India put AI inside the network stack. Test the fallback before calling it intelligent

THE SIGNAL IN ONE SENTENCE

A network can be intelligent right up until it needs to make a phone call. India's Department of Telecommunications led a session on intelligent networks and quantum communication at the Emerging Science, Technology and Innovation Conclave in New Delhi on October 28. The official account describes a future stack that places AI-native networks, AI-driven cyber defence, quantum-safe communication, zero-trust design and indigenous telecom technology beside one another. The session included presentations on AI-enabled medical imaging carried over secure networks and AI-assisted optimization in radio access networks. It also pointed toward testbeds, commercial deployment and a national quantum-secure backbone as longer-term priorities. That is a direction, not a deployment report. The government release does not name a production AI controller, publish a network benchmark or show that an autonomous system is currently running critical traffic. Good. This is the moment to define the failure test. An intelligent network should fail boringly. When the model drifts, the data arrives late, the optimization loops disagree or the management layer disappears, essential service should return to a simpler known state that operators understand. That sounds less exciting than an AI-native 6G vision. It is also the part that keeps a clinic connected. Start by separating three promises that are easy to blend together. AI can help a network predict traffic, tune radio resources, detect anomalies, prioritize maintenance and recommend responses. Quantum-safe cryptography can protect information against attacks that may become practical as quantum computers improve. Zero-trust architecture can require every user, device and service request to prove its identity and authorization instead of trusting location alone. Those technologies can complement one another. They do not validate one another. An AI controller can make a bad routing decision inside a quantum-safe connection. A zero-trust policy can be applied inconsistently by an automated system. Strong encryption can protect the confidentiality of a message that the network delivered too late. Test each claim on its own. Then test the combined system when one layer fails. The Department of Telecommunications says India has about 1.2 billion mobile connections and 5G service in almost every district. At that scale, a small automated mistake can become a very large public event. A model that shifts traffic away from congestion may improve average throughput. It may also push a rural cell, disaster area, railway corridor or hospital link below the service level people depend on. Do not begin with a national average. Publish the service classes. Identify emergency calls, health facilities, public warnings, payment links, government services, ordinary consumer traffic and machine connections with physical consequences. For each class, define latency, availability, restoration time, capacity, geographic coverage and acceptable degradation. Then state which metric the AI is allowed to optimize. Optimization without a declared objective is just a machine moving pain around. The first practical artifact should be a control map. List every input the model can read. That may include traffic load, signal quality, handover records, equipment alarms, weather, power state, maintenance history, location, device class and security events. For each input, publish its collection interval, owner, freshness limit, missing-data rule and privacy classification. A network model should know when it is looking at yesterday. Operators should know when it does not. Next, list every action the system may take. Can it recommend a parameter change? Can it allocate spectrum, adjust power, alter a handover threshold, quarantine a device, reroute traffic, restart equipment or open a maintenance ticket? Can it do those things automatically? Can one action trigger another model? Every permission needs a boundary, a named owner and a condition that returns authority to a person. The difference between an assistant and an autonomous controller is not the marketing name. It is write access. Then build the deterministic fallback. This does not mean recreating the entire network without software. It means preserving a tested operating mode whose rules are explicit, whose dependencies are known and whose behavior does not change because a model updated. The fallback may use fixed thresholds, approved routing tables, rate limits, static priorities, regional isolation or the last known good configuration. It should protect essential service even if performance is lower. The operator must be able to enter that mode without asking the failed system for permission. The trigger cannot live only on the same management platform it is meant to replace. Test the physical and organizational path. Who notices the problem? Who has authority to switch modes? Which teams are notified? Which regions change first? How is customer harm communicated? How does the network return to normal? An emergency lever that nobody is allowed to touch is decorative hardware. The test plan should begin with data failures. Feed the controller stale load measurements. Remove a region from the telemetry stream. Corrupt a device class. Delay security events. Create conflicting records from the radio, transport and core layers. The system should identify uncertainty, limit its actions and escalate. Silence is not confidence. Next, test model drift. A network changes with new devices, software releases, spectrum use, weather, construction, festivals, emergencies and customer behavior. The data distribution that made an optimization safe in a lab can move beneath it. Publish drift indicators by geography and service class. Set thresholds that reduce authority before performance becomes dangerous. Keep the current model, previous model and deterministic baseline available for comparison. Then test interacting loops. One model may optimize radio capacity while another flags suspicious traffic and a third predicts equipment failure. Each can be reasonable alone and unstable together. The capacity model moves traffic. The security model sees the movement as anomalous. The maintenance model interprets the resulting load shift as failing hardware. Congratulations, the network has invented a meeting. Run conflict scenarios before deployment. Define which controller wins, when actions are paused and what evidence the operator sees. Medical traffic makes the stakes concrete. The ESTIC session included a presentation on AI-powered medical imaging over quantum-secure intelligent health networks. The official account describes the promise of high-bandwidth, low-latency and secure connectivity for scalable diagnostics beyond major urban centres. That is a useful application to test. It is not proof of clinical benefit or network readiness. Separate the image model from the communication service. Measure upload completion, end-to-end latency, packet loss, image integrity, retry behavior, encryption state, diagnostic turnaround and the consequence of failure. A scan that arrives securely but incompletely is not a successful transmission. A network that prioritizes one large imaging job while starving routine care messages may have optimized the wrong queue. When connectivity degrades, the clinical workflow needs a local path. Can the image be stored safely? Can a lower-bandwidth version be sent? Can the case be routed to another specialist? Can staff see that the result is pending rather than complete? The fallback belongs to the service, not just the router. Radio access networks need a field receipt. The official session also discussed Indian-developed chips, radio units and base-station systems alongside AI-enabled network optimization and adaptation. For every field test, publish the hardware, software, model, spectrum conditions, terrain, device mix, traffic profile and power environment. Report the median and the ugly tail. Show what happened at the edge of coverage, during handovers, under power loss, with older devices and when the backhaul was constrained. The network experienced by the worst-served district matters more than the smooth demonstration beside the conference hall. Security testing should assume the model is a target. Attackers may manipulate telemetry, trigger costly reconfiguration, hide traffic inside normal patterns, learn thresholds or flood operators with plausible alerts. Red teams should test poisoned inputs, prompt-like instructions hidden in operational data, model extraction, privilege escalation and malicious updates. Record whether the system can distinguish an attack from an ordinary failure. Record how quickly a human can contain it. Record whether the fallback remains trustworthy after the primary controller is compromised. The audit log needs to survive the incident. For each automated decision, preserve the inputs, model and policy version, proposed action, authority used, affected elements, operator response, outcome and rollback. Logs should be available to an independent reviewer and protected from the system that generated the action. Do not let the defendant edit the courtroom transcript. India is already funding telecom development. The Department of Telecommunications says its Telecommunications Technology Development Fund has supported more than 130 projects worth about 500 crore rupees, and a separate technology development and investment scheme has an outlay above 200 crore rupees through 2031. Attach evidence requirements to that money. Fund common failure scenarios, shared telemetry formats, reproducible testbeds and independent evaluation. Require projects to publish what they cannot yet do. Make rural performance, energy use, accessibility, maintenance skills and supply-chain independence part of the score. Indigenous technology is more than where a component was designed. It is whether local operators can inspect, repair, replace and safely run it when the original team is unavailable. Finally, rehearse restoration. A rollback is not complete when the old software starts. It is complete when service stabilizes, queued actions are reconciled, security state is verified, customers are informed, incorrect automated changes are reversed and the incident record explains what happened. Run the drill without warning. Time every handoff. Include a region with limited staff. Include a failure during another emergency. Include an update that makes the new controller incompatible with the old configuration. Then publish the result. Intelligence is useful when it improves a network that people can still operate. If the network becomes mysterious, fragile or impossible to restore without the model that broke it, the clever layer has eaten the infrastructure. India's next telecom stack can be ambitious. The fallback should be gloriously dull.