Find similar content

AI Act's Foundation Model Provision Faces Obsolescence Experts warn that a crucial provision in the AI Act, aimed at assessing risks posed by foundation models like ChatGPT, could become obsolete within a year due to rapid technological advancements. According to Dragoş Tudorache, an MEP involved in the legislation, by the time the rules become applicable, either a few dominant models will meet the criteria or new technological breakthroughs will redefine the landscape. The provision in question, set to come into effect in 12 months, may struggle to keep pace with the evolving AI landscape. As the industry witnesses rapid advancements, including the potential for more efficient models or the emergence of new technologies, regulatory frameworks must adapt to ensure relevance and effectiveness. This development underscores the challenges regulators face in keeping up with the dynamic nature of AI technology. With the potential obsolescence of key provisions on the horizon, policymakers will need to remain vigilant and agile in crafting regulations that balance innovation with accountability. #AIAct#FoundationModels#AIRegulation

Rethinking Foundation Model Evaluations: A Call for More Rigorous Standards In a recent article, Elliot Jones, Mahi Hardalupas from the Ada Lovelace Institute, and William Agnew, Carnegie Bosh Fellow at Carnegie Mellon University, critically examine the current approaches to evaluating foundation models in AI. While global policy efforts, such as the EU's AI Act, have emphasized the importance of evaluating these models to mitigate risks, Jones, Hardalupas, and Agnew highlight significant gaps in this process. They argue that without standardized terminology, consistent methods, and mandatory enforcement, evaluations alone cannot guarantee the safety of AI systems in real-world applications. The authors caution that the selective and often voluntary nature of these evaluations may allow unsafe AI products to enter the market, undermining the goal of ensuring robust and responsible AI development. #AI#FoundationModels#AIEthics#AIRegulation

🌐📖Stanford HAI: Foundation Models Pose “Unprecedented” Privacy Risks The Stanford Institute for Human-Centered Artificial Intelligence (HAI) published a paper assessing privacy risks associated with foundation models and potential governance responses. The study finds these systems create “unprecedented and largely unaddressed” risks across the entire lifecycle: large-scale scraping of personal data during training, memorization and reproduction of sensitive information in outputs, and disclosure of intimate data through user interactions with chatbots. The paper also highlights technical vulnerabilities, including prompt injection, data poisoning, and model inversion, which can bypass safeguards and expose personal data. It concludes that existing frameworks such as the GDPR are structurally misaligned with how foundation models are developed, while neither the EU nor the U.S. has adopted comprehensive rules to address these risks. In the absence of clear regulation, privacy protection largely depends on voluntary actions by developers, prompting calls for stricter governance, including data minimization, transparency, privacy-by-design, and limits on harmful outputs. #AIregulation#Privacy#FoundationModels#GDPR#AIGovernance

Navigating General-Purpose AI Requirements: Insights from Stanford's Analysis Stanford’s Center on Research of Foundation Models has published a comprehensive overview of requirements for general-purpose AI, with a focus on the implications of the EU's AI Act. The analysis identifies 25 key requirements, primarily centered on disclosure obligations for developers to governments or downstream companies. Public transparency remains limited, with only one requirement advocating for a summary of training data to be disclosed to the public. Significantly, the report highlights the stringent requirements for models deemed to pose systemic risks, such as mandatory risk mitigation, incident reporting, and cybersecurity measures. Eight major models, including those from Google, Meta, and OpenAI, currently meet the systemic risk criteria. The analysis underscores the importance of these regulatory frameworks, especially as other regions like the US consider similar policies. #AIRegulation#StanfordAnalysis#AIAct#FoundationModels

EU AI Act Faces Uncertainty in 2023 Hello, everyone! The much-anticipated EU AI Act is in uncertain territory. European lawmakers are struggling to reach a consensus on how to regulate foundational AI models, and it's looking unlikely that the act will be passed into law before December. Currently, Spain, holding the EU presidency, is advocating for more rigorous vulnerability assessments and a tiered regulatory system based on the user base of AI models. So far, there have been several trilogues—discussions involving the European Parliament, the Council of the European Union, and the European Commission—regarding the AI Act. Ans so far there are concerns that decision-making on this legislation might be postponed until next year. European lawmakers had initially aimed to pass the AI Act by year-end. One of the AI Act's draft proposals suggests that developers of foundational AI models must assess potential risks, subject models to rigorous testing during development and post-release, scrutinize training data for biases, validate data, and publish technical documents before market release. In response, some open-source companies are calling for consideration of smaller enterprises in these discussions. They argue that compliance with these regulations might pose challenges for certain developers, emphasizing the need for distinctions between for-profit foundation models and those used by hobbyists and researchers. Stay tuned for further developments on this critical legislation! #EUAIAct#AIRegulation#FoundationModels#EuropeanLaw

Research Group Demands Global Shutdown of Foundation Model Development The Machine Intelligence Research Institute (MIRI) calls for a global halt on the development of foundation models, fearing they could "destroy humanity" without proper safeguards. Foundation models, capable of a broad range of applications, may evolve to be smarter than humans. MIRI urges a complete shutdown of attempts to build any system smarter than a human. This extends beyond the previous calls by tech leaders like Elon Musk and Steve Wozniak, who sought a pause on models more powerful than OpenAI’s GPT-4. MIRI stresses the need for urgent and sweeping legislation, including an "off switch" for AI systems to prevent malevolent behaviors. The group emphasizes the importance of addressing AI existential risks seriously and ensuring safe AI development in the future. #AI#ArtificialIntelligence#AIEthics#FoundationModels#MIRI

Research on compliance with the AI Act Stanford University researchers have conducted a thorough evaluation of major foundation model providers, including OpenAI and Google, to assess their compliance with the European Parliament's version of the AI Act. The findings reveal that these providers currently do not fully meet the Act's requirements, but the researchers believe that it is possible for them to do so in the future. One key observation from the analysis is the lack of adequate information disclosure by foundation model providers. Important details regarding data, compute, deployment, and key characteristics of their models are often not transparently shared. This raises concerns about transparency and accountability in the AI ecosystem. To address these challenges, the researchers suggest that EU policymakers consider additional factors to ensure that foundation model providers become more transparent and accountable. They emphasize the need for policymakers to apply these requirements selectively to influential providers, while avoiding excessive burden on smaller companies. Furthermore, it is crucial to allocate the necessary technical resources and expertise to the agencies responsible for enforcing the AI Act. Can policymakers ensure transparency and accountability in the rapidly evolving field of AI, while also fostering innovation and supporting smaller companies? #AIAct#FoundationModels#Transparency#Accountability#Compliance#Innovation

UK: CMA Releases Principles for Responsible AI Development Hello, everyone! The Competition and Markets Authority (CMA) has unveiled a set of principles to ensure the responsible development and use of foundation models (FMs) in AI. Foundation models, like ChatGPT and Office 365 Copilot, are versatile AI systems poised to revolutionize various sectors. The CMA's report lays out these guiding principles, with a focus on safeguarding consumer protection and fostering healthy competition in the AI industry. As AI rapidly integrates into our daily lives, the CMA recognizes the need for proactive intervention. These principles aim to strike a balance between promoting AI's potential for innovation and ensuring consumer safety. The CMA's proposed guiding principles focus on crucial areas like accountability, access, diversity, choice, flexibility, fairness, and transparency. These principles aim to guide FM developers and deployers toward responsible AI development and use. The CMA plans to engage extensively with stakeholders to refine these principles further. Stay tuned for updates in early 2024 as they continue shaping AI markets for the betterment of all. #ResponsibleAI#AIRegulation#ConsumerProtection#Competition#FoundationModels#CMAAIprinciples

The Complexity of Regulating Foundation Models in the AI Act Hello, AI & Law community! Kai Zenner, the Head of Office and Digital Policy Adviser at the Office of MEP Axel Voss, shared his opinion on the OECD website about regulating foundation models in the AI Act. 🔹 The Existing Gap: The proposed AI Act by the European Commission, created before foundation models gained prominence in AI, doesn't explicitly cover these versatile models. Their potential for diverse, unforeseen purposes makes it tricky to fit them into the current product safety approach. The Act's use case approach, limiting AI systems to specific risk classes, is too inflexible for the latest foundation models that can handle various tasks. This creates a regulatory gap that needs to be addressed. 🔹 Positive Progress: The European Parliament has taken a proactive step to tackle this issue by introducing Article 28b, which adds a regulatory layer specifically for foundation models. This article outlines nine essential obligations for developers, including identifying risks, testing, evaluation, and thorough documentation. These measures aim to strike a balance between ensuring safety and fostering innovation in the AI landscape. 🔹 Targeted Approach: A crucial consideration is to avoid putting too much burden on smaller providers while still effectively regulating foundation models. Zenner proposes adopting a systemic approach, targeting only a small number of highly capable and relevant foundation models under the AI Act. This strategy could be similar to how Very Large Online Platforms are designated under the Digital Services Act, ensuring a balanced and efficient regulatory framework. #AIRegulation#FoundationModels#AIAct#AIInnovation#AICommunity#TechLaw#OECDInsights