面向通信系统的可信人工智能白皮书

Content Abstract 4 Introduction Trustworthiness for traditional/generic AI/ML Trustworthiness for foundation models and LLMs Example use cases Conclusion Authors Abstract Artificial intelligence (AI) is becoming integral to next-generation telecom systems, but itbrings risks. The recent AI advancements in large language models (LLMs) and agentic AIintroduce new dimensions to that risk. To trust AI-enabled systems, we must be able to trust The key message of the paper is that for AI to be integrated into the telecom domain,including 5G and 6G networks, it must move beyond mere performance metrics to a holisticframework of trustworthiness. These capabilities should be embedded by design in thetelecom domain. Trustworthiness is a core requirement for adopting AI-based systems, Introduction As networks evolve into AI-native architectures, AI moves from a recommended trait to afoundational requirement, embedding trust at the core. In 5G, 6G, and autonomous networkmanagement, trustworthiness encompasses safety, security, transparency, reliability,and ethics. This integration of AI introduces a complex paradox: while AI is essential formanaging the scale and complexity of modern traffic, its black box nature and susceptibility At Ericsson, we build trust into our AI portfolio by design. Our solutions respect humanvalues and aim for outcomes that are both optimal and sustainable. This vision aligns with the EU AI ethics guidelines[1], which Ericsson adopted in 2019and which set out requirements for human agency and oversight, transparency, privacyand data governance, diversity, non-discrimination and fairness, technical robustness and Our previous whitepaper on trustworthy AI outlined this strategy and presented how toaddress trustworthiness requirements in AI-enabled systems. In Europe, the regulatoryprocess culminated in 2024 with the EU AI Act[2], which defines rules for both providersand deployers of AI and applies a risk-based approach with stricter obligations for higher-risk systems. Provisions of the act are now coming into force, and supporting materials such For creators and users of AI in telecom, trustworthiness is now a must, for compliance andfor market confidence. It also opens space for innovation in a trustful manner. Operatorsand subscribers expect it, and competitors will offer it. In addition, the recent explosion in the use of LLMs, generative AI, and agentic systemsbrings a new era for AI, which puts a new spin on trustworthiness. AI now interprets andgenerates natural language, reasons, operates autonomously, and can act in the real world.Thus, AI trustworthiness as a moving target calls for new techniques. The ethics guidelines For Ericsson, this means ensuring that our current and future networks—5G, 6G, andbeyond—become more autonomous, resilient, predictable, aligned with the regulations,and worthy of our customers’ trust. Trustworthinessfor traditional/generic AI/ML As networks progress towards higher levels of autonomy, AI/ machine learning (ML)functions are being integrated across mobile network architecture, from lower-level layersto high-level management [3]. It becomes essential to ensure these AI entities can be Here, “traditional” refers to AI models that require structured data with a singular modalityand rely heavily on feature engineering. These systems are often thought of as ML ratherthan AI. Trustworthiness requirements are addressed through the development andapplication of technical methods specific to telecommunication use cases. Transparency/explainable AI Transparency in AI-based systems provides interpretable, relevant, and understandablereasoning for why a decision was made and/or how AI arrived at this decision, usingexplainable AI (XAI) techniques. This builds customer confidence and trust and supports XAI gives developers and testers a window into the model's internal logic and helpsverify correctness. These techniques help engineers better understand the AI model andoptimize it with efficient feature engineering and data collection. Explainability can be applied to ensure correctness and enable trust at the model,component, or system level. At the model or component level, it provides model-specificinsights in terms of input features, model parameters, or internal logic. This is useful forvalidating the black-box models and improving them by identifying corner cases.It can also be applied to discover and describe unusual or new behavior exhibited by the System-level explainability makes sure the entire system behaves as expected. Itenhances traceability for the entire flow from inputs, through the model and othercomponents, to outcomes delivered to the user. These include dynamic, model-agnostic explainabilitytechniques such as decision flow trees or graphs, system traces, and case-basedsummaries. Surrogate model techniques can also be applied to explain other models. Both external users—customers and non-expert stakeholders—and intern