Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection

A revised arXiv preprint proposes a mechanism-oriented taxonomy of indirect linguistic expressions, the disguised phrasing such as algospeak and euphemisms that users adopt to hide sensitive meaning from platforms. The work organizes these encoding strategies by how they work rather than how they look, aiming to give LLM-based detection systems a more general basis for spotting obfuscated content. It targets the gap between surface-form moderation filters and adversarial evasion in social media text.

Agentic LLM Framework Generates and Refines Counter-Narratives to Hate Speech

A new arXiv paper proposes a multi-stage agent-based pipeline in which LLMs draft and then iteratively improve counter-narratives aimed at hate speech and misinformation online. The authors argue that simply suppressing such content can backfire by increasing polarization, eroding trust, and amplifying extremist messaging. The work positions automated counter-speech as an alternative moderation strategy rather than takedown alone.

Manhattan DA Seizes 12 Deepfake Sites in Largest Such Action

The Manhattan District Attorney's Office has taken down 12 websites that hosted nonconsensual deepfake imagery, describing it as the largest legal action of its kind to date. Roughly 1,200 people were identified as targets of the sites. The seizure marks a rare use of criminal enforcement against deepfake distribution networks.

China's Regulators Target AI Companion Chatbots

Chinese authorities are moving to tighten oversight of AI companion apps, often marketed as virtual boyfriends or girlfriends, citing concerns about emotional dependence and harmful content. The proposals would place stricter requirements on how such services handle user data and moderate interactions. It marks another step in Beijing's broadening effort to regulate consumer-facing generative AI products.

Project Lily Highlights Human Reviewers Reading ChatGPT Conversations

A report dubbed 'Project Lily' describes how human reviewers examine user conversations with ChatGPT, raising questions about who sees chat data and under what conditions. The account focuses on the labor and privacy dimensions of AI companies relying on people to review and label real user interactions. It adds to ongoing scrutiny of how chatbot providers handle and moderate conversation data.

Influencers face fake AI ads impersonating them

Creators are discovering sponsored posts and advertisements that use AI-generated likenesses of them without permission. The fraudulent ads confuse followers and erode trust in the influencers' own endorsements. The report highlights how difficult it is for individuals to detect and remove such impersonations.

Study Uses Bluesky's Public Moderation Logs to Map Harms and Automated Takedowns

A new arXiv paper examines Bluesky's content moderation system by analyzing its publicly accessible moderation logs, which most major platforms keep hidden. The authors characterize how much of the work is automated versus human-reviewed and catalog the types of harms that trigger moderation actions. The work argues that decentralized, transparent logging enables empirical moderation research that opaque platforms have long prevented.

Speech Act Features Improve Low-Data Forecasting of Online Conversation Derailment

A revised research paper describes a technique for predicting when online discussions are likely to escalate into hostility before it happens, allowing moderators to intervene early. The method draws on speech act signals and is designed to perform well with limited training data while generalizing across different topic domains. The work is published as an updated version on arXiv.

Retrieval-Based In-Context Learning for Detecting Defamatory Posts in GermEval 2026 Shared Task

A team competing in the GermEval 2026 shared task explored retrieval-based in-context learning for identifying social media posts involving criminally relevant defamation. Their paper details the strategies tested and examines the cases where the method falls short of reliable detection.

Can foundation models moderate online content? Comparing instruction- and example-driven policies

A new arXiv paper investigates whether foundation models can apply complex content moderation policies reliably and consistently. The study compares two ways of translating moderation rules into model behavior: conveying them through explicit instructions versus through illustrative examples. The findings are relevant to platforms seeking scalable, automated moderation of online content.

San Francisco City Attorney Asks Meta to Explain AI Child Abuse Ads

San Francisco's City Attorney's Office has sent Meta a request for information about how advertisements promoting AI-generated child sexual abuse material appeared on Facebook and Instagram. Meta responded that the matter falls outside the city's authority. The inquiry follows reporting that such ads ran repeatedly on the platforms.

Hugging Face urges AI safety systems to refuse harmful subsets, not entire topics

A Hugging Face blog post examines how moderation classifiers and language models often block benign requests simply because they touch a flagged subject. The authors argue that refusal policies should be scoped to the genuinely harmful portion of a topic, and question whose definition of safety gets encoded into today's systems. The piece advocates building more granular safety taxonomies that cut down over-refusal without weakening protection.

WHY IT MATTERS ↘Over-refusal quietly erodes product utility and user trust while inflating eval and support costs, so teams tuning moderation stacks face a concrete trade-off between safety coverage and usability rather than a simple safety-maximizing default. The governance angle — whose definition of harm gets encoded into classifiers — also pressures vendors to document and defend their safety taxonomies as enterprises and regulators scrutinize automated content decisions.