5G bolted AI onto the network. 6G is designed around it. An AI-native air interface means machine learning is part of the physical layer from the first specification — not an optimisation added later. Here is how 3GPP is getting there.
The starting point: AI/ML in 5G-Advanced
3GPP Release 18 introduced the first study and work on AI/ML for the NR air interface, focusing on three high-value use cases:
| Use case | What AI improves |
|---|---|
| CSI feedback | Compress channel state information with autoencoders → less overhead, better accuracy |
| Beam management | Predict the best beam in time/space → faster, lower-overhead beam tracking |
| Positioning | ML-aided positioning, especially in tough multipath (indoor/industrial) |
Release 19 expands this — more use cases, and crucially the life-cycle management (LCM) of models: how models are trained, identified, activated, monitored and rolled back across the network and devices.
What 'AI-native' actually means in 6G
- Designed-in, not bolted-on: waveform, reference signals and procedures co-designed with ML in mind.
- Two-sided models: coordinated models split across UE and gNB (e.g., encoder in the device, decoder in the base station).
- Model life-cycle management: standardised training data, model delivery, performance monitoring and fallback.
- AI for and on the network: AI optimises the air interface, and the network carries AI workloads (federated learning, edge inference).
The engineering challenges
- Generalisation: a model trained in one city must work in another — robustness and dataset diversity are critical.
- Interoperability: multi-vendor two-sided models need agreed interfaces and reference data.
- Complexity & energy: inference must fit device power and latency budgets.
- Monitoring & fallback: detect when a model degrades and revert safely to classical methods.
How to skill up
AI-native 6G rewards engineers who understand both the NR physical layer and machine learning. CafeTele pairs deep RAN training with AI-for-telecom material — start with the 6G & Release-20 course, the 6G complete guide and our browser 5G labs.
Frequently asked questions
What is an AI-native air interface?
It means machine learning is designed into the 6G physical layer from the start — the waveform, reference signals and procedures are co-designed with AI, rather than AI being added as a later optimisation.
Where did 3GPP start AI/ML for the air interface?
In Release 18, with AI/ML for CSI feedback, beam management and positioning. Release 19 expands the use cases and adds model life-cycle management.
What is a two-sided AI model?
A model split across the device and the base station — for example, a neural encoder in the UE and a decoder in the gNB for CSI feedback — which requires standardised interfaces to interoperate.
What is model life-cycle management (LCM)?
The standardised process of training, delivering, activating, monitoring and rolling back AI models across the network and devices — the hardest part of making AI-native practical.
Do I need to be a data scientist to work on 6G AI?
No, but you need both sides: a solid grasp of the NR physical layer plus practical machine-learning fundamentals. CafeTele's RAN and AI-for-telecom material covers both.
Learn to build and optimise these networks
CafeTele turns the spec into skills — cinematic video courses, 10+ browser-based 5G labs, and 40+ free engineer tools.
Browse courses Try a 5G lab Free tools