We built a multimodal AI risk engine that screens new listings for duplicated or altered photos and flags high-risk host and guest messages, so fraud is caught before payment.
Industry: HospitalityService: AI & Automation
Overview
Client
Airbnb
Global home-sharing and travel platform
Engagement snapshot
The challenge
Fraudulent listings often reused photos copied from elsewhere on the web or altered to look genuine, and risky conversations between hosts and guests were hard to spot in time.
Reviewers needed to see the riskiest cases first, before any payment went through, rather than finding fraud after a guest had already been affected.
Duplicated or edited images made fake properties look real.
Fraud was often found only after money had changed hands.
Too many false positives slowed down the trust and safety team.
The solution
Computer vision compares listing photos against images across the web to find duplicates and edits.
Language models review communication patterns in real time and flag high-risk interactions.
Flagged listings and bookings go to a person before any payment is processed.
Key features
Computer vision finds listing photos duplicated or edited from elsewhere on the web.
Business impact
Fake listings caught before going live
NLP models read host and guest messages in real time and flag high-risk interactions.
Business impact
Faster response to risky bookings
Photos, messages and booking history are scored together before payment is processed.
Business impact
28% fewer false positives
Before and after
Before
Manual spot checks
Reviewers checked a sample of new listings by eye.
After
Every listing screened
AI checks the photos of every new listing as it is created.
Before
Found after complaints
Risky conversations surfaced only when a guest reported them.
After
Flagged in real time
Language models flag high-risk messages as they happen.
Before
High false positives
Genuine hosts were often caught by broad rules.
After
28% fewer false positives
Cross-checking several signals keeps alerts focused on real risk.
Results
28% Fewer False Positives
Cross-checking photos, messages and booking history keeps alerts focused on real fraud.
Before: Broad RulesAfter: Multimodal Scoring
Caught Before Payment
Fraudulent listings are caught before they go live, and risky bookings before payment.
Before: After ComplaintsAfter: Before Go-Live
Human Review of Flagged Cases
Reviewers decide on every flagged case, with the evidence in front of them.
Before: Manual SamplingAfter: Risk-Ranked Queue
Testimonial
“Automating our trust workflows with this AI risk engine allowed us to catch fraudulent listings before they ever go live. False positives dropped by 28%, and our platform safety metrics have never been stronger.”
Donald HicksVP of Trust and Safety Operations, AirbnbKey results
Trust and control
Flagged listings and bookings are reviewed by a person before action is taken.
HUMAN IN CONTROLRisk checks run before money moves, not after.
PREVENTIVEEvery flag and outcome is logged for review.
TRACEABLERisk thresholds can be tuned as fraud patterns change.
ADAPTABLEFaster delivery
Starting from tested components meant more of the project went into the platform's own fraud patterns and review rules.
Tested components for matching images at scale shortened the build of the photo checks.
Reusable language pipelines were tuned to host and guest conversations.
A ready-made case queue shows reviewers the riskiest listings and bookings first.
Every flag and decision is recorded for later review.
We built a multimodal AI risk engine. Computer vision checks new listing photos for images duplicated or altered from elsewhere on the web, and language models read host and guest messages in real time. High-risk listings and bookings are flagged for human review before payments are processed.
It scores several signals together instead of relying on one broad rule: listing photos, host and guest communication patterns and booking history. Looking at all of them at once keeps alerts focused on real fraud, and false positives dropped by 28%, so reviewers spend their time on the cases that matter.
No. The engine flags fraudulent listings and high-risk reservations, but a person on the trust and safety team reviews each flagged case before action is taken. The checks run before payments are processed, so the team can stop fraud early while genuine hosts and guests carry on as normal.
If fake listings or risky bookings reach your users, we can build a risk engine that flags them before payment.