Juan Montojo, a leading expert in 3GPP TSG Radio Access Networks (RAN) and rapporteur for the work item Artificial Intelligence/Machine Learning for NR air interface (NR_AIML_air), has spoken during the ETSI AI Conference on February 11, 2025 - delivering an overview of the technology’s progress in 3GPP to date and the future prospects as we consider 6G priorities.

AI models have been in use by vendors and network operators for some time, as an implementation choice, enhancing some conventional methodology in areas such as network management and automation (SON, etc) as well as various processing algorithms at the network and device sides.  

3GPP Priorities

Juan Montojo told the conference that 3GPP AI/ML projects have been chiefly to help improve those specific areas, given that there are no plans in sight to specify AI/ML models themselves. The priority in 3GPP groups has been to ensure some basic – but key – advances in the areas of:

• Infrastructure/operator control.

• AI/ML model performance monitoring.

• AI/ML model activation and deactivation.

• Air interface extensions specific to AI/ML implementations.
• Establishing early standards-based approaches for devices’ data collection, as well as AI/ML model transfer and delivery.
• Testing, interoperability and consistent device behaviour.

Looking at the industry engagement in AI/ML, Montojo gave a personal assessment of the pro’s and con’s of current day AI/ML in cellular networks. He said that at this early stage in its development, the technology is very good for bringing a lot of information together for localized data-driven decision making. On the other hand, it is not so good for solving complex problems with analytical solutions at hand.

AI/ML based implementations are expected to shine in complex non-linear problems with a high degree of if/then/else conditions to consider, which are currently solved from heuristics from multiple generations of wireless technologies. Another byproduct of that complexity has been that AI/ML processing is not good for energy consumption in the network or on the device side.

Montojo acknowledged that there is a broad demand that AI should be everywhere…and soon. He described how 3GPP is working on offering standards- based mechanisms for operators to collect data and manage AI/ML models throughout their network. This standards-based approach complements having ‘over the top’ services for e.g., data collection for training AI/ML models and for AI/ML models delivery.

Juan Montojo described AI/ML model training and the concept of On-line and Off-line learning in his presentation noting that all 3GPP work on AI/ML for NR air interface has assumed thus far off-line training, i.e, AI/ML models are deployed in commercial networks when they have already been fully trained.

He returned to that in his summing up – Identifying the potential for some form of On-line training of AI/ML models to grow from the start of 6G. He noted that 3GPP WG SA2 has already looked at some forms of Federated Learning as part of their work on enhanced network automation.

3GPP has an established Release based progression for features and projects. In his presentation, Juan Montojo explained the AI/ML work split within 3GPP working groups since Rel-16. He looked at the options for data collection being considered and the requirements for AI/ML use cases over 3GPP systems.

Principles for AI work in 3GPP 

The agreed principles, driven by operators and resembling some of the principles of the European Commission’s AI Act, are:

1. Data must be secured and data integrity & confidentiality ensured.
2. Data privacy, anonymity and user consent respected.
3. Operators to retain control of standardized data collection transfer process & to manage data transfer to the server for UE-side data collection, without the need for SLA (includes initiating, terminating & managing data transfer).
4. Operator to have full visibility for standardized data.
5. Design to be futureproof & extendable.

As part of the study of data collection solutions, 3GPP will investigate the handling of non-standardized data collection (denoted as partial visibility) and on the overall ‘controllability’ of growing data collection.

To avoid discussions being overly abstract, 3GPP has concentrated on finding tangible AI/ML use cases to focus on. Juan Montojo looked at both RAN and architecture examples and then moved to some concluding thoughts on the limits of standards work on AI/ML. As a counter balance to that, he also stressed the potential standards hold for making a positive impact on the performance of both cellular networks generally and on the AI/ML eco-system itself.

The main take-aways of this presentation were that AI/ML is pervasive and that its reach into 3GPP’s work is increasing – expected to touch virtually every aspect of the system. Although AI/ML models are not standardized, there are significant drivers to encourage standardization around AI/ML in 6G, with AI/ML to be in Rel-21 specifications from day 1. In addition, there is a great potential for specifications to be less ‘watertight’ in the 6G releases, allowing for AI/ML based implementations using the best data-driven parameterization for some functionalities.  

The ETSI Artificial Intelligence Conference – “How Standardization is Shaping the Future of AI” took place from 10-12 February 2025, in ETSI - Sophia Antipolis, France.

The presentation material is online.

Further Reading

• See the 3GPP presentation: Overview of AI/ML related Work in 3GPP, Juan Montojo, Qualcomm, 3GPP Rapporteur NR_AIML_air.
• See the Conference website for more information and links to speakers details and presentations.