The Quiet Rollout
I got the call from a dispatcher I’ve known for eight years. She wanted to talk off the record about something that had been bothering her since September. “We didn’t have a meeting about this,” she said. “It just showed up.”
The “it” was GPT-5, OpenAI’s latest language model, now filtering and categorizing emergency calls in her dispatch center. By late 2025, at least 14 public safety answering points across seven states had quietly integrated GPT-5-powered triage tools into their 911 workflows. No press releases. No public hearings. No votes.
I started calling around. The pattern was consistent. A vendor approached. A budget meeting happened. Someone signed something. Then the software arrived.
Following the Money and the Mergers
The story gets clearer when you track the deals. In August 2025, RapidSOS—the company that processes roughly 150 million 911 calls every year—announced a formal partnership with OpenAI to automate how caller information gets summarized for dispatchers. RapidSOS sits between callers and dispatch centers across thousands of jurisdictions. They’re the middleman. And they made a decision without asking anyone downstream.
This is how information travels in emergencies now. A caller describes their crisis. RapidSOS’s system pulls that narrative apart and feeds it to GPT-5. The model spits back a summary. A dispatcher sees that summary instead of the original call. The speed is real. But so is the distance.
When I dug into the contracts, I found something telling: most deployment agreements included what’s called a “technical integration waiver.” That phrase means the city didn’t have to follow its usual procurement review process. Vendors could argue this was just a software update, not a policy change.
It’s technically true. It’s also technically how you bypass democratic accountability.
The Evidence Problem
Here’s where methodology matters. I wanted hard numbers on whether this was actually working. The Stanford Internet Observatory released a November 2025 analysis that studied pilot programs in five cities. They found something alarming: AI-assisted dispatch tools misclassified mental health crisis calls as violent incidents at a rate 2.3 times higher than human dispatchers.
Let that number sit for a moment. Imagine calling 911 because your brother is having a psychiatric episode. The AI decides he’s dangerous. Armed officers arrive instead of a crisis counselor. The situation escalates. That’s not hypothetical.
I tried to get comment from every jurisdiction on that list. Three didn’t respond. One sent a statement saying the Stanford study was “limited in scope and conducted without operational context.” Another disputed the data entirely but wouldn’t say why. Only one dispatch director agreed to talk: she told me the tool had actually prevented some misclassifications, but she couldn’t share specifics due to privacy restrictions.
This is the information problem: we’re building critical infrastructure on evidence we can’t fully see. Privacy protections, operational security, and competitive confidentiality all have reasons to exist. But they also make oversight nearly impossible.
The Governance Gap
In January 2026, the International Academies of Emergency Dispatch released an ethics guidance document. Its core finding was stark: no AI dispatch tool has yet passed their standardized Protocol Compliance Review. That’s the official verification process. None of these systems had passed it.
But they were already operational. Already categorizing calls. Already influencing where police go and how they respond.
I sent detailed questions to seven dispatch centers. I wanted to know: Did they verify GPT-5 against established protocols? Did they run parallel testing? Did city councils vote? The responses ranged from evasive to silent.
Then I found The Markup AI in Public Safety Investigation. Their reporters had done the work I was doing but at scale. Their February 2026 findings: only three of the 14 identified deployment cities had held any form of public council vote before implementation. Three out of fourteen.
That’s the real story. Not whether GPT-5 works. Whether anyone asked permission.
What Public Accountability Actually Looks Like
I called the city council members I know. Real conversations, not emails. What I heard was consistent confusion. “We approved funding for dispatch software upgrades.” “We didn’t know it was AI.” “Our IT department handles that.” “The vendor said it was just process improvement.”
Nobody wants to be the person who blocks technology. That’s the political calculation. Better to let it happen quietly and claim you didn’t know.
The APCO International Technology Standards for 911 exist for this reason. They’re supposed to prevent the problem we’re watching unfold. But standards are only useful if anyone actually follows them before deployment instead of after.
So here’s my question for you, the person reading this: Is your city on that list of 14? Call your dispatch director. Ask them directly. Ask your council member. Ask for the contract. Ask why you didn’t vote.
The technology isn’t going anywhere. But the window to demand transparency about how it got here? That’s still open.
