Did OpenAI hide evidence in the NYT copyright case?
++ Apple Sues OpenAI Over Trade Secret Theft and Contract Breach Claims, Netherlands launches global AI strategy, Google adds AI labels to ads, Meta removes Instagram AI photo tool & more..
This week’s highlights:
The New York Times and The Daily News have asked a judge to punish OpenAI in their ongoing copyright lawsuit, claiming the company hid evidence about how it tracked copyrighted content in ChatGPT training data and user chats. The publishers say OpenAI wrongly claimed it could not search its training datasets or chat logs, even though internal tools and a large database of de-identified conversations were allegedly already being used for that purpose. They also argue OpenAI gave the court an unusable sample of chat logs and may have deleted or replaced large amounts of data after the case began. OpenAI has denied the allegations, saying the claims are false and that it is trying to protect user privacy while defending its fair use position.
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
Apple Sues OpenAI Over Alleged Trade Secret Theft and Contract Breach
Apple has sued OpenAI in a California federal court, accusing the company of trade secret theft and breach of contract tied to its growing hardware efforts. The lawsuit claims former Apple employees at OpenAI, including senior hardware executive Tang Tan, misused Apple’s confidential project details, sought unreleased product information, and tried to bypass Apple’s security rules. Apple also alleges another former employee kept an Apple laptop and downloaded sensitive technical files before joining OpenAI. OpenAI said it has no interest in other companies’ trade secrets and is focused on building its own technology, while Apple is asking the court to stop any use of its confidential information and force the return of any materials.
Netherlands Sets International AI Strategy for Safe Use and Competitiveness
The Dutch government has launched a new international AI strategy aimed at making AI safe, responsible, and future-ready while protecting public values and boosting the Netherlands’ and Europe’s position in the field. The plan treats AI as a key strategic technology that affects the economy, security, and global influence, and says the country must reduce dependence on other nations. It focuses on three main goals: improving competitiveness, increasing AI sovereignty, and ensuring safe and responsible use of AI. The strategy also calls for closer international cooperation through joint investment, better market access, stronger partnerships across the AI chain, and common global rules to address risks such as misuse of advanced AI and human rights violations.
Google Adds AI Disclosure Labels to Ads Across Search, YouTube
Google is adding a new disclosure feature to show when an ad was created or edited using AI. The label will appear in the “My Ad Center” panel on ads shown through Google Search, YouTube, and Google Discover, helping users understand when images may not be real product photos. Ads made with Google’s own generative AI tools will get the disclosure automatically, while advertisers using outside tools will have to self-report AI use. Google said some ads may also carry AI labels in certain markets where local laws require it.
Meta Removes Instagram AI Photo Tool After Strong User Backlash
Meta has removed a new Instagram AI feature that let people generate images using photos from public accounts after strong backlash from users and talent agencies. The tool was part of Meta’s new Muse Image rollout and allowed public Instagram profiles to be referenced without notifying the account owner. Meta said the feature “missed the mark” and has now been taken down. The criticism centered on fears that the tool could be misused, especially for harmful or inappropriate AI-generated images, a problem that has already affected social media platforms.
Reddit Uses AI Tools to Detect and Reduce AI-Generated Spam
Reddit says it is using large language models to fight a growing spam problem that large language models have also helped worsen across the internet. The company said its updated AI-powered tools now detect subtle and coordinated fake activity better than older systems, helping block about 23 million spam views and catch around 25,000 new spam posts and comments each day. Reddit also said user exposure to spam fell 20% from January to March compared with the previous three months. The development highlights how online platforms are increasingly using AI to police AI-made content, though experts say human moderation is still needed for the best results.
Amazon to Stop Taking New Mechanical Turk Customers Starting July 2026
Amazon will stop accepting new Mechanical Turk customers from July 30, 2026, while existing customers will still be able to use the service. AWS said it will keep supporting the platform with security and availability updates, but no new features are planned. Launched in 2005, Mechanical Turk became known for paying workers small amounts to do tasks such as labeling data, solving CAPTCHAs, and reviewing content, later becoming part of AI training workflows. In recent years, the platform faced criticism over labor practices, fraud, and falling trust, with reports also showing many workers were using large language models to complete tasks.
Discord Admits AI Moderation Bug Wrongly Banned Thousands Over Harmless Images
Discord has admitted that a bug in its AI moderation system wrongly banned more than 8,000 users over the past two months after harmless images such as spreadsheets, chessboards, game textures, and plain transparent backgrounds were mistakenly flagged as harmful. The company said the issue had been affecting accounts since May, with about 200 more users banned over the weekend before the bug was found and fixed, and it is now restoring affected accounts. Discord explained that its system compares uploaded content with databases of known harmful material, but false matches can happen, and a bug caused bans to be applied immediately. The incident has renewed concerns about the risks of AI-based moderation, as users said sudden account bans can seriously disrupt work, gaming communities, and personal connections.
Google Deepfake Detector Helps Debunk Fake McConnell Hospital Image Online
Google’s SynthID watermarking system helped expose a fake image that falsely appeared to show Senator Mitch McConnell in severe distress in a hospital bed. The image spread widely on social media, but fact-checkers found it carried Google’s invisible SynthID marker, confirming it was AI-generated. The case is seen as a notable success for anti-deepfake technology because the watermark remained detectable even after the image was shared across platforms. SynthID was introduced in 2025 and works only with image tools that support the system, including Google’s Gemini and, since May 2026, OpenAI’s image models.
Anthropic Adds Claude Reflect Dashboard to Encourage Ongoing Everyday AI Use
Anthropic has added a beta feature called Reflect to Claude that shows users charts and summaries of how they use the AI, including topics, habits, and common tasks. The tool is designed not only to track usage but also to make Claude feel like a regular part of daily work while encouraging more thoughtful AI use through reminders, quiet hours, and break nudges. It can also suggest better ways to use Claude’s features, which may help the company keep users more tied to its platform instead of switching to rivals. Anthropic said sensitive chats may appear only in broad form, health-related conversations are excluded, and the insights data is not used for other purposes.
Questions Grow Over How OpenAI’s Sol Won Government Release Approval
The U.S. government appears to have allowed OpenAI’s new frontier model, Sol, to reach the public without a clearly explained review process, and even experts say they do not know exactly how that decision was made. Reports indicate there were private talks between OpenAI and senior officials, and the company pointed to outside safety testing, but key details about who evaluated the model, what standards were used, and which agency is in charge remain unclear. The broader system for approving powerful AI models is still being built, with no settled rules on which models need review or how that review should happen. Critics say this secretive and ad hoc approach risks giving too much influence to personal connections and too little to independent safety experts.
Hugging Face Says Open Source AI Is Becoming More Important
Hugging Face’s CEO said open-source AI is becoming more important as many companies move away from expensive frontier AI APIs and adopt open models as they grow. Hugging Face, now widely used by large businesses including about half of the Fortune 500, has become a major platform for sharing AI models and datasets. He also said it is a concern that most open models downloaded in the U.S. are coming from Chinese labs, but argued this should be addressed by building stronger alternatives rather than by rejecting open source. The discussion also highlighted worries that a small number of big companies could gain too much control over AI, and said open, transparent AI is especially important in robotics because robots could have access to sensitive parts of home and family life.
Supreme Court Rejects AI-Fake Judgments, Warns of Zero Tolerance in Courts
India’s Supreme Court has warned that AI-generated fake case citations have no place in courts and set aside an NCLT order that relied on such false precedents in the Essel Infraprojects insolvency matter. The court said both lawyers and judges must follow a zero-tolerance approach and independently verify any AI-generated material before using it in legal proceedings. It held that citing fake AI judgments is professional misconduct, and any ruling influenced even partly by hallucinated material cannot be treated as a valid judicial decision. At the same time, the court clarified that it is not against the proper use of AI in the legal field, but against passing off fabricated material as genuine law.
OpenAI Research Finds Major Flaws in SWE-Bench Pro Coding Evaluations
OpenAI said a new audit found major problems in SWE-Bench Pro, a widely used benchmark for testing AI coding ability, and warned that its scores may not reliably show how capable models really are. The company estimated that about 30% of the benchmark’s 731 public tasks are broken, with issues such as unclear prompts, overly strict tests, weak test coverage, and prompts that point models toward the wrong answer. Its review used automated checks, AI investigator agents, and human software engineers, who found flaws in a large share of the dataset. Based on these findings, OpenAI said developers should treat SWE-Bench Pro results carefully and withdrew its earlier recommendation to adopt the benchmark.
🚀 AI Breakthroughs
OpenAI launches GPT-5.6 model family with stronger coding and cybersecurity tools
OpenAI has released a new AI model family called GPT-5.6, with three versions: Sol, Terra, and Luna, aimed at enterprise work, coding, scientific research, and lower-cost use cases. The company says the models are more efficient than earlier versions, with Sol described as its best coding model and its strongest cybersecurity model so far, built to support defensive tasks such as threat modeling, code review, patching, and blue team testing. OpenAI also rolled out ChatGPT Work, a workplace assistant for desktop, web, and mobile that helps teams create documents, spreadsheets, and presentations. GPT-5.6 is now available through ChatGPT, Codex, and the OpenAI API, with pricing starting at $1 per million input tokens and $6 per million output tokens for Luna.
OpenAI Releases New Voice Models for More Natural Live Conversations
OpenAI has released two new voice models, GPT-Live-1 and GPT-Live-1 mini, that are designed to make ChatGPT sound more natural and handle live back-and-forth conversation better, including interruptions and real-time translation. The company is making GPT-Live-1 mini the default voice mode in ChatGPT, while paid users will get access to the larger GPT-Live-1 model, which can also connect to newer GPT systems for search, reasoning, and visual responses during a conversation. OpenAI said the new voice mode is built for longer, hands-free conversations and added that more than 150 million people already use ChatGPT’s voice and dictation features. However, the live translation demo still showed limits, with Hindi speech sounding unnatural and heavily accented, and the company did not clarify which languages are best supported.
OpenAI Shuts Atlas as It Expands ChatGPT Browser Features Elsewhere
OpenAI is shutting down Atlas, the AI browser it launched in October, but it is continuing to expand its AI tools for web browsing. Instead of keeping Atlas as a separate browser, the company is moving its main features into ChatGPT’s desktop app and a new Chrome extension. The Chrome extension can read the page a user is viewing, answer questions, summarize content, and help with longer tasks, putting it in direct competition with Google’s Gemini tools. At the same time, the desktop app is getting stronger browser features, including the ability to open websites, sign in, download files, and let AI agents complete tasks through a cloud-based browser.
Meta Launchs Muse Image as Users Raise Privacy Concerns Over Photos
Meta has launched Muse Image, a new AI image generator from Meta Superintelligence Labs, now available for free in the Meta AI app, Instagram Stories, and WhatsApp. The tool can create and edit images, suggest ready-made prompts, and support uses such as ads, interior design ideas, and social media effects. But the launch is already facing backlash because it lets users turn photos from public Instagram accounts into AI-generated images by tagging those users, unless they manually switch the setting off. The company says people can control this feature, but critics say the opt-out system raises fresh privacy concerns, especially given Meta’s long history of scrutiny over how it handles user data.
Meta Launches Muse Spark 1.1 to Challenge AI Coding Rivals
Meta has launched Muse Spark 1.1, a new multimodal AI model built for agentic coding, joining a fast-growing market already led by OpenAI and Anthropic. The company says the model can handle multistep reasoning, manage digital workflows, fix bugs, support large code migrations, and help deploy features in enterprise systems. Pricing is set at $1.25 per million input tokens and $4.25 per million output tokens, making it broadly competitive with rival offerings. The release adds to a busy week in AI, as Meta also unveiled a new image-generation model while other major AI firms rolled out fresh model updates.
Google Photos Adds AI Video Remix Tool for Fast Video Editing
Google Photos has added a new AI video editing feature called Video Remix that can transform clips in seconds using Google’s Gemini Omni model. The tool, available in the app’s Create tab, lets users brighten dark videos, change backgrounds, and apply styles such as watercolor, sketchbook, and oil painting effects with just a few taps. Google said the feature is part of its wider effort to bring generative AI tools into consumer apps and make video editing easier without professional software. Video Remix is rolling out to eligible Google AI Plus, Pro, and Ultra subscribers in the U.S. and several other countries, alongside other recent AI upgrades in Google Photos.
SpaceXAI Releases Grok 4.5 as Faster Lower Cost Opus Class Model
SpaceXAI has released Grok 4.5, its first new AI model since going public, describing it as a general-purpose system for coding, research, writing, office work, and other routine knowledge tasks. The company said the model is more token-efficient than rival systems and priced it at $2 per million input tokens and $6 per million output tokens, making it cheaper than several competing premium models. Company benchmark results suggest Grok 4.5 is competitive with other leading AI models, though not clearly the top performer. The release comes during a busy week for the AI industry, with another major model launch from OpenAI also expected.
CEO Says Video Games Beat Internet Data for Training AI Models
General Intuition, a New York startup backed by Jeff Bezos, says video game data could be better than internet text for building AI that understands movement, space, and time. The company argues that while large language models like ChatGPT and Claude are strong at text, they are weaker at learning how the physical world works, which is important for broader artificial intelligence. General Intuition is developing world models trained on gaming data and says this approach could help advance physical AI. The startup, which grew out of gaming platform Medal TV, recently raised $320 million at a reported $2.3 billion valuation, with investors including Coatue, Eric Schmidt, and researchers linked to MIT and Google DeepMind.
Sony Pictures India Ties Up With Google Gemini for KBC Season
Sony Pictures Networks India has partnered with Google Gemini for the upcoming season of Kaun Banega Crorepati, adding an AI-based feature to support viewers and contestants. Google Gemini will act as an interactive knowledge companion, helping users with KBC registration as well as general knowledge preparation through the app. The tie-up is aimed at improving audience engagement and adding a new layer to the show’s content experience. The collaboration combines the long-running popularity of KBC with Google Gemini’s artificial intelligence tools.
MeitY Empanels Six Firms Including TCS for Government AI Projects
The Ministry of Electronics and Information Technology’s National e-Governance Division has selected six technology firms, including TCS and CoRover, to help roll out AI projects across central and state government departments as well as public sector units. The other empanelled companies are NEC Corporation India, Innefu Labs, Kyndryl Solutions, and Cactus Technology Solutions, chosen through a competitive process involving nearly 80 bidders. The move is meant to speed up government AI adoption by letting departments directly hire these firms for consulting, development, deployment, automation, analytics, and technical support without issuing separate tenders each time. The empanelment will remain valid for two years, with an option to extend it by one more year, and supports the broader Digital India and IndiaAI Mission goals across sectors such as healthcare, education, agriculture, taxation, and citizen services.
Government, Nasscom Develop AI Curriculum for Undergraduate Programmes Across India
The government is working with IT industry body Nasscom to create a new artificial intelligence curriculum for undergraduate programmes in India. The updated course is expected to take about six months to finalize, as it will need review from bodies such as AICTE and UGC. The move comes as AI becomes more important across industries and raises demand for reskilling and upskilling among graduates. The development follows the government’s earlier plan to introduce AI education in schools from Class 3 onward.
🎓AI Academia
Early Review Highlights Governance Challenges for Fast-Rising Agentic AI Systems
A new academic review says agentic AI is quickly moving beyond generative AI into systems that can plan, adapt, and carry out tasks with little or no human oversight, creating fresh governance and ethics concerns. The paper describes 2025 as a major turning point for agentic AI, pointing to fast adoption by businesses, rising investment, and strong market growth forecasts. It argues that these systems need more targeted rules than traditional AI because of their higher autonomy and ability to act on behalf of people. The review maps out early priorities for governance, including oversight, accountability, and the roles of different stakeholders, and presents itself as a starting point for a broader roadmap for responsible agentic AI governance.
Study Highlights Major Gaps in AI Incident Governance and Reporting
A new research paper says AI systems can still fail after launch in ways that testing before release does not catch, including from unexpected behavior, misuse, or attacks. It finds that current AI incident governance is fragmented, with no common standard for how incidents are defined, classified, monitored, reported, or analyzed. That lack of consistency makes it harder to compare cases, understand causes, and prevent similar harms in the future. The paper argues that the biggest gap is the absence of standardized monitoring and reporting rules, and it proposes practical guidelines and a reporting template to improve accountability and AI safety.
Framework Guides Banks on Governing Generative AI Under SR 26-2
A new research paper says U.S. banking rules have been updated with SR 26-2, replacing SR 11-7 with a more risk-based approach to model oversight, but leaving generative and agentic AI outside the formal framework. The paper argues this creates a governance gap because AI tools can still shape important banking processes such as monitoring, policy review, and the drafting of adverse-action explanations. To address that, it proposes a Generative AI Control Framework, or GAICF, designed to help financial institutions manage AI use in regulated workflows even when those systems are not treated as formal risk models. The goal is to give banks a practical way to align generative AI controls with the supervisory expectations set by SR 26-2.
New AI Safety System Proposes Early Warnings and Sectorwide Risk Response
A new research paper argues that frontier AI needs oversight similar to post-2008 banking rules, because checking one model at a time is not enough to catch risks spreading across the whole industry. It proposes a two-layer system for internal AI used inside labs, with the first layer sending reports about dangerous capabilities, autonomy signals, and security breaches to a government clearinghouse and expert response groups. The second layer would tie safety controls to measurable risk indicators such as compute exposure, red-team testing effort, and model robustness, so that stronger safeguards kick in automatically as systems become more powerful. The paper says this approach could help spot industry-wide risk build-ups earlier and create clear response steps before failures spread across multiple frontier AI labs.
Enterprise AI Governance Model Sets Reference for Trusted Agentic Systems
A new June 2026 paper describes AGL-1, a vendor-neutral reference model for enterprise AI governance. It presents the Enterprise AI Governance Layer as a control plane designed to help companies manage key parts of AI systems, including retrieval, memory, policy enforcement, provenance, observability, and agent-driven execution. The goal is to support trusted enterprise intelligence by giving organizations a structured way to oversee how AI systems access data, follow rules, track sources, and operate in real-world business settings. The paper positions AGL-1 as a governance framework rather than a product, aimed at improving control, transparency, and reliability across enterprise AI deployments.
CAGE-1 Framework Sets Trust and Control Checks for Enterprise AI Agents
A new independent technical report outlines CAGE-1, a framework designed to test whether enterprise AI agents are safe and ready for real-world deployment. The report says companies are moving beyond basic AI chat and search systems toward agents that can plan tasks, use tools, update systems, and work across business apps, creating much higher operational risk. CAGE-1 focuses on whether an agent’s actions are properly authorized, policy-compliant, traceable, and stoppable before they cause business impact, not just whether the agent gives correct answers. It also adds the idea of “Prebind Assurance,” which means proving an action is under control before it becomes effective or binding, while checking areas such as memory, retrieval quality, tool safety, human oversight, and safe failure.
Study Outlines Risk Framework for AI-Native Engineering Teams and Governance
A new research paper says traditional software risk controls do not work well for teams building agentic AI systems, because these systems are probabilistic, can take multi-step actions on their own, and can change risk behavior quietly between deployments. The paper proposes an organizational framework to help engineering managers assign responsibility, detect failures, and escalate problems more effectively in AI-native teams. Its main finding is that risk coverage gets worse as teams move from standard software to AI-native operations, with some of the most serious failures happening at the boundary where AI outputs are used by systems that expect fixed, predictable behavior. The paper argues that managers should focus less on only owning software components and more on owning risk surfaces, monitoring semantic behavior, and preparing rollback plans when AI actions affect other teams or systems.
Study Proposes Audit Framework for LLM Use and AI Content Governance
A new research paper argues that large language models can give polished, well-structured answers that make users feel they understand a topic, even when they lack real-world experience or evidence to judge it properly. The paper calls this risk “pseudo-rational cognition” and says AI systems mainly remix large amounts of human knowledge into responses that only appear deeply reasoned. It also warns that unchecked AI-generated content can create wider problems, including false confidence, errors in AI-to-AI exchanges, weak detection systems, and “memory pollution” when such content is stored and reused later. To reduce these risks, the study proposes an auditing framework that asks users to define the task clearly, check evidence sources, test outputs in practice, log changes, and keep ways to review or reverse decisions.
Australian Government AI Transparency Rules May Miss People Most at Risk
A new study of 92 AI transparency statements from Australian government agencies finds that many agencies are meeting official disclosure rules on paper, but that does not always mean the public is getting meaningful transparency. The research says statements are usually better suited to people inside government or those with more control over AI systems, while people most affected by AI decisions often get less useful information. It describes this gap as a “transparency illusion,” where compliance creates the appearance of openness without equally serving all stakeholders. The paper argues that transparency should be judged by whether it meets the needs of different groups, not just by whether an organisation has published the required documents.
Study Benchmarks Open-Weight AI Models for Global Governance Bias Accuracy
A new study examines whether open-weight language models are less accurate when answering questions about countries that are less represented in training data, especially in the Global South. It tests four publicly available models against a verified global AI governance dataset covering 227 countries, using nearly 3,000 country-year observations from 2010 to 2023. The research also uses a more detailed scoring system to separate accurate answers from confident false claims, refusals, vague hedging, and wrong source attribution. The goal is to produce a more reliable and repeatable benchmark for measuring geographic bias in AI systems used for policy and governance work.
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.
