What Happens When AI Joins the Cabinet? Albania Just Appointed One.
In a world-first, Albania appointed an AI "Minister," Diella, to autonomously handle public procurement in its fight against corruption..
Today's highlights:
You are reading the 128th 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
World’s First AI Minister Launched in Albania
In a groundbreaking move, Albanian Prime Minister Edi Rama unveiled Diella, the world’s first AI-driven “virtual minister,” on September 11, 2025. Named after the Albanian word for "sun," Diella will oversee public procurement- a domain historically riddled with corruption in Albania. This AI entity, presented as a young woman in traditional attire, will autonomously evaluate bids and award government contracts via the e-Albania digital platform. Rama claims this step is designed to ensure all tenders are “100% free of corruption,” and the move is part of Albania’s broader push to join the European Union by 2030.
From Virtual Assistant to Decision-Maker
Initially launched as a digital assistant in January 2025, Diella helped citizens access documents and services. She has now been promoted from a support role to the official “Minister for Public Procurement,” responsible for deciding tender outcomes without human interference. Her algorithms will assess bids based on pre-set criteria, aiming to make the procurement process transparent and entirely traceable. Rama’s government argues that Diella’s incorruptibility will eliminate bribery and favoritism from state contracting.
Mixed Reactions and Legal Gray Zones
While some in Albania praised Diella’s appointment as a revolutionary digital transformation, skepticism abounds. Critics, including opposition leader Gazmend Bardhi, argue the move is unconstitutional, as the Albanian legal framework doesn’t recognize non-human ministers. President Bajram Begaj avoided calling Diella a “minister,” and no parliamentary vote was held to confirm her role. Legal scholars stress the need for clearer laws to define Diella’s authority and ensure accountability for her decisions—especially since AI systems can’t testify or be prosecuted.
Global Context and Precedents
Albania isn’t the first to experiment with AI in politics. Japan, Denmark, and New Zealand have each explored AI for symbolic or advisory political roles, such as Tama City's “AI Mayor” campaign or Denmark’s “Synthetic Party.” However, none of these projects placed AI in official government posts. Governments like Estonia and the U.S. have also deployed AI for administrative purposes, such as fraud detection and legal claims processing, but never at the ministerial level. Albania’s Diella is thus the first AI to formally hold such authority in a national cabinet.
Promise and Pitfalls of AI in Governance
If successful, Diella could help Albania curb corruption, streamline procurement, and impress EU observers. AI could bring speed, objectivity, and auditability to traditionally opaque processes. Yet, the experiment comes with real risks: lack of legal clarity, potential data manipulation, system bias, and ethical concerns over AI decision-making in high-stakes governance. Diella’s performance, and the Albanian government’s ability to manage and oversee her responsibly, will determine whether this pioneering move becomes a global model or a cautionary tale
🚀 AI Breakthroughs
OpenAI Reveals Massive Study on ChatGPT Usage Trends and Economic Impact
A comprehensive study by OpenAI and the National Bureau of Economic Research highlights how ChatGPT usage has expanded, closing demographic gaps and creating economic value through its broad adoption for personal and professional tasks. The research, based on 1.5 million conversations, reveals that the gender gap among users has narrowed significantly, with widespread adoption in low- and middle-income countries outpacing that in higher-income regions. The majority of interactions with ChatGPT involve everyday tasks such as information seeking and writing, underlining the tool's role as a productivity enhancer and valued advisor in both work-related and personal contexts.
OpenAI Releases GPT-5-Codex With Dynamic Thinking for Enhanced Coding Performance
OpenAI has introduced GPT-5-Codex, an enhanced version of its AI coding agent, Codex, which exhibits dynamic "thinking" times ranging from seconds to hours, improving performance on coding benchmarks. The update, aimed at making Codex more competitive, is available to ChatGPT Plus, Pro, Business, Edu, and Enterprise users, with plans to expand availability to API customers. The model has demonstrated superior performance in conducting code reviews, with evaluations showing it submits fewer incorrect comments and offers more impactful feedback. The market for AI coding tools is increasingly competitive, driven by high demand, as noted by significant developments within the sector.
Google Research Releases VaultGemma: A Groundbreaking Differentially Private Large Language ModelIMAGE CREDITS:
Google Research has unveiled VaultGemma, a language model trained entirely with differential privacy (DP), boasting a 1-billion parameter scale. Through novel "scaling laws" developed in partnership with Google DeepMind, the research addresses the trade-offs involved in DP training, which involves adding noise to protect data privacy, while maintaining model utility. The VaultGemma model, which offers privacy guarantees at a sequence level, demonstrates performance comparable to non-private models from five years ago and highlights potential future improvements for privacy-focused AI development. The weights and technical details are openly available on platforms like Hugging Face and Kaggle.
Cursor Enhances Developer Productivity with New AI Model Reducing Erroneous Suggestions
Cursor has introduced a new system, Cursor Tab, designed to enhance developers' productivity by predicting their next actions while coding. The Tab model processes over 400 million requests daily using online reinforcement learning to refine its suggestions, aiming to reduce unnecessary interruptions caused by incorrect predictions. Unlike typical models trained on static data, Cursor frequently updates the Tab model using policy gradient methods, optimizing it to show fewer yet more accurate suggestions. The latest iteration of Tab has achieved a 28% increase in the accept rate of its suggestions while issuing 21% fewer suggestions than previous versions, indicating a significant improvement in coding efficiency.
NextEra Partners with ServiceNow to Drive Digital Transformation Across MENA Region
NextEra, a joint venture between LTIMindtree and Aramco Digital, has partnered with ServiceNow to drive digital transformation across Saudi Arabia and the MENA region. As part of the collaboration, a proximity centre will be established at Imam Abdulrahman Bin Faisal University in Dammam, and a Centre of Excellence will train the local workforce. The partnership aims to leverage ServiceNow's AI platform and NextEra's services to enhance enterprise environments with generative AI and localized solutions, focusing on sectors like energy and the public sector. This initiative reflects a commitment to modernization and creating job opportunities while leveraging Arabic user interfaces to meet regional needs.
Adobe Launches AI Agents to Transform Customer Experience and Marketing Strategies
Adobe has launched AI agents designed to transform how businesses build and optimize customer experiences and marketing campaigns. These agents, powered by the Adobe Experience Platform (AEP) Agent Orchestrator, allow for management and customization across Adobe and third-party ecosystems, enabling context-aware interactions and multi-step actions. The platform is already in use by over 70% of eligible AEP customers, with companies like The Hershey Company, Lenovo, and Merkle leveraging these tools to enhance capabilities and drive personalized experiences at scale.
Qwen3-Next Model Boosts Training and Inference Efficiency with Hybrid Attention
Qwen3-Next, developed by the QwenTeam, presents a significant advancement in AI model efficiency by introducing a cutting-edge architecture that combines hybrid attention mechanisms, ultra-sparse Mixture-of-Experts (MoE) structures, and multi-token prediction capabilities. The Qwen3-Next-80B-A3B-Base model, at its core, activates only a fraction of its 80 billion parameters during inference, drastically reducing computational cost while outperforming its dense counterparts. It achieves substantial efficiency in both training and inference, particularly excelling in tasks requiring long context handling, evidenced by its performance with context lengths up to 256,000 tokens. Released on platforms like Hugging Face and ModelScope, Qwen3-Next aims to set a new standard in large-scale language model performance.
Gemini CLI Expands with Security and Cloud Extensions for Seamless Terminal Integration
Gemini CLI has unveiled two new extensions that integrate security analysis and cloud deployment directly into the terminal, streamlining the development process. The security extension, /security:analyze, automates vulnerability scans in local repositories and provides detailed feedback, with GitHub integration on the horizon. The Cloud Run extension, /deploy, allows for quick application deployment to a serverless platform. These tools are the first in a new extensibility framework for Gemini CLI, aimed at enhancing workflow efficiency in development and deployment phases.
Anthropic's Blueprint for Building Better Tools for AI Agents
Anthropic unveiled a detailed guide on how to build high-quality, agent-optimized tools using Claude and the Model Context Protocol (MCP). The post outlines a tool development lifecycle that includes prototyping, evaluating, and refining tools in collaboration with Claude itself. Key contributions include best practices like namespacing, token-efficient responses, prompt-engineered specs, and meaningful context returns. Anthropic also introduced evaluation methods and automation workflows to help developers test tool performance, identify weak points, and continuously improve functionality using Claude Code. This positions Claude not just as an AI assistant, but as a co-developer of its own capabilities.
Claude App Enhances Work Productivity with Optional Memory and Incognito Chat Features
Claude has introduced a new memory feature designed to enhance productivity by remembering users’ projects, preferences, and professional contexts, thereby reducing the need for repetitive explanations. This feature, available now for Team and Enterprise plan users, includes granular user controls to manage what is remembered and offers Incognito chats for discussions that need to remain off the record. Enterprise admins have the option to disable memory. Memory is structured to keep project details distinct and is intended to support work processes across different roles. Incognito chats provide a fresh slate for sensitive conversations, ensuring they aren't saved or affect ongoing memory. The feature aims to streamline workflow while maintaining privacy and security, though it can be adjusted or disabled as needed.
Gartner Highlights Technologies Driving AI Adoption in Government Sector by 2025
Gartner's latest insights reveal that sovereign AI and AI agents are predicted to drive the adoption of artificial intelligence in the public sector over the next two to five years, having reached the Peak of Inflated Expectations on the 2025 Gartner Hype Cycle for Government Services. Sovereign AI empowers national governance through enhanced automation and improved citizen engagement, whereas AI agents are anticipated to automate a significant portion of government service interactions by 2029. Additionally, the report highlights the importance of prompt engineering and machine customers in future AI-driven government operations, forecasting the latter's rise to necessitate new regulatory models.
⚖️ AI Ethics
Perplexity Faces Lawsuit for Allegedly Infringing on Britannica and Webster Content
Perplexity, an AI web search company, faces a lawsuit from Encyclopedia Britannica and Merriam-Webster in New York federal court, accusing it of copyright and trademark infringement. The plaintiffs allege that Perplexity's "answer engine" illicitly scrapes content from their sites, redirects traffic, and produces results identical to Merriam-Webster's definitions, alongside claims of improper use of their trademarks. Despite these legal challenges, Perplexity continues to collaborate with media companies through ad revenue sharing and publisher programs, recently partnering with World History Encyclopedia to launch a chatbot.
Google Sued by Penske Media for Alleged Unauthorized Use of Publisher Content in AI Summaries
Google is facing a lawsuit from Penske Media Corporation, which accuses the tech giant of illegally using publishers' content to generate AI summaries that harm their business. Penske, owner of major publications like Rolling Stone and Variety, claims Google's AI Overviews exploit its content without proper consent, reducing site traffic and ad revenue. Google, in turn, argues that AI Overviews enhance search utility and broaden traffic to a variety of sites. The lawsuit underscores ongoing tensions over content use in AI and parallels other copyright challenges faced by AI companies.
xAI Lays Off 500 Amid Strategic Shift to Specialist AI Tutor Focus
Elon Musk's AI startup, xAI, recently laid off 500 employees from its 1,500-person data-annotation team, according to internal messages seen by Business Insider. The move aligns with a strategic shift to focus on expanding and prioritizing specialist AI tutor roles over generalist ones. Subsequently, xAI intends to grow its Specialist AI tutor workforce significantly, hiring across areas such as STEM, finance, and medicine. This decision comes as the company aims to optimize its resources for the development of its chatbot, Grok.
CEO Accuses Google of Using Same Bot for Search and AI Crawling
Neil Vogel, CEO of People, Inc., accused Google of unfair practices by using the same crawler for both indexing websites for search and for its AI products, claiming this leads to unauthorized use of content. At the Fortune Brainstorm Tech conference, Vogel revealed that Google's contribution to the company's traffic dropped significantly from 65% to around 20% over three years. In response, People, Inc. has begun blocking AI crawlers that don't pay, leveraging Cloudflare's solutions to push for content deals. Vogel argued that Google's strategy leaves publishers at a disadvantage, as their websites rely on Google's search traffic. The situation has led to increased calls for new regulations, as legal solutions like copyright law may not adequately address these challenges.
🎓AI Academia
OpenAI Releases GPT-5-Codex with Enhanced Safety for Coding Environment Applications
OpenAI has released GPT-5-Codex, an advanced version of GPT-5 designed for agentic coding tasks in Codex. This model, enhanced with reinforcement learning on real-world coding scenarios, aims to closely emulate human coding styles and follow precise instructions. Available through various platforms such as Codex CLI, IDE extensions, GitHub, and the ChatGPT mobile app, GPT-5-Codex includes robust safety measures like specialized safety training and network access controls to ensure secure usage.
AI Academy Program Boosts Generative AI Literacy Among University Professors and Instructors
Research conducted at the University of Notre Dame highlights the urgency of equipping higher education instructors with generative AI literacy as such technologies increasingly impact academia. The study introduces the AI Academy, a faculty development initiative blending AI exploration with pedagogical strategies and collaborative learning among peers. Instead of assessing course outcomes, the program offers insights into how educators engage with AI tools, policies, and institutional supports. Results from the study indicate significant improvement in AI literacy among 25 instructors, emphasizing their role in designing responsible AI practices. The research proposes a viable model for professional development that adapts to rapidly advancing AI tools while fostering ethical discussions.
Survey of Time Series Reasoning Systems with Large Language Models Explored
A recent survey examines the role of large language models in reasoning with time series data, categorizing methods based on reasoning topology: direct reasoning, linear chain reasoning, and branch-structured reasoning. The study highlights challenges and opportunities within time series analysis, focusing on the capabilities and limitations of each approach in terms of robustness, computational cost, and ability to handle uncertainty and dynamic data streams. It underscores the importance of developing benchmarks that link reasoning quality to practical utility and advocates for systems capable of explanation and decision-making in real-world scenarios. A repository of related resources is provided to support further research.
Understanding the Role of AI Bots as Manager Clone Agents in the Future Workplace
Researchers from Carnegie Mellon University and Emory University are exploring the potential and challenges of AI-powered Manager Clone Agents in workplaces, which replicate managerial communication and decision-making styles. These AI agents could transform pivotal roles such as presence, information dissemination, productivity, and leadership. The research features discussions and speculative scenarios from workshops with managers and workers, highlighting prospective advantages and significant concerns about adoption, trust, and the impact on workplace dynamics. This study aims to guide the responsible integration of AI Manager Clones by emphasizing employee perspectives, interpersonal connections, and adaptable configurations.
Survey Examines Large Language Models in Enhancing Security Operations Centers' Efficiency
A recent survey highlights the transformative potential of Large Language Models (LLMs) in enhancing Security Operations Centers (SOCs) that are vital for cybersecurity. These models can automate log analysis, streamline triage, and improve detection accuracy, addressing challenges such as high alert volumes and resource constraints faced by SOCs. The study systematically explores how integrating LLMs into SOC workflows could lead to more efficient incident detection and response, marking the survey as a pioneering exploration of LLM applications in cybersecurity operations.
Assessing Safety and Security Risks of Large Language Models Amid Rising Threats
A recent study highlights significant safety and security vulnerabilities in Large Language Models (LLMs), stressing their susceptibility to adversarial manipulation and exploitation for harmful purposes. Researchers from academic institutions analyzed nine prominent LLMs across 24 different security and safety categories, revealing that existing safety filters are inadequate against adversarial attacks and misuse. They introduced the Risk Severity Index to evaluate and compare the security profiles of these models. The findings underscore the need for enhanced model governance and responsible deployment practices, particularly for open-access models.
Survey Highlights Impact of Large Language Models on Legal Artificial Intelligence Development
A recent survey highlights the integration of Large Language Models (LLMs) into Legal Artificial Intelligence, detailing their significant advancement of legal tasks like judgment prediction and document retrieval. The study comprehensively reviews 16 legal LLM series and 47 frameworks, in addition to analyzing 15 benchmarks and 29 datasets designed for evaluating legal applications. Despite their promising achievements, challenges remain, and the paper discusses these alongside future research directions. This serves as a foundational resource for those new to the field and aims to encourage continued exploration and development in legal AI. More information on this research can be found on their GitHub repository.
AI Agentic Programming: Addressing Challenges and Opportunities in Software Development with LLMs
A new survey on AI agentic programming highlights a transformative shift in software development through the use of large language model (LLM)-based coding agents. These agents, unlike traditional code generation tools, autonomously plan, execute, and interact with various development resources such as compilers, debuggers, and version control systems. The report explores a taxonomy of agent behaviors and system architectures, emphasizing advanced capabilities that enable agents to decompose tasks, coordinate multi-step processes, and adapt based on feedback. By examining techniques for planning, context management, tool integration, and execution monitoring, the survey draws attention to both challenges and opportunities in building more reliable, transparent, and collaborative AI coding agents for end-to-end software development.
Integration of AI with Web3 Analyzed: Landscape, Governance, Security, and Trust Explored
The integration of Web3 technologies and AI agents is emerging as a transformative force in decentralized ecosystems, according to a comprehensive analysis. The study examines the intersection of these technologies across five key areas: landscape, economics, governance, security, and trust mechanisms, revealing how AI agents can optimize decentralized finance, enhance governance and security, and establish reliability frameworks via Web3’s trust infrastructure. Despite demonstrating significant market traction with projects valued at over $6.9 billion, the integration faces challenges in scalability, security, and ethics, necessitating further research and development for robust and trustworthy decentralized systems.
New AI Safety Framework Aims to Minimize Harmful Content in Language Models
Researchers at Carnegie Mellon University, among other institutions, have developed a new pretraining framework to enhance the safety of large language models (LLMs), especially when deployed in sensitive environments. The framework incorporates a comprehensive approach with four primary components: safety filtering, where a classifier distinguishes between safe and unsafe content; safety rephrasing, which recontextualizes harmful data into safer forms; native refusal training, using datasets aimed at helping models actively refuse unsafe inputs based on moral reasoning; and a harmfulness-tagging system that flags potentially unsafe material during pretraining to guide the model's responses during inference. Preliminary results demonstrate a significant reduction in attack success rates on standard benchmarks, from 38.8% to 8.4%, without compromising performance on general tasks.
Survey Highlights Evolution and Impact of Large Language Model-based Data Agents
A recent survey from The Hong Kong Polytechnic University examines the role of data science agents powered by Large Language Models (LLMs), showcasing their potential to revolutionize traditional data analysis by simplifying complex tasks and lowering barriers for users lacking expertise. The study highlights key features such as planning, reasoning, reflection, and multi-agent collaboration, which enable these agents to tackle data-centric problems with minimal human input. Through various case studies, the survey demonstrates practical applications while also identifying challenges and future research directions aimed at advancing data agents into sophisticated statistical analysis tools.
Survey Explores Challenges of Combating Hallucinations in Large Language Models
A recent survey from researchers associated with Tencent AI Lab and several universities investigates hallucination in large language models (LLMs), a phenomenon where these models generate outputs that misalign with user input or factual knowledge. The study categorizes hallucination patterns and analyzes approaches to detect and mitigate them, highlighting the unique challenges posed by the massive training datasets, versatility, and imperceptibility of errors in LLMs. It underscores the urgency of addressing these issues to improve the reliability of LLMs, especially given their potential to produce convincing yet misleading information in critical areas like medical advice.
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.
