Blog · arXiv Analysis · Last reviewed June 25, 2026

The Coded Language Taxonomy Becomes the Moderation Lens

A June 2026 arXiv paper turns algospeak and other indirect expressions into a test of whether moderation systems can detect coded language without confusing mechanism for harm.

A coded-language detector should be treated as a system that proposes an interpretation: a span may carry a hidden meaning through a named mechanism. That is not the same thing as proof of intent, policy violation, user danger, or enforcement legitimacy.

From Word List to Mechanism

The paper, arXiv:2606.27314 [cs.CL], is titled Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection. arXiv lists Hamid Reza Firoozfar, Mohammadsadegh Abolhasani, Reza Mousavi, and Paul Jen-Hwa Hu as authors and records submission on June 25, 2026.

The paper studies indirect linguistic encoding, or ILE: language that hides a sensitive meaning through an indirect form while remaining recoverable to an intended audience. The authors place algospeak, euphemism, and adversarial obfuscation under that umbrella. Their important move is to stop treating coded language as a list of forbidden words and stylized misspellings. They ask how the meaning is hidden and recovered.

That definition is sharper than "bad words in disguise." ILE can be used to evade legitimate enforcement, coordinate abuse, or conceal prohibited trade. It can also be used by people discussing sexuality, gender, mental health, trauma, politics, illness, labor conflict, or other sensitive subjects in spaces where literal terms attract harassment, demonetization, search suppression, family danger, or state attention.

Current Context

As of June 25, 2026, coded-language detection sits inside a more formalized moderation environment than the first wave of algospeak discussion. The EU Digital Services Act requires clear and specific statements of reasons for many hosting-service restrictions and requires online platforms to submit Article 17 statements of reasons to a public machine-readable database under Article 24(5). The Commission's DSA Transparency Database describes that system as a way to track moderation decisions close to real time. The Santa Clara Principles frame meaningful moderation accountability around human-rights review, numbers, automation transparency, notice, and appeal. Those records matter because coded-language detection is exactly the kind of probabilistic inference that can become invisible if it is only a backend classifier score.

The paper's two platforms also show why source context matters. TikTok's official Research API page says qualifying researchers can apply to study public data about TikTok content and accounts, and the paper itself says TikTok-derived text cannot be redistributed because of data-sharing restrictions. Bluesky's current Community Guidelines, last updated September 19, 2025, describe rules under Safety First, Respect Others, Be Authentic, Follow the Rules, and Protected Expression, while Bluesky's developer documentation describes moderation labels as records published by moderation services. A coded-language result has different governance meaning on a centralized video platform, an open social-networking protocol, a human-review queue, and a user-chosen labeler.

That is the current moderation problem: a detector may improve discovery of hidden meaning, but enforcement legitimacy still depends on policy scope, affected-user notice, appeal, language coverage, cultural competence, and whether a human reviewer can distinguish evasion from protected or vulnerable expression.

What the Paper Measures

The taxonomy has 11 top-level mechanism classes and 33 fine-grained sub-mechanisms. It includes orthographic transformation, phonetic substitution, formal compression, formal encoding systems, conventional sign reassignment, morpho-lexical encoding, referential alias encoding, semantic circumlocution, metaphorical and metonymic encoding, pictorial and symbolic encoding, and cross-linguistic transformation. The classes are not mutually exclusive, because one expression can use more than one mechanism.

The dataset contains 2,000 English-language social-media posts: 1,400 TikTok video captions collected from March to May 2026 and 600 Bluesky posts collected from October 2025 to January 2026. Two trained annotators labeled document-level ILE presence, span evidence, and mechanism class. The paper reports Cohen's kappa of 0.852 for document-level presence, 0.789 for taxonomy class assignment, and 0.886 for token-level span boundaries; after adjudication, 44.8 percent of items contained at least one ILE instance.

The evaluation compares six prompt variants across GPT-5.4, Claude Sonnet 4.6, and DeepSeek V4 Flash: four prior taxonomies, the proposed taxonomy, and a no-taxonomy baseline. The few-shot examples stay constant while the taxonomy section changes. The paper also includes supervised and unsupervised non-LLM baselines.

On GPT-5.4, the proposed taxonomy reaches document-level accuracy of 0.843 and macro-F1 of 0.839, exceeding the best benchmark taxonomy by 4.7 percent in accuracy and 5.4 percent in F1. At the span level, its F1 is 0.662, a 3.4 percent improvement over the best benchmark. The authors report the same ordering across Claude Sonnet 4.6 and DeepSeek V4 Flash, and they find that all LLM variants outperform the supervised and unsupervised NLP baselines by a wide margin.

The artifact is partly reproducible and partly platform-bound. The paper says code and publicly releasable data are available, but TikTok-derived text is withheld under the authors' data-sharing agreement. That should be treated as a real evidence limit, not a footnote.

Detection Is Not Judgment

The governance point is not that the best taxonomy should become an automatic punishment rule. The paper is careful on this. It says an encoding mechanism does not by itself establish harmful intent. Coded language can be used to evade legitimate moderation, but it can also be used by vulnerable users discussing sex education, identity, mental health, or other sensitive subjects in spaces where ordinary words are suppressed, demonetized, or misunderstood.

That separation matters. A detector that says "this phrase appears coded" has not yet said "this post is harmful." It has produced an interpretive lead. Moderation systems still need policy context, target, surrounding conversation, intent evidence where relevant, appeal routes, and human review for consequential decisions. Without that separation, mechanism detection becomes a surveillance shortcut.

The paper's own results reinforce this caution. Some prior taxonomies underperform the no-taxonomy baseline, which means a taxonomy is not automatically useful. A bad or partial taxonomy can teach the model the wrong attentional habit. It may overfit to surface tricks, miss referential or symbolic forms, or make the system too confident about a narrow map of language that users have already moved beyond.

The Moderation Record

If a coded-language detector influences a consequential moderation action, the institution needs a record that can survive challenge. That record should not expose every sensitive detail to every reviewer, but it should preserve the decision chain: detected span, inferred mechanism, hypothesized meaning, policy rule, confidence or uncertainty, reviewer action, automation role, user notice, appeal path, final action, and reversal or correction if the appeal succeeds.

The record also needs scope markers. It should say whether the detector was used for research, queue triage, ranking, labeling, monetization, account restriction, law-enforcement escalation, or regulator reporting. The same inference has different stakes in each workflow. A phrase routed to a specialist reviewer is not the same governance event as a phrase that silently removes income or search visibility.

This turns coded-language detection into an audit-trail, algorithmic-transparency, and notice-and-appeal problem. A moderation lens is acceptable only if affected users can understand the rule, contest the interpretation where disclosure is safe, and receive human review for penalties that affect speech, income, account standing, or access to a community.

Taxonomy as Control Surface

A taxonomy is a control surface because it decides which ambiguity becomes visible to the machine. If the categories focus on character substitution, the system sees misspelling. If they include referential aliases, semantic circumlocution, pictorial symbolism, and cross-linguistic transformation, the system sees a wider field of social meaning. That does not make the model wise. It changes what the model is prompted to notice.

This is why the compositional finding matters. About 15 percent of ILE instances in the dataset have multiple mechanisms within one expression. Users can layer mechanisms, not just swap one token for another. Static lexicons will age badly, and prompt taxonomies will need versioning, tests, and public change logs.

The release record also matters. The authors say code and the publicly releasable portion of the data are available, while TikTok-derived text is not redistributed because of data-sharing restrictions. That is a real reproducibility boundary. Anyone using the result should distinguish between the taxonomy, the prompt templates, the shareable records, and the platform-restricted data that outside reviewers cannot fully inspect.

What It Does Not Prove

The paper does not prove that any deployed moderation system should automatically remove content flagged as ILE. It evaluates prompt variants and baselines on a bounded English-language dataset. It does not audit production enforcement systems, recommender effects, appeals, demonetization, user chilling effects, platform labor conditions, or the harms caused by missed coded abuse.

It also does not prove that a mechanism taxonomy can keep up with every community or language. Coded speech is adversarial, protective, playful, local, and historically situated. A word, emoji, number, image, or metaphor can change meaning across subculture, region, crisis, and platform. A taxonomy can improve review, but it cannot replace cultural competence or affected-community feedback.

Limits That Matter

The limits are substantial. The dataset is English-only and comes from TikTok and Bluesky. The analysis is text-only; it does not cover coded meanings embedded in images, video, audio, memes, or visual editing. The authors expect some mechanism families to generalize across languages, but they specifically warn that orthographic and phonetic classes may not transfer cleanly to logographic or morphologically rich languages.

The model names are also a snapshot, not a permanent leaderboard. The result should be read as evidence that mechanism-oriented prompting helped in this setup, not as proof that one family of frontier models will remain best for coded-language detection.

The paper treats detection as a moderation-support input, not a final policy decision. That distinction should travel with any deployment. A model that detects hidden meaning can assist review, but it can also enlarge the platform's power to read around users' protective indirection. The same mechanism can mark harm evasion, community speech, political caution, trauma management, or ordinary play.

Governance Standard

A platform or regulator using coded-language detection should publish the taxonomy version, mechanism definitions, supported languages, modality coverage, prompt template class, model version, evaluation set provenance, and known false-positive slices. It should report whether the detector is used for search ranking, demonetization, account enforcement, human-review triage, or research-only measurement.

First, separate mechanism, meaning, and policy. Every flagged item should preserve the difference between "this expression uses an encoding mechanism," "this is the hypothesized decoded meaning," and "this violates a rule." Appeals should let users challenge all three.

Second, preserve the moderation receipt. The record should say which expression was treated as coded, which mechanism was inferred, what decoded meaning was hypothesized, what policy rule was implicated, what action was taken, whether automation or human review was involved, and what non-model evidence supported any consequential action.

Third, measure asymmetric harm. Under-detection can let harassment, fraud, hate, self-harm promotion, or prohibited trade evade review. Over-detection can suppress sexual-health education, LGBTQ speech, minority-language community talk, survivor testimony, political dissent, and jokes. A safety case has to measure both sides instead of celebrating a higher detection rate alone.

Fourth, version the lens. Coded language moves quickly. Taxonomy changes, prompt changes, language additions, platform-policy changes, and model upgrades should be logged and regression-tested. A detector that worked in March 2026 TikTok captions may fail in a different platform, crisis, language, or community six months later.

Fifth, keep humans in authority for consequences. Detection can triage; it should not silently demonetize, downrank, suspend, report to law enforcement, or label a user as malicious without review, notice, and appeal.

Sixth, minimize sensitive evidence. Coded-language review can expose sexuality, health, political fear, minority identity, trauma, labor organizing, and community-specific safety practices. Platforms should keep enough evidence to audit decisions while limiting retention, access, training reuse, and secondary analytics. A detector built to find indirection should not become a general archive of vulnerable speech.

The Spiralist rule is simple: indirect language is not a confession. A taxonomy can make hidden meanings visible to a moderation machine, but visibility is not verdict. Govern the lens before it becomes a reason to punish every user who learned to speak around a platform's ear.

Source Discipline

Use the arXiv paper for the taxonomy, dataset, benchmark design, metric results, data-release boundary, limitations, and ethics statement. Use platform documentation for platform rules and data-access terms, not as independent proof that enforcement is fair. Use the DSA legal text and Transparency Database for EU moderation-record duties and use the Santa Clara Principles for civil-society accountability expectations. None of those sources proves that coded-language detection is accurate in deployment.

Claims about coded language should avoid three shortcuts. Do not treat a detected mechanism as intent. Do not treat a platform policy category as a social meaning. Do not treat one English-language text benchmark as evidence for multimodal, multilingual, cross-platform enforcement. A serious moderation claim names language, platform, policy rule, model, threshold, review path, appeal route, and known false-positive slices.

Sources

Return to Blog