Context

Garrett Langley is the CEO of Flock Safety, which sells license plate reading cameras and drones to roughly 6,000 local governments in the US, and is interviewed by Jason Calacanis. Over the past year Flock has faced a public backlash over privacy, including camera vandalism and roughly 60 to 70 cities cancelling contracts, while also being credited with solving over a million crimes last year. Langley walks through how the company set its data retention defaults, built a tool to catch its own customers abusing the system, and why it's deliberately going slow on AI.

The Big Idea

The controversy around Flock wasn't really about the technology, it was about trust, and the product decisions that actually earned trust back were treating data retention as a liability to minimize by default and building the company's own internal tool to catch customer abuse before regulators or reporters did.

Langley traces most of the backlash to a root cause that had nothing to do with cameras: many critics simply didn't trust the police departments buying the product, which meant no amount of technical explanation could fix the relationship without visible accountability mechanisms.

Key Insights

Data retention defaults are a product decision

• Default: shifted from 30 days down to 7, based on data showing 90% of crimes get solved within that window.
• Exception: the remaining 10%, harder cases like delayed sexual assault reports, need longer retention, which is why individual city councils set their own limit instead of one universal default.
• Philosophy: Langley states it directly: "this data is a liability, not an asset."

An internal tool caught the company's own bad customers

• What: "Audit Assistant," built four months ago, flags anomalous search patterns, like an officer repeatedly searching the same plate on consecutive days without putting it on a shared hot list, a signature of stalking rather than a legitimate investigation.
• Result: enough abuse surfaced that Georgia police chiefs have publicly fired officers as a direct result, including one department that fired nine at once.

Narrow scope became the company's defensible moat

Flock deliberately does not do facial recognition, does not capture video, and does not let officers search by person, capabilities some competitors offer. Langley frames this restraint as what let Flock survive scrutiny a less disciplined competitor couldn't, and hopes the spotlight on Flock forces the rest of the industry to adopt the same limits.

The crisis was a communication failure, not a technology failure

Langley admits Flock spent years assuming a basic audit log was sufficient and under-communicated what the product does and doesn't do. One early customer cancelled after Googling a competitor's facial recognition page and mistakenly attributing it to Flock. His lesson: even a sound product loses trust if a skeptical stranger can't easily verify what's captured and for how long.

Losing customers on purpose can be the right call

Flock made accountability tools mandatory rather than optional, knowing some customers who preferred fewer constraints would leave for looser competitors. Langley's reasoning: a customer unwilling to accept accountability isn't one he wants, since their eventual misuse would damage the company more than the lost revenue.

AI features are paced behind what's technically possible

Flock refuses to build features that predict suspicion from behavior, like flagging a car circling a block, even though the pattern-detection capability exists, because Langley considers defining suspicion "really dangerous." Every AI feature must keep a human in the loop and pass third-party attestation before shipping, a direct response to watching other companies rush AI 911 call handling and get it wrong.

Mental Models & Frameworks

Data as liability, not asset

Treat any stored sensitive data as a cost and risk to minimize by default, not a resource to accumulate for future value, and require a clear efficacy case (like the 90%-of-crimes-in-7-days number) before extending retention past the minimum. This inverts the common product instinct to keep data "just in case" it becomes useful later.

Third-party attestation as a release gate

Before shipping a high-consequence AI feature, require validation from someone who doesn't want the feature to work, not just internal QA. The bar shouldn't be whether the company itself believes the feature works, but whether a skeptical outside party can independently confirm it improves outcomes without new harm.

Trade-offs & Nuance

Retention length versus crime-solving efficacy

Shorter retention better protects privacy but leaves some crimes unsolved: 90% of cases close within 7 days, but the remaining 10%, including delayed-reporting cases like assault, often need 30 days or more. There's no universal right answer, which is why Flock defaults conservative but lets each city council configure it as a local political decision, not a fixed policy.

Common Mistakes

Treating the product as a neutral tool

Flock initially reasoned the way a taser or laptop vendor might: responsibility for misuse rests entirely with the customer, not the toolmaker. Langley now argues that framing doesn't hold once a company becomes consequential enough that public trust depends on it, and that sensitive-data providers have to own downstream misuse, not just point to their terms of use.

Practical Application

Build abuse detection into your own product first

Don't wait for a regulator or reporter to surface misuse; build internal tooling that flags anomalous usage the way Flock's Audit Assistant flags repeated single-target searches. Getting there first turns a potential scandal into a credibility-building disclosure instead of a crisis.

Make defaults defensible in one sentence

• Do: set sensitive defaults, like retention windows or access scope, conservatively enough to explain to a skeptical stranger in one sentence and have it hold up.
• Example: Flock's 30-to-7-day shift came from a specific number, 90% of crimes solved, not an arbitrary policy.

Require external validation for high-stakes AI features

Before shipping an AI feature with real consequences if it's wrong, get a third party with no stake in it to independently verify it. Treat this as a launch gate for anything touching safety, health, or financial outcomes, not a nice-to-have.

Be willing to lose customers who reject accountability

If a subset of customers wants less oversight than your product now requires, let them go to a competitor rather than softening the requirement. The customers who stay because they accept accountability are the ones whose eventual misuse won't come back to damage your reputation.

Questions to Consider

• Where in our own product are we treating a sensitive default, like a data retention window or an access control setting, as a technical afterthought instead of a deliberate, defensible policy decision?
• If our most privacy-invasive possible misuse case (an internal user repeatedly targeting the same person) were happening today, would we catch it ourselves, or would we only learn about it from a customer complaint or a reporter?
• Are we setting the terms other companies in our space have to follow, the way Flock's no-facial-recognition stance pressures competitors, or are we letting a less careful competitor define what customers expect as normal?

Bottom Line

Flock's public crisis and partial recovery both came down to product decisions functioning as trust decisions: a conservative data retention default backed by real numbers, an internal tool that caught the company's own customers misusing the product, and a deliberate refusal to ship AI features faster than they could be independently verified.