Who do you trust with your personal data? And what happens when that “who” becomes a “what”, as AI models are increasingly entrusted with high-stakes decisions?. Every time an AI system determines how data should be routed across a network, it is making decisions that directly affect privacy, security, and reliability.
Yet the skepticism many of us feel is entirely rational. AI models that make these decisions are inherently opaque. They optimize for performance in ways that are often difficult – if not impossible – to explain, operating at speeds and scales far beyond human oversight..
At the same time, the benefits of AI-driven decision-making in mission-critical environments, like healthcare, essential transport services, and emergency communications, are impossible to ignore. In these contexts, delaying adoption risks foregoing significant gains in efficiency and operational performance..
Trust is not a feature you add to an AI system at the end; it is built through rigorous, verifiable testing and optimization at every stage Sameh Yamany, Viavi. The challenge, then, is not whether operators should trust AI, but how they can trust it. Investing in infrastructure that rigorously validates decisions before they reach the live network – and continuously verifies their behavior once in operation – provides a key opportunity here.
With comprehensive testing and optimization throughout the AI lifecycle, operators can deploy autonomous networks with far greater confidence.. For more than a century, Viavi has sought to bring distinctive value to this challenge, building a foundation of high-fidelity, highly granular datasets that reflect the realities of network behavior.
As Sameh Yamany, CTO and chief AI officer at Viavi, puts it:. “Trust is not a feature you add to an AI system at the end; it is built through rigorous, verifiable testing and optimization at every stage of development, deployment, and operation.”. A question of accountability.
Operators’ concerns about who is testing the AI making decisions across their networks are well founded. Across much of the industry, the same vendor that develops the AI is responsible for proving that it works, creating an inherent conflict of interest.. “That is not independent testing – that is self-certification,” says Yamany.
“When a vendor tests their own gear to their own specifications, the performance in the lab rarely maps cleanly to performance in the live network.”. AI raises the stakes because its behavior is far more dynamic than traditional software. Rather than following fixed rules, AI models learn from data, adapt over time, and make probabilistic decisions.
Their effectiveness therefore depends not only on the quality of the model itself, but also on the accuracy and representativeness of the data used to train and validate it.. A model trained on inaccurate data does not fail gracefully – it fails with confidence. That is the most dangerous failure mode in any automated system Sameh Yamany, Viavi.
If that data is incomplete, biased, or fails to reflect real-world network conditions, AI can make flawed decisions at machine speed and at unprecedented scale.. A poorly trained model could trigger a cascade of incorrect decisions across the RAN, transport, IP, and service layers before operators even become aware of the problem.
By the time the issue is detected, it may have influenced hundreds of operational decisions.. “A model trained on inaccurate data does not fail gracefully – it fails with confidence. That is the most dangerous failure mode in any automated system,” Yamany explains, adding:.
“Operators need the confidence to say: this AI was tested independently, against real network data, by a trusted third party whose job is to find failures, not validate assumptions. That is the role Viavi is built to play.”. Training to be trustworthy.
The performance of any AI system is fundamentally limited by the quality of the data from which it learns. If trustworthy AI is the goal, trust must begin with the data used to train, test, and validate it.. In telecommunications, that challenge is particularly acute, as modern networks are among the most complex cyber-physical systems ever built, spanning multiple vendors, technologies, radio environments, traffic conditions, and operational scenarios.
According to Yamany:. “Generic/synthetic datasets are insufficient for telecommunications. Without high-fidelity data, AI simply learns the wrong behaviors faster.”.
Generic/synthetic datasets are insufficient for telecommunications. Without high-fidelity data, AI simply learns the wrong behaviors faster Sameh Yamany, Viavi. So, what does high-quality training data actually look like?
Drawing on decades of measurement expertise and data collected from laboratories, field deployments, and live operational networks, Viavi has identified three essential data characteristics:. Real: The data is collected from live networks under operational conditions, not generated synthetically in a lab and not derived from a single vendor’s test environment..
Correlated: Data sets from different layers of the network – radio, transport, core, and application – are timestamped and aligned so that cause-and-effect relationships can be observed and modelled.. Longitudinal: The data reflects how the network behaves across time, load cycles, weather events, subscriber density changes, and other variables that influence network performance..
Together, these characteristics provide AI models with a far more representative understanding of real-world network behavior, improving both the accuracy of their decisions and operators’ confidence in deploying them.. Validity: Grounded in truth. Training AI on trusted data is only part of the equation.
Equally critical is ensuring that its decisions remain valid once deployed, through continuous testing and verification.. This is where AI-powered digital twins become indispensable. By combining a live digital representation of the network with real-time telemetry, operators can evaluate AI decisions, predict their likely outcomes, compare alternative actions, and continuously verify that autonomous systems remain aligned with operational objectives – before they impact customers..
In effect, the digital twin acts as an independent validation environment. Rather than relying on AI to simply trust its own judgment, it offers a mechanism to verify that decisions are safe, effective, and consistent even as circumstances evolve.. Designing for data that doesn’t exist yet.
Building trustworthy AI is challenging enough when decades of real-world network data are available. But what happens when the technology being designed must anticipate traffic patterns, applications, and network behaviors that do not exist yet?. That is precisely the challenge facing the telecommunications industry as it looks toward 6G.
Unlike previous generations, 6G will embed AI as a native component of the network itself. As a result, engineers must design and validate autonomous systems against future workloads that cannot yet be observed in production.. As Yaman writes in a recent blog, “6G will not be designed in a lab.
It will be designed against AI traffic that does not yet exist, for applications that are not yet built, by operators who cannot afford to find out in production whether the design works.”. The scale of this challenge is growing rapidly. By 2030, AI traffic is expected to account for almost 40 percent of all wireless and network traffic, fundamentally changing how networks are designed and operated.
Unlike conventional network traffic, AI workloads exhibit non-stationary traffic patterns that traditional modelling techniques struggle to represent. As Yamany explains:. 6G will not be designed in a lab.
It will be designed against AI traffic that does not yet exist, for applications that are not yet built, by operators who cannot afford to find out in production whether the design works Sameh Yamany, Viavi. “The autonomous networks conversation has concentrated on the RAN, but a modern IP/MPLS core runs BGP, IGP, SR-TE, and RSVP simultaneously, with routing policies layered across hundreds of nodes.
Designing AI for 6G without accounting for how AI-driven decisions propagate through that transport topology is designing half a network.”. Waiting until a live network becomes the first production-scale experiment is simply too risky for infrastructure that underpins critical services..
Instead, the solution lies in using decades of trusted network telemetry to build generative models capable of simulating future AI traffic patterns before they exist at scale. This is precisely the role Viavi’s generative reality digital twin (GRDT) is designed to fill.. The realistic digital twin.
As established, AI models trained solely on synthetic data often fail the moment they encounter the complexity of a live network.. Building a truly representative digital twin, however, requires far more than mirroring the current state of the network. To provide meaningful assurance, the twin must be continuously calibrated against real operational data, synchronized with live network telemetry, capable of generating entirely new scenarios, able to validate AI-driven decisions with AI itself, and federated across every domain of the network..
Its generative capability is particularly important. While replaying historical scenarios is valuable for regression testing, it offers little insight into how AI will behave under conditions the network has never experienced. Instead, the twin must be able to create plausible future scenarios – from emergency incidents and widespread coverage failures to original congestion events and entirely new classes of AI-generated traffic – allowing operators to stress-test before those situations arise in production..
Equally, the test environment must evolve as quickly as the systems it is validating. As autonomous networks become more adaptive, AI models must be challenged by AI-generated scenarios that match the pace, complexity, and unpredictability of continuously changing network behavior..
The twin must also span the entire network. As Yamany explains: “Validating AI across the transport and IP layer requires purpose-built tooling: specifically, an algorithmic twin that computes what the network will do – routing table state, MPLS and SR-TE path selection, utilization projections – combined with an emulation twin that recreates how the network will behave, including actual protocol convergence timing and BGP and IGP state machines.
Each covers failure modes the other would miss.”. To achieve this level of realism, the GRDT is calibrated using three complementary classes of data:. High-volume, high-granularity operational telemetry collected from live networks over time..
3D spatial intelligence, allowing the twin to model behavior across diverse physical environments, such as dense urban centers, ports, railway corridors, rural low-coverage areas, and non-terrestrial network deployments.. Correlated AIOps and location intelligence telemetry, connecting network behavior directly to subscriber experience and application performance..
These capabilities are what clearly differentiate a conventional digital twin from a generative one. The GRDT can simulate interactions between AI inference workloads and conventional subscriber traffic at the scales projected for 2030 and beyond – crucial for 6G capacity planning..
This forward-looking capability is equally important for emerging technologies such as quantum-safe networking. With 6G expected to incorporate post-quantum security from the outset, operators cannot afford to wait until standards are finalized before understanding the operational implications.
Vivai expands:. Viavi is building quantum-safe scenarios into the twin today, so that when 6G is deployed, the validation work is already done Sameh Yamany, Viavi. “The testing for that requirement cannot start after the standard is finalized – it must start now, in the twin, so that operators understand the performance implications of post-quantum cryptographic protocols before they are committed to the network architecture.
Viavi is building quantum-safe scenarios into the twin today, so that when 6G is deployed, the validation work is already done.”. Trust is still earned. As AI models become more capable and networks more autonomous, it is tempting to assume that human oversight will become less important.
In reality, the opposite is true. The more responsibility AI assumes, the greater the consequences when it fails – and the harder those failures become to detect, explain, and attribute. The validation required should therefore become more rigorous..
As Yamany emphasizes, there is also a broader, more philosophical point at stake. AI systems are validated against objectives defined by people, within environments designed by people, using success criteria chosen by people. Even in a fully autonomous network, determining what constitutes correct behavior is itself a human judgment.
Autonomy doesn’t remove accountability; it raises the standard required to demonstrate it.. “What we do expect to change is the nature of validation. As AI systems become more capable, the validation tools must keep pace.
Manual testing processes will be replaced by AI-driven testing environments.”. Ultimately, trust in AI will be earned the same way trust has always been earned in telecommunications: through validation that is independent, repeatable, and evidence-based.. By combining decades of expertise in network testing, validation, observability, and AI-driven digital twins, Viavi is helping operators deploy AI with confidence – not simply because it is autonomous, but because every decision has been independently tested before it reaches the live network..
Where AI is concerned, trust is never assumed – it is proven.. For more information, please visit Network Digital Twin | VIAVI Solutions Inc.. More from Viavi
