Insights

Public Safety

What a 911 center should publish before AI listens to live calls

The Seattle Times reports that since 2023 AI has listened to every Seattle 911 medical call and prompted dispatchers on which callers can go to a nurse line, with no caller disclosure or public review. A 911 center can publish three things before AI touches live calls: a disclosure, an evaluation, and override data.

November 11, 2026 · 6 min read

What a 911 center should publish before AI listens to live calls

Since 2023, AI has been listening to all 911 medical calls in Seattle and prompting dispatchers on which callers do not need a rapid response, according to The Seattle Times. The paper reports that the Fire Department began using the system without disclosure to callers, without public review, and without showing that it treats diverse callers equitably. Callers did not know about it until a Times report in June.

The Nov. 4 story, by Daniel Beekman, adds a record from 2019. Seattle's then chief privacy officer, Ginger Armbruster, reviewed the Fire Department's plans for AI from Corti, the Danish company building its new 911 call-taking platform. Her privacy risk analysis flagged equity concerns, including that decisions could be affected by language, accents, and disabilities. She recommended a race and social justice assessment, working with public leaders on the use of the technology, and preparing documentation for public inquiries. She told the Times she never heard back about the equity concerns.

I am not going to argue about whether the system is good or bad. Nobody outside the department can tell yet, and that is the problem worth writing about.

What the department says, and what it has not shown

The department's position, through spokesperson Kristin Hanson, is that dispatchers, who are firefighters, remain the decision-makers on 911 medical calls even when they receive AI prompts. They use protocols developed with oversight by the department's medical director, receive state-certified training, and follow state policies for callers with disabilities and callers who speak languages other than English.

Officials have cited a 2023 Danish study showing Corti's AI could detect stroke in emergency calls better than human dispatchers, across age and gender. The Times notes that the department's medical director co-authored it, and that the department has not shared research on Corti's AI and nurse line triage, on equity, or on attributes such as ethnicity or accents. UW Medicine is conducting an evaluation of the department's use of AI and its nurse line, with no specifics or completion date shared. The department has asked the city's IT department to work on retroactive privacy reviews of the Corti platform. The mayor's office promised a public update on all city AI use in early 2027.

Every one of those is a reasonable step. They are also all happening after the system went live on real calls, and none has produced a document the public can read today.

Why "a human decides" is not enough on its own

Human-in-the-loop is the right design for 911. It is not evidence by itself, because a prompt changes the decision it informs. University of Washington professor Emily Bender told the Times the automated prompts are "going to impact" dispatchers' choices. University of Michigan professor Shobita Parthasarathy said research has shown that when AI gives a prompt, the human is less likely to trust their own judgment.

There is also a specific technical concern. Bender's worry is that a speech model works less well for callers whose speech differs because of disabilities, accents, or stress. The Times cites a 2022 UW study of Microsoft speech recognition that found much higher error rates for Black and Native speakers than for white speakers in Washington. That study was not about Corti. It is a reason to measure, not a finding about this system.

So the useful question for any 911 center is not whether a human is in the loop. It is how often the human follows the prompt, and whether that rate differs across callers.

Three things to publish before go-live

If I were advising a 911 center about to put AI on live calls, I would ask for three public documents before the switch is turned on. None of them requires releasing call audio or patient data.

  1. A disclosure. A plain-language page that says which calls the AI listens to, what it produces (a transcript, a prompt, a suggested call type), what the dispatcher sees, and what decisions it can and cannot influence. It should name the vendor, say where call data is stored and for how long, and say who in the agency owns the system. If callers are not told on the call, the page should explain why.
  2. An evaluation. A short report on how the system was tested before go-live and on whose calls. Give performance by the groups the agency can measure: language, interpreter use, caller age where known, and call type. Bender's suggested approach in the Times was to collect demographic information from consenting participants and report how the technology works across groups. Where a group could not be measured, say so.
  3. Override data, on a schedule. A recurring report on how often dispatchers accept, change, or reject AI prompts, broken out the same way as the evaluation. For any prompt that routes a caller away from an immediate response, report it separately. A rising acceptance rate with no change in outcomes can mean the prompts are good, or that dispatchers stopped questioning them. Publishing the number lets someone ask.

I covered a similar disclosure question on the health care side in what Texas health system customers will ask about AI disclosure. Texas law there puts the duty on the provider, not the vendor. The same logic fits a 911 center: the agency, not the vendor, owes the public an account of how the tool is used.

The pilot that became production

The detail I keep returning to is how the system moved. Armbruster told the Times she was told in 2019 that the project was a pilot for quality assurance on completed calls, not live emergencies. According to the city's IT department, because it was a pilot, the project did not go through the city's normal process for buying services or receive a comprehensive technology review. She believes she should have been consulted again in 2023 when AI prompts moved onto in-progress calls. "Things like this end up in no-man's-land," she said.

That is the control I would add first. Any change from reviewing completed calls to prompting on live calls is a new deployment, and it should trigger the same review as a new purchase. Write that rule down now, before the next pilot.

If your center already has AI on live calls, start with the third document. Pull last month's prompt acceptance rate. If nobody can produce it, that is the first finding.