AI Fraud Alerts: Connected Doesn't Mean Convinced

The article opens with a hypothetical scenario: a phone rings showing 96110, an AI voice delivers a fraud warning, the recipient says "understood" and hangs up, yet returns to the original chat window. This illustrates the core problem: call connection, warning delivery, and a person actually pausing a transfer are three distinct events.

Legal Basis and Local Deployments

Article 34 of China's Anti-Telecom Fraud Law mandates a warning and dissuasion system with timely measures. Public records show AI already embedded in this chain:

Henan Province (Dec 2025): The 96110 intelligent call center uses "AI pre-screening + human review." Unanswered or doubtful calls are transferred to human agents. Trial data separately tracks answer rate and effective information collection rate — a meaningful distinction because connection is only the start of communication.

Yancheng City (Jan 2026): A disclosed case shows an intelligent voice warning followed by the person visiting a police station to verify and stopping further operations. The two local practices differ but jointly indicate that warnings need space for verification and follow-up when feedback is uncertain.

Answer Rate Measures Reach, Not Belief

Intelligent voice can dial massive volumes quickly, but the recipient must cross a trust threshold. A sudden risk alert conflicts with an ongoing chat, transaction, or supposed "customer service guidance." If the warning only repeats "you may be scammed" without helping the person identify which specific step to halt, the recipient may respond politely while continuing the risky behavior.

This is not an evaluation of any specific region's effectiveness, but a product-level observation from public processes: the system knows the call was answered, but not whether the person understood, believed, or stopped the high-risk action.

The Moment That Most Needs a Human: "I'll Think About It"

Suppose someone is about to transfer money under a stranger's guidance. The AI call urges a pause; the person replies "I'll think about it" and hangs up. This phrase could mean newfound alertness or simply a desire to end the call. A single voice response cannot resolve the ambiguity.

Henan's "AI pre-screen + human review" design addresses this: unanswered or doubtful cases flow to human agents. In the Yancheng case, the person only connected the warning to their own situation after visiting the police station. Both practices highlight that warning systems must leave room for verification and continuation when feedback is uncertain.

The human element is not merely making another call. Crucially, the hand-off must preserve context — the risks already explained, the person's questions, and the unconfirmed status — avoiding repetition and preventing "answered" from being misrecorded as "successfully dissuaded." This is the author's process observation from public materials, not a feature list of any deployed system.

A Good Loop Records Whether the Risk Actually Stopped

If only outbound volume and answer rates are measured, intelligent systems easily produce impressive workload numbers. The real questions are: after the warning, did the risk status change? Can doubtful people verify through trusted channels? Is there follow-up for those still proceeding?

These are not a single "success rate" but distinct states: reached, understood, pending verification, risk action stopped. States cannot be automatically equated; especially, a polite response must not be mistaken for a safety outcome. For product design and governance assessment, separating states provides more explanatory power than collapsing every call into "success/failure."

Practical Guidance for the Public

Official channels remain the starting point for judgment: 96110 is the official anti-fraud warning line. Take warnings seriously; if doubtful about the call, actively dial 96110 to verify. Do not transfer money, provide verification codes, or install unknown software at a stranger's direction. This advice is based on public government materials and does not replace specific handling guidance with the article's hypothetical scenario.

AI can make warnings arrive earlier. The harder next step is giving a hesitant person the chance to stop, ask clearly, and letting the system honestly record the "not yet confirmed" portions. The value of anti-fraud warning ultimately lies in whether the risk behavior changed, not merely whether the call connected.

Sources and References

Anti-Telecom Fraud Law of the PRC (Ningxia Communications Administration): Article 34 on warning systems and measures.

Henan Communications Administration: Henan 96110 Intelligent Call Center launch — local "AI pre-screen + human review" mechanism and trial distinction between answer rate and effective information collection rate.

Yancheng Public Security Bureau: Yancheng police develop intelligent call dissuasion AI robot — local public case showing phone warning linked to offline verification.

Wuhan Public Security Bureau: Anti-fraud tool introduction — 96110 purpose and answering reminders.