Flock Gunshot Detection Is Not 'Just Gunshots,' Despite Its CEO's Claim

Flock gunshot detection became harder to describe on October 14, when CEO Garrett Langley insisted that the company’s microphones detect “just gunshots.” Flock’s own website says the system also detects fireworks, car crashes, sideshows, and other disruptive events. That contradiction matters because each added sound category expands what public microphones must analyze.

Langley made the statement during TechCrunch Disrupt 2026 in San Francisco. He was responding to questions about a discontinued feature that detected sounds of human distress, including screaming. Flock had tested that feature with a small number of customers before withdrawing it following public criticism.

The dispute is larger than a poorly chosen phrase. Flock sells audio sensors alongside license plate readers, video cameras, and connected investigation software. Critics including the ACLU and Electronic Frontier Foundation argue that this combination creates opportunities for surveillance to expand beyond its original purpose.

SoundThinking’s ShotSpotter offers the clearest historical comparison. Acoustic detection can help police respond to unreported gunfire, but its alerts also raise questions about accuracy, police deployment, and incidental recordings. Flock now faces the same questions while navigating broader opposition to its nationwide surveillance network.

Flock Gunshot Detection Does More Than Detect Gunshots

The central conflict is visible in Flock’s own product descriptions, which identify several event categories beyond gunfire.

Langley repeatedly narrowed the system’s purpose while speaking with TechCrunch editor-in-chief Connie Loizos. He explained that Flock had ended its human-distress experiment after critics raised concerns. He then said the system would handle gunshots, “plain and simple.”

When asked directly about broader claims in Flock’s marketing, Langley answered, “It’s just gunshots.” Yet the company’s audio detection page presents a different product scope.

The page includes separate sections for gunshot, sideshow, crash, and fireworks detection. It promises alerts for car crashes, fireworks, street takeovers, and other disruptive community events. Flock also says its detection library continues expanding around threats relevant to customer needs.

Those statements do not establish that Flock secretly records conversations. They do establish that its system classifies more acoustic events than Langley acknowledged onstage. That distinction should remain clear.

Flock spokesperson Paris Lewbel later gave TechCrunch a broader explanation. Lewbel said the technology primarily detects gunfire but can also identify fireworks, sideshows, and potential car crashes. He said Langley intended to distinguish event detection from monitoring conversations or analyzing speech.

That clarification addresses intent, but it does not erase the contradiction. “Just gunshots” describes a single-purpose detector. Flock’s website describes an extensible audio-analysis product connected to cameras, license plate data, and public-safety workflows.

Flock’s current explanation says sensors continuously analyze approximately three-second rolling audio windows on the device. A processor compares incoming sounds with supported event signatures. If no qualifying event appears, Flock says the audio is discarded without being recorded or transmitted.

When the system identifies a qualifying event, it saves an encrypted clip and estimates where the sound originated. It then sends an alert to authorized customers. This process is commonly called edge processing, meaning analysis occurs on the deployed device instead of a remote server.

However, edge processing does not mean that a sensor hears nothing until a gun fires. The microphone must process ambient sound continuously to determine whether an acoustic pattern matches a supported category. Flock itself acknowledges this difference between processing and recording.

The company also says a qualifying clip can contain incidental words spoken near the triggering event. That could happen when someone is talking as a gunshot, crash, or other recognized sound occurs. Flock says the product cannot understand, transcribe, identify, or search those words.

These protections matter, but their effectiveness depends on technical design, customer permissions, retention settings, and enforceable policies. The public cannot evaluate those safeguards from a simplified “just gunshots” description. Accurate language is essential when cities place microphones near homes, businesses, and public gathering spaces.

The original TechCrunch account therefore exposed more than an awkward interview moment. It revealed a gap between an executive’s narrow public assurance and the capabilities advertised to potential customers.

That gap creates the article’s core tension. Flock wants residents to judge the product by its limited stated purpose. Its sales materials invite agencies to imagine a broader and expanding range of uses.

Why the Human-Distress Pilot Still Shapes the Debate

Flock withdrew human-distress detection, but the pilot demonstrated how an installed microphone network can acquire new categories through software.

Flock introduced the audio product under the Raven name as an acoustic gunshot-detection system. In 2025, its marketing began promoting detection of human distress. Early material referred more directly to screaming before the company changed that language.

The proposed feature would have alerted customers when its model recognized sounds associated with distress. That represented a meaningful shift. The system was no longer limited to explosive or mechanical events such as gunfire, fireworks, crashes, or screeching tires.

Human distress is more subjective than a sharp acoustic impulse. A scream can indicate an assault, an argument, a celebration, a medical emergency, or a frightened child. Context changes its meaning, while an automated classifier receives only a short acoustic signal.

The EFF’s initial criticism focused on civil-liberties risks and unanswered technical questions. It asked how the feature worked and how it would interact with state laws governing recorded conversations.

Flock says it offered the feature only to a small number of customers during a limited pilot. The company later removed it after community consultation and public criticism. Its current system does not detect screaming or human distress, according to Flock.

That reversal is meaningful. A company abandoned a planned capability instead of deploying it across its customer base. Privacy advocates can reasonably treat that outcome as evidence that public pressure changed a product decision.

Yet the episode also supports their concern about mission creep. Mission creep occurs when a system built for one narrow purpose gradually gains broader uses. The underlying sensors did not require physical replacement before Flock could test human-sound classification.

The expansion occurred through changes to the detection model and supported event library. Flock’s current marketing continues to promote an expanding library, even after it removed the most controversial category. That leaves the boundary dependent on company policy and future customer demand.

Flock argues that the present boundary is firm. Its technical explanation says the system cannot understand speech, analyze language, identify speakers, or provide a live audio feed. It also says most sound is discarded immediately after local processing.

These are important distinctions. Detecting a scream is not the same as transcribing a conversation, while capturing incidental speech is not the same as systematically monitoring language. Public debate becomes less useful when every form of audio processing is treated as identical.

Still, capability boundaries can change faster than public contracts or municipal oversight. An agency might approve microphones for gunfire without anticipating later detection of crashes, street takeovers, or human behavior. Residents may never receive a fresh hearing when software changes the device’s function.

The withdrawn pilot therefore remains relevant even if nobody currently receives distress alerts. It shows that Flock considered human-generated sounds a plausible extension of its product. It also shows that customers requested the capability for locations such as parking lots.

Langley described those requests as attempts to detect abductions or people being harmed. The public-safety rationale is understandable. Faster awareness might help responders reach someone who cannot call emergency services.

The same scenario illustrates the risk. A model would decide which voices sound sufficiently distressed, send information to authorities, and potentially direct armed officers toward a location. False positives and missing context could carry consequences beyond an inconvenient notification.

Flock says alerts support human decisions rather than automatically initiating enforcement. However, the speed and apparent precision of automated systems can affect how people interpret an alert. An officer told that a system detected violence may approach differently from someone answering an uncertain noise complaint.

The disagreement is therefore not about whether technology can classify sounds. It concerns who defines qualifying events, what evidence validates each classifier, and what oversight accompanies future additions. “Just gunshots” avoids those questions instead of answering them.

The Broader Flock Network Raises the Stakes

An audio alert becomes more consequential when it connects automatically with nearby video and license plate records.

Flock’s audio system does not operate only as a collection of isolated microphones. The company promotes integration with its FlockOS platform, license plate readers, video cameras, and computer-aided dispatch systems. A detected event can surface nearby vehicles and footage for investigators.

That integration offers an obvious operational benefit. Police responding to gunfire can receive a location, review nearby images, and identify vehicles entering or leaving the area. Investigators may gain leads even when no witness calls emergency services.

Flock cites an Albany, Georgia, customer who said more than 30 percent of local gunfire had gone unreported. That figure is a customer statement rather than an independent national measurement. It illustrates the gap Flock says acoustic detection can fill.

The connected platform also enlarges the consequences of an incorrect classification. A false gunshot alert does not merely produce a mistaken label. It can focus attention on people, vehicles, or homes captured near the reported location.

Flock says an audio alert does not create an automatic enforcement action. A law-enforcement officer or security professional must review the information and decide what to do. Human review can reduce harm, but it does not guarantee that the initial alert will receive sufficient skepticism.

The company’s scale intensifies those concerns. According to an Associated Press investigation, thousands of law-enforcement agencies across 49 states can search and share Flock data. Flock says its nationwide network includes more than 120,000 license plate readers.

Police supporters credit that network with locating missing people and identifying suspects in violent crimes. Privacy advocates see a system that records routine travel and makes movement searchable across jurisdictions. Both descriptions capture real aspects of the same infrastructure.

The audio controversy arrives while that broader system faces mounting resistance. Cities and agencies have canceled contracts amid concerns about immigration enforcement, unauthorized tracking, and searches conducted for personal reasons. Texas and Florida have also moved against parts of Flock’s camera deployment.

Amazon’s Ring ended its partnership with Flock in 2026. The integration had allowed police to request footage from Ring customers through Flock’s platform. Its termination removed one source of privately collected video from that connected environment.

Flock responded to criticism by announcing new safeguards for its law-enforcement customers. It said customers would have to connect searches with case records and use an audit tool designed to identify abnormal activity. Emergency overrides would receive additional review.

The company also announced a shorter default retention period for license plate data. The standard window would fall from 30 days to seven days, with longer preservation available for evidence connected to a case. Agencies could still apply retention periods required by local or state policy.

These changes show that misuse is not merely theoretical. Guardrails became necessary because authorized users could search a large database for inappropriate reasons. Adding audio-derived events creates another entry point into the same investigative environment.

Flock says customers own their collected data and control how outside agencies use it. Yet a distributed network can remain difficult for residents to understand. Different jurisdictions may choose different retention policies, sharing arrangements, and definitions of permissible use.

The ACLU argues that local control does not solve the structural problem. Its assessment of Flock says shorter retention is helpful only if evidence exceptions remain narrow. It also questions whether revised safeguards sufficiently limit nationwide tracking.

That criticism provides context for the microphone dispute. People skeptical of Flock are not evaluating the audio product in isolation. They are judging it through earlier disputes involving access controls, data sharing, public consent, and shifting product descriptions.

For enterprise buyers and city officials, this is a procurement lesson. Individual specifications cannot substitute for reviewing how a product interacts with the vendor’s entire platform. A narrowly described sensor can produce much broader effects when linked with cameras, databases, and alerts.

Acoustic Detection Has an Accuracy and Accountability Problem

The unresolved question is not whether microphones can detect gunfire, but how reliably their classifications guide high-stakes decisions.

Flock says its sensors use multiple detections to estimate where an event occurred. The system attempts to distinguish gunshots from fireworks and other similar sounds. Customers then receive an alert with associated information for review.

The company does not claim that every alert is correct. Its product explanation explicitly states that no acoustic system is perfect. Flock says mistaken event clips can become feedback for improving its model.

That acknowledgment is necessary because urban soundscapes are difficult to classify. Fireworks, vehicle backfires, construction equipment, and other impulsive sounds can resemble gunfire. Buildings, weather, traffic, and distance can change the signal reaching each microphone.

Flock’s competitor SoundThinking has faced similar scrutiny over ShotSpotter. That system uses distributed acoustic sensors and human review to identify potential gunfire. It has operated for years, providing a substantial precedent for evaluating both benefits and risks.

An AP examination found that ShotSpotter reviewers sometimes reversed algorithmic decisions. According to a 2021 company account, human reviewers disagreed with the machine approximately 10 percent of the time.

SoundThinking presented human review as a quality-control measure. Independent experts told the AP that identifying gunfire can remain difficult, particularly under time pressure. The example shows why an apparently precise alert can involve uncertain judgments.

Flock uses a different product and workflow, so ShotSpotter’s figures cannot be assigned to Flock. The comparison identifies questions that Flock should answer with its own evidence. Those include false-positive rates, false-negative rates, review procedures, and performance across different environments.

Flock’s product page says alerts are filtered and verified, but it provides limited public detail about what verification means. It is unclear whether every alert receives human review before reaching a customer. Public materials also do not provide complete performance measurements for each supported sound category.

Adding categories complicates evaluation. A model that distinguishes gunshots from fireworks addresses one classification problem. Detecting severe crashes or street takeovers introduces different sounds, operating conditions, and definitions of success.

A sideshow, for example, can include screeching tires, engines, music, shouting, or ordinary traffic noise. A severe crash may resemble dropped construction materials or another explosive event. Each category needs evidence appropriate to its real-world deployment.

Independent testing would help separate marketing language from measurable performance. Evaluators would need representative samples, known ground truth, and results broken down by event type. Aggregate accuracy alone can hide poor performance on rare or difficult categories.

Public agencies should also ask what happens after an alert. A system can perform well statistically while still creating serious harm in the cases it gets wrong. Dispatch policy, officer training, and confirmation procedures matter alongside model accuracy.

The Flock microphone privacy debate includes another verification gap. Flock says ordinary audio is processed locally and discarded almost immediately. That design would limit exposure, but outsiders need technical and contractual evidence to evaluate whether every deployment follows it.

Customers also need clear records of software changes. If the detection library expands, agencies should know which categories became active, when they changed, and whether local approval was required. Residents should not have to monitor marketing pages to discover new capabilities.

Flock’s withdrawal of human-distress detection suggests that category-level controls are technically possible. The company could strengthen trust by making those controls visible and auditable. Public dashboards could identify active event types without exposing sensitive investigation data.

The company should also publish retention behavior for event clips. A short recording connected to a qualifying event differs from continuous surveillance, but it still contains public-space audio. Retention length, access logs, sharing rules, and deletion verification determine the practical privacy risk.

None of these concerns prove that Flock records conversations at scale. The available evidence supports a narrower conclusion. Its sensors continuously process ambient sound, retain selected clips, and can classify several events beyond gunfire.

That conclusion is enough to challenge Langley’s wording. High-stakes systems deserve descriptions based on actual capabilities, not the narrowest possible summary. Precision becomes more important as a platform expands into additional data sources and automated analysis.

Privacy Safeguards Cannot Depend on Executive Reassurance

Flock’s credibility problem stems from changing capabilities and contested oversight, not from the existence of a microphone alone.

A microphone that processes three-second windows is not automatically a wiretap. Public acoustic sensors can provide useful information without offering live listening or searchable speech. Flock says it designed its system around that narrower architecture.

The company’s safeguards include local processing, limited event clips, encryption, and restricted access. It says customers cannot open a live feed. It also says the current model cannot identify voices, understand language, or transcribe speech.

Those claims deserve accurate reporting. So does Flock’s decision to remove the human-distress pilot. Criticism becomes less credible when it ignores meaningful limitations or treats every audio sensor as a continuous conversation recorder.

However, residents should not have to rely entirely on vendor assurances. Product architecture can change, and contractual promises can contain exceptions. The party selling the system has an incentive to emphasize benefits while minimizing uncertainty.

Independent audits can test whether devices behave as described. Contracts can forbid speech recognition, require deletion, document model updates, and mandate public approval before enabling new categories. Access logs can show who retrieved event clips and for what purpose.

Local governments also need transparent placement policies. A microphone outside a nightlife district raises different concerns from one near a school, clinic, protest route, or residential window. Community risk depends on location as much as technical capability.

Flock’s scale creates another accountability challenge. Cities may approve individual deployments without understanding cross-jurisdictional connections. Residents can encounter the system while traveling through places where they had no opportunity to participate in procurement decisions.

The company frames its technology as a response to constrained police resources and unreported crime. Langley has argued that society should not pause technological progress because a system remains imperfect. That argument emphasizes comparison with older practices rather than an unattainable ideal.

Critics answer that scale changes the comparison. Traditional police observation was limited by staffing, time, and geography. Networked sensors can collect and connect information automatically across many jurisdictions.

Both sides therefore use different baselines. Flock compares an alert with the absence of information after an unreported shooting. Privacy advocates compare the network with a society lacking persistent, searchable records of ordinary public activity.

The most useful policy approach must address both realities. Cities can evaluate whether acoustic alerts improve emergency response while imposing limits on integration, retention, and secondary use. They do not need to accept every available feature as one indivisible package.

Flock’s imprecise public messaging makes that evaluation harder. If the CEO describes a multi-event system as “just gunshots,” officials may overlook the need for category-specific rules. Clear procurement requires a complete inventory of current and planned capabilities.

The distinction between detection and recording also deserves careful language. Flock says the system continuously analyzes sound but records only qualifying clips. Its product FAQ, however, says the system is not always “recording or listening” because it activates for acoustic signatures.

Those statements can confuse ordinary readers. Continuous audio processing reasonably feels like listening, even when sound is not stored or available to a human. Flock should explain the technical distinction without asking the public to abandon common meanings.

Trust also depends on whether safeguards survive pressure from customers. The distress pilot reportedly began because customers wanted alerts for abductions and assaults. That is a compelling request, but compelling cases often drive surveillance systems beyond their original boundaries.

A future emergency could produce similar demands. Agencies might request detection of breaking glass, chanting crowds, explosions, or other sounds. Each addition might appear reasonable alone while collectively transforming the system’s purpose.

Binding rules matter because executives, products, and political priorities change. A promise made during a contentious interview cannot control future software. Technical restrictions and public contracts provide stronger protection than statements about present intent.

Three Signals Will Show Whether Flock Can Rebuild Trust

Flock’s next steps will matter more than another verbal assurance about what its microphones supposedly hear.

The first signal is whether Flock publishes category-specific performance data. Buyers need false-positive and false-negative measurements for gunshots, fireworks, crashes, and sideshows. They also need testing across different neighborhoods, weather conditions, and sensor layouts.

Transparent measurements would strengthen Flock’s argument that its system delivers useful, bounded event detection. Vague accuracy claims would weaken it. Independent validation would carry more weight than customer testimonials or internally selected examples.

The second signal is how Flock governs changes to its detection library. Public agencies should receive advance notice before new categories become available. Activating a materially different category should require an explicit decision, not a silent software update.

Flock could publish change logs listing supported events, release dates, validation methods, and default settings. Contracts could require renewed authorization for categories involving human vocalizations or behavior. Such controls would address the mission-creep concern directly.

If Flock continues expanding its library without public documentation, Langley’s “just gunshots” claim will look less like simplification and more like misdirection. A transparent approval system would support the company’s narrower privacy narrative.

The third signal is how governments and major partners respond. Contract cancellations, state restrictions, court cases, and local procurement debates will test whether Flock’s policy changes satisfy officials. New independent audits or statutory limits would also reshape deployment conditions.

Flock’s August safeguards provide a partial test. Mandatory audit controls, case-linked searches, and shorter retention can reduce abuse if implementation matches the announcement. Public reporting should show whether those controls detect misuse and produce consequences.

The same standard should apply to audio alerts. Agencies should report alert volumes, confirmed incidents, mistaken classifications, clip access, and complaints. Without those figures, communities cannot compare public-safety value against operational and privacy costs.

Flock gunshot detection sits at an important boundary in public technology. It can notify authorities about events that witnesses never report. It can also connect a machine-generated sound classification with cameras and searchable vehicle records.

That dual use does not produce an easy verdict. It does require precise claims, verifiable safeguards, and meaningful community control. Flock’s website currently offers a more accurate description than its CEO gave onstage.

Readers should watch for evidence that resolves this discrepancy. Does Flock narrow the website, revise Langley’s statement, or acknowledge the full multi-event system plainly? More importantly, will it let outsiders verify what each deployed sensor detects, retains, and shares?

Until those questions receive documented answers, “just gunshots” is not an adequate description. It obscures the system’s existing categories and its ability to change through software. For public microphones linked to a national surveillance platform, that distinction is the story.