Flock's CEO says just gunshots. Its website says more.

Flock's CEO says just gunshots. Its website says more.

There is a specific kind of product problem where the person selling the thing and the documentation describing the thing are no longer telling the same story. Flock Safety walked straight into it on stage at TechCrunch Disrupt.

CEO Garrett Langley was asked about the company's acoustic sensors, which it markets as Gunshot Detection (recently renamed Audio Detection in its own navigation). The sensors have been the target of sustained criticism from privacy advocates. His defense was narrow and blunt: the system detects "just gunshots."

Flock's own website disagrees with him.

Two descriptions of one product

Pull up the product page and the scope is wider than a single class of sound. Flock advertises that the system can "detect street takeovers early to stop chaos before it escalates," "identify severe crashes in real time so help arrives faster," and "expand awareness beyond gunfire with alerts for car crashes, fireworks, and sideshows."

The FAQ goes further. "In addition to gunshots, Flock Gunshot and Audio Detection can detect fireworks, sideshows, and other disruptive community events." Then the sentence that should make any engineer pause: the company's "detection library is continually expanding to include emerging threats that are relevant" to a customer's needs.

When a reporter put that contradiction to Langley directly, his answer was three words: "It's just gunshots."

The spokesperson smooths it over

Flock's spokesperson Paris Lewbel later gave a more careful statement to TechCrunch. The technology is "primarily used to detect gunfire but is also capable of detecting fireworks, sideshows, and potential car crashes." Langley's remarks, Lewbel said, were meant to emphasize that the system "is not designed to monitor conversations or analyze speech and does not currently detect human screaming or distress."

Read those two positions side by side and the disagreement is not about capability. It is about framing. The CEO narrows to the one category that sounds defensible. The documentation lists four categories and promises more. Both are describing the same device.

This is not the first time the scope moved

Flock's earlier attempt at a broader audio feature is the reason this story has context. When the feature was announced in 2025, the company said the sensors would listen for "screaming." After an EFF report and public backlash, the language was softened to "sounds of human distress." Then the company pulled the distress-detection feature from the handful of customers piloting it. Today, the official line is that the system "does not currently detect human screaming or distress," and Flock's own product blog states plainly that the sensors "cannot understand speech, analyze language, or identify voices."

That is the pattern worth naming: the capability list has moved outward, inward, and outward again, and each time the change showed up first in marketing copy rather than in a technical specification. For anyone who builds systems that classify real-world events, that is the whole lesson. If the list of things your product detects lives in prose that marketing can edit, you do not actually have a defined scope. You have a pitch.

What Flock's own docs say about the mechanics

To be fair to the engineering, Flock's product blog is candid about how the system works. Sensors analyze audio in rolling windows of about three seconds, on the device:

• If a sound does not match a qualifying event, that audio is "never recorded, stored, or sent anywhere."
• If it does match, the system records, encrypts, and saves only that short clip, estimates a location, and sends an alert.

The blog describes a single on-device question: does this sound match a qualifying event such as "a gunshot, vehicle crash, or tires screeching?" The privacy-relevant part is what happens next. A confirmed event "can surface the vehicles seen entering or leaving the area around the time of the sound, using nearby Flock license plate reader cameras."

So the audio sensor does not exist in isolation. It is a trigger wired into a separate, much larger data collection system.

The developer version of this problem

Strip away the vendor and the dispute is a familiar engineering failure: a detection system whose behaviour changes without a matching change to its contract. The fix is to make the allowed event set explicit, versioned, and enforced at the boundary instead of trusting a sentence on a marketing page.

Here is a small pattern that turns "the detection library is continually expanding" into something an operator can actually audit:

from dataclasses import dataclass

# The contract: only these classes may ever raise a user-facing alert.
# Changing this set must be a reviewed, versioned change -- not a copy edit.
ALLOWED_EVENT_CLASSES = {
    "gunshot",
    "firework",
    "vehicle_crash",
    "tire_screech",
}
SCHEMA_VERSION = "2026-10-15"

@dataclass(frozen=True)
class Detection:
    event_class: str
    confidence: float
    device_id: str

def emit_alert(det: Detection) -> dict | None:
    if det.event_class not in ALLOWED_EVENT_CLASSES:
        # Fail loud. A new class must be added to the contract on purpose.
        raise ValueError(
            f"undeclared event class {det.event_class!r}; "
            f"allowed={sorted(ALLOWED_EVENT_CLASSES)}"
        )
    if det.confidence < 0.8:
        return None
    return {
        "schema": SCHEMA_VERSION,
        "class": det.event_class,
        "device": det.device_id,
    }

The point is not the exact code. The point is that the set of events the system is allowed to act on is a constant that a reviewer can diff, not a paragraph a copywriter can quietly extend. If a vendor's own FAQ says the classification list is "continually expanding," the honest engineering response is to demand that expansion arrive as a versioned schema change you can audit and roll back.

Why this matters beyond one vendor

Flock is not a fringe company. It says its microphone count is comparable to ShotSpotter's, which Wired reported at roughly 25,000 sensors as of 2024. The scrutiny is real and mounting: the ACLU has pressed the company, some members of the public have physically destroyed its cameras, Amazon's Ring ended its partnership, and in September 2026 both Texas and Florida moved to end their relationships with Flock. The company also cut staff in early October 2026.

None of that changes the technical question. A sensor that classifies ambient sound is a classifier, and every classifier has a label set. If the label set is described one way to a journalist and another way to a customer, at least one of those audiences is being misled, and neither has a stable specification to reason about. For teams building anything adjacent to this space, whether it is an AI-assisted alerting tool or a police-facing dashboard, the takeaway is the same: define the contract in code, surface it to users, and make widening it a deliberate act with a paper trail. The mark of a mature detection system is not that it detects only one thing. It is that everyone, including the sales team, can point to the exact same list of what it detects.

If you want a related case of AI systems making consequential claims in a law-enforcement context, our earlier piece on the false Anthropic-generated homicide tip is a useful companion.

Sources

• TechCrunch: Flock CEO says its microphones detect 'just gunshots' — its website says different
• Flock Safety: How Flock's Gunshot Detection Works, and the Concerns We Hear Most
• Flock Safety: Flock Gunshot Detection product page