Can anyone actually explain what an AI native network is and why we need it?

One of the key concepts behind 6G is that it should be “AI native”. The intent is that AI can breathe some magic onto the network, resulting in greater capacity and lower operating costs; and the inference is that this cannot be fully achieved in 5G because it was not designed to allow AI-driven optimisation. But quite what AI native means and what AI might do for networks remains very unclear.

Ericsson helpfully dedicated a long blog[1] to discussing what AI native might mean. It started by noting that “[the] AI native concept still does not have a clear definition”. Then it went on to define AI native itself as:

"AI native is the concept of having intrinsic trustworthy AI capabilities, where AI is a natural part of the functionality, in terms of design, deployment, operation, and maintenance. An AI native implementation leverages a data-driven and knowledge-based ecosystem, where data/knowledge is consumed and produced to realize new AI-based functionality or augment and replace static, rule-based mechanisms with learning and adaptive AI when needed.”

Does that leave you any the wiser? It feels like the industry is trying to find problems for AI to solve, rather than identifying the problems and then coming up with the optimal solution. AI is the answer, now what is the question?

Before delving deeper it is worth recalling how AI works, and therefore where it may add value. AI systems are broadly pattern recognisers. They are shown many examples of the pattern (eg pictures of cats) and eventually learn to recognise cats in unlabelled pictures. Or more usefully, many examples of weather data sequences and then learn to predict how the start of a sequence will evolve, hence forecasting the weather. AI is great where the underlying physics is difficult to understand or model and the patterns too complex for humans to recognise. It is not useful where the physics is easily understood. For example, it would be pointless to implement a calculator using AI – the mathematical rules are well known and simply implemented whereas learning the patterns would take vast effort and still likely make errors when presented with new forms of questions.

AI can be valuable but has costs. The energy consumption needed to train and operate models is well documented and growing. AI also makes “mistakes”, sometimes called hallucinations, where the best pattern match is not the right answer. In some cases, like ChatGPT and weather forecasting, mistakes may be tolerable, in others such as running mobile networks, they may not be.

There are clear cases in telecoms networks where AI is useful. Customer service is top of the list, using chatbots to provide customers with responses to questions. This is beneficial not because AI is better than human customer service agents (it generally isn’t) but because it automates their role resulting in cost savings. Fraud detection is another where AI can detect unusual patterns of activity and flag them for evaluation by a fraud expert, although algorithms have been doing this for many years. AI may have a role in preventative maintenance, for example by learning patterns of change in power consumption that occur in amplifiers before they fail, although this requires a large amount of data to be gathered on nodes prior to failure which may be hard to assemble.

But none of these cases require “AI native” networks. Indeed, they are independent of the network, working alongside 4G and 5G networks rather than within them.

The ideas put forward as to where AI might be valuable within networks fall into two areas:

1.??? Improving traffic management by predicting congestion and prioritising valuable traffic.

2.??? Enhancing network capacity by making networks more technically efficient.

Before taking a closer look at these, it is worth noting that neither may be important. As I set out in “The End of Telecoms History” growth in data usage is slowing and will soon reach zero. In this case, we will not be seeing any new congestion, nor will we need increased network capacity. If this is what native AI is for in telecoms networks, then it will have no value but likely significant cost.

Even if it were needed, it is far from clear whether it will work.

Regarding traffic management, forecasting congestion is of little value if there is nothing that can be done about it. Networks do not have spare resource ready to bring online in an instance when needed. The best that can be done is to prioritise “important” traffic and, by implication, block or throttle lower priority traffic. But this has been done for decades, and AI will be no help in that respect since it is very much a human judgement as to what data is valuable.

Regarding network efficiency AI has been suggested for three main areas (and others may be identified):

1.??? Reducing the overhead in sending channel state information (CSI).

2.??? Enabling better beam pointing, enabling higher gain beams and less time spent working out where to point the beam.

3.??? Enhancing error correction coding by predicting where error rates may increase and changing the coding scheme in advance.

For all of these, clever minds have been applying themselves for a decade or more to ways to improve things. CSI is already compressed and encoded often using Fourier transforms. Many beam pointing algorithms have been invented. And error correction systems have been intensely studied and are highly adaptive, changing as soon as they detect the error rate changing. They are all, already, highly optimised.

For all of them, the thinking behind using AI is that there is much variation in things like channel state, but that the variation might be less on a very localised basis. For example, across a whole network the variation in the CSI could be huge, but within a single cell it might be much less. If AI could learn the variation within each cell, for example, then it might be able to effectively develop an optimised algorithm for CSI, beam forming and channel coding on a cell-by-cell basis.

But for all of these areas, the implications of inaccuracy are significant. If the CSI information is imperfect then channel capacity will fall rapidly and if beams do not point in the right direction then mobiles will lose their connection (and probably have to fall back to 4G or similar). An AI algorithm that reduced the volume of CSI by 20%, but made the information 10% less accurate would likely reduce overall network capacity.

Patterns are hard to spot in radio parameters. All it takes is a bus to drive by, a tree to be in leaf, or even a metallised window to be opened to create a quite different radio environment. Different handsets see different signals depending on their antenna placement and changing the angle that the handset is held by just a few degrees can change the CSI. An AI system that thought it had learned a pattern could be confused by small changes in the environment or new models of cellphones.

Even if those issues could be overcome, for most solutions the handset needs to adapt as well as the network. That could mean the handset having to be informed of the model to use as it enters each cell, which could consume more resource than it saves. And it requires all handsets to work to common standards and use the same AI tools.

All told, it is far from clear that AI can make material gains in these areas, and even if it does they may have little value. This should not stop academics and researchers exploring the role of AI in telecoms networks – there may be benefits that can be found, perhaps in other areas such as power reduction. But making “AI native” one of the core reasons for a new generation of cellular technology looks like a very weak, and very vague, justification.

[1] https://www.ericsson.com/en/reports-and-papers/white-papers/ai-native