An AI answered the first two 911 calls in the Bourbon Street attack. Callers weren't told.

For roughly 20 seconds on the morning of January 1, 2025, the first voice that callers to the Bourbon Street attack heard was an AI. The Orleans Parish Communication District had designed the tool to pick up only motor-vehicle accident calls, only when every human call taker was busy, and only when the caller was within about 200 meters of a crash already in the computer-aided dispatch system. The attack killed 14 people, the FBI confirmed as terrorism, and the tool's stated guardrails should have kept it silent. They didn't.

Times-Picayune reporting from April 8, 2026 made that 20-second window public for the first time. The account comes from OPCD Executive Director Karl Fasold, who told the paper that an AI agent answered the first two 911 calls from the scene and that the incident was initially classified as a hit-and-run before a human call taker took over and reclassified it as a mass-casualty event. The 911 call recordings are sealed pending the federal probe into the attack, so the 20-second figure rests on Fasold's account to the paper, not on released audio.

The tool itself is from Carbyne, the vendor whose software connects to OPCD's computer-aided dispatch system. OPCD's public framing of the deployment, laid out in an August 6, 2026 statement, describes a narrow, last-resort fallback: motor-vehicle accidents only, all call takers busy, caller within roughly 200 meters of a crash already reported. OPCD says the tool has been in use for about two years and that it was built to absorb a chronic staffing problem at the city's 911 call center, a problem the agency describes as structural and nationwide rather than local.

The 200-meter geofence is the load-bearing piece. A caller standing 250 meters from a reported crash does not reach the AI. A caller standing 200 meters from any reported crash does, regardless of what their own call is about. On a quiet Tuesday morning, that scope is defensible. On a New Year's morning in the middle of a Bourbon Street crowd, the geofence is a radius drawn around a live mass-casualty event that has not yet been classified, and the tool has no way to know what the call is about before the call connects.

Three design questions follow directly from that gap, and none of them are answered in the public record.

First, disclosure. OPCD's public materials describe an automated voice at the start of the call. They do not say the call is coming from an AI. A caller who has just watched a vehicle plow into a crowd is told to wait for a human, then handed off, and the human takes the same call back from a system that has already classified the incident as a car crash.

Second, handoff latency. The 20-second figure is the gap between the AI's initial classification and a human telecommunicator's reclassification. OPCD has not published the distribution of handoff times for motor-vehicle accident calls that the tool does process, so it is not possible to say whether 20 seconds is an outlier or a typical handoff for a non-MVA call that slips through the geofence. A reader cannot tell from the public record whether the same 20-second window would recur on the next false-positive call, or whether it was a worst case.

Third, audit. The 911 audio from the Bourbon Street attack is being withheld by the FBI pending the federal investigation. OPCD has not said whether the AI's audio and classification logs are being preserved for any audit body, or which oversight body would review the tool's decisions after the federal case closes. The vendor, the local PSAP, and the federal investigators all have plausible reasons to want the audio, and none of them have said publicly who is holding it.

Other cities are running similar deployments. The Bourbon Street attack is the first attested public-safety failure of a narrow-design AI 911 tool, and the load-bearing questions for any city considering the same procurement are: does the PSAP publish a written policy on caller disclosure, handoff latency, geofence behavior during non-MVA mass-casualty events, and audio retention for audit? The public answers to those questions are not yet on the record.

OPCD has not said when, or whether, it will publish any of those answers. The federal investigation is still open.