Wireless testing has traditionally followed a well understood playbook: apply a known input, measure the output and determine whether a device performs as expected. AI is complicating that dynamic. As intelligence moves deeper into the radio access network, wireless systems are becoming more adaptive, making decisions autonomously and adjusting their behavior in ways that are not always predictable.

That transition is already underway with AI-RAN, and 6G is expected to take it much further by embedding AI and machine learning directly into the cellular network architecture. We spoke with Middle Wen about the progression from AI-enabled 5G to AI-native 6G, what happens when intelligence reaches the radio itself, and why testing must evolve, too.

We already hear a lot about AI-RAN and 5G. How will 6G be different?

Middle: The way AI is used in today’s 5G networks is largely additive. In most cases, operators deploy AI to improve efficiency through the RAN Intelligent Controller. That’s good for load balancing, traffic steering and massive MIMO beamforming, and it helps with energy consumption and resource allocation in these dense operating environments.

But 5G wasn’t built for AI. Today, it’s mostly found in the upper software layers, like Layer 2, Layer 3 and network management planes. That means data has to travel from the physical radio layer and through the stack for AI processing before instructions can even be sent back to the radio. That’s a pretty long round trip. And it’s why we see so much data latency throttling back AI-RAN performance.

If we look at 6G, AI and machine learning move directly into the PHY and media access control layers. AI models won’t operate above the radio; they’ll be housed within it. That’s going to shorten the decision-making loop and optimize compute all the way out to the network edge.

What becomes possible when AI moves that deeply into the network?

Middle: I think AI will make 6G networks much more responsive, because they’ll have enough intelligence to react quickly to changing conditions while continuously learning and adapting to their environments.

That’s going to benefit things like wearables, autonomous robots, and other physical AI systems that depend on low latency and reliable connections. And that’s possible because the radio itself will become so much more adaptive. 

Just think about it. If an AI radio can adapt to changing traffic, then it can adjust modulation parameters, beamforming strategies and spectrum allocation in real time. That means we can design dynamic networks that can anticipate traffic bursts and assign resources before congestion has a chance to degrade service quality.

You’ve described an evolution from “AI-for-RAN” to “AI-on-RAN.” What does that mean?

Middle: “AI-for-RAN” is what the industry is doing today when it tries to optimize network operations. Over the next three to four years, this will give way to “AI-and-RAN,” which is a way for operators to make better use of their 5G infrastructure investments.

“AI-on-RAN” takes things a step further, because the RAN itself begins to function as a distributed AI computing platform that can sense, process and communicate. That’s important because it allows a 6G radio unit to do more than transmit and receive data. It could function as an environmental sensor or mini radar system by gathering information about its surroundings. And that information could be used for things like predictive maintenance or dynamic power optimization.

What does this mean for the computing architecture behind the RAN?

Middle: In this scenario, computing becomes much more distributed. AI inference engines that exist today inside centralized cloud data centers can begin to move into radios, edge servers and end-user devices.

And that’s going to have a knock-on effect. I expect we’ll see chipset designers begin to augment their DSP architectures by incorporating AI acceleration directly into wireless silicon. Telecom equipment will start dispersing distributed AI workloads across radio units, as well as distributed units and centralized units.

And this will significantly impact hyperscale cloud providers. Today, cloud computing and cellular PHY operate under very different timing assumptions. An AI-enhanced 6G environment will require cloud and radio processing to meet at the edge while still adhering to extremely tight radio network timing requirements.

How does an AI-native network change wireless test?

Middle: Today, engineers apply a known input, expect a known output and compare the result against well-understood 3GPP criteria. But when AI models continuously adapt through training and inference, these same conditions may not always produce the same response. In many cases, more than one response may even be acceptable.

So, while a test engineer will still need to determine if a radio meets certain specifications, they’ll also need to understand how AI models respond across a wide range of dynamic conditions. And that moves us from a test environment defined by fixed pass-fail thresholds into the realm of probabilistic outcomes.

Does that change what needs to be tested at the device and system levels?

Middle: Absolutely. Beamforming, network slicing, scheduling and spectrum allocation all become more dynamic.

And there’s a system-level dimension. Because AI-RAN is highly distributed, multiple AI models can potentially operate simultaneously. That means engineers have to evaluate how independent AI systems interact and collaborate as part of a unified network.

You can see that the result begins to look more like continuous system validation than isolated device certification, because you’re evaluating individual components and the collective behavior that emerges when those components operate together.

What does this mean for testing uplink performance?

Middle: Historically, the industry has placed a lot more emphasis on downlink traffic, because applications like video streaming move lots of data from the network to the user.

AI networks are different, because they generate information from billions of smart devices – like cars, industrial sensors, health monitors and robots – that continuously transmit data back into the network for processing and inference. That makes uplink performance much more important.

What should equipment and device makers be preparing for?

Middle: They should be thinking about validation as a continuous process rather than something that ends when a product passes conformance testing. OEMs will need to support continuous over-the-air updates that remain synchronized with changing AI algorithms – even after a product has entered the market. 

6G also brings together two very different operating models. Cloud providers come from an IT environment based on rapid iteration and frequent software updates. They can tolerate a little downtime. Telecom, on the other hand, is an operational technology discipline that’s built around strict hardware compliance and “five nines,” or 99.999%, reliability.

I see a challenge bringing those models together at a time when AI is already expanding the test surface and becoming more dependent on smarter radios and faster networks and adapting to constantly changing environments.

Any additional considerations on where AI-RAN and 6G are headed?

Middle: It’s fascinating to see where wireless test is evolving as we shift from deterministic networks to adaptive AI. The job of validating individual devices with fixed specifications is changing, and it’s going to require an understanding of how intelligent systems behave and how they interact and adapt over time. For anyone preparing products for the 6G market window, I think it’s important that they consider how their wireless test strategies need to evolve alongside the network.