AI-native air interface could play a key role in 6G
When people talk about AI in telecommunications, the conversation typically gravitates toward network management, such as AI handling traffic flows, making routing decisions, and allocating resources more intelligently. However, these are examples of bolting AI into existing infrastructure. The AI native air interface is a little different. Refers to the use of machine learning to design wireless signals at the physical layer.
With AI-native air interfaces, AI-based systems fundamentally change the way signals are encoded, modulated, and transmitted. As 6G research accelerates, this approach could potentially replace decades of hand-crafted waveform design with neural networks that learn the best signal patterns for specific hardware and environments. Here’s how it works:
What is AI-native air interface?
The key to how an AI-native air interface works is the “AI-native” part. This distinguishes it from other “AI-enhanced” networks, where machine learning handles things like routing optimization and resource allocation, but leaves the underlying signal design intact. AI-native air interfaces operate at a deeper level. Machine learning designs the signals themselves at the physical layer and replaces traditional mathematical models for learned representations.
Wireless communications have relied on waveforms such as OFDM (Orthogonal Frequency Division Multiplexing) for decades. These are signals developed through rigorous mathematical theory and standardized across the industry. Engineers handcraft them based on theoretical models of radio wave propagation, interference behavior, and ideal hardware performance. But AI-native air interfaces turn that on its head. Neural networks learn optimal signal designs by training how real hardware actually behaves under real-world environmental conditions.
This amounts to a fundamental overhaul of encoding, modulation, and transmission. Rather than applying a predetermined signal structure, the system learns what is best for a particular deployment scenario. This is especially useful in niche environments where the network may behave a little differently than the average urban environment. These are characteristics that the AI can detect and adjust to, rather than having to be specified up-front by the engineer.
Deep learning at the PHY layer
Technological advances towards AI-native air interfaces are occurring in stages. Early efforts focused on using machine learning for encoding, symbol mapping, equalization, or decoding to replace individual processing blocks in traditional digital signal processing chains. The next task is to replace multiple connected blocks. The most advanced implementations replace the entire physical layer.
At this stage, both the transmitter and receiver are implemented as deep neural network modules and act as autoencoders. The transmitter learns to encode information into signals, and the receiver learns to decode those signals back into data. The most important point is the fact that the system trains end-to-end and optimizes both jointly rather than individually. Traditional systems optimize each component individually, so if those components don’t perfectly complement each other, overall performance can be less than ideal.
The goal moves from minimizing bit errors under an ideal channel model to minimizing “semantic loss” under real channel constraints. AI-native approaches learn the actual flaws in equipment, rather than designing systems with theoretical hardware performance in mind.
Improved performance
While research and early field trials indicate meaningful improvements in some aspects, these results are preliminary and primarily derived from controlled environments or pilot deployments.
Increased spectral efficiency comes from AI-designed waveforms that create bespoke constellations and pilot signals that adapt to available spectrum conditions. Rather than a fixed modulation scheme, the system learns a representation that is optimized for the current channel characteristics. Some studies suggest that compression effectiveness may be up to three times greater than traditional approaches, but such numbers need to be validated under different conditions.
Energy savings is another claimed benefit. Research shows that transmission power can be reduced by up to 50% compared to 5G for comparable bandwidth and data rates. Field trials with AI-optimized scheduling demonstrated a 34% reduction in network energy in real-world deployments. Although these savings are significant for both operational costs and environmental impact, the computational overhead associated with training the AI model can partially offset the transmission energy savings.
Latency improvements have been demonstrated in large-scale operator trials across over 5,000 gNodeBs. These deployments were found to reduce air interface delays by 25-34% in urban and vehicular environments. In one specific example, we observed that streaming latency for short videos was reduced from 43.0 ms to 32.0 ms. That said, these results were obtained from a specific carrier with an incentive to publicize successful pilots, and generalization across global networks has not yet been established.
Real world applications
Private networks for factories and warehouses appear to be the most promising application in the near term. These environments prioritize flexibility over standardization, and the closed nature of private deployments avoids the interoperability concerns that complicate public networks. The learning network can be reconfigured from supporting low-bandwidth industrial sensors to high-throughput video surveillance to latency-critical robot control without manually retuning radio parameters.
In high-interference environments, especially in dense urban areas, there are situations where traditional signals degrade, and AI-native approaches can provide new solutions. Learning from real interference patterns rather than theoretical models may result in better performance when traditional waveforms are difficult.
Latency-sensitive services could also benefit from an AI-native approach. For example, autonomous vehicles that require vehicle-to-everything (V2X) communication really need an optimized air interface. Dynamic spectrum scenarios where conditions change rapidly due to changing weather or usage patterns may also benefit from systems that can adapt on the fly.
However, consumer mobile broadband is less clear-cut. The global mobile ecosystem relies on standardization across vendors and interoperability across borders. Whether an AI-native approach will work within that framework or be limited to specialized deployments is still an open question.
Interoperability and standardization
Of course, without some standardization none of this really makes sense. 5G works around the world because manufacturers and carriers have agreed to 3GPP specifications. Everyone uses the same waveform, the same modulation scheme, and the same protocol structure. Communication between equipment from different vendors becomes a bit of an issue as each AI system learns its own optimal waveform. When user devices from one manufacturer communicate with base stations from multiple vendors or roam between networks operated by different companies, they rely on shared signaling standards.
Some researchers have proposed “dynamically generated control interfaces,” potentially enabled by large-scale language models, that can negotiate signal parameters between incompatible systems. This is still very speculative. Others have suggested that the 3GPP standards process itself needs to be fundamentally changed, moving from fixed specifications to a framework that supports learned behavior. Neither approach has reached consensus.
Validation and testing are also somewhat difficult. While traditional networks can be verified against mathematical specifications, AI-native air interfaces require new testing approaches such as hardware-in-the-loop testing, black-box evaluation, and advanced simulation environments. Regulators, carriers, and vendors have also yet to decide on a standardized protocol.
The energy and computational costs of AI training must also be considered. Optimizing the network may reduce transmit power, but training models at scale or implementing federated learning at the edge requires large amounts of computation. It remains to be seen whether net energy savings will be seen over the entire lifecycle of the system, including training, deployment, operation, and updates.
