You see a TikTok comment: "I'm so corn 🌽 tired of this seggs 🥵". Humans instantly understand — "corn" is algospeak for "porn," and "seggs" is a phonetic variant of "sex." But an algorithmic moderation system sees two unrelated words.
This is the world of Indirect Linguistic Encoding (ILE) — an ever-evolving system of coded language invented by social media users to bypass algorithmic moderation, throttling, and demonetization. These codes live for weeks: once algorithms learn the old ones, new ones emerge.
A paper from the University of Utah (arXiv:2606.27314) does something never systematically done before: instead of organizing by *why* users encode (evasion, politeness, group identity), it organizes by *how* — dividing all coded language into 7 mechanism classes and building a mechanism-oriented taxonomy.
Why a Mechanism Taxonomy?
Previous ILE taxonomies had several problems:
- Narrow: covering only some mechanisms (e.g., only surface perturbations like
coc@!ne→cocaine) - Conflated: mixing encoding paths with decoding paths
- Inconsistent abstraction levels: sometimes organized by intent ("evasion"), sometimes by form ("homophone")
- 🍑 (iconic resemblance — C1)
- 🌽 (attribute-based alias — C7)
- ⭐ (name-based pictorial mapping — C5)
- New taxonomy vs. 4 existing taxonomies + no-taxonomy baseline: the new taxonomy wins across 3 LLMs
- +4.7% accuracy, +5.4% F1 (relative to the strongest baseline)
- Largest gains: C4 (formal encoding systems) +62.2%, C6 (morpho-lexical) +24.3%, C7 (referential alias) +20.8% — all categories poorly covered or entirely missing from prior taxonomies
- Ablation: removing any class degrades performance, showing all seven are necessary
- Title: Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection
- Authors: Hamid Reza Firoozfar, Mohammadsadegh Abolhasani, Reza Mousavi, Paul Jen-Hwa Hu
- Institutions: University of Utah, University of Virginia
- arXiv: https://arxiv.org/abs/2606.27314
- Code: https://github.com/hmdfiroozfar/mechanism-oriented-ile-taxonomy
- Data: 2,000 human-annotated TikTok + Bluesky posts (TikTok portion not publicly released due to data agreements)
The authors' key insight: intents and content vary, but encoding mechanisms are finite. Users may encode discussions of drugs, self-harm, or sex — but the encoding operations are limited. Abstracting the mechanism makes the taxonomy stable.
The Seven Mechanisms
The paper proposes 7 top-level classes (built via the iterative Nickerson et al. 2013 method):
C1: Orthographic Transformation
Modifying linguistic form while preserving human readability, recovered via visual similarity. Examples:coc@!ne → cocaine, f@ck → fuck. Key: humans recover via perceptual normalization, while algorithm tokenizers get disrupted.C2: Phonetic Substitution
Encoding via phonetic similarity, recovered through sound reconstruction. Examples:seggs → sex, unalive → suicide (a TikTok classic). Especially effective in multilingual and code-mixing contexts.C3: Formal Compression
Shortening into compact forms, recovered via community consensus. Examples:SA → sexual assault, SH → self-harm. Meaning cannot be inferred from the form itself; one must know the community convention.C4: Formal Encoding Systems
Hiding meaning through explicit, reversible rules independent of linguistic structure. Examples: Leet speak (1337 → leet), reversed spelling. Decoding requires procedural rule inversion — the class least covered by prior taxonomies.C5: Conventional Sign Reassignment
Surface form unchanged, but the community assigns new meaning. Examples: 🌽 (corn emoji) →porn, 🍑 → ass. Recovery relies on retrieving stored sign–meaning associations, not computation.C6: Morpho-Lexical Encoding
Constructing new forms via productive word-formation processes. Example:pornstarring → being a porn star (verbing), embedding sensitive semantics through morphology.C7: Referential Alias Encoding
Replacing a name with an alternative label for a specific entity. Examples: referring to a politician as "the Florida man," or a region as "that island." Recovery depends on shared cultural or reputation knowledge.One Post Can Combine Three Mechanisms
A striking example from the paper: a TikTok creator wanted to express "porn star," but searching "pornstar" returned nothing. So they combined three emoji:
Three mechanisms layered together: instantly legible to humans, invisible to algorithms.
Experimental Results
Experiments on 2,000 human-annotated TikTok and Bluesky posts:
A counterintuitive finding: using an incomplete taxonomy performs worse than using no taxonomy at all. Giving an LLM a partial framework appears to constrain its inherent recognition ability — either give the full framework or none.
Why This Matters
The paper's deeper contribution is not "yet another taxonomy" but a paradigm shift:
From detecting specific coded terms to understanding encoding mechanisms.
Coded dictionaries go stale weekly — once algorithms learn "seggs," next week it becomes "segs" or "s3ggs." But the mechanism of phonetic substitution persists forever. Once an LLM understands the mechanism, it generalizes to never-before-seen coded language.
This parallels adversarial example research in AI safety: you cannot enumerate every attack, but you can understand the attack mechanism space. The paper offers content moderation a more stable scaffold — not chasing individual codes, but guarding the exits of the mechanism space.