A Comparative Look At The EU And UK Strategies

A Comparative Look At The EU And UK Strategies Date: June 2023 Agenda 1. Definitions and Al categories 2. Obligations, standards/sandboxes, regulatory powers 3. Take home message Definitions - Technology neutral - Software + techniques and approaches listed in Annex I: 1. Machine learning approaches 2. Logic- and knowledge-based approaches 3. Statistical approaches - Commission's delegated acts for amendments. to Al technique and approaches - System Designed to operate with elements of autonomy / Based on machine &/or human provided data and inputs - Machine-based system / Designed to operate with varying levels of autonomy - General purpose Al system means an Al system that can be used in and adapted to a wide range of applications for which it was not intentionally and specifically designed - Foundation model means an Al model that is trained on broad data at scale, is designed for generality of output, and can be adapted to a wide range of distinctive tasks - For a given set of human-defined objectives - Infers to achieve a given set of objectives through machine learning &/or logic- and knowledge- based approaches  Commission's implementing acts to update approaches - For explicit or implicit objectives - Generate outputs such as: -  Content  predictions  Recommendations  Decisions Influencing (physical or virtual) environments of interaction Other Definition Life cycle of an Al system, Provider, small and medium-sized enterprise, User, Authorized representative, Product man area Importer, Distributor, Operator, Intended purpose, Reasonably foreseeable misuse, Substantial modification, Harmonized standard, Common specifications, Biometric data, Emotion recognition system, Biometric categorization system, Remote biometric identification system, Real time remote biometric identification system, Personal data, Non personal data, Post remote biometric identification system Risk, Significant risk Deployer Affected person, Biometric-based data Biometric identification, Biometric verification, Special categories of personal data, Profiling, Deep fake, etc. Defining Al by 2 characteristics that lead to the necessity for a bespoke regulation: 1. Adaptivity 2. Autonomy a. Novel risks b. Regulatory implications Al: products and services that are adaptable and autonomous Domain-specific definitions of Al formulated by regulators based on interpretation and application of Al's characteristics: 1. Future proof framework 2. Adapting the Al definition 3. Ongoing monitoring and iteration of the regulatory framework  Al supplier any organization or individual who plays a role in the research, development, training, implementation, deployment, maintenance, provision or sale of Al systems  Al user any individual or organization that uses an Al product  Al life cycle all events and processes that relate to an Al system's lifespan, from inception to decommissioning, including its design, research, training, development, deployment, integration, operation, maintenance, sale, use and governance  Al ecosystem the complex network of actors and processes that enable the use and supply of Al throughout the Al life cycle (including supply chains, markets, and governance mechanisms Impacted third party an individual or company that is impacted by the outcomes of the Al systems that they do not use or supply themselves. Al Categories From top to bottom: 1. Prohibited Practices (Unacceptable risk) 2. High Risk Al a. Mandatory requirements + ex ante conformity assessment b. Obligations for providers, users, value chain operators c. Sustainability: environmental footprint records 3. General Purpose Al Foundation Models a. Mandatory requirements (design, environmental) + Risk assessment and mitigation b. Transparency + registration in the EU database c. Technical documentation 4. Low Risk Al a. Transparency obligations 5. Minimal Risk Al a. Free use Requirements and obligations GP-AI/FM ≠ high risk Al systems unless directly integrated in a high-risk Al system No blanket new rules for specific technologies or applications of Al but Foundation models: 1. General purpose 2. Transformative potential 3. Opacity 4. Limitations Foundation Models: Definition: a type of Al model that is trained on a vast quantity of data and is adaptable for use on a wide range of tasks. Foundation models can be used as a base for building more specific Al models. - Foundation Model Taskforce - Life cycle accountability Obligations - Effect-driven outcomes under the EU Al Act based on requirements for high-risk Al systems: 1. Risk management 2. Data governance 3. Human oversight 4. Accuracy, robustness and cybersecurity 5. Transparency and system information disclosure; - Quality management: 1. Embedded 2. Ex ante conformity assessment - internal or third- party 3. Fundamental rights impact assessments by users 4. Post-market monitoring - Transparency vis-à-vis natural persons, for all AΙ systems interacting with them - Incidents reporting Obligations & Standards - Al supply chain: developers, users - Non-statutory obligations - Al governance compliance with Al principles - Risk assessment and management 1. Assurance techniques 2. Al technical standards Tools for trustworthy Al - Interaction and collaboration with authorities 1. Central Government; 2. Central Risk Function, 3. Regulators IDENTIFICATION ENFORCEMENT MONITORING Report new risk to Comply with guidance and Flag to Regulators/Central Regulators/Central Function. adopt a risk-based Function when mitigation approach by design. measures are ineffective Standards, Sandboxes & Regulatory Powers - Compliance supported mainly through: 1. Standards to be adopted by the European standardization organizations 2. Possibly in coordination / cooperation with international standardization bodies 3. European Commission's technical specifications 4. Regulatory sandboxes for development, testing and oversight in controlled environment 5. Industry / sector codes of conduct - Regulatory oversight: 1. At EU level -- EU AI Board v/s EU AI Office?  Methodological guidance and coordination among national authorities  How much oversight? 2. National supervisory authorities, acting as an accreditation ("notifying") authority and market surveillance authority 3. Accredited quality assurance entities ("notified bodies"), conducting ex ante thirdparty conformity assessments and certifying certain high-risk Al systems Take home message Similarities G7 Hiroshima Leaders' Communiqué: advance international discussions on inclusive artificial intelligence (Al) governance and interoperability to achieve our common vision and goal of trustworthy Al, in line with our shared democratic values. 1. Compliance with societal values 2. Protection of individuals' fundamental rights and safety 3. Risk assessment 4. Principles incorporated by design in the development and use of Al 5. Sandboxes 6. Standards 7. Future proof mechanisms (delegated/ implementing acts vs monitoring and evaluation the interpretation and application of cross-cutting principles) 8. International Interoperability Differences 1. Legislation: Regulation vs Policy: White Paper 2. Risk-based approach vs Principled-based approach 3. Functional definition of AI (UK) 4. Al categories vs No categories (except for Foundation Model) 5. Regulation of Al development, use and outcome vs Regulation of Al outcome 6. Allocation of regulatory powers (a limited number of regulators vs 7. all regulators concern with Al application in their remits) Roles of the Regulators (issuance of guidance in the application of 8. Al principles UK) 9. Regulation of the Al supply chain responsibilities and liabilities (at the moment only in the EU)