A newly published survey tackles a question the telecom industry has largely avoided spelling out: why do the machine learning techniques that now permeate 5G network management break down when applied to 6G?

The survey, highlighted by Bioengineer.org, sets out a systematic account of the mismatch between today's machine learning assumptions and the demands that sixth-generation networks are expected to place on them. Its core argument is straightforward. The statistical foundations of most deployed ML models — stable data distributions, abundant labeled training sets, and centrally available compute — do not hold in the environments 6G is being designed for.

That matters commercially. Operators and vendors have spent the past decade embedding AI into radio access networks, traffic forecasting, anomaly detection and customer experience management, generally on top of 5G architectures. Standards bodies, including 3GPP and the AI-relevant work streams around network automation, have progressively formalized those techniques. The survey's contribution is to catalog where that accumulated practice stops scaling.

Three structural shifts drive the problem.

First, heterogeneity. 6G research targets span sub-terahertz spectrum, reconfigurable intelligent surfaces, integrated sensing and communication, and non-terrestrial links. Each introduces channel behavior that differs sharply from the sub-6 GHz propagation on which most existing wireless ML models were trained. Models tuned for one band or topology degrade when conditions change — the classic distribution-shift problem, aggravated by the sheer diversity of 6G candidate technologies.

Second, scale and timing. The envisioned densities of devices and the latency budgets of 6G use cases push inference toward the network edge, where compute, energy and data availability are constrained. Training regimes that assume large, centralized datasets sit uneasily with that reality. The survey points to federated and distributed learning approaches as necessary responses, while noting their own unresolved issues around communication overhead, heterogeneity of participating devices and robustness.

Third, the data itself. Labeled network data remains scarce and expensive to produce, and the arrival of AI-native network functions in 6G raises the prospect of models that must adapt continuously rather than retrain periodically. The survey frames this as a move away from static, benchmark-driven ML toward continual and online learning under non-stationary conditions — a substantially harder problem than the one the industry solved for 5G.

The commercial implication is a gap between vendor roadmaps and research readiness. AI-native 6G is a recurring theme in operator and vendor positioning, but the survey makes clear that the supporting ML methodology is not yet settled. Standards timelines compound this: with 6G standardization work gathering pace and first commercial deployments targeted toward the end of the decade, the methods that survive the research phase will need to be stable enough to be specified, interoperable and auditable — requirements that current experimental techniques largely do not meet.

The survey positions itself as a map for researchers rather than a solutions document. For network operators and equipment makers, its value lies in the discipline of the accounting: which assumptions of modern machine learning hold, which fail, and what the candidate replacements cost in complexity and reliability.

The industry's working assumption remains that 6G will be AI-native by design. Whether the machine learning stack matures in step with the radio and standards work now underway will help determine whether 6G's promised capabilities arrive on the late-decade timeline that operators and vendors have sketched.