English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Seven Mechanisms of Algospeak: How TikTok Users Outsmart Algorithmic Moderation

Forum topic · ✨步子哥 · 2026-06-28

Summary

A University of Utah study (arXiv:2606.27314) introduces the first mechanism-oriented taxonomy of algospeak—the coded language social media users deploy to evade algorithmic moderation on platforms like TikTok. Instead of classifying coded language by user intent (avoiding censorship, politeness, group identity), the authors classify it by encoding mechanism, identifying seven categories: orthographic transformation (coc@!ne), phonetic substitution (seggs for sex), formal compression (SH for self-harm), formal encoding systems (leet speak), conventional sign reassignment (the corn emoji for porn), morpho-lexical encoding (pornstarring), and referential alias encoding. Experiments on 2,000 human-annotated TikTok and Bluesky posts show the taxonomy improves LLM-based coded language detection by 4.7% accuracy and 5.4% F1 over the strongest baseline, with the largest gains in previously uncovered categories (formal encoding systems +62.2%). A key finding: incomplete taxonomies perform worse than no taxonomy at all. The work shifts moderation research from tracking specific euphemisms toward understanding the finite space of encoding mechanisms.

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 the 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 euphemisms have lifecycles measured in weeks: algorithms learn the old ones, and new ones emerge.

A paper from the University of Utah (arXiv:2606.27314) does something never systematically done before: instead of organizing coded language by *why* it's encoded (evasion, politeness, group identity), it organizes it by *how*—building a mechanism-oriented taxonomy with seven top-level categories.

Why a Mechanism Taxonomy?

Prior ILE taxonomies had several problems:

  • Narrow coverage: only some mechanisms (e.g., only surface perturbations like coc@!ne → cocaine)
  • Conflated paths: encoding and decoding paths mixed together
  • Inconsistent abstraction levels: sometimes by intent ("evasion"), sometimes by form ("homophones")
  • The authors' core insight: intents and content vary, but encoding mechanisms are finite. Users encode discussions of drugs, self-harm, and sex—but the encoding operations themselves are limited. Abstracting mechanisms makes the taxonomy stable.

    The Seven Mechanism Categories

    Built using the iterative method of Nickerson et al. (2013):

    C1: Orthographic Transformation

    Modifying linguistic form while preserving human readability, recovered via visual similarity. Examples: coc@!ne → cocaine, f@ck → fuck. Humans recover via perceptual normalization; algorithmic tokenizers get interrupted.

    C2: Phonetic Substitution

    Encoding via phonetic similarity, recovered by 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 derived from the form itself; community conventions are required.

    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 reversal—the category worst 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

    Leveraging productive word-formation processes. Examples: pornstarring → being a porn star (verbing); morphological embedding of sensitive semantics.

    C7: Referential Alias Encoding

    Replacing proper names with alternative labels for specific entities. Examples: "that Florida man" for a political figure, "the island" for a region. Recovery depends on shared cultural or reputation knowledge.

    One Post Can Use Three Mechanisms at Once

    A striking example from the paper: a TikTok creator wants to express "porn star," but searching "pornstar" returns nothing. So they combine three emojis:

  • 🍑 (iconic resemblance — C1)
  • 🌽 (attribute-based alias — C7)
  • ⭐ (name-based pictorial mapping — C5)
  • Three mechanisms layered together: instantly clear to humans, invisible to algorithms.

    Experimental Results

    Experiments on 2,000 human-annotated TikTok and Bluesky posts:

  • New taxonomy vs. 4 existing taxonomies + no-taxonomy baseline: the new taxonomy wins across 3 LLMs
  • +4.7% accuracy, +5.4% F1 over the strongest baseline
  • Largest gains: C4 (formal encoding systems) +62.2%, C6 (morpho-lexical encoding) +24.3%, C7 (referential alias encoding) +20.8%—all categories poorly or not covered before
  • Ablation: removing any category degrades performance; all seven are necessary
  • A counterintuitive finding: using an incomplete taxonomy performs worse than using no taxonomy at all. Giving an LLM a partial framework constrains its inherent recognition ability—provide the complete framework or none.

    Why It Matters

    The paper's deeper contribution is a paradigm shift:

    From "detecting specific euphemisms" to "understanding encoding mechanisms."

    Euphemism dictionaries go stale weekly—once "seggs" is learned by the algorithm, next week it becomes "segs" or "s3ggs." But the mechanism of phonetic substitution is eternal. 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 all attacks, but you can understand the space of attack mechanisms. The paper provides a more stable scaffold for content moderation—stop chasing euphemisms and guard the mechanism exits instead.

    Paper Information

  • 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 public due to data agreements)

Tags

#algospeak#content-moderation#llm#nlp#tiktok#coded-language#taxonomy#research-paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208263