EU AI Act BIG UPDATE: Council gives final green light to simplify and streamline rules
++ Trump lifts export curbs on Anthropic Mythos and Fable, Cloudflare tells AI crawlers to pay publisher, Alibaba reportedly bans Claude Code for staff, Midjourney pushes Hollywood to reveal AI use...
This week’s highlights:
The Council of the European Union has given final approval to a new law that simplifies parts of the EU AI Act and delays some key compliance deadlines. Rules for high-risk AI systems, originally due from 2 August 2026, will now apply from 2 December 2027 for stand-alone systems and 2 August 2028 for AI built into products. The law also adds a ban on AI tools used to create non-consensual sexual deepfakes and AI-generated child sexual abuse material, with some image-based bans starting in December 2026. It further pushes back the deadline for national AI regulatory sandboxes to August 2027, shortens the timeline for transparency measures on AI-generated content to 3 months, and clarifies how the AI Act will work alongside sector-specific product laws.
Typical EU lawmaking steps for conceptual understanding: the Commission first proposes a law, then the European Parliament and Council act as co-legislators under the ordinary legislative procedure; a law is adopted only when Parliament and Council agree on the same text, and after this Parliament vote the next step is formal Council adoption, followed by signature/publication before entry into force. Digital Omnibus means a bundle/package of EU amendments used to simplify existing digital laws. Why it is important for the EU AI Act: it changes the practical compliance timeline and some scope details.
At the School of Responsible AI (SoRAI), we help both individuals and organizations build practical, real-world AI literacy and Responsible AI capability through structured, engaging, and action-oriented programs. For individuals, this includes AI Literacy, globally relevant certification training such as AIGP, RAI, and AAIA, as well as career transition and advisory support for professionals moving into AI governance roles. For organizations, we offer customized enterprise AI literacy training, Responsible AI strategy and governance setup, and AI assurance support to help teams understand, operationalize, and validate AI responsibly. At the core of SoRAI is a progressive three-layer approach: first helping people understand AI, then build the right governance foundations, and finally validate readiness through assurance and audit-focused thinking. Want to learn more? Explore our AI Literacy programs, certification trainings, and career support offerings, or write to us for customized enterprise solutions.
⚖️ AI Ethics
Trump Lifts Export Curbs on Anthropic Mythos and Fable Models
The U.S. government has removed a rule that required Anthropic to get a license before offering its Mythos and Fable AI models outside the country, allowing public access to return from July 1. The restriction, added on June 12, had effectively forced Anthropic to shut access because checking every foreign user was too difficult at scale. Officials said the decision followed talks in which Anthropic agreed to help detect security risks, coordinate with the government on standards, and report malicious use. The move comes as pressure grows to keep U.S. AI firms competitive, especially as Asian companies release increasingly capable rival models, though the wider AI industry still faces uncertainty over future U.S. rules.
Cloudflare Sets 2026 Deadline for AI Crawlers to Pay Publishers
Cloudflare has set a new rule aimed at pushing AI companies to pay for publishers’ content and clearly separate search crawlers from bots used for AI training and agents. Starting September 15, 2026, its default settings will block “mixed-use” crawlers from ad-supported pages unless site owners choose otherwise, affecting new customers, new sites, and existing free users. The move could limit how AI companies collect web data for training and services, while giving publishers more control over how their content is used. Cloudflare is also expanding its payment tools from “Pay Per Crawl” to “Pay Per Use,” starting with partners Ceramic.ai and You.com, so publishers can earn money when AI services use or show their content.
Alibaba Reportedly Bars Employees From Using Claude Code Starting July 10
Alibaba will reportedly ban employees from using Anthropic’s Claude Code starting July 10, according to several reports. The move comes as Anthropic already blocks Chinese companies and foreign firms owned by them from accessing its AI models, while also tightening controls on loopholes that let some users in China use Claude. Reports also say an earlier version of Claude Code was used in tests to help detect Chinese users as part of efforts to stop misuse and model distillation. Alibaba has reportedly labeled Claude Code as high-risk software and is directing staff to use its own Qoder coding tool instead.
Midjourney Presses Hollywood Studios to Disclose Their Own AI Practices
Midjourney is asking a court to make Disney, Universal, and Warner Bros. share more details about how they use generative AI in their own work as part of an ongoing copyright lawsuit. The studios say Midjourney’s image tools can create well-known characters they own, while Midjourney argues that training on copyrighted images can be protected as fair use. A judge had already said the studios must hand over some records tied to consumer-facing AI-made images and videos, but Midjourney now wants that limit removed. The company says the withheld records could show the studios are also using AI in similar ways internally, while the studios argue Midjourney is asking for overly broad information not central to the case.
GeneBench-Pro Tests Whether AI Can Handle Real Biology Research Judgment
OpenAI has released GeneBench-Pro, a new benchmark designed to test whether AI models can handle difficult, judgment-based biology research rather than simply follow fixed steps. The benchmark includes 129 synthetic but realistic problems across genomics, genetics, proteomics, cancer research, and clinical medicine, with tasks built to measure how well models deal with messy data, uncertainty, changing assumptions, and analysis choices. In tests, the company’s top model, GPT-5.6 Sol, reached a pass rate of 28.7%, rising to 31.5% in Pro mode, suggesting progress but also showing that current AI systems still fail on most expert-level scientific problems. OpenAI said the benchmark was reviewed by outside specialists and partly open-sourced, with the goal of helping track whether AI can eventually support or speed up high-value scientific analysis that now takes human experts many hours to complete.
Centre Signals Separate AI Law as Export Curbs on Mythos Ease
India’s electronics and IT ministry may begin work on a separate law to regulate artificial intelligence, as the government now sees the timing as right for a dedicated framework. So far, AI-related issues such as deepfakes and content labelling have been handled through existing laws and IT Rules, but officials said a draft AI law could now be prepared, though no timeline was given for its rollout. On access to advanced AI models, the government said export curbs on Anthropic’s Mythos model have been eased, while restrictions remain on newer models such as Mythos 5 and Fable 5. Any wider access for Indian entities to those newer models would still need approval from the US government.
Can India Regulate Foreign AI Models Despite RBI’s Tough New Rules?
The Reserve Bank of India has issued new model risk management guidelines for banks and financial firms, making them responsible for checking the accuracy, bias, safety and suitability of all decision-making models, including third-party AI tools. The rules require firms to have board-approved oversight, technical documentation, audit rights and exit plans, marking a strong step toward AI governance in finance. But the article argues that India’s control remains limited because most advanced AI models are foreign-owned, opaque and trained on data that may not fit Indian conditions. It says this dependence could make compliance difficult, especially for smaller firms, and raises a larger question about how far India can regulate AI systems it does not build or own.
Samsung SK Hynix Micron Sued Over Alleged RAM Supply Cuts
A new class action lawsuit filed in a US court accuses Samsung, SK Hynix, and Micron of cutting DRAM supply to push up RAM prices, while shifting production from older DDR3 and DDR4 memory to higher-priced HBM used in AI servers. The complaint says the three companies control nearly all of the global DRAM market, making it hard for any rival to step in because building memory factories costs billions, takes years, and requires advanced trade secrets and equipment. Plaintiffs argue that, in a normal market, high prices would bring more supply, but that did not happen here, which they say points to coordinated behavior. The case also highlights that these companies have faced similar scrutiny before, including past US and Chinese investigations into DRAM pricing.
UN Report Warns Policymakers Struggle to Keep Pace With AI Growth
A new UN report says AI is advancing so fast that governments and current rule-making systems are struggling to keep up. The panel said AI can bring major benefits in areas like medicine, vaccine research, disease detection and food security, but it is also creating serious risks such as deepfakes, false information, cybercrime and harmful behavior reinforcement. It warned that more powerful and autonomous AI systems are becoming harder to monitor, while the growth of AI data centers could also hurt nearby communities. The report called for stronger independent testing, international cooperation and shared standards, saying that without safeguards, AI could worsen inequality, harm human rights and leave most benefits concentrated in a few countries and companies.
Ford Rehires 350 Engineers After AI Quality Systems Miss Critical Defects
Ford said its big push to use AI and automation for vehicle quality control did not work as well as expected, so the company brought back 350 experienced engineers to catch problems the systems were missing. The automaker had installed 900 AI-powered inspection cameras and expanded automated quality checks, but later said human expertise was still needed to spot failure risks early and help improve the tools. The company says the move is already helping, with Ford ranking first among mainstream brands in the latest JD Power Initial Quality Study and expecting about $1 billion in lower costs this year, mainly from fewer warranty and recall expenses. The case highlights a wider industry reality: AI can support factory work, but it still struggles to replace deep engineering judgment built over many product cycles.
🚀 AI Breakthroughs
Google Releases Faster, Cheaper Nano Banana 2 Lite Image Generator
Google has released Nano Banana 2 Lite, a new AI image generator designed to be faster and cheaper than its earlier versions. The company said the model can create images in about four seconds and costs $0.034 per 1,000 images, making it suited for high-volume work such as rapid image drafting and editing. Nano Banana 2 Lite is now available through Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform, and it replaces the original Nano Banana as Google’s older model. Alongside this, Google expanded access to Gemini Omni Flash for video generation at $0.10 per second and showed a new app, Omni Product Studio, that turns static AI-made images into e-commerce videos.
Microsoft Starts AI Deployment Unit With $2.5 Billion Enterprise Investment
Microsoft has launched a new business unit called Microsoft Frontier Company to help enterprises put AI tools into real-world use. The company is backing the effort with a $2.5 billion commitment and a team of 6,000 industry and engineering experts. The move is similar to recent AI deployment efforts from rivals such as Amazon Web Services, OpenAI, and Anthropic, even though Microsoft says it goes beyond the usual forward-deployed engineering model. Microsoft may have an early advantage because it already works closely with many Fortune 500 companies, with early partners including the London Stock Exchange Group, Unilever, Land O’Lakes, and Accenture.
Meta Quietly Launches Pocket App for AI-Generated Games and Interactive Experiences
Meta has quietly launched a new app called Pocket that lets users create and share small interactive apps and games using simple AI text prompts. The app appears to build on technology from Gizmo, a vibe-coded gaming platform whose team Meta acquired earlier this year, and it includes a feed for discovering and playing creations made by others. App data shows Pocket went live on the App Store and Google Play on June 29, 2026, although Meta has not officially commented on the release. The launch adds to Meta’s broader push to bring AI creation tools into mainstream consumer apps, while the app still appears to be in an early testing stage.
Meta Shares Brain2Qwerty v2 for Noninvasive Real Time Brain to Text
Meta has shared Brain2Qwerty v2, an AI system that can turn non-invasive brain signals into text in real time without requiring surgery. The system was trained on about 22,000 sentences from nine volunteers using MEG brain recordings collected while they typed, and it uses deep learning and language models to turn noisy signals into readable sentences. Meta said the model reached 61% word accuracy overall, far higher than the roughly 8% reported for other non-invasive methods, with its best participant reaching 78% accuracy. The company has also released the training code for Brain2Qwerty v1 and v2, while a research partner is making the v1 dataset available, aiming to support work on communication tools for people who cannot speak because of brain injuries or lesions.
Anthropic launches Claude Sonnet 5 as a lower-cost AI agent model
Anthropic has released Claude Sonnet 5, a new midsize AI model designed to handle agent-style tasks such as planning, using tools, coding, and completing longer jobs with less human help. The company says the model delivers performance close to its stronger Opus 4.8 model at a much lower price, starting at $2 per million input tokens and $10 per million output tokens through August 31, before rising to $3 and $15. Sonnet 5 will become the default model for Claude’s free and Pro users, and Anthropic says it improves on Sonnet 4.6 in reasoning, coding, tool use, and knowledge work, while also reducing hallucinations, deception, and other unsafe behaviors. The launch highlights a broader industry shift as AI companies compete not just on agentic features, but on how cheaply and reliably those systems can run.
Anthropic Discusses New Custom AI Chip Partnership With Samsung
Anthropic is in talks with Samsung about possibly building a custom AI chip, showing it may be moving ahead with plans first reported earlier this year. According to recent reporting, the company has not yet decided exactly what the chip will do, how it will fit into servers, or how powerful it will be. Anthropic said its computing strategy will still rely on a mix of chips from Google, Amazon, and Nvidia, and it did not provide more details on the Samsung discussions. The move reflects a wider trend in the AI industry, where companies are developing their own chips to reduce dependence on Nvidia and improve efficiency.
California Secures Half-Price Anthropic Claude Access for State Government Agencies
California has reached a deal with Anthropic that lets state agencies and local governments use Claude at about half the usual price, along with training and support. The state said the AI chatbot will help public workers draft documents, review information, and improve efficiency without replacing human jobs. The agreement follows California’s broader push to expand AI use in government while keeping safety standards in place. The move also stands in contrast to the federal government’s recent clash with Anthropic, which lost a Pentagon deal after disagreements over limits on surveillance and autonomous weapons use.
Anthropic Targets Scientists With Claude Science Workflow Instead of New Model
Anthropic has released Claude Science, a beta AI workbench for researchers that focuses on improving the scientific workflow rather than offering a new model. It uses the same Claude models already available today, but brings databases, tools, code, figures, and citations into one place so scientists do not have to switch between many systems. The platform connects to more than 60 scientific databases, can create helper AI agents for different tasks, and includes a fact-checking step to review citations and calculations, though it still relies on the same underlying model. Anthropic says the tool can also run on a lab’s own infrastructure for data control, and early users at research groups have reported faster analysis work.
Picsart Builds Creator Communities and Opens Monetization Beyond Follower Counts
Picsart says much of the creator economy still rewards only a small top group, while most creators struggle to get visibility, feedback, and income. The company, which says it has 130 million monthly active users, has built more than 12,000 interest-based communities called Spaces to help creators share work, get support, and improve skills in a moderated environment. It has also launched Earn with Picsart, a monetization program open to creators regardless of follower count, where payouts are tied to content performance such as views, shares, and reach. Picsart says the goal is to give newer and smaller creators a clearer path from learning and community participation to earning money from their content.
Google Ad Reimagines Declaration of Independence Drafted With AI Tools
A new Google commercial imagines the Declaration of Independence being created with modern Google Workspace tools, showing the Founding Fathers using Docs, Calendar, Meet, e-signatures, and some AI features in a playful 1776 setting. The ad keeps the AI angle fairly light, with tools used for things like meeting notes, image ideas for the national seal, and handling a document access request, rather than writing the Declaration itself. The campaign appears to have received mostly positive reactions on YouTube and Instagram, while criticism on Bluesky called it awkward and tone-deaf, especially for its use of AI. Some critics also pointed out that, despite the theme, very little of the commercial actually depends on AI.
KPMG: 92% of Tech Executives See AI Management Skills as Vital
A KPMG report says 92% of tech executives believe managing AI agents will become a vital job skill by 2031 as companies rely more on automated decision-making. Based on a survey of 2,500 tech leaders across 27 countries, the report found that 88% of organizations are already investing in agentic AI, and digital assistants are expected to make up 36% of core tech teams by 2027, up from 28% in 2025. The report says this shift is pushing companies to redesign teams, adopt flatter structures, and focus on broader business change rather than only individual productivity. It also found that 90% of executives plan to deepen outside partnerships for AI expertise, even as concerns grow around security, governance, data protection, and future risks such as quantum-related encryption threats.
India, Japan Expand AI Partnership Across Tech Stack, Focus on Safe Ecosystem
India and Japan have expanded their AI partnership, agreeing to build a safe, secure, trustworthy, inclusive and human-centric AI ecosystem across the full technology stack. The two sides said they will work together on AI governance, infrastructure, model development, research, talent exchange and real-world applications, while supporting global coordination through forums such as the G20, OECD, GPAI and the United Nations. They also agreed to strengthen cooperation in data centres, GPUs, semiconductors and AI compute, and to promote multilingual, open-source and domain-specific AI models through industry and academic partnerships. The partnership also includes a goal to bring 500 highly skilled Indian AI professionals to Japan by 2030 and a shared focus on using AI for public good across the Indo-Pacific and the Global South.
Mukesh Ambani, Sunil Mittal Join Global AI Panel as Founding Members
Mukesh Ambani and Sunil Mittal have been named founding members of the AI for Good Global Commission, a new panel set up by the International Telecommunication Union. The group brings together global leaders and industry executives to promote artificial intelligence that is trustworthy, inclusive, and useful for the public good. One of its main goals is to narrow the digital divide so that AI benefits more people, especially in underserved communities. The commission will also focus on using AI to address major global challenges without deepening existing inequalities.
Portugal Launches Amália Open-Source AI Model for European Portuguese Use
Portugal has launched Amália, a government-backed open-source AI model built specifically for European Portuguese. The model is designed as a foundation for public institutions, companies, universities, and researchers to create AI tools, rather than serve as a direct consumer chatbot. It is expected to support uses in public services, education, culture, defence, and research while giving the country a local alternative to foreign AI systems. The move also places Portugal within Europe’s wider effort to build homegrown AI infrastructure focused on local languages, public-sector needs, and greater regional control over key technology.
🎓AI Academia
ContextNest Adds Verifiable Governance Layer for Autonomous AI Agent Knowledge
A new research paper describes ContextNest, an open system designed to help autonomous AI agents use external knowledge in a more trustworthy and traceable way. Instead of replacing retrieval systems like RAG, it acts as a governance layer that checks whether documents are approved, current, verifiable, and valid for AI use at the time they are accessed. The paper says the system uses version tracking, secure hashes, structured metadata, and audit logs so organizations can later verify exactly which information shaped an AI agent’s output. In two controlled tests, ContextNest outperformed a BM25 retrieval baseline in a stale-information attack and also showed fully stable repeat results, while a dense retrieval setup produced inconsistent results on many queries.
AI Red Teaming Gives Boards a New Way to Test Strategy
A new arXiv technical report says companies should not treat board approval of an AI strategy as proof that the plan is sound. It argues for “strategic red teaming,” a board-level process that stress-tests the key assumptions behind major AI decisions, especially when AI use is consequential, expensive, dependent on outside providers, hard to explain, or likely to create regulatory or accountability risks. The report says normal risk reviews often focus only on known risks and controls, instead of challenging whether the strategy itself is built on weak assumptions. It proposes a six-part governance model to test mission fit, dependencies, economic fragility, regulatory exposure, and accountability, while noting that the framework is a governance design idea and not a legally validated or universally required standard.
Study Finds AI-Native Engineering Teams Need New Risk Governance Frameworks
A new research paper says traditional software risk management does not work well for teams building agentic AI systems, because these systems behave probabilistically, can take multi-step actions on their own, and can change risk patterns between deployments. The paper proposes an organizational framework to help engineering managers assign ownership, monitor failures, and escalate problems in AI-native teams. It finds that risk coverage gets weaker as teams move from standard software to AI-native operations, with some of the most serious gaps appearing where AI outputs are used by other systems that expect deterministic behavior. The study argues that managers should focus less on only components and more on boundaries, actions, and tool-use surfaces, while adding semantic monitoring and rollback authority to better contain failures.
Study Proposes Audit Framework for LLM Content Risks and Governance
A new arXiv paper argues that large language models can give users polished, structured answers that look rational and reliable even when the user has not fully understood the real-world problem. It says this can create “pseudo-rational cognition,” where people mistake AI-generated wording for genuine understanding, and warns that AI content can also feed back into future systems, causing memory pollution, false detections, and self-reinforcing errors. The paper describes this as a result of models recombining large amounts of human experience into outputs that seem practical without actual lived practice or verification. To reduce these risks, it proposes a practice-auditing framework that checks problem boundaries, evidence sources, validation steps, logging, version control, and rollback before AI-generated content is trusted or reused.
Study Finds AI Interview Follow-Up Tools Raise Ethics and Privacy Concerns
A new study examined how an AI tool could help interviewers ask better follow-up questions during live semi-structured interviews, using GPT-4o with a human reviewer still in control. In tests with 17 interviewers, the system showed potential to reduce mental effort and support deeper questioning, but it also raised major ethical and social concerns. Participants worried about biased or harmful AI-generated questions, reduced respect for interviewees when attention shifts to the tool, unequal access for some users, unclear accountability when problems happen, and privacy risks when sensitive conversations are recorded or processed by AI. The paper says AI-assisted interviewing may be useful, but only with strong human oversight, clearer responsibility, and better safeguards for fairness, respect, and data protection.
Study Finds Safety Scores Can Mislead in Large Language Model Evaluations
A June 2026 paper says many common tests for large language models may give a false sense of progress, because scores can improve even when real capability, safety, or reliability do not. The study reviews research from 2018 to 2026 and compares problems in model evaluation with problems in AI safety, arguing that both can fail when systems are optimized for the metric instead of the real goal. Its audit of 10 models found no clear statistical link between stronger capability and better long-term resistance to adversarial attacks, while the gap between open and closed models on safety looked small and depended heavily on governance and disclosure practices. The paper says current public evidence is too mixed and inconsistent to support firm rankings, and calls for more dynamic testing, clearer reporting, repeated attack-based safety checks, and more auditable alignment methods.
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.
