Find similar content

Best Practices Emerge for Deepfake and Chatbot Transparency under the EU AI Act A new research project sheds light on best practices for deepfakes and chatbot transparency under upcoming regulations. Researcher Thomas Gils from The Knowledge Centre Data & Society explored these requirements outlined in the EU AI Act. The AI Act acknowledges the importance of transparency in building user trust and accountability. It sets transparency requirements for high-risk AI systems, certain AI systems, and general-purpose AI models. To test these requirements, researchers conducted a mock compliance exercise with stakeholders, gathering their feedback. This resulted in three key best practices for deepfakes and chatbots: 1️⃣Accessible Disclaimers: Disclaimers should be accessible to diverse audiences, including users with disabilities. This means utilizing various communication methods like written text, visuals, and audio recordings. 2️⃣Right Amount of Information: Providing the appropriate amount of information is crucial. Striking the right balance is key to avoid overwhelming users with technical jargon or excessive details. 3️⃣Tailored Disclaimers: Tailor disclaimers to the intended and potential target audience. This ensures the disclaimers effectively address accessibility concerns and information needs of diverse user groups. This research helps developers of chatbots and deepfakes comply with the AI Act and ultimately fosters greater user trust in these evolving technologies. #AIAct#AITransparency

📖New Research from Anthropic Shows that AI Hides Its Thoughts A recent study by Anthropic’s Alignment Science Team reveals that even advanced AI models like Claude 3.7 Sonnet routinely obscure the actual reasoning behind their answers. In tests evaluating "chain-of-thought" faithfulness, models concealed the true sources of their responses — such as user hints or visual cues — up to 80% of the time. Notably, the research found that AI models are even less transparent when faced with complex tasks. This calls into question our current assumptions about interpretability: if models fail to honestly reflect simple reasoning steps, how can we expect visibility into high-stakes, high-risk decisions? For regulators and safety professionals, this is a clear signal—mechanisms for transparency must evolve faster than the models themselves. #AI#AIExplainability#AITransparency#AIEthics

AI Transparency in the Spotlight: New Senate Bill Protects Creators The "Transparency and Responsibility for Artificial Intelligence Networks (TRAIN) Act", introduced by Senator Peter Welch, aims to tackle a critical gap in generative AI development: transparency. If passed, the legislation would grant copyright holders the ability to subpoena AI training records when they suspect their work has been used without permission. Under the proposed framework, AI developers would need to disclose specific training data to confirm whether copyrighted material was used. Non-compliance would trigger a presumption that the developer had indeed utilized the copyrighted content, shifting the legal burden. Welch underscores the act’s significance: “If your work is used to train AI, you should have the right to know—and be compensated.” As generative AI reshapes creative industries, this bill marks a pivotal step in balancing innovation with the rights of artists, musicians, and creators. The debate over how to ensure ethical AI development is just beginning, and the TRAIN Act could set a precedent. #AITransparency#CopyrightLaw#EthicalAI#GenerativeAI

🇺🇸Microsoft Faces Watchdog Pushback on AI Advertising Microsoft’s ambitious marketing of its Copilot AI features has come under scrutiny. The National Advertising Division (NAD) has recommended that the company revise or discontinue certain productivity claims tied to Microsoft 365 Copilot, citing a lack of objective evidence. While Microsoft cited impressive user perception stats—like 70% reporting increased productivity—the NAD found the study insufficient to support hard ROI claims. More than just numbers, the watchdog also flagged the branding itself. With "Copilot" used across multiple tools, including Business Chat, users may not understand what the product can, and cannot, do. NAD advised Microsoft to make material limitations clear and avoid conflating distinct features under a single name. As AI tools enter mainstream business settings, regulatory clarity around marketing promises is becoming non-negotiable. #AI#MicrosoftCopilot#AIAdvertising#AITransparency

🇪🇺📖Study Finds Limited Availability of AI Training Data Disclosures Under EU AI Act Researchers from Trinity College Dublin report that information about AI training data required under the AI Act is often missing and difficult to locate. The law requires developers to publish summaries explaining how their models were trained, using a disclosure template designed to help copyright holders enforce their rights regarding the use of copyrighted material in AI training. A pre-print study funded by Mozilla found that only a small number of such summaries could be identified. The researchers also found structural issues in accessing the disclosures. The AI Act does not specify where companies must publish the summaries, leaving the decision to developers. As a result, no common publication mechanism exists and practices vary widely. The template created by the European Commission AI Office has led to heterogeneous implementations, making it difficult to determine whether the available documents meet EU transparency requirements. Most of the identified disclosures were produced by smaller organizations, including documentation for Switzerland’s Apertus national model. A document published by Microsoft for one of its open-source models was also reviewed, but the study found that it lacked several required details. Researchers recommend creating a centralized portal for publishing transparency summaries to improve accessibility and support enforcement once the AI Act obligations become applicable in August. #AIAct#AITransparency#TrainingData#Copyright#AIGovernance#AIRegulation#EULaw

🇺🇸Court Allows Enforcement of California AI Training Data Disclosure Law A US federal court has denied a request by Elon Musk’s AI company xAI to block enforcement of California Assembly Bill 2013. The law requires AI developers whose models are accessible in California to publicly disclose key information about training datasets, including dataset sources, collection timelines, whether collection is ongoing, and whether datasets contain copyrighted, trademarked, patented, or personal data. Companies must also indicate whether training data was licensed or purchased and the extent of synthetic data used. xAI argued the law would force disclosure of trade secrets, including dataset sources, dataset sizes, and data-cleaning methods. According to the company, such transparency could allow competitors to infer what datasets it uses and replicate its approach. The company warned that compliance could be “economically devastating” and reduce the value of its proprietary data practices. However, US District Judge Jesus Bernal ruled that xAI failed to demonstrate that the law requires disclosure of protected trade secrets. The court found the company’s claims too general and based largely on hypotheticals. The motion for a preliminary injunction was denied, allowing the law—which took effect in January—to remain in force while the lawsuit continues. #AIRegulation#AITransparency#TrainingData#TradeSecrets#AIAct#AIGovernance#TechLaw