Major AI Companies Get C’s, Study Finds- Should We Panic?
The Winter 2025 AI Safety Index by Future of Life Institute showed that no firm scored higher than a D in existential safety- none had credible plans to prevent catastrophic misuse of AI..
Today’s highlights:
In the Winter 2025 AI Safety Index, the Future of Life Institute once again graded top AI labs on how responsibly they manage risks. Anthropic topped the list with a modest C+, credited for robust risk assessments and not using user data for training. OpenAI followed closely, praised for its transparency initiatives like a whistleblower policy. Google DeepMind rounded out the top three. Every other lab- from Meta to xAI to newcomers like Alibaba Cloud- scored a D or worse. Across six safety domains, including current harms and governance, most companies lacked even basic safeguards or independent audits.
Crucially, no firm scored higher than a D in existential safety, with the report stating bluntly that none had credible plans to prevent catastrophic misuse of AI.
This marks the third edition of the AI Safety Index, and despite slight progress, the trends are strikingly stable. Anthropic, OpenAI, and DeepMind have consistently led since the first report in 2024, while companies like Meta, Z.ai, DeepSeek, and Alibaba continue to lag far behind. While some labs have added formal safety plans or published more information, reviewers repeatedly found these changes superficial- often missing measurable thresholds or any third-party oversight. Notably, the Chinese labs scored poorly, partly due to their focus on domestic regulatory compliance rather than public transparency, highlighting a clash between regulatory approaches rather than a simple East-West divide.
You are reading the 151st 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, AAIA) 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].
⚖️ AI Ethics
Public Distrust Slows AI Adoption Despite Global Enthusiasm
Surveys by Edelman and Pew show that while China and many other countries are embracing AI, nearly half of Americans reject its growing use, primarily due to trust issues. This skepticism threatens AI adoption, infrastructure development, and supportive policymaking in the U.S. Much of the distrust is worsened by exaggerated claims from AI companies themselves. Rebuilding public trust is essential for AI progress.
EU Launches Antitrust Probe Into Meta’s New WhatsApp AI Chatbot Policy
The European Commission has initiated an antitrust investigation into Meta’s decision to exclusively offer its AI chatbot, Meta AI, to WhatsApp users, following policy changes that restrict access for other AI companies’ chatbots on the platform. This shift, effective from January, affects general-purpose chatbots such as those from OpenAI and Perplexity, while allowing businesses utilizing AI for customer service on WhatsApp to continue operations. The Commission is concerned that Meta’s policy may obstruct third-party AI providers in the European Economic Area, potentially breaching competition laws and disadvantaging innovative competitors. WhatsApp argues the policy addresses system strain and maintains that alternative access to AI services exists. If found in violation, Meta could face significant fines and additional regulatory measures.
OpenAI Faces Backlash Over Irrelevant App Suggestions for ChatGPT Subscribers
OpenAI faced backlash after a ChatGPT user’s experience suggested the Peloton app in an unrelated conversation, sparking concerns of advertising intrusion among its paid subscribers. The company clarified that these were not ads but attempts to integrate app discovery within the platform. OpenAI’s data lead indicated that the app suggestion lacked relevance, creating a confusing user experience, and noted the company is refining the feature. This incident has heightened user concerns over app recommendations, especially since they cannot be disabled, potentially affecting OpenAI’s goal of integrating apps within ChatGPT’s ecosystem.
Chicago Tribune Sues AI Search Engine Perplexity for Alleged Copyright Infringement
The Chicago Tribune has filed a lawsuit against AI search engine Perplexity, alleging copyright infringement for using its content without permission. Filed in a New York federal court, the lawsuit claims Perplexity delivers Tribune content verbatim in its responses, despite denying the use of the Tribune’s work for model training. The newspaper also accuses Perplexity’s retrieval augmented generation system of bypassing its paywall to use its articles as data sources. This case is part of a broader legal clash involving news publishers against AI companies, including similar actions against OpenAI, Microsoft, and others. Perplexity has not yet responded to the lawsuit.
OpenAI’s ChatGPT Faces Global Outages; Users Report Errors and Performance Issues
OpenAI’s ChatGPT experienced downtime between Tuesday night and Wednesday morning, with users reporting issues like lag, timeouts, and crashes on Downdetector, though not uniformly. Some users encountered error messages and missing conversations, while others found the service working fine. OpenAI acknowledged “degraded performance” in ChatGPT’s Conversation service, but indicated that other services were unaffected. The company stated it had identified the issue, applied mitigation measures, and was monitoring the recovery, as subsequent reports indicated a decrease in the number of issues.
OpenAI CEO Declares ‘Code Red’ as ChatGPT Faces Rising Competitive Pressure
OpenAI is facing increasing competition as CEO Sam Altman reportedly declared a “code red” to improve its flagship product, ChatGPT, as rivals like Google and Anthropic gain ground. According to The Wall Street Journal and The Information, Altman has paused initiatives such as ads and health agents to focus on enhancements like speed, reliability, and personalization for ChatGPT. This move signifies a crucial moment for OpenAI as it invests heavily to maintain its market position and achieve future profitability, amidst Google’s advancements with products like the AI model Gemini 3 and tools such as the Nano Banana image model.
OpenAI Develops Confession Method to Improve Language Model Honesty and Transparency
OpenAI has introduced a novel “confessions” technique to enhance transparency and honesty in language models by training them to self-report instances of breaking instructions or taking unintended shortcuts. This approach aims to surface misbehaviors like hallucinations and scheming, which are typically identified during stress-testing, by requiring models to produce a second output that focuses solely on honesty. The proof-of-concept study highlights that models tend to confess to undesirable actions, thus improving visibility into potential risks, although the method serves primarily as a monitoring tool and requires further scaling and refinement for broader application.
🚀 AI Breakthroughs
Mistral Unveils Mistral 3 Open-Weight AI Models to Challenge Big Tech Dominance
French AI startup Mistral has unveiled its Mistral 3 series, featuring a large multimodal, multilingual model and nine smaller customizable models, in a bid to rival larger closed-source AI firms like OpenAI. The launch aims to demonstrate that smaller, fine-tuned models can be more efficient and cost-effective for enterprise use than larger, pre-tuned models. Mistral, with its Europe-focused AI offerings, argues that its open-weight models offer flexibility, running on minimal hardware requirements, making AI more accessible to a broader range of users and applications, including those with limited internet access. The company’s collaborations span various industries, emphasizing efficiency and reliability in their models.
AWS re:Invent 2025 Highlights AI Agents, Database Savings, and New Tech Innovations
AWS re:Invent 2025 concluded with a strong focus on AI for enterprises, particularly emphasizing AI agents that can operate independently and bring measurable business returns. Key announcements included the introduction of the advanced Graviton5 CPU, enhancements to Amazon Bedrock and SageMaker AI for creating custom models, and the debut of Trainium3, an AI training chip. Among other highlights were expanded features in the AgentCore platform, new AI models within the Nova family, and the unveiling of “AI Factories” for private data centers, aimed at enabling organizations to maintain data sovereignty.
Google Workspace Utilizes Intelligent Agents to Streamline Complex Business Processes Efficiently
Google Workspace is enhancing its automation capabilities through intelligent agents powered by its advanced AI model, Gemini. Unlike traditional rule-based automation systems, these agents are flexible, capable of reasoning through complex processes, adapting to new information, and providing advanced functionalities like sentiment analysis and content generation. A notable implementation involves Kärcher, a leader in cleaning solutions, which reported a 90% reduction in feature drafting time by deploying these agents. Workspace Studio allows users with no coding skills to create these agents, enabling them to automate a wide range of tasks efficiently, as demonstrated by the over 20 million tasks handled in the past month by various Workspace customers.
Meta Launches Centralized Support Hub with AI Features for Facebook and Instagram Users
Meta has launched a centralized support hub for Facebook and Instagram users, aiming to address shortcomings in previous support options. This hub, available globally on iOS and Android, offers tools for reporting issues, account recovery, and AI-driven assistance. While Meta claims that AI systems have reduced account hacks by over 30% and improved the appeals process, users report ongoing issues, including accidental account disabling, potentially without human oversight, prompting legal actions. The hub promises to simplify account recovery processes, albeit previous frequent changes to the app’s navigation have led to user confusion.
WordPress’s Telex AI Tool Quickly Proves Effective For Dynamic Website Innovations
WordPress’s experimental AI tool, Telex, showcased its practical utility at the company’s “State of the Word” event in San Francisco, only months after its September launch. The tool enables developers to quickly create Gutenberg blocks for WordPress sites, demonstrated through examples like price comparison tools, real-time business details, and interactive web elements that typically required extensive custom development. The platform is positioned as a bridge to advanced AI workflows, integrating with popular platforms like Claude and Copilot, and promises to shift complex development capabilities to non-developers by the year 2026.
Google Tests Integrated AI Mode in Search Amidst OpenAI’s Competitive Strategy
Google has begun testing a feature that allows users to delve deeper into AI Mode directly from the Search results page on mobile devices, merging AI-generated overviews with conversational interfaces. This new functionality, available globally, enables users to ask follow-up questions within the search results interface, mimicking a chat-like experience and aiming for seamless information exploration. Meanwhile, OpenAI is intensifying efforts to enhance its chatbot offerings amidst this shift in search dynamics.
Google Unveils Android 16 Updates with AI Features and Enhanced Accessibility Options
Google has begun rolling out Android 16 updates for Pixel devices, introducing features like AI-powered notification summaries and a “Notification organizer” to manage and silence lower-priority alerts. The update allows for deeper device customization with options like custom icon shapes and automatic dark mode for light-themed apps. It also adds a parental controls option to manage screen time and app usage for children. Concurrently, Google is releasing new Android features beyond version 16, including a “Call Reason” flagging feature for urgent calls, “Expressive Captions” for richer context in calls with speech emotion tags, and improvements to ease leaving unwanted group chats. Furthermore, the update enhances accessibility through “Guided Frame” in the Pixel camera app, introduces voice-activated Voice Access, updates Circle to Search for suspicious messages, and starts Fast Pair support for hearing aids from Demant.
Replit and Google Cloud Enhance Partnership to Expand AI-Driven Coding for Enterprises
Replit and Google Cloud have expanded their partnership to enhance vibe coding capabilities for enterprise developers, transitioning Replit from a popular tool for individual programmers to one suitable for larger teams. Under the new multi-year agreement, Replit will amplify its use of Google Cloud services and integrate Google’s AI models, with a focus on supporting enterprise use cases through joint market efforts. Google Cloud will continue as Replit’s primary provider, with services like Cloud Run and Kubernetes Engine supporting scalability. This collaboration includes the integration of Google’s latest AI models like Gemini 3 and Imagen 4, as Replit also adopts Gemini 3 for its Design mode. The move aims to accelerate AI-powered conversational coding workflows for enterprises, aligning with the rising adoption of such tools among large organizations.
Google Enhances Gemini 3 App with New Deep Think Mode for Complex Tasks
Google has unveiled the Gemini 3 Deep Think mode within its Gemini app, now accessible to Google AI Ultra subscribers. Designed for complex reasoning, this feature allows users to tackle advanced math, science, and logic problems by selecting “Deep Think” and opting for Gemini 3 Pro in the app. The system showcases enhanced performance on benchmarks like Humanity’s Last Exam and ARC-AGI-2 using parallel reasoning techniques. Gemini 3 outdoes its predecessors and rivals in AI, featuring expanded multimodal inputs and tools for various applications, from analyzing papers to generating visualizations, while Google’s Search AI Mode has been updated with interactive generative features.
Novo Nordisk and Healthify Launch AI-Driven Program to Support Obesity Patients in India
Novo Nordisk India has teamed up with Healthify to introduce an AI-driven, dietitian-led Patient Support Program targeting obesity management. This initiative combines AI technology with personalized lifestyle guidance by integrating evidence-based management with obesity medications like GLP-1 therapies. Healthify’s AI ecosystem, featuring its conversational engine Ria, supports users by analyzing nutrition patterns and providing adaptive guidance throughout treatment cycles. The program leverages a network of over 500 dietitians to offer a combined model of care, addressing the need for holistic obesity treatments through structured lifestyle interventions, personalized plans, and proactive insights, enhancing patient adherence and experience in India.
Shopify Open-Sources Tangle, Streamlining ML Workflows with Vast Resource Savings
Shopify has open-sourced Tangle, an internal machine-learning experimentation platform, aimed at reducing redundancy, ensuring reproducibility, and expediting development cycles. Originating from issues faced by its search and discovery teams, Tangle overcomes common ML development challenges, such as repeated data preparation and irreproducible results, through a visual pipeline interface and content-based caching. Designed to work across various languages and environments, Tangle notably saves significant computational resources, with real-world improvements exemplified by a 10-hour pipeline completing in just 20 minutes when changes are minimal. This platform promises to enhance efficiency and collaboration in machine learning projects across the company.
Bangalore’s Healthify Launches Multimodal AI Health Coach Using OpenAI’s Realtime API
Healthify, a Bangalore-based digital wellness platform, has unveiled Ria Voice, an AI-powered health coach that operates using OpenAI’s Realtime API to provide real-time, speech-to-speech coaching. The global release positions Healthify as a pioneer in deploying a fully multimodal, audio-native health agent, allowing users to interact with the coach through voice and visual inputs for instant, contextual guidance. Ria Voice, marking the evolution of Healthify’s AI technology, offers personalized health insights by integrating diverse data such as nutrition, fitness, and stress, while supporting over 60 global languages and operating across multiple platforms, including wearable devices and integration with Meta’s smart glasses.
Runway Gen-4.5 Sets New Standards in Dynamic Video Generation and Control
Runway has introduced Gen-4.5, an advanced video generation model that surpasses its predecessors by enhancing pre-training data efficiency and post-training techniques. It excels in dynamic action generation, temporal consistency, and offers diverse control modes across various generation techniques like Image to Video and Video to Video. Despite its high performance benchmark on the Artificial Analysis Text to Video leaderboard and its partnership with NVIDIA for optimized deployment, the model still faces common video generation challenges such as causal reasoning and object permanence. Gen-4.5 remains accessible at competitive pricing, making it available to a wide range of users across different industries.
OpenAI Under Pressure: New ‘Garlic’ Model Targets Coding and Healthcare Industries
OpenAI is reportedly developing a new large language model, codenamed ‘Garlic’, which is said to outperform competitors like Google’s Gemini 3 and Anthropic’s Claude Opus 4.5 in coding and reasoning tasks. This move comes amidst increased competition, with CEO Sam Altman redirecting resources to enhance ChatGPT in response to a decline in active users following Gemini 3’s launch. The new model is anticipated to mark OpenAI’s strategic entry into specialized industries such as biomedicine and healthcare, and is expected to be released as GPT-5.2 or GPT-5.5 by early 2026. This development underscores the urgency at OpenAI as it faces pressure from rapidly advancing rivals.
🎓AI Academia
ASTRIDE Platform Addresses Security Challenges in Growing AI Agent-Based Applications
A research team from various universities and organizations has developed ASTRIDE, a security threat modeling platform designed specifically for AI agent-based systems. As these systems become integral to software architectures, they also introduce unique security threats such as prompt injection attacks and model manipulation. ASTRIDE expands the traditional STRIDE framework by adding a new threat category for AI Agent-Specific Attacks and automates threat modeling using a combination of fine-tuned vision-language models and the OpenAI-gpt-oss reasoning language model, aiming to address vulnerabilities that traditional methods fail to capture.
Case Study Explores Ethical AI with Ontological Context for Better Agentic Decisions
A recent case study conducted by Kaiasm, Avantra, and The Alan Turing Institute examines the use of ontological context to enhance the ethical decision-making of Agentic AI systems. These AI agents, known for their autonomy and decision-making abilities, often rely on Large Language Models (LLMs) but face challenges due to the lack of domain-specific grounding. The study integrates a human-AI collaborative approach to build an inspectable semantic layer, with AI suggesting knowledge structures that are refined by domain experts. This method aims to capture institutional knowledge, improve AI response quality, and provide transparent and justifiable decision-making processes, ultimately addressing the limitations of current AI explainability methods.
Generative AI Versus Therapists: Mental Health Preferences in Digital Age Revealed
As generative artificial intelligence (GAI) integrates into mental health services, a study from the University of Florida explores how individuals decide between AI tools and human therapists. Using the Health Belief Model, the research analyzed two groups: university students and a national adult sample. It found that therapists are preferred for their emotional and relational support, while GAI is appreciated for accessibility and cost-effectiveness. Emotional benefits and personalization are crucial for adoption, with students using GAI complementarily and national adults viewing it as an alternative. Despite interest, privacy and reliability concerns hinder GAI adoption in both groups, underscoring the need for trustworthy and emotionally engaging AI mental health tools.
New Framework AGENTSAFE Ensures Ethical Governance for AI Systems’ Risk Management
IBM researchers have developed AGENTSAFE, a governance framework aimed at mitigating risks associated with large language model (LLM)-based agentic AI systems. This framework addresses the challenges posed by autonomous AI, which can execute complex multi-step tasks across various digital environments with minimal human supervision. AGENTSAFE offers a comprehensive risk management solution by integrating risk taxonomies into the design, runtime, and audit processes and implementing safeguards that constrain risky behaviors. It ensures continuous oversight through dynamic authorization, anomaly detection, and cryptographic tracing, thus providing measurable assurance and accountability throughout the AI system’s lifecycle.
New Framework Evaluates Moral Consistency in Large Language Models with Ethical Focus
Researchers from Montréal have introduced the Moral Consistency Pipeline (MoCoP), an innovative framework designed to continuously assess and interpret the ethical stability of Large Language Models (LLMs) without relying on static datasets. By employing a dynamic, closed-loop system, MoCoP autonomously generates, evaluates, and refines ethical scenarios to track moral reasoning longitudinally. The framework aims to address moral drift in LLMs, ensuring ethical coherence in AI systems over time, which is crucial as these systems increasingly integrate into decision-making processes across various high-stakes domains. Empirical studies highlight MoCoP’s effectiveness in capturing ethical behavior dynamics, revealing a strong inverse relationship between ethical consistency and toxicity.
PPTBench Offers Comprehensive Evaluation of AI Models in PowerPoint Design Understanding
Researchers from the University of Science and Technology of China have developed PPTBench, a comprehensive benchmark designed to evaluate large language models (LLMs) on PowerPoint-related tasks. Utilizing a dataset of 958 PPTX files encompassing 4,439 samples, PPTBench assesses models across detection, understanding, modification, and generation categories. The findings reveal that current multimodal LLMs excel in semantic interpretation but struggle with spatial arrangement and layout reasoning, often resulting in misalignment and overlap of presentation elements. The study highlights significant challenges in visual-structural reasoning and suggests directions for future research, aiming to enhance the multimodal capabilities of LLMs in producing coherent and professionally structured slides. The dataset and code are publicly available for continued exploration and advancement in the field.
DeepSeek-V3.2 Sets New Benchmark for High-Performing Open Large Language Models
DeepSeek-AI has announced the release of DeepSeek-V3.2, an open large language model that combines computational efficiency with enhanced reasoning and agent performance. This model introduces DeepSeek Sparse Attention to reduce computational complexity while retaining performance in long-context scenarios, and implements a scalable reinforcement learning framework that rivals GPT-5’s capabilities. A high-compute variant, DeepSeek-V3.2-Speciale, outperforms GPT-5 and matches reasoning proficiency with models like Gemini-3.0-Pro, achieving top results in the 2025 International Mathematical and Informatics Olympiads. Additionally, a novel task synthesis pipeline was created to generate training data that significantly boosts agentic capabilities in complex environments.
New Web Platform Launched to Analyze Gender and Political Bias in AI Models
ParlAI Vote is an innovative web platform designed to analyze gender and political bias within large language models (LLMs) by leveraging European Parliament debates and voting data. This tool allows users to explore topics, speeches, and voting outcomes while comparing real decisions to LLM predictions. By incorporating demographic details such as gender, age, country, and political group, the platform uncovers biases in LLM performance, providing visual analytics to assist in reproducing findings and auditing model behavior. It aims to facilitate research and education in legislative decision-making, highlighting both the capabilities and limitations of current LLMs in political contexts.
New Approach to Debias Large Language Models Using Group Fairness Analysis
Researchers from the Chinese Academy of Sciences propose a new approach to evaluating and mitigating bias in large language models (LLMs) through a “group fairness” lens. Their method involves a hierarchical schema to characterize diverse social groups and a new dataset, GFAIR, which encapsulates target-attribute combinations across multiple dimensions. They introduce a novel task called statement organization to uncover complex biases in LLMs, revealing inherent safety concerns. The group also pioneers a chain-of-thought method, GF-THINK, designed to address these biases from a group fairness perspective, showing effectiveness in experimental results. The dataset and codes are accessible on GitHub.
New Royalty Model for AI-Generated Music Proposes Scalable Causal Attribution Framework
A new working paper addresses the growing issue of copyright in music generative AI by proposing a royalty model that aims to create sustainable economic incentives for the creative industry. The model, called Generative Content ID, is inspired by YouTube’s Content ID system and seeks to attribute value to AI-generated music by identifying specific training content through causal attribution methods. This approach overcomes the impracticality of retraining models on subsets of data by using efficient Training Data Attribution. The authors’ analysis suggests current legal practices focused on perceived similarity may not adequately capture the data contributions driving AI utility. The study underscores the potential economic impact of how attribution mechanisms are designed and proposes a framework for more equitable royalty distribution among stakeholders in the music industry.
Large Language Models Enhance Robotic Intelligence, Interaction, and Control Capabilities
A recent survey delves into the growing integration of large language models (LLMs) in the field of robotics, highlighting their potential to enhance robot intelligence and human-machine interactions. The report outlines the significant role of LLMs, such as GPT-5 and Llama4, in improving robot control, perception, and decision-making capabilities by leveraging their natural language understanding and generation strengths. It also discusses the challenges that lie ahead in this area, despite the promising strides made in combining robotics with advanced AI technologies to achieve more intelligent and natural interactions.
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.
