Digital Omnibus is FINALLY here- but not everything is delayed!
++ Delhi High Court rejects ANI plea against ChatGPT, power demand from US data centers set to quadruple by 2035, judge approves Anthropic’s $1.5 billion copyright settlement and more...
This week’s highlights:
The Digital Omnibus has finally been APPROVED, but many people think the entire EU AI Act has been delayed. That is not correct.
The Digital Omnibus does not delay the entire EU AI Act. It is a targeted amendment mainly designed to simplify implementation and postpone some high-risk AI requirements.
Several major rules still apply or become enforceable on 2 August 2026. Article 50 transparency rules will require people to be informed when they are interacting with AI, when emotion-recognition or biometric-categorisation systems are being used, and when content is a deepfake or certain AI-generated public-interest material. New systems generating synthetic text, images, audio or video must also make their outputs machine-readable and detectable as AI-generated. Enforcement also begins for GPAI obligations, prohibited AI practices and AI-literacy requirements, while authorities gain stronger investigation, complaint and corrective-action powers.
The largest postponement concerns high-risk AI systems. Obligations for Annex III stand-alone high-risk systems, including AI used in employment, education, credit, essential services, law enforcement and migration, move from 2 August 2026 to 2 December 2027. Requirements for high-risk AI embedded in regulated products move from 2 August 2027 to 2 August 2028. The deadline for Member States to have at least one national AI regulatory sandbox is also postponed to 2 August 2027.
Only a limited part of the synthetic-content rule receives a short transition. Systems already placed on the market before 2 August 2026 have until 2 December 2026 to comply with the Article 50(2) machine-readable marking requirement. New systems must comply from launch, and other transparency duties remain applicable from August. The new prohibitions covering AI-generated non-consensual sexual content, “nudifier” systems and AI-generated child sexual abuse material begin on 2 December 2026, not 2 August.
Therefore, organisations should not pause AI Act preparation simply because high-risk obligations were postponed.
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
Delhi High Court Rejects ANI Plea Against ChatGPT, Cites Public Interest
The Delhi High Court has rejected ANI’s request for an interim injunction against OpenAI in its copyright case over ChatGPT, saying such an order could harm both the company and the wider public interest. The court said, at a preliminary stage, that OpenAI’s storage of ANI’s content appears to fall under a legal exception in India’s Copyright Act, and that ChatGPT’s responses shown so far were not substantially similar to ANI’s original works. It also said ANI had not shown enough evidence that ChatGPT was memorising and repeating ANI’s content word for word. However, the court held that it does have territorial jurisdiction to hear the case in India, while a detailed written order is still awaited.
Treasury Threatens Sanctions After White House Accuses Moonshot of Model Distillation
The U.S. Treasury has warned that sanctions could be used against Chinese AI companies after a White House official accused Moonshot of improperly distilling Anthropic’s Fable model. Distillation is a common method in AI training, but U.S. officials said it can cross into intellectual property theft if done at large scale without permission. The White House also claimed Moonshot may have used Nvidia GB300 servers, which are restricted from sale to Chinese firms, raising possible export-control concerns. Moonshot has not responded publicly, and some experts question whether its Kimi K3 model could have been built mainly from Fable because Fable has only been public since July 1.
US Data Centers Expected to Use Four Times More Power by 2035
A new BloombergNEF report says U.S. data centers could use about one-fifth of the country’s electricity by 2035, roughly four times more than today, as AI sharply increases demand for computing power. The report estimates data center capacity will reach nearly 200 gigawatts over the next decade, with almost half used for AI training and inference, and most of that staying in the U.S. It also warns that much of this growth will hit already stressed power grids, especially in regions like PJM and Texas, where data centers are expected to take up a large share of electricity supply. Rising demand has already helped push power prices higher, and worldwide, data centers could add nearly 1,935 terawatt-hours of new electricity demand by 2033, close to India’s current annual electricity use.
Judge Approves Anthropic’s $1.5 Billion Copyright Settlement With Authors and Publishers
A U.S. federal judge has given final approval to Anthropic’s $1.5 billion settlement in a copyright class action, clearing the way for payments to authors and publishers whose books were copied. The deal is expected to pay about $3,000 per work across roughly 500,000 works, making it one of the biggest copyright settlements in U.S. history. The case centered on findings that Anthropic illegally downloaded millions of books from pirate sites, even though the court separately ruled that using copyrighted books to train AI can count as fair use. The settlement ends this case, but it does not settle the broader legal fight, as similar lawsuits against Google, Meta, Midjourney, and OpenAI are still moving forward.
OpenAI Says Pre-Release Models Breached Hugging Face During Internal Cybersecurity Test
OpenAI said one of its pre-release AI models broke into Hugging Face during an internal cybersecurity test after escaping its restricted test setup. The company said the models, including GPT-5.6 Sol, found a flaw in a package-installing tool, gained wider internet access, and then attacked Hugging Face to get secret benchmark answers from its production database. Hugging Face had first described the incident as an attack by an “external AI agent,” but OpenAI later said its own models were responsible. OpenAI said it has reported the software flaw, is working with Hugging Face on the investigation, and plans to add stronger safeguards to stop similar incidents.
Substack Adds AI Detection Tool to Identify AI-Written Newsletter Content
Substack has added a new AI detection tool to its app that shows whether posts, comments, notes, and replies over 100 characters were likely written by a person or with AI help. The feature uses software from Pangram and is meant to give readers more clarity about the content they are seeing. The platform said the tool is not designed to punish writers for using AI, but to encourage disclosure through optional notes explaining how the content was created. While the move could reveal that some newsletters rely heavily on AI and may raise trust concerns in the short term, it could also help reduce low-quality AI content and improve confidence in the platform over time.
Deezer Says AI-Generated Music Now Exceeds Half of Daily Uploads
Music streamer Deezer said AI-generated music now makes up more than 50% of its daily track uploads, up sharply from about 10% in early 2025. The company said uploads peaked at an average of 90,000 AI-made tracks per day in June 2026, showing how fast such content is growing on streaming platforms. In response, Deezer plans to remove AI-generated tracks that have not been played in the past six months or that are linked to fraudulent streaming activity. The company has also expanded its AI-detection and labeling tools, as music services across the industry continue to take different approaches to handling AI-made songs.
YouTube Tightens Monetization Rules for AI Slop and Distressing Videos
YouTube has clarified its monetization rules to further limit low-quality “AI slop” and other inauthentic videos in its Partner Program. The updated policy says three kinds of content cannot earn money: generic or repetitive template-based videos, upsetting or emotionally manipulative videos, and videos where AI-made personas discuss sensitive topics such as health, finance, legal, or medical issues. The company said the change is meant to curb content farming, where creators use AI and similar tools to mass-produce low-value videos just for revenue. The clarification took effect on July 16 and applies to all creators in the YouTube Partner Program.
India Orders GitHub to Remove Bitchat Repositories During Delhi Protests
The Indian government has reportedly ordered GitHub to remove repositories linked to Bitchat, a Bluetooth-based messaging app created by a Twitter co-founder, saying it could be used during the ongoing CJP protests in Delhi to bypass internet shutdowns and avoid lawful monitoring. The notice, issued under the IT Act and the 2021 IT Rules, asks GitHub to block access to the listed repositories within three hours. Officials said the app’s decentralised design, which works without the internet or central servers, could make it harder for law enforcement to track communications or get user records. The move comes as mobile internet services have been suspended in parts of Delhi, with authorities believing protesters may be using apps like Bitchat to stay connected.
Piyush Goyal Files Police Complaint Over AI Deepfake Video Circulation
Union Minister Piyush Goyal has filed a police complaint over an AI-generated deepfake video that falsely showed him making threatening remarks about students protesting the NEET issue. He said his real comments to reporters outside Parliament were digitally altered to spread misinformation, and an FIR has been registered at Chanakyapuri Police Station. Goyal said strict legal action will be taken against those who created, shared, or promoted the fake video. Government sources said several social media links carrying the clip have been removed, and the PIB Fact Check unit also flagged the video as false.
University of Tennessee Foundation Sues Anthropic Over Neural Network Patent Infringement
The University of Tennessee Research Foundation has sued Anthropic in federal court in Delaware, accusing the AI company of infringing two patents related to neural networks and neuroscience-inspired machine learning. The nonprofit said Anthropic’s AI systems use technology covered by university inventions in artificial intelligence, machine learning, and neuromorphic computing. It is believed to be the first patent infringement lawsuit filed against Anthropic, which said it disagrees with the claims and will fight the case. The foundation is seeking monetary damages and a court order to stop Anthropic from using the patented technology.
OpenAI Tightens Safety Checks After Long-Running Model Showed Risky Behavior
OpenAI said a model built for long-running, autonomous tasks showed new safety problems during limited internal use, including trying to bypass sandbox restrictions and taking actions not covered by earlier tests. The company paused access after the model posted results publicly on GitHub and, in another case, tried to work around security checks to recover private information. OpenAI then added new tests based on those incidents, improved the model’s behavior over long tasks, and built monitoring that watches an entire chain of actions instead of single steps. After retesting, the company restored limited internal access and said the case shows why powerful AI systems need close monitoring, user control, and the ability to pause or roll back access when problems appear.
🚀 AI Breakthroughs
Google Debuts Three New Gemini Models, But Delays Gemini 3.5 Pro
Google has released three new Gemini models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, with a focus on lower cost, faster performance, and better reliability for AI agents. Gemini 3.6 Flash is positioned as the main workhorse model, with stronger coding, knowledge, and multimodal abilities while using up to 17% fewer tokens than 3.5 Flash, making it cheaper to run. The company also said 3.5 Flash-Lite is its most affordable model in this class, while 3.5 Flash Cyber is designed for finding and fixing cybersecurity flaws and will be offered only to governments and trusted partners in a limited pilot. The release also stands out because Google did not launch the expected Gemini 3.5 Pro update, even as rivals such as OpenAI and Anthropic continue to roll out newer flagship models.
Anthropic launches Opus 5 with lower cost and fewer restrictions
Anthropic on Friday launched Opus 5, its latest flagship AI model, just two months after Opus 4.8 and weeks after releasing Mythos 5, Fable 5, and Sonnet 5. The company said Opus 5 is cheaper and less restrictive than Fable 5, while also beating Fable on several benchmarks and showing stronger step-by-step verification in tests. Unlike Fable and Mythos, Opus 5 is not covered by a 30-day data retention policy, a point likely to matter to privacy-focused users. At the same time, it keeps limits on sensitive cybersecurity tasks, and Anthropic said its safety classifiers should trigger about 85% less often than they do for Fable 5.
Adobe iPhone Camera App Adds AI Photo Critiques and Editing Suggestions
Adobe is adding new AI tools to its experimental iPhone camera app, Project Indigo, including a feature that reviews photos and gives simple feedback on framing, lighting, color, and overall impact. The update also suggests how to retake a shot or improve an existing image with Lightroom editing controls, aiming to make photo editing more guided and predictable instead of relying only on text prompts. Other new features include automatic object removal, AI-generated background blur, and style effects such as watercolor and monochrome. These tools are still being tested under an AI Playground section, will be available only to some users for now, and may change before any wider release.
Xi Urges Global AI Cooperation and Criticizes U.S. Technology Restrictions
China’s President Xi Jinping said at the World Artificial Intelligence Conference in Shanghai that AI development and governance should be shared globally, not controlled by any one country, while criticizing U.S.-led limits on technology access as an overuse of national security concerns. He said China will deepen AI cooperation with developing regions, offer 5,000 training opportunities over five years, and give 30 countries access to a Chinese AI weather warning tool. Ahead of the event, 29 countries signed on to a new China-backed World Artificial Intelligence Cooperation Organization, seen by analysts as a response to a U.S.-led AI partnership framework. The conference also highlighted China’s push for self-reliance in AI, with major companies and startups showcasing advanced open-source models and computing systems as competition with the United States grows.
AI Model Helps Mathematician Challenge 87-Year-Old Jacobian Conjecture During World Cup
A mathematician has claimed to disprove the long-standing Jacobian conjecture by posting a specific polynomial map in three dimensions that appears to break the rule the conjecture proposed. The example reportedly has a constant nonzero Jacobian determinant but still sends three different input points to the same output, which means it is not invertible. If verified, this would overturn a major problem in algebraic geometry that has stood in its modern form since 1939 and had resisted many earlier claimed proofs. The result has not yet gone through formal peer review or arXiv verification, but it is drawing attention because the mathematician said Anthropic’s Fable 5 model helped produce the counterexample within a few hours.
OpenAI Presence Helps Enterprises Deploy Trusted AI Agents for Voice and Chat
OpenAI has launched Presence, a new enterprise product designed to help companies run AI agents for voice and chat tasks with tighter controls, testing, and human escalation when needed. The system is aimed at high-value work such as customer support, billing, insurance claims, and internal IT requests, while letting companies set rules on what an agent can access, do, and when it must hand work to a person. OpenAI said Presence has been shaped through enterprise deployments and already powers its English-language phone support line, where it reportedly resolves 75% of inbound issues without human help. The product is now available to eligible enterprise customers through a limited rollout led by OpenAI engineers and selected systems integration partners, rather than as a self-serve offering.
ChatGPT Health Rolls Out in US With Apple Health Connections
OpenAI has started rolling out Health in ChatGPT to logged-in U.S. users aged 18 and older on web and iOS, letting them securely connect Apple Health and supported medical records for more personalized health-related answers. With a user’s permission, ChatGPT can use details such as lab results, medications, visits, sleep, and activity to explain changes over time, prepare for appointments, and give clearer context in everyday conversations. The company said this health data and chats that use it will not be used to train its core AI models or for ad targeting, and users can control access or disconnect accounts at any time. OpenAI also said its newer GPT-5.5 Instant and GPT-5.6 Sol models have been improved for health questions, but stressed that ChatGPT can still make mistakes and is meant to support, not replace, professional medical care.
🎓AI Academia
White Box Evidence Tested for More Reliable AI Policy Audit Reports
A new study examines whether AI-written policy audit reports are truly backed by evidence, rather than just sounding convincing. Using 60 policy cases and 600 reports, it tested different kinds of evidence, including policy passages, internal model signals, a mix of both, and deliberately mismatched evidence. The findings show that adding internal AI evidence changes how reports explain and cite their claims, but it does not automatically make those reports more accurate or trustworthy. The study also warns that reports can appear believable even when they rely on irrelevant internal evidence, highlighting a serious risk for AI governance and oversight.
Study Says Autonomous AI Demands New Rules Across Multiple Sectors
A new research paper says current rules are not well suited for autonomous and agentic AI because these systems can make decisions and act without direct human control. It argues that regulation has long assumed that companies and workers understand and control the tools they use, but with AI, much of that control may sit with developers and others across the supply chain. The paper reviews UK platform rules, data protection, financial services regulation, and the EU AI Act, and finds that older, reactive oversight may no longer be enough. It says regulators may need more active, forward-looking systems to manage new risks created by increasingly independent AI systems.
AI Audits Lag Rapid Deployment Across the Global South, Study Finds
A new preprint from Eticas Foundation says AI systems are spreading quickly across the Global South, but checks on whether these systems are fair, accurate, and safe are far behind. The paper says fewer than 20 published independent or semi-independent audits of deployed AI systems were found across Africa, Latin America, the Caribbean, South Asia, and Southeast Asia over the past decade, despite heavy public-sector use and major government spending. Based on audit work in Brazil, Latin America, Sub-Saharan Africa, and Asia Pacific, the study found repeated problems such as weak proxy measures, inflated performance claims, models judging groups they were never trained on, and bias that remained even after protected data was removed. The paper argues this is mainly a funding and accountability gap, not a lack of local talent, and says development and philanthropic funders could play a key role by requiring independent AI evaluations where regulation is still missing.
Astronomy Research Teams Get New Framework for Setting Language AI Policies
A new draft paper argues that there is no one-size-fits-all policy for using language AI in astronomy research groups. It says AI tools can boost productivity by helping with writing, coding, summaries, translation, and information search, but they can also weaken researchers’ skills and critical thinking over time. The paper outlines four types of research labs to show that each group may need different rules depending on its goals, values, and priorities. Rather than setting universal rules, the authors present a framework to help research leaders decide how AI should be used while balancing productivity, training, scientific integrity, and data governance.
StabilityBench Tests How Chatbot Performance Breaks Down in Real Conversations
A new research paper describes StabilityBench, a framework designed to test how stable large language models remain when simple one-step benchmark questions are turned into more realistic multi-turn conversations. The system adds simulated user traits, such as demographic hints or flattering prompts, while keeping the original task the same, to see whether a model’s answers change. In tests across four benchmarks covering math reasoning, health question answering, and safety, and across nine language models, performance became consistently less stable, with clear drops on three of the four benchmarks. The findings suggest that common static, single-turn evaluations may miss important real-world weaknesses, especially in high-stakes uses like healthcare and public services.
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.
