Key Issues for AI Integration in 6G

Artificial Intelligence is expected to play a much bigger role in 6G than it does in today's mobile networks. Almost every major 6G vision now refers to concepts such as AI-native networks, AI for Network, Network for AI and, increasingly, AI agents.

The terminology can, however, make the subject sound more mature than it really is.

There is broad agreement that AI will become deeply integrated into future mobile systems, but much less agreement about what an AI-native network should actually look like. Should AI capabilities be embedded within individual network functions? Should there be a separate AI plane or domain? Where should models execute? How should AI agents interact with network functions and with each other? And how can operators retain control when increasingly autonomous systems are making decisions about live networks?

These are some of the questions now being considered as the industry studies the architecture of the 6G system.

The ITU's IMT-2030 framework already gives AI a prominent role. AI and Communication is one of the six IMT-2030 usage scenarios, while AI integration is also identified among the new capabilities expected from future systems.

Within 3GPP, Release 20 is the main study phase for 6G architecture, while Release 21 is expected to introduce the first normative 6G specifications.

One of the important activities in SA2 is the Study on Architecture for 6G System, documented in TR 23.801-01.

A recent Omdia white paper, AI Native 6G System Architecture, provides a useful snapshot of this work and highlights four SA2 Key Issues that are particularly relevant to AI integration:

  • KI #18 – AI for 6G Architecture
  • KI #19 – 6G Network for AI
  • KI #21 – Data Framework
  • KI #22 – 6G Computing Support

Together, these illustrate why AI integration in 6G is not simply about inserting machine-learning algorithms into the network.

The first of these, AI for 6G Architecture, looks at how AI capabilities and AI agents could become part of the network itself.

AI is already used in mobile networks for tasks such as optimisation, prediction, anomaly detection and automation. In 6G, the expectation is that AI could take a much more active role in network operation.

This raises important architectural questions. AI functionality could potentially be embedded within existing network functions. Another approach would be to introduce dedicated AI functions or even a separate logical AI domain alongside the traditional control and user planes. Hybrid approaches are also possible.

Current proposals therefore include distributed AI capabilities embedded within network functions, AI-capable network functions, agent frameworks, hierarchical multi-agent approaches and separate AI domains. There is no single agreed definition today of what an AI-native 6G architecture actually means.

The question becomes even more interesting as networks start using AI agents. An AI agent may receive an objective or intent, collect information from several sources, reason about possible actions and interact with other agents or network functions to achieve that objective.

This is very different from traditional network automation based primarily on predefined rules and workflows.

The 6G architecture may consequently need to consider intent-based and non-intent-based requests, dynamic procedure composition, interaction between AI agents and network functions, AI capability exposure and mechanisms for monitoring AI performance.

Most importantly, operators will need mechanisms to determine how autonomous AI is allowed to become.

Allowing AI to recommend a configuration change is one thing. Allowing an autonomous agent to reconfigure a live network without human intervention is quite another.

Control, authorisation, auditability and fallback mechanisms will therefore be fundamental parts of any practical AI-enabled network architecture.

The second Key Issue, 6G Network for AI, approaches the problem from the opposite direction.

Instead of asking how AI can improve the network, it asks what the 6G network needs to provide for AI applications and AI agents.

This distinction between AI for Network and Network for AI is likely to become increasingly important.

Much of today's telecom AI activity falls into the AI for Network category. Examples include RAN optimisation, traffic prediction, automated assurance, intelligent orchestration, intent-based management and energy optimisation.

Network for AI is different. During the 6G timeframe, a growing proportion of network activity may originate from AI-enabled devices and autonomous systems rather than directly from humans using smartphones. Examples could include AI assistants, robots, vehicles, drones, cameras, XR devices and industrial systems.

AI agents on different devices may also need to discover each other, communicate with other agents or interact with AI capabilities located in the network, edge or cloud. This could produce traffic characteristics that look very different from today's consumer mobile broadband traffic.

Multimodal AI systems, cameras, machines and autonomous devices may create substantially more uplink traffic. AI agents might communicate continuously or trigger large bursts of data without following familiar human busy-hour behaviour. At the same time, increasingly capable on-device AI could reduce the amount of information that needs to leave the device.

Predicting future AI traffic is therefore extremely difficult. Rather than designing 6G around one assumed AI traffic model, networks will need enough flexibility to accommodate very different combinations of local processing, network-edge processing and cloud AI. The network may also need better mechanisms for AI-agent discovery and communication, exposure of network capabilities, AI-service provisioning and differentiated handling of AI-generated traffic.

This raises another important point. Network for AI does not simply mean giving AI applications more bandwidth. An AI application may also require particular latency, reliability, policy, compute-location or data-access characteristics. Connectivity and computing therefore increasingly become part of the same problem.

The third Key Issue is the Data Framework, and this may eventually prove to be one of the most important parts of AI integration in 6G. AI depends on data, but telecommunications networks contain enormous amounts of data distributed across different domains, systems and vendors. The challenge is not simply collecting more of it.

A 6G data framework may need to support functions including data discovery, collection, labelling, processing, storage and exposure. It may need to support network AI, UE AI and application-level AI while simultaneously dealing with privacy, consent, security and governance. Data quality also matters.

A model is only useful if the information on which it bases a decision is sufficiently accurate and recent for the task being performed. Network data can become stale very quickly. A radio condition measured several seconds ago may already be irrelevant for a real-time mobility or optimisation decision. Data provenance will therefore matter as well. Operators need to understand where information came from, how it was processed, whether it can be trusted and whether a particular AI system is authorised to use it.

The challenge becomes greater when AI models and network functions come from different vendors. A model developed by one supplier may consume information generated by equipment from another supplier and then make recommendations affecting yet another part of the network. For AI integration to work properly in a multi-vendor environment, the industry therefore needs not only interoperable network interfaces but also interoperable approaches to data, context and AI lifecycle management.

Models themselves will need managing too. They may need to be trained, validated, deployed, monitored, updated and occasionally rolled back. A model that performs well when initially deployed may degrade over time as network conditions, traffic patterns and customer behaviour change. Operating AI inside telecommunications infrastructure therefore cannot be treated as a one-off software deployment.

The fourth Key Issue is 6G Computing Support. This recognises that communication and computing resources may need to be managed much more closely together in future networks. 

Consider an AI application running on a mobile device. Inference might be performed directly on the device. This can provide very low latency and help preserve privacy, but it consumes device processing resources and battery power.

The same task could instead be moved to nearby edge infrastructure. That can provide greater computing capacity while retaining relatively low latency. Alternatively, a workload could execute in regional or centralised cloud infrastructure where substantially greater compute and storage resources are available. 

Each option involves different trade-offs. Future networks may therefore need to decide not only how to route packets, but also where a workload should execute.

This leads to the idea of a compute continuum. Different AI workloads may execute in different places depending on factors including latency requirements, available computing resources, device and network energy consumption, privacy and data-governance requirements, radio coverage and capacity, mobility and operational cost.

A real-time industrial control workload may favour nearby compute because latency is critical. A privacy-sensitive personal AI application may prefer local inference on the device. Training a very large model may require centralised infrastructure with access to much larger amounts of compute and storage. The optimum location may also change dynamically.

A device may normally perform inference locally but offload a more complex request when sufficient radio capacity and nearby compute are available. This means 6G may eventually have to coordinate communication, computing and AI workload placement rather than treating them as independent layers.

These four Key Issues also expose several wider challenges that will determine whether AI integration in 6G works in practice.

One of the biggest is interoperability. Telecommunications networks are inherently multi-vendor environments. Operators will not want AI agents that can communicate only with agents, tools and models supplied by the same vendor. Agents may therefore require standardised ways of discovering capabilities, exchanging information and invoking functions across vendor boundaries.

There is also a wider question about how much telecom-specific AI standardisation is actually necessary. The broader AI ecosystem is already developing mechanisms for agent communication, tool invocation and model interaction. Where suitable, the telecom industry may benefit from reusing widely adopted approaches rather than developing completely separate equivalents for everything.

Another fundamental issue is predictability. Telecommunications networks have traditionally been engineered around highly deterministic behaviour. Given the same state, configuration and message sequence, engineers generally expect a network function to produce the same outcome.

AI models, particularly generative AI and autonomous agents, do not necessarily behave this way. That may be acceptable when AI is recommending optimisation changes. It becomes much more problematic if an AI system directly controls availability-critical network functions.

This suggests an important principle: AI-native should not mean AI-only.

Critical network procedures are likely to continue requiring deterministic mechanisms, guardrails and fallback behaviour even when AI performs much of the optimisation, orchestration or decision support. Security also becomes considerably more complicated.

An AI system that can observe network information is useful. An AI agent authorised to change radio parameters, network policies, routing or resource allocation is much more powerful, and potentially much more dangerous if compromised. Authentication and authorisation may therefore need to apply not only to users and network functions but also to AI agents, tools, models and data sources.

AI introduces additional threats including poisoned training data, malicious inputs, manipulated model outputs and attempts to persuade autonomous agents to misuse legitimate tools or permissions. Trustworthiness therefore becomes as important as intelligence.

Energy and sustainability are another consideration. AI can help reduce network energy consumption through improved traffic prediction, resource management and network optimisation. But AI itself consumes energy. Running inference continuously across large numbers of base stations, edge servers, cloud platforms and devices could create a significant new source of power consumption. The value of an AI solution therefore needs to be considered against the additional compute, hardware, operational complexity and energy required to deliver it.

Finally, there is the question of how much of today's architecture should actually change. The term AI-native can imply that 6G will replace today's mobile architecture with something fundamentally new. That seems unlikely, particularly during the early stages of deployment.

The 5G Service-Based Architecture provides an important foundation that can evolve into 6G, and operators will also need smooth coexistence and migration between generations. The first 6G systems may therefore combine evolved 5G architectural concepts with new 6G network functions and additional AI, data and computing capabilities. That may ultimately be much more practical than attempting to redesign the entire network around AI from day one.

Artificial Intelligence will undoubtedly play a major role in 6G, but the difficult part is no longer demonstrating that AI can optimise networks. The real challenge is integrating AI into a telecom system in a way that is interoperable, secure, trustworthy, controllable and economically worthwhile.

As the Release 20 studies progress towards normative 6G work, some of the most important questions will therefore be where intelligence should reside, how autonomous AI agents should become, how data and models should be managed across multiple vendors, how communication and computing resources should be coordinated, and how workloads should be placed across devices, edge and cloud infrastructure.

Perhaps most importantly, the industry still needs to determine where AI provides sufficient benefit to justify introducing additional complexity into an already complex mobile network.

The answers to these questions will ultimately determine whether AI-native 6G becomes a genuine architectural transformation or simply another label attached to the next generation of mobile technology.

Related Posts: 

Posted by on

Comments

Post a Comment