Yokosuka put listening AI at the school disclosure door. Keep the human rescue line visible

THE SIGNAL IN ONE SENTENCE

A child trying to tell someone about sexual harm does not owe the system a perfectly organized story. They may start with a fragment. They may type at midnight because speaking aloud feels impossible. They may test the listener with something small, change the subject, return later or use words that do not match an adult's neat risk categories. That is the difficult space Yokosuka is asking an AI service to enter. On November 9, Tsunagari AI said Yokosuka municipal schools will use the city's existing listening service, Yokosuka AI Friend Ai-chan, as one consultation route for sexual-harm concerns under Japan's Child Sexual Violence Prevention Act. The law takes effect on December 25. The service itself is not new. Yokosuka announced it on July 31 after an earlier pilot. It offers an AI chat entrance around the clock and can connect users to professional counselors or city support offices. The city says ordinary staffed counseling runs on weekdays from 8 AM to 9 PM, while emergencies are handled around the clock. The service is available through LINE and a browser-based version. The November announcement expands how the existing doorway is presented to schools, children and parents. It does not establish that the AI can reliably detect a disclosure, assess danger or improve safeguarding outcomes. That distinction matters more here than it does in almost any other kind of chatbot. A listening tool can lower one barrier. Text can feel safer than a phone call or a conversation across a desk. It can be used outside school hours. It can give a child a private moment to begin. Japan's national guidance recognizes that children and parents need accessible internal and external routes for consultation, including text-based options. But a doorway is not a rescue. The useful question is not whether the chatbot sounds warm. It is whether a child reaches a capable person quickly, safely and with as little retelling as possible. Yokosuka's procurement specification requires high-risk alerts, human escalation and written procedures for emergencies, including nights and holidays. It also assigns responsibility for city follow-up and, when necessary, coordination with police or other agencies. Those are important structural requirements. The public material does not show how well the risk system performs. It does not publish sensitivity, false-negative rates, false-positive rates, subgroup results, response-time distributions, staffing outcomes or an independent safeguarding evaluation. The provider says its system does more than look for keywords, but it does not disclose a validated performance measure on child sexual-harm disclosures. No one should fill that gap with confidence borrowed from a friendly interface. The first job of the AI is to avoid making things worse. It should not investigate. It should not ask leading questions, decide whether a child is credible or push for vivid detail. It should not suggest that the child caused the harm. It should not promise secrecy that the service cannot legally or operationally keep. It should not turn a distressed person's words into a customer-service puzzle. If the conversation indicates immediate danger, the system needs plain, age-appropriate safety language and a rapid human alert. If the danger is not immediate, the child still needs a clear choice: continue by text, speak with a counselor, contact another trusted adult or use a separate official support route. The handoff should be warm. That means the human receives the relevant conversation, the child knows who is joining, and the system explains what will happen next. The child should not have to begin again from zero unless they choose to. The counselor should acknowledge what has already been shared without treating an AI summary as the child's own words. The exact transcript and any machine-generated summary should be kept distinct. A summary can omit hesitation, uncertainty or a crucial phrase. It can also add an interpretation the child never made. For safeguarding work, the original disclosure is evidence. The machine's compression is only an aid. Privacy needs the same honesty. The provider says consultation data is encrypted in a domestic cloud environment, access is limited, and conversations are not used to train the underlying generative model. Yokosuka's procurement specification requires access logs, configurable retention and deletion, role controls, transfer of agreed data at contract end and certified deletion afterward. Those are useful controls, not a complete child-facing privacy explanation. Before a child begins, the service should say in simple language what information is collected, who may read it, when a human may be alerted, when information may be shared with the city or another agency, and how long records are kept. A child cannot make an informed choice from a policy written for procurement lawyers. The service also needs a plan for the device around the conversation. A school-issued computer may keep browser history, notifications or screenshots. A family device may be shared. A LINE account may reveal message previews. A browser-based option can widen access, but it does not automatically create privacy. The safe-exit design should be tested on the actual devices children use, including accessibility tools and managed school hardware. Language is another failure point. Yokosuka's original specification requires Japanese and English, with additional languages encouraged. That does not prove the model understands a child's phrasing, slang, disability-related communication, mixed language or indirect disclosure. Testing should involve children and safeguarding specialists, with careful consent, and should measure whether the system recognizes meaning without demanding adult vocabulary. The system should also admit when it is uncertain. A risk score is not a verdict. A low score must never close the door. A high score must not make an accusation. The score should trigger a defined human review path, not become a hidden label attached to the child or an employee. This is where the miss list belongs beside the success story. Publish how many conversations were escalated, how many high-risk cases humans confirmed, how many urgent disclosures the system failed to recognize, how often children abandoned the exchange before a handoff, and how long each stage took. Break results down carefully enough to detect unequal performance without exposing individuals. Count failed handoffs too. An alert that fires into an empty queue is not a safety feature. A referral to an office that is closed is not support. A phone number a child cannot safely call is not an exit. A counselor who receives the transcript but lacks authority to act is not the human in control. For every risk level, the service needs an accountable person, a response target, an alternate route and an escalation path when the first route fails. Staff need training, supervision and enough capacity to respond. The city needs incident review that includes the child-facing experience, not only whether a technical alert appeared. Independent review is especially important because the service provider has a commercial interest in expansion. Tsunagari AI says it will share information with other municipalities considering similar uses. That makes Yokosuka more than a local deployment. It may become a reference case. Before anyone copies it, the city should publish an evaluation designed with child-safeguarding experts, counselors, disability advocates, privacy specialists, parents and young people. The review should test real operational failure modes without exposing children to experimental harm. Start with scripted scenarios and trained reviewers. Include indirect disclosures, misspellings, joking language, silence, sudden exits, grooming concerns, immediate danger, historic harm, a disclosure about a school employee and a false alarm. Check every branch from the first message to the final human action. Then examine live operations with strict privacy protections. Measure the door, the handoff and the help separately. The door asks whether children can find and use the service. The handoff asks whether the right human receives the right information in time. The help asks whether the child reaches appropriate support and understands what happens next. Do not collapse those stages into one cheerful engagement number. Japan's new law requires more than criminal-record checks. Yokosuka's public explanation says covered institutions must create consultation mechanisms, train staff, detect warning signs, investigate with respect for children's rights and provide protection and support when harm occurs. An AI chat window can be one part of that system. It cannot carry the institution's duty. Schools still need trusted adults, visible external channels, staff training, careful evidence handling, protection from retaliation and a culture in which a child can speak without first satisfying a model. The AI route should add a choice, never replace the other doors. There is a humane idea at the center of Yokosuka's plan: make it easier to begin. Keep that idea. Then build the rest of the rescue line in public. Show the privacy promise before the first message. Show the person behind the escalation. Show how long help takes. Show what happens when the model is wrong. Show the child a way out of the chat and into the care of someone who can act. The best listening AI in this setting is not the one that keeps a child talking the longest. It is the one that knows when to get out of the way.