BRENNEN

Industries / Construction and property / Real estate / Short-term rentals

Trust and safety

This is the work of making it safe to let a stranger into a home or to sleep in one: checking who people are, rating them, paying for damage, stopping parties and hidden cameras, and catching discrimination and fake listings.

6

Milestones achieved, from 2011 to 2024.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2011

    Damage guarantee after EJ

    In mid-2011 a San Francisco host known as EJ reported that a guest had ransacked her home. On 1 August 2011 Airbnb's chief executive Brian Chesky apologised ("we let her down") and announced that from 15 August Airbnb would cover hosts for up to $50,000 of loss or damage from vandalism or theft by a guest, backdated to earlier reports. By November 2022 Airbnb's host cover, AirCover for Hosts, listed $3 million of damage protection, guest identity checks and reservation screening.

    LessonAirbnb said it "should have responded faster" and taken "more decisive action"; the platform, not each host, came to carry the risk of damage.

    techcrunch.comnews.airbnb.com

    Credit

    Brian Chesky and Airbnb

  2. 2017

    Discrimination measured by experiment

    Three Harvard economists sent about 6,400 booking requests to Airbnb hosts in five US cities in July 2015 from made-up guests who differed only in name. Guests with distinctively African American names got a positive reply about 42% of the time and guests with distinctively white names about 50%, so they were about 16% less likely to be accepted. The paper appeared in the American Economic Journal: Applied Economics in April 2017. Airbnb began measuring booking success by perceived race in 2020 (Project Lighthouse) and reported in December 2024 that the largest gap it found in 2021, between guests perceived as Black and as white, had been cut almost in half; that figure is Airbnb's own.

    LessonThe authors found that only a subset of hosts discriminated, most of all hosts who had never had an African American guest, and said Airbnb's design choices made it easier.

    benedelman.orgdoi.orgnews.airbnb.com

    Credit

    Benjamin Edelman, Michael Luca and Dan Svirsky

  3. 2020

    Ratings bunch near the top

    Collecting ratings for millions of Airbnb properties, three researchers found that nearly 95% averaged 4.5 or 5 stars and almost none fell below 3.5, while about 700,000 hotels on TripAdvisor averaged 3.8 stars; the pattern held in data from both 2015 and 2018. A separate experiment on Airbnb, published in 2021, hid each review until both host and guest had written theirs: people wrote more reviews, retaliated less and gave lower ratings.

    LessonWhen two people review each other, fear of retaliation pushes ratings up; hiding reviews until both are in brought them down.

    link.springer.comdoi.org

    Credit

    Georgios Zervas, Davide Proserpio and John Byers; Andrey Fradkin, Elena Grewal and David Holtz

  4. 2022

    Party ban made permanent

    After five people died at a Halloween party in a California Airbnb rental in 2019, Airbnb promised to check all its roughly 7 million listings by 15 December 2020, open a 24-hour line for neighbours and screen high-risk bookings worldwide. It banned parties for the pandemic in 2020 and made the ban permanent on 28 June 2022, saying party reports had fallen 44% in a year and that more than 6,600 guests were suspended in 2021 for trying to break it; both figures are Airbnb's own.

    techcrunch.comthegrio.com

    Credit

    Airbnb

  5. 2023

    Identity checks for every booking

    From the end of June 2023 Airbnb blocked bookings by guests, and the calendars of primary hosts, who had not completed its identity check: matching a government ID, sometimes with a selfie, or matching name, address and other details against outside databases. Since November 2022 the check had applied to guests travelling to Airbnb's top 35 countries and regions.

    LessonAirbnb itself says "no identification process is foolproof".

    news.airbnb.com

    Credit

    Airbnb trust team under Naba Banerjee

  6. 2024

    Indoor cameras banned

    On 11 March 2024 Airbnb banned security cameras inside listings worldwide from 30 April 2024, even disclosed ones, replacing a rule that had allowed disclosed cameras in common areas. Outdoor cameras must be disclosed before booking and may not view indoor spaces, and noise monitors that measure only decibel levels, recording no sound, are allowed in common spaces if disclosed.

    news.airbnb.com

    Credit

    Airbnb

Watch for

  1. Party screening by machine

    Airbnb says its system for spotting bookings at risk of an unauthorised party, which weighs signals such as trip length, how far the guest is from the listing and whether the booking is last-minute, redirected more than 20,000 people in the US away from booking whole homes over the July 4 weekend in 2025. It earlier said the party ban "correlated with" a 55% drop in the rate of party reports and that 0.039% of reservations in 2022 led to a party allegation.

    Achieved whensomeone outside Airbnb, such as a city police department or a noise-complaint service, publishes records showing fewer party incidents at short-term rentals after the screening began.

    news.airbnb.comnews.airbnb.comnews.airbnb.com

    Claimed by

    Airbnb, June 2026

Open

  1. 01

    Test hidden-camera finders

    A public bench test of the methods travellers use to find hidden cameras: phone apps that look for lens glint, handheld radio-frequency and lens detectors, Wi-Fi network scans, and the open-source LAPD phone app as a research baseline. Cameras are planted in a room the testers own or rent for the purpose, and volunteers search for them with each tool.

    The brief
    Credit · unclaimed
  2. 02

    Open party-noise dataset

    An open dataset of decibel-only sound levels, one reading per second with no audio kept, logged during staged gatherings of known size (2 to 30 people) and during ordinary loud events (television, vacuuming, a dinner, a barking dog), plus a simple published model that tries to tell a party from everyday noise.

    The brief
    Credit · unclaimed

Open 01 · the brief

Test hidden-camera finders

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public bench test of the methods travellers use to find hidden cameras: phone apps that look for lens glint, handheld radio-frequency and lens detectors, Wi-Fi network scans, and the open-source LAPD phone app as a research baseline. Cameras are planted in a room the testers own or rent for the purpose, and volunteers search for them with each tool.

Why it is still open

The one large measured study, LAPD's, tested its own research app against the naked eye, finding 88.9% of hidden cameras with the app and 46.0% by eye across 379 participants; it did not rank the detectors and apps sold to travellers. Carnegie Mellon's Lumos (USENIX Security 2022), written with hotel rooms and Airbnb rentals in mind, identified hidden Wi-Fi devices with 95% accuracy and located them to a median 1.5 m within 30 minutes across 44 devices, but it finds only devices that use Wi-Fi and is a research system, not a retail tool. A search found no independent test of the detectors sold to travellers. Airbnb's 2024 ban on indoor cameras leaves guests to check for themselves.

github.comdoi.orgusenix.org

Built with what exists

Off-the-shelf mini cameras in common disguises (clock, charger, smoke-alarm housing); consumer detectors and apps bought at retail; the LAPD Android app (MIT licence) on a phone with a time-of-flight sensor; nmap for Wi-Fi scans; a spreadsheet for the tally.

  1. First step

    Buy three cameras and three detectors, and run 20 timed searches in one room with two volunteers, with the cameras pointed at a wall and their footage deleted.

  2. Done looks like

    Ten camera models in five hiding places, at least 20 timed searches per tool by different volunteers, and a published table of find rate, false alarms and search time for each tool, with photos of every placement.

Open 02 · the brief

Open party-noise dataset

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open dataset of decibel-only sound levels, one reading per second with no audio kept, logged during staged gatherings of known size (2 to 30 people) and during ordinary loud events (television, vacuuming, a dinner, a barking dog), plus a simple published model that tries to tell a party from everyday noise.

Why it is still open

Hosts may use decibel-only monitors under Airbnb's rules, and vendors such as Minut train sound-recognition models on their own library of about 50 sound classes without publishing the data or the model. The nearest research, Crowdotic (2023), estimated how crowded a hospital waiting room was from audio with speech removed and did better than a thermal camera, but it used sound recordings rather than decibel readings alone, and its paper does not say the data is released. A search of the web and GitHub found no open dataset of indoor sound levels labelled with the number of people present.

minut.comarxiv.org

Built with what exists

A Raspberry Pi with a calibrated USB measurement microphone that computes A-weighted levels on the device and discards the audio; Python and scikit-learn for the baseline; Hugging Face Datasets or Zenodo to publish.

  1. First step

    Log one home for a week, including a dinner and a birthday party, with every person present agreeing and a written log of how many people were in the room each hour.

  2. Done looks like

    At least 100 hours of labelled readings from 10 or more rooms, including 30 or more gatherings with head counts, released under an open licence, with the baseline's precision and recall at flagging gatherings of 10 or more people within 15 minutes.