Social media and AI chatbots fail to deliver promised digital safety

Researchers at the International Computer Science Institute (ICSI) have released two studies examining critical gaps between technological promises and real-world impacts on vulnerable users. Presented at the 29th ACM Conference on Computer-Supported Cooperative Work and Social Computing in Salt Lake City from Oct. 1–14, 2026, the work highlights deficiencies in how social media platforms communicate youth safety measures and how AI chatbots support family caregivers. Principal research scientist Pamela Wisniewski led the investigations, which underscore the urgent need for transparent platform accountability and nuanced AI design. The first study analyzed over 350 press releases and safety-related blog posts from YouTube, TikTok, Meta, and Snapchat over a five-year period. Researchers found that platforms frequently employed selective, vague, and unverified language when describing youth safety features. Communications often emphasized that tools were being tested or were region-specific, obscuring actual availability. Furthermore, claims of improved efficacy lacked baselines or measurable outcomes. By framing safety primarily as a matter of individual behavior and parental oversight, platforms effectively shifted responsibility away from structural design flaws and algorithmic practices. The findings advocate for stricter guidelines requiring platforms to clearly disclose feature functionality, access protocols, and independent verification of effectiveness. In a separate investigation focused on health technology, researchers evaluated how AI chatbots compared to human clinicians and peer caregivers when addressing real questions from Alzheimer’s disease caregivers. Drawing from 85 queries submitted to ALZConnected, the team assessed responses across linguistic, emotional, and clinical dimensions. While chatbots scored highly on politeness and emotion-support markers, their outputs remained formal, verbose, and narrowly advisory. Notably, fine-tuning a model on peer-support discussions did not yield a meaningful improvement in human-like interaction. In contrast, human responders provided context-rich narratives, addressed broader emotional, legal, and financial concerns, and demonstrated lived experience that AI systems consistently missed. The research warns against mistaking polished AI responses for substantive support, particularly in high-stakes caregiving scenarios where emotional nuance and contextual awareness are paramount. Both studies converge on a central theme: the growing disconnect between technology design and user vulnerability. As digital tools become increasingly embedded in social and health ecosystems, researchers emphasize that accountability must be measured by actual user outcomes rather than marketing narratives. Wisniewski will further address these themes at an upcoming National Academies of Sciences, Engineering, and Medicine workshop on digital technologies and brain development, scheduled for Oct. 26. The proceedings from both studies are expected to inform regulatory discussions, platform policy revisions, and the development of more transparent AI systems tailored to complex human needs.