What AI-native means
The phrase AI-native appears often in 6G discussions. It generally means that machine learning is a built-in part of network design, used for tasks such as radio resource management, channel estimation, energy saving and fault prediction, instead of an add-on analytics tool. There is no single agreed definition, and standards bodies are still deciding which AI functions belong in specifications. Readers should treat the term as a direction of research and not as a settled feature set.
Why Open RAN is a natural testbed
Open RAN already includes pieces that AI in the radio network requires. The RAN Intelligent Controller offers a place to run models close to the network, the E2 and A1 interfaces provide structured ways to collect data and apply policy, and the SMO framework supports model management at longer timescales. Because the interfaces are documented, third parties can write applications, which broadens who can contribute algorithms and encourages comparison.
The AI workload question
Two related ideas are often confused. The first is AI for the RAN: using models to improve how the radio network performs. The second is AI on the RAN: sharing the same infrastructure to run other AI workloads at the network edge. The AI-RAN Alliance was formed to study how these ideas can work together. Each raises its own questions about compute sizing, energy use, isolation between radio and AI tasks and commercial value. Claims that edge AI revenue will offset RAN costs should be backed by an operator's own business case and not accepted as a general rule.
Governance of models in a live network
- Data quality: models trained on unrepresentative data can behave badly on live traffic.
- Validation: test models in simulation or a safe environment before they influence real users.
- Guardrails: set limits on what any model-driven application may change, with automatic rollback.
- Explainability and logging: record inputs and decisions so operators can investigate failures.
- Lifecycle management: version models, monitor drift and retire models that degrade.
Limits and honest expectations
Published studies of AI in the radio network report benefits in specific scenarios, but results depend on the network, the data and the baseline. An improvement against a weak baseline may not appear against a well-tuned one. Operators should ask for results on networks like their own and for the full test conditions. Be cautious about vendor claims that AI will deliver a fixed percentage gain across all networks.
Regulation matters too. AI rules in some jurisdictions are developing, and network functions may be treated differently from consumer applications. Consult qualified advisers for obligations that apply to your market.
What to watch
Look for operator reports naming specific RIC applications in commercial use, for 3GPP decisions on AI functions in the 6G air interface and architecture, and for work on testing and certifying AI-driven applications. Those signals will show how much of the AI-native vision becomes standardized practice.
A cautious first project
A sensible first project is an energy-saving or traffic-prediction application in a small area. Define the objective, the baseline, the allowed actions and the rollback rule before starting. Run the model in advisory mode first, comparing its recommendations with what operators would have done. Only when the recommendations prove useful over a fair period should the application be allowed to act, and even then with limits. This sequence builds trust and produces evidence that is useful for wider decisions.
Skills and organization
AI in the radio network needs a blend of radio engineers, data scientists and software operations staff. Many operators find that the hard part is organizational, not technical: agreeing who owns the data, who approves a model for production and who responds when it misbehaves. Settle those questions in writing early, and give a single team accountability for each deployed application.
