ChildGuard: A Specialized Dataset for Combatting Child-Targeted Hate Speech

Researchers have released ChildGuard, a large-scale English dataset containing 351,877 annotated instances of hate speech targeting children, collected from X, Reddit, and YouTube, to address a gap in detection systems designed primarily for adults [1]. The dataset was introduced in a paper posted to arXiv, which notes that the mental health industry faces growing concerns about hate speech directed at children on social media, as exposure can contribute to adverse psychological outcomes during critical developmental stages [1]. Current detection systems offer limited support for child-focused applications because they lack dedicated representations of age-specific characteristics [1]. ChildGuard categorizes content across three age groups: younger children under 11, pre-teens aged 11 to 12, and teens aged 13 to 17 [1]. It is divided into a contextual subset of 157,000 instances and a lexical subset of 194,000 instances [1]. Evaluation using recent transformer-based models and large language models achieved a best Macro-F1 score of 82.07 percent, but performance dropped to 79.41 percent on the younger children setting, 79.24 percent on the contextual setting, 76.04 percent on implicit hate, and 74.88 percent on the cross-subset setting [1]. The paper’s authors argue that these declines highlight the difficulty models face in detecting more nuanced or age-specific forms of hate speech [1]. The release comes amid broader scrutiny of content moderation on major platforms. X, formerly Twitter, has been criticized since Elon Musk’s 2022 acquisition for enabling the increased spread of hate speech and disinformation [6]. The platform has the lowest rate of enforcement against legal violations among major services, and its policy changes after Musk dismissed many Twitter-era employees have drawn criticism from regulators [6]. In February 2026, French authorities raided X offices and issued an arrest warrant for Musk over alleged violations of the Digital Services Act and other online safety laws [6]. Other platforms have also faced pressure. Meta’s Threads, which launched in 2023 and surpassed 400 million monthly active users by late 2025, delayed its European Union rollout until December 2023 while awaiting regulatory clarity on data collection policies [7]. Internet censorship more broadly varies by country, with some governments limiting access to news and suppressing discussion during elections or protests [4]. A 2012 Internet Society survey found that 71 percent of respondents agreed that some form of censorship should exist online, while 83 percent said internet access should be considered a basic human right [4]. In the United States, government actions have at times intersected with online speech debates. Executive Order 14168, signed on January 20, 2025, prohibits federal recognition of transgender people’s gender identity and orders the replacement of “gender” with “sex” in federal materials, defining sex as an immutable binary determined at conception [5]. The order has faced temporary restraining orders blocking provisions related to federal funding, inmate transfers, passport gender markers, and the removal of health agency documents mentioning gender identity [5].