current moderation relies mainly on keyword matching and classifiers. But the essence of ILE is bypassing keywords — "seggs" isn't on the sexuality keyword list, and "🍃" is just an emoji to most moderation systems.
If moderation systems don't understand encoding mechanisms, they're stuck in a reactive loop: users invent an encoding, the system labels data, trains a classifier, deploys it — then users invent a new encoding. A cat-and-mouse game.
The new taxonomy offers a mechanism-level framework: whatever the specific encoding, it must fall into one of 11 paths. Moderation systems can design detection strategies per path rather than per encoding instance. A paradigm shift from "whack-a-mole" to "filling the holes."
7.2 Implications for AI Safety
LLMs themselves can be used to generate ILE — bypassing safety filters, producing veiled harmful content, spreading in-group codes. If an LLM can recognize ILE, it can also generate it. Understanding encoding mechanisms is the first step of defense.
A deeper issue: LLMs' understanding of ILE depends on the taxonomy framework provided. This means the LLM's "understanding" is not general-purpose pattern recognition but constrained by the supplied framework. Swap frameworks and the same LLM's performance can swing by 20 percentage points. That's an important data point in the debate over whether LLMs truly understand language.
7.3 A Methodological Contribution to Linguistic Classification
The work also demonstrates a methodology for computational linguistics: building computationally usable taxonomies via Nickerson's iterative method. Many NLP tasks still use ad hoc classifications ("I think it should be split into these categories") rather than systematic iterative construction. This paper shows how taxonomy building itself can be a reproducible research process.
---
8. An Honest Assessment: The Work's Limitations
1. English-only data. ILE strategies in Chinese, Russian, or Arabic may differ entirely. Zhang et al.'s 5 Chinese categories overlap only partially with the new 11. Cross-lingual generalization needs more validation. 2. Two platforms. ILE on Reddit, Twitter/X, Telegram, or Discord may differ. TikTok and Bluesky are both short-text platforms; long-form ILE may be more complex. 3. The 44.8% ILE rate may be inflated due to seed-word biased sampling; the true distribution is likely lower. The authors acknowledge this. 4. Few-shot LLM setup. All experiments used 4-shot. Fine-tuning or longer contexts might change results. 5. Only two annotators. Agreement was high, but two people's subjectivity remains; more annotators would strengthen credibility. 6. The taxonomy is static. ILE evolves constantly — new encodings appear daily. The paper doesn't discuss a mechanism for keeping the taxonomy updated.
---
9. Closing: McLuhan's Prophecy
In 1964, Marshall McLuhan said "The medium is the message" — the mode of transmission itself shapes meaning.
ILE is a perfect illustration. When someone writes "seggs" instead of "sex", she isn't just evading moderation — she's declaring "I belong to this community," "I know the rules," "I'm willing to pay an extra character cost for this expression." The encoding itself is a social signal.
The Firoozfar team's 11-category taxonomy essentially says: humans use 11 basic ways to hide sensitive meaning inside surface forms. Whatever the platform, language, or culture, these 11 paths are stable.
And for AI to truly understand human language, it must understand these 11 paths — not as keyword lists, but as encode-decode mechanisms.
Because language isn't just about what you say, but how you say it. And "how" is often more important than "what."
---
Paper: arXiv:2606.27314 Code and data: github.com/hmdfiroozfar/mechanism-oriented-ile-taxonomy Authors: Hamed Firoozfar, Sina Firoozfar, Amir Karami (University of Virginia & University of South Carolina)