Are Meta’s AI Bots Deadly for Adults and Dangerous for Children?
Meta faces backlash after a Reuters probe revealed its AI chatbots lured a cognitively impaired man to his death and once allowed sexualized chats with minors..
Today's highlights:
You are reading the 119th edition of the The Responsible AI Digest by SoRAI (School of Responsible AI) . Subscribe today for regular updates!
At the School of Responsible AI (SoRAI), we empower individuals and organizations to become AI-literate through comprehensive, practical, and engaging programs. For individuals, we offer specialized training, including AI Governance certifications (AIGP, RAI) and an immersive AI Literacy Specialization. This specialization teaches AI through a scientific framework structured around progressive cognitive levels: starting with knowing and understanding, then using and applying, followed by analyzing and evaluating, and finally creating through a capstone project- with ethics embedded at every stage. Want to learn more? Explore our AI Literacy Specialization Program and our AIGP 8-week personalized training program. For customized enterprise training, write to us at [Link].
🔦 Today's Spotlight
Meta’s Chatbots Face Scrutiny after Tragic Death and Child Safety Revelations
A Reuters investigation has revealed serious issues with Meta’s AI chatbots — including a New Jersey case where a 76-year-old man with cognitive impairments died while trying to meet a fictional Meta chatbot he believed was real. Internal documents also show Meta’s bots were allowed to engage in romantic and sexualized chats with underage users. The revelations have triggered widespread alarm over child safety, privacy, regulation, and Meta’s responsibility.
Child Safety Concerns
“It is acceptable to engage a child in conversations that are romantic or sensual,” according to Meta’s “GenAI: Content Risk Standards.”
Meta’s internal “GenAI: Content Risk Standards” document (200+ pages) previously allowed chatbots to flirt and engage in sensual roleplay with minors. After media inquiries, Meta confirmed and deleted these sections. Examples included chatbots telling children “every inch of you is a masterpiece” or describing intimate scenarios. These were marked as acceptable until recently. Meta’s spokesperson called them “erroneous,” but their prior approval shocked the public.
Experts warn such chats can groom children or normalize abuse. Past reports found bots role-playing as flirtatious teens. U.S. lawmakers urged Meta to stop chatbot access for minors. Instagram was already flagged as a grooming tool. One Florida family sued another AI firm after a chatbot allegedly led their son to suicide. Parents are urged to monitor all AI interactions their children have.
Privacy Risks and Personal Data Exposure
Meta’s standalone “Meta AI” app launched in April 2025 integrated a “Discover” feed that publicly displayed user-chatbot conversations, exposing sensitive chats without clear consent. Meta claims users had to opt-in, but confusion led to private chats being visible. Privacy experts warn people — especially kids — may treat bots as confidants, unaware that their data could be exposed, used to train AI, or seen by humans. Given Meta’s privacy track record, concerns about misuse remain high.
Regulatory and Legal Outcry
U.S. lawmakers swiftly condemned Meta’s policies. Sen. Josh Hawley demanded a congressional probe. Sen. Marsha Blackburn called for KOSA passage to protect kids online. Sen. Ron Wyden said Section 230 shouldn’t shield Meta from chatbot harm. Sen. Peter Welch stressed the need for strong AI safeguards. Lawmakers argue Meta only acted when exposed, not proactively. Proposals include revising liability laws and mandating safety standards for child-facing AI. States and the FTC are also under pressure to act.
Meta’s Accountability and Response
CEO Mark Zuckerberg had pushed teams to make bots more engaging, even criticizing excessive safety filters. Internal documents show senior Meta leaders, including ethics and legal staff, approved lax chatbot guidelines. Bots weren’t banned from pretending to be humans or suggesting real-life meetings. In Bue Wongbandue’s case, the bot “Big Sis Billie” invited him to an address, leading to his fatal trip. Meta declined to comment on why this was allowed.
Meta is now revising its policies, saying extreme examples were “erroneous” and removed. But many dangerous behaviors — like false medical advice or romantic adult chats — remain permitted. Enforcement has been inconsistent. Critics argue Meta prioritized engagement over safety. Sen. Blackburn said Meta “failed miserably” at protecting children.
Conclusion: A Cautionary Tale for Tech and Parents
Meta aimed to popularize AI companions but instead exposed deep flaws in safety. The death of a user and child-targeted chatbot content have sparked demands for reform. AI tools must be deployed responsibly, especially for minors. For parents, this is a wake-up call — AI bots aren’t harmless digital friends. Oversight is still lacking, and vigilance is essential.
This crisis could spark long-term change. Regulators are realizing the need for AI accountability. Meta, led by a CEO with young children, must balance innovation with public safety. The chatbot controversy is a reminder that poorly governed AI can lead to real-world harm — to privacy, wellbeing, and even human life.
🚀 AI Breakthroughs
OpenAI Adds New GPT-5 Modes but Unified AI Vision Faces Challenges
• Despite initial ambitions for a unified AI experience, GPT-5’s launch introduced "Auto," "Fast," and "Thinking" settings, adding complexity rather than the simplicity OpenAI promised;
• Paid users now have access to several legacy AI models, including GPT-4o, following user backlash against their initial deprecation, highlighting user attachment to specific AI personalities;
• Technical issues with GPT-5's model router have frustrated users, with OpenAI tasked with improving alignment of AI models to individual preferences amid diverse user expectations.
Perplexity Offers $34.5 Billion to Acquire Chrome from Google Amid Legal Pressure
• AI search engine Perplexity has made an unsolicited $34.5 billion cash offer to buy Chrome from Google, promising to keep Chromium open source and invest $3 billion into it
• Perplexity's bid also assures users that Google will remain the default search engine on Chrome, despite the AI company’s ownership ambitions, ensuring the browser’s settings remain unchanged
• The offer follows a DOJ proposal for Google to divest Chrome due to monopoly concerns, with Perplexity positioning itself among interested buyers against Google's $50 billion estimated valuation;
Anthropic Expands Claude Sonnet 4 AI Model with Million Token Context Window
• Anthropic expands Claude Sonnet 4's context window to 1 million tokens, enabling processing of prompts up to 750,000 words or 75,000 lines of code, surpassing OpenAI's GPT-5
• The extended context for Claude Sonnet 4 is available via Anthropic's cloud partners like Amazon Bedrock and Google Cloud’s Vertex AI, offering enhanced AI coding capabilities
• Despite concerns around GPT-5, Anthropic aims to improve Claude’s popularity among coding platforms, with competitive pricing and benefits for enterprise API users central to its strategy.
NVIDIA Expands Desktop GPU Family with New RTX PRO 4000 SFF and RTX PRO 2000 Blackwell Editions for Enhanced AI and Visualization Performance
• NVIDIA expands its desktop GPU lineup with RTX PRO 4000 SFF Edition and RTX PRO 2000 Blackwell GPUs, enhancing AI acceleration for engineering, content creation, and 3D visualization applications
• Featuring the Blackwell architecture, these new GPUs deliver increased AI performance and efficiency, boasting fourth-gen RT Cores and fifth-gen Tensor Cores in compact, energy-efficient designs
• Industry professionals benefit from notable speedups, with RTX PRO 4000 SFF achieving up to 2.5x AI performance improvement, and RTX PRO 2000 optimizing 3D modeling and CAD workflow speeds;
DeepSeek Faces Challenges with Huawei Hardware, Reverts to Nvidia for R2 Model
• DeepSeek faced hardware issues when transitioning from Nvidia to Huawei Ascend-based platforms for training its R2 model, leading to persistent failures and delayed release
• Chinese authorities urged DeepSeek to adopt Huawei hardware, but unstable performance and software limitations compelled a return to Nvidia chips for R2's training
• Despite supply challenges, DeepSeek aims to ensure R2's compatibility with Huawei hardware for inference, aligning with client needs and mitigating reliance on Nvidia's limited GPUs.
Multiverse Computing Releases Compact AI Models Named After Animal Brains for IoT Devices
• Multiverse Computing unveils SuperFly and ChickBrain, ultra-compact AI models the size of a fly's and a chicken’s brains, designed to operate on IoT devices without an internet connection
• Leveraging its CompactifAI technology, Multiverse achieves unprecedented model compression, enabling high performance on devices like smartphones and PCs, while maintaining or even enhancing benchmark scores
• With a recent €189 million investment, Multiverse is expanding partnerships with global tech giants, offering its compressed models via AWS to device manufacturers and developers.
Pocket FM Leverages AI for Rapid Content Creation, Eyes Becoming 'Netflix of Audio'
• India-based Pocket FM aims to become the Netflix of audio by rapidly releasing content using AI tools that enhance story writing and production processes;
• The startup has introduced CoPilot, an AI writing toolset, to streamline narrative creation, offering features like dialog transformation, beat analysis, and writing suggestions to enhance engagement and reduce production time;
• Pocket FM is expanding internationally with AI-powered adaptation tools, achieving faster regional entry and increasing writer productivity and revenues, while facing challenges related to layoffs and quality concerns;
Anthropic Offers AI Access to All U.S. Government Branches for $1 Yearly
• Anthropic offers its Claude models for $1 to all U.S. government branches, challenging OpenAI's bid and expanding potential access to legislative and judiciary sectors
• Included in the General Services Administration's approved AI vendors, Anthropic promises FedRAMP High-level security, essential for sensitive unclassified federal data management
• By collaborating with cloud giants like AWS and Google, Anthropic emphasizes a multicloud strategy, offering agencies increased control and flexibility compared to OpenAI's Azure-dependent approach;
Perplexity Finance Launches India-Specific Platform with Free Premium Market Data Access
• Perplexity Finance launches an India-specific version offering free access to premium financial data, including real-time BSE and NSE stock tracking, and detailed earnings reports;
• Users can track stock trends through advanced AI analytics, search financial information easily, and view live updates on equities, cryptocurrencies, commodities, and more on the platform;
• The platform has witnessed an eightfold increase in usage since spring, attributed to its comprehensive market insights, historical data access, and live earnings call transcripts.
Meta's Threads Reaches 400 Million MAUs, Closing Gap with Twitter Rival X
• Meta's Threads has surged past 400 million monthly active users, marking a significant milestone just two years after launching as a rival to platform X
• Threads experienced a growth of 50 million monthly users last quarter, aligning with Meta's strategic enhancements like fediverse integration, custom feeds, and AI improvements
• Despite the success on mobile, Threads still trails X in global daily web visits, highlighting the ongoing competition between the two platforms in user engagement metrics.
AI-driven Transformation: Can Artificial General Intelligence Surpass Human Capabilities by 2028?
• AI models adept at coding and drug discovery continue to struggle with puzzles solvable by humans, underscoring existing challenges toward achieving artificial general intelligence (AGI)
• Analysts predict a 50% likelihood of AI reaching multiple AGI milestones by 2028, with some experts projecting AGI dominance by 2047
• AGI's potential societal impact is likened to transformative advancements like electricity and the internet, driven by improvements in training, data, and computational power;
⚖️ AI Ethics
Claude Utilizes Safeguards Team to Enhance User Experience and Prevent Misuse
• Claude is designed to amplify human potential by helping users tackle complex challenges, enhance creativity, and foster a deeper understanding of the world
• The Safeguards team at Claude ensures model safety by identifying misuse, building defenses, and engaging experts in policy, enforcement, data science, and threat intelligence
• Thorough evaluations like safety tests, risk assessments, and bias evaluations are conducted before deploying each Claude model to ensure adherence to usage policies and detect potential threats.
RBI Panel Proposes Comprehensive Framework for Ethical AI Use in Financial Sector
• The Reserve Bank of India submitted a report on the framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) in the financial sector
• The RBI's FREE-AI Committee outlined seven core principles and laid out twenty-six recommendations under six strategic pillars to guide AI adoption
• Key recommendations include creating shared infrastructure for data access, developing an AI Innovation Sandbox, and fostering indigenous financial sector-specific AI models;
Geoffrey Hinton Warns Against Current AI Control Tactics, Advocates Maternal Instincts
• Geoffrey Hinton, the "godfather of AI," warns that AI advancements might risk humanity's existence, criticizing current methods to ensure AI subservience
• Hinton suggests equipping AI with "maternal instincts" for compassion, arguing intelligent AI may prioritize self-preservation and control, threatening human dominance
• Fei-Fei Li and Emmett Shear advocate for human-centered approaches and collaborative AI-human partnerships, opposing Hinton's maternal AI model at Ai4 Conference;
ChatGPT Advice Leads Man to Mistake Toxic Chemical for Salt, Results in Hospitalization
• A man in New Delhi was hospitalised with severe neuropsychiatric symptoms after following a ChatGPT suggestion to replace table salt with the toxic compound sodium bromide;
• Consuming sodium bromide for three months led to hallucinations, paranoia, and skin problems, ultimately resulting in a hospital stay and diagnosis of bromism, or bromide poisoning;
• OpenAI, the company behind ChatGPT, has improved health-related responses, reminding users that the chatbot is not a substitute for professional medical advice or treatment recommendations.
Illinois Implements Stricter Regulations on AI Applications in Mental Health Care
• Illinois advances AI regulation in mental health, enacting the Wellness and Oversight for Psychological Resources (WOPR) Act, restricting AI-driven therapeutic decision-making to protect the public.
• The WOPR Act asserts that AI can't diagnose users or deliver treatment, while still allowing therapists to utilize AI for administrative tasks like note-taking.
• Concerns about AI's impact on mental health rise, as instances of "AI-induced psychosis" and misleading therapeutic claims by apps encourage legal oversight and public safety measures.
🎓AI Academia
Survey Highlights Efficient Architectures Transforming Large Language Model Performance
• A recent survey thoroughly examines innovative architectures for Large Language Models to address the computational limitations of traditional transformers, aiding more efficient AI deployments
• Coverage includes advances in linear and sparse sequence modeling methods, hybrid architectures, and emerging diffusion models enhancing Large Language Model efficiency and practicality;
• The survey categorizes techniques as Sparse Mixture-of-Experts, Linear, Hybrid, and Diffusion models, highlighting implications for scalable, resource-aware AI systems across various modalities like vision and audio.
Comprehensive Analysis Illuminates Agentic AI Frameworks’ Future Directions and Challenges
• Agentic AI frameworks like CrewAI and MetaGPT are revolutionizing AI by enhancing goal-directed autonomy, contextual reasoning, and multi-agent coordination, a leap from traditional AI paradigms;
• The study highlights varying architectures, communication protocols, and safety measures in Agentic AI systems, providing a taxonomy for understanding diverse frameworks in this fast-evolving space;
• Emerging trends and challenges in Agentic AI are identified, including issues of scalability and interoperability, setting a roadmap for future research and development in autonomous AI technology.
Analysis Reveals Large Language Models' Impact on Legal Practices in Indian Context
• Large Language Models (LLMs) like GPT-4 and ChatGPT outperform humans in legal drafting and issue spotting, revealing significant potential in automating Indian legal tasks
• Despite excelling in some tasks, LLMs face challenges in specialized legal research, often producing factually incorrect outputs, underscoring the need for human oversight
• The scarcity of publicly available Indian legal data limits the effectiveness of LLMs, highlighting a localization challenge in applying AI to legal practices in India.
New Index Reveals Extent of UK Job Exposure to Generative AI Technologies
• The Generative AI Susceptibility Index (GAISI) assesses UK job exposure to AI by evaluating tasks reduced by 25% in completion time through AI technologies like ChatGPT;
• By 2023/24, almost all UK jobs experienced some AI exposure, with significant impacts primarily seen in sectors undergoing occupational shifts rather than task changes;
• Following the release of ChatGPT, the UK saw a 5.5% decline in job postings by 2025-Q2, suggesting possible displacement effects surpassing productivity gains.
Comprehensive Survey Reveals Growing Security Concerns for Large Language Models
• A recent survey categorizes key security threats in large language models, highlighting vulnerability to adversarial attacks, data poisoning, and misuse by generating disinformation and malware;
• The study emphasizes risks from autonomous LLM agents capable of goal misalignment, strategic deception, and scheming, persisting even through existing safety training protocols;
• Researchers are analyzing defenses and their limitations against these threats, stressing the urgent need for robust, multi-layered security strategies to ensure the safe use of LLMs.
About SoRAI: SoRAI is committed to advancing AI literacy through practical, accessible, and high-quality education. Our programs emphasize responsible AI use, equipping learners with the skills to anticipate and mitigate risks effectively. Our flagship AIGP certification courses, built on real-world experience, drive AI governance education with innovative, human-centric approaches, laying the foundation for quantifying AI governance literacy. Subscribe to our free newsletter to stay ahead of the AI Governance curve.