Latest / The CTO Podcast with Fexingo: Technical Leadership, Architecture, and Engineering Org / How Airbnb Rebuilt Its Trust and Safety Team for 150 Million Guests
Transcript
- Lucas: So Luna, Airbnb just spent the last eighteen months rebuilding their entire trust and safety stack from the ground up, and the numbers are frankly wild. They cut fraudulent bookings by 40 percent in six months after the new system went live. Luna: Forty percent — that's not a small win. What was driving that fraud in the first place? I mean, Airbnb has been around for over a decade now. Lucas: Right, but the scale of the problem changed. By 2024, they were dealing with 150 million guests and 8 million listings, and the fraud rings got sophisticated. They were creating fake accounts en masse, using stolen identities, and then listing properties that didn't exist or didn't belong to them. Luna: So it's not just the classic 'fake host' scam — it's the whole lifecycle, from account creation to the actual booking. Lucas: Exactly. And the old system, which was mostly rules-based — think flagging certain email domains or IP addresses — just couldn't keep up. The fraudsters were using residential proxies and ai generated photos, so the signals that used to work were dead. Luna: So what did they actually build? Give me the architecture, not just the buzzwords. Lucas: The core is a graph database. Every user, device, payment method, and listing becomes a node, and the edges are relationships. When a new account signs up, the system runs a real-time risk score by traversing that graph — checking if the device has been seen before, if the payment card is linked to other accounts, that kind of thing. Luna: That's the classic 'fraud cluster' approach, but the key is the real-time part. How fast is this scoring actually happening? Lucas: They got it down to under fifty milliseconds per request. That's the critical number, because anything slower would add noticeable latency to the booking flow. And they're doing this for every single action — not just sign-ups, but also logins, listings, and booking requests. Luna: Fifty milliseconds is impressive, but what about the false positives? If you're blocking legit users, that's a trust problem in itself. Lucas: That's where the human-in-the-loop comes in. The model doesn't just say 'block' or 'allow'. It outputs a risk score, and if it's in a gray zone, the case gets routed to a human review team. They also built a feedback loop where those human decisions feed back into retraining the model. Luna: So it's a hybrid system — automation for the obvious cases, humans for the ambiguous ones. How big is that review team? Lucas: They scaled it to about 2,000 people globally, but they're not reviewing individually. They built a case management system that clusters related accounts and surfaces the highest-risk ones first. So the team's time is focused where it matters most. Luna: That makes sense. But I imagine the really hard part isn't the tech — it's the trade-offs. You mention false positives, but what about privacy? You're collecting device fingerprints, biometric data, maybe even social graph info. That's a minefield. Lucas: Privacy was actually a huge consideration. They had to balance the need for data with GDPR and other regulations. For example, they moved away from storing raw biometric data. Instead, they use on-device processing to generate a hash, and that hash is what gets compared. So the actual biometric never leaves the user's phone. Luna: That's a smart approach — privacy-preserving verification. But hashing isn't perfect, right? If the hash is deterministic, a stolen hash could be replayed. Lucas: Right, but they added a liveness check — the user has to blink or turn their head during the scan — so a static image or a video replay won't pass. And the hash is salted with a session-specific value, so it can't be reused across sessions. Luna: Okay, so they've built this impressive system. But I'm curious about the human side. How did they get the user community to accept this? I mean, extra verification steps are a pain for guests and hosts. Lucas: That's actually the part I find most interesting. They rolled it out gradually, starting with high-risk regions and account types. They also communicated transparently — explaining why the checks exist and how they protect the community. And they saw a measurable drop in scam reports, which they could point to as proof that it was working. Luna: So the trust that comes from safety outweighs the inconvenience. That's a good lesson for any platform that's thinking about adding friction. Lucas: And it ties into a broader point about the show — how operators and builders need to think about these trade-offs together. The business side has to accept that some friction is necessary, and the engineering side has to minimize that friction as much as possible. Luna: Yeah, it's a classic partnership. And speaking of partnerships — I know we don't run ads on this show, but that's a choice we make because we want these conversations to be purely useful for people building and running systems like this. If you find value in that, and you want to support it, the link is buy me a coffee dot com slash fexingo. No pressure, just an option. Lucas: Exactly. It keeps us independent and ad-free, which is how we like it. But back to Airbnb — what's the one takeaway you'd want our listeners to remember from this? Luna: For me, it's that trust infrastructure is not a one-time project. It's an ongoing process that requires constant adaptation. The fraudsters are always evolving, so your defenses have to evolve too. Lucas: And I'd add that the human element is non-negotiable. You can't just throw AI at the problem and expect it to work. You need that feedback loop, that judgment, to make the system both effective and fair. Luna: So, looking ahead, what's next for Airbnb's trust and safety? Are they planning to expand this to other areas? Lucas: They're already using it to detect fake reviews and to flag suspicious activity during the stay itself. And there's talk of using the same infrastructure for their experiences product. So the platform is becoming a general-purpose trust engine. Luna: That's a great example of how a well-built system can be reused across different parts of the business. It's not just a siloed solution. Lucas: Exactly. And that's what makes it a cto level success story. It's not just about the tech, it's about the organizational design and the culture that supports it. Luna: Well, I think we've covered a lot of ground here. Any final thoughts? Lucas: Just that if you're building a marketplace or any platform with user-generated content, trust and safety should be a first-class citizen from day one. You can't bolt it on later. Luna: That's a good note to end on. Thanks for listening, everyone.