6G’s Next Reality Check: Validating AI, Marrying Sensing with Semantics, and Designing What We Can Test

As the wireless industry races toward 6G, much of the spotlight falls on artificial intelligence (AI), sensing, and spectrum. But beneath the buzzwords lies a more sobering truth: 6G will only succeed if we can validate what we build, especially when it comes to AI in the physical layer (PHY). Francisco “Frankie” Garcia, Master Scientist at Keysight and honorary professor at the University of Edinburgh, offers a grounded view of what it will take to move from concept to reality.

AI: From Hype to Hardware

For decades, wireless engineers could specify “every single bit” over the air. AI disrupts that determinism. Garcia believes the real breakthrough will come when we can train, benchmark, and test AI models with the same rigor we apply to traditional signal processing.

“When we're comfortable with using AI and we can validate it… that will create those first breakthroughs in pushing these specs forward.”

The challenge? We don’t yet have abundant, representative 6G data. Today’s models rely on synthetic datasets and 5G testbeds. Garcia emphasizes the need for repeatable workflows that quantify the total cost and benefit of AI, including training data provenance, compute budgets, and real over-the-air gains.

“If it takes too much training, too much computational complexity, and you only get a marginal improvement… why bother?”

Sensing + Semantics: A Surprising Pairing

One idea Garcia didn’t expect to like, but does now, is combining joint communications and sensing (JCAS) with semantic communications. At first, making the PHY “application-aware” sounded risky, but fusing codewords at the physical and semantic levels could unlock extreme efficiency for mission-specific scenarios.

“At the physical layer, we eventually transmit codewords. At the semantic level, we’re also trying to transmit codewords. Could we marry the two?”

This opens the door to generative AI-designed physical layers tailored to specific sensing tasks, transmitting only the task-relevant information, not a flood of redundant bits.

Spectrum: Use What We Have, Smarter

While frequency range 3 (FR3) is gaining attention, Garcia cautions that FR2 still has a role to play, especially in high-density and fixed wireless scenarios. He sees spectrum choices being driven by concrete, safety-critical use cases, not just bandwidth.

“We want bands whose propagation and bandwidth support both communication and sensing at short ranges.”

The message: spectrum must serve real-world needs, and hardware constraints will shape how these bands are used.

AI in the PHY: Start Where It Matters

Garcia sees AI landing first in areas like channel state information (CSI) feedback, beam management, and adaptive modulation and coding, where learning has a clear path to measurable gains. But he’s clear: we’re still in the early stages, often relying on 5G infrastructure and open testbeds to train and benchmark models.

“We’re still very much at the learning stages… using what we have available to us around 5G technologies.”

Sensing Use Cases That Count

The sensing stories Garcia finds most compelling are the ones with immediate operational value:

“Factory automation, especially the interaction between robots and humans, needs millimeter precision. That’s where sensing really matters.”

These are precisely the scenarios where JCAS + semantics could shine: tight loops between perception, inference, and minimal-but-sufficient communications.

Debunking the Myth: “AI Solves Everything”

Garcia is refreshingly blunt: many AI papers tout “X-times better” without accounting for the full bill that includes training cost, inference latency, power draw, and hardware footprint.

“We don’t look at the overall cost of that AI model. That’s a problem.”

In 6G, those omissions are showstoppers. The bar should be system-level value, reproducible results, and traceable data.

Security: AI Cuts Both Ways

AI is both a tool and a threat. The same techniques we hope to use for smarter PHY/RAN can be turned to probe protocol edges, exploit reserved fields, and poison models.

“AI and ML could also be used to exploit the holes that exist in the communications fabric.”

Add JCAS and pervasive data capture, and the privacy surface expands. Garcia’s advice: security and AI must co-evolve, with specialists anticipating how models learn from and attack each other.

What to Watch in 2026

Conclusion: Build What You Can Measure

Garcia’s perspective is a timely reminder: 6G won’t be defined by how much AI we can sprinkle on the stack, it will be defined by how much validated, testable, and secure AI we can productize.

“Build what you can measure and defend. Everything else is a lab curiosity.”

Listen to the full interview here.

limit
3