AI系统

legal or similarly significant effect on an individual,regulators will need to consider the suitability ofrequiring AI system operators to provide an appropriatejustification for that decision to affected parties.•AI systems should comply with regulatory requirementsrelating to vulnerability of individuals within specificregulatory domains.Regulators will need to considerhow use of AI systems may alter individuals’vulnerability,pursuant to their existing powers andremits.•Consider the role of available technical standardsaddressing AI fairness,bias mitigation and ethicalconsiderations(e.g.ISO/IEC TR 24027:2021,ISO/IEC12791*,ISO/IEC TR 24368:2022)to clarify regulatoryguidance and support the implementation of risktreatment measures.Accountabilityand governanceWe anticipate that regulators will need to:•Determine who is accountable for compliance withexisting regulation and the principles.In the initial stagesof implementation,regulators might provide guidance onhow to demonstrate accountability.In the medium tolong term,government may issue additional guidance onhow accountability applies to specific actors within theecosystem.•Provide guidance on governance mechanisms including,potentially,activities in scope of appropriate riskmanagement and governance processes(includingreporting duties).•Consider how available technical standards addressingAI governance,risk management,transparency and

1.42.Below,we provide some illustrative examples of AI systems to demonstrate their autonomous and adaptive characteristics.While many aspects of the technologies described in these case studies will be covered by existing law,they illustrate how AI-specific characteristics introduce novel risks and regulatory implications.Figure 1:Illustration of our strategy for regulating AIcharacteristics ensure any current or future AI system that meets this criteria will be within scope.See A guide to using artificial intelligence in the public sector,Government Digital Service and Office for Artificial Intelligence,2019.23A pro-innovation approach to AI regulationCase study 3.1:Natural language processing in customer service chatbotsAdaptivity:Provides responses to real-time customer messages,having been trained on huge datasets to identify statistical patterns in ordinary human speech,potentially increasing personalisation over time as the system learns from each new experience.Autonomy:Generates a human-like output based on the customer's text input,to answer queries,help customers find products and services,or send targeted updates.Operates with little need for human oversight or intervention.Illustrative AI-related regulatory implication:Unintentional inclusion of inaccurate or misleading information in training data,producing harmful instructions or convincingly spreading misinformation.Case study 3.2:Automated healthcare triage systemsAdaptivity:Predicts patient conditions based on the pathology,treatment and risk factors associated with health conditions from the analysis of medical datasets,patient records and real-time health data.Autonomy:Generates information about the likely causes of a patient’s symptoms and recommends potential interventions and treatments,either to a medical professional or straight to a patient.

When AI systems are not sufficiently explainable,AI suppliers and users riskinadvertently breaking laws,infringing rights,causing harm andcompromising the security of AI systems.At a technical level,the explainability of AI systems remains an importantresearch and development challenge.The logic and decision-making in AIsystems cannot always be meaningfully explained in a way that is intelligibleto humans,although in many settings this poses no substantial risk.It is alsotrue that in some cases,a decision made by AI may perform no worse onexplainability than a comparable decision made by a human.98 Futuredevelopments of the technology may pose additional challenges to achievingexplainability.AI systems should display levels of explainability that areappropriate to their context,including the level of risk and consideration ofwhat is achievable given the state of the art.Principle FairnessDefinitionandexplanationAI systems should not undermine the legal rights of individuals ororganisations,discriminate unfairly against individuals or create unfair marketoutcomes.Actors involved in all stages of the AI life cycle should considerdefinitions of fairness that are appropriate to a system’s use,outcomes andthe application of relevant law.Fairness is a concept embedded across many areas of law and regulation,including equality and human rights,data protection,consumer andcompetition law,public and common law,and rules protecting vulnerablepeople.Regulators may need to develop and publish descriptions and illustrations of