en

The VICE Channels

    Uber's Phantom Cabs

    Written by

    Alex Rosenblat

    When Heather*, a driver who has been working for Uber for about eight months, opened up the passenger app a few weeks ago from her residence, she noticed something peculiar. The app’s map showed four drivers on the streets immediately by her pick-up location. Yet, the estimated wait time for the closest car was 17 minutes, and there were no other drivers in sight.

    Screenshot of the four phantom UberX cars Heather saw on her screen. Note how close the cars appear to the passenger pick-up location pin, yet the closest available vehicle appears to be 17 minutes away, by Uber’s estimate.

    There are two versions of Uber’s app: one for drivers to use to find passengers, and one for passengers to use to hail a ride. Frequently, drivers login to the passenger app to see where other drivers are so they don’t sit unknowingly in the same one-mile stretch as the competition.

    What the passenger app shows can be deceptive, however. The discrepancy Heather noticed wouldn’t have been obvious in a busy location with a shorter wait time. But in more remote areas, the app clearly shows drivers where there are none.

    Over a six month period, my colleague Luke Stark and I have been studying how Uber drivers interact with the Uber app as part of a research project funded by Microsoft FUSE Labs. Our research was conducted primarily in Uber driver forums, and through interviews with Uber drivers. We’ve observed that drivers across multiple forums discuss the fake cars they see on their own residential streets. One driver, who makes regular broadcasts advising drivers about using the Uber system, even made a YouTube video to show other drivers how the app sometimes displays cars that aren’t there.

    Some drivers on the forums noted that it could be a 30 to 60 second lag in the system that causes the number of cars to be misrepresented, but others report regularly seeing a group of phantom cars wherever the passenger pin is placed.

    "I know this seems a misleading to you but it is meant as more of a visual effect more than an accurate location of drivers in the area."

    There is speculation that it’s more likely that this is intentional on Uber’s part, rather than a bug in the system. If a potential passenger opened up the app and saw no cars around, she might take another cab service. But if she saw a cluster of cars seemingly milling around on the same street, she’s more likely to request a ride.

    When asked about this seeming discrepancy, an Uber spokesperson said that on the rider app, “The number of cars and their location are generally accurate.”

    When Heather asked an Uber Help staff member, however, she was told that the rider map was just a “screen saver.”

    “The app is simply showing there are partners on the road at the time,” the staffer wrote in an email. “This is not a representation of the exact numbers of drivers or their location. This is more of a visual effect letting people know that partners are searching for fares.”

    “I know this seems a misleading to you but it is meant as more of a visual effect more than an accurate location of drivers in the area. It would be better of you to think of this as a screen saver on a computer. Once a rider request a trip there will be actual information about the partners [sic] location showing up in the app."

    Uber makes no distinction, visually, between the little black sedans that are available to passengers, and those that may be part of a “visual effect.”

    I wrote to an Uber Help staffer as a passenger in the same month as Heather, but in a different region, asking the same question. I got a different answer: the staffer explained any discrepancy as a function of zooming in or zooming out.

    “The map is as accurate as possible in the close vicinity of your location,” the representative wrote.

    The real-time data challenge

    Uber has extensive data on historical and real-time passenger and driver movements, but it can be a challenge to relay relevant information back to the passengers and drivers using the app.

    “From a purely engineering problem, trying to fit a lot of fairly complex information fairly intuitively into an iPhone screen [is] without a doubt one of our biggest issues,” said Uber data scientist Kevin Novak in a 2014 presentation [relevant part starts at 38:30] to a meetup group.

    In theory, if Uber displayed the real-time positions of drivers, that information could be used by competing driver services, who might send their drivers to fill any gaps. But another Uber Help staff who advised Heather on the phantom car issue added, “The Uber Engineers are working hard in trying to improve the software for us to be able to have a real-time view of the availability of Uber partners in the area.”

    Yet, by keeping the app’s users in the dark about the accuracy or the limitations of the information it relays currently, Uber may be able to leverage more control over how drivers and passengers interact with the system.

    Outsmarting the algorithm

    Other anomalies in Uber’s visual representations emerged from the drivers’ experiences, particularly around surge pricing.

    Uber connects drivers with passengers through direct matches between a specific driver and a specific passenger, and by clustering drivers in areas where Uber indicates there is high passenger demand. It uses text messages, push notifications, emails, heat maps, and dynamic pricing to mobilize drivers to roam in high-demand areas at certain times.

    Uber says that its surge pricing algorithm adjusts the cost of a trip to reflect supply and demand: prices go up when there are more passengers requesting rides, like during peak hours or when it starts to rain. “Dynamic pricing is triggered by data streams from the rider and driver apps that provide information regarding supply and demand,” an Uber spokesperson said in an email. Uber shows drivers heat maps to visualize this supply and demand.

    Sample of a heat map that drivers see. Yellow areas indicate rising passenger demand, orange means pre-surge, and red means surge is in effect. The magnitude of surge can vary across a given region.

    While surge pricing is represented as a reflection of the marketplace, our research suggests that Uber’s algorithms are also predictive: they forecast supply and demand so that drivers can be pre-positioned to meet predicted demand, but they don’t always reflect an accurate picture in real time.

    The predictive nature of the algorithm is also explained in Uber’s patent application for its dynamic supply positioning system. The application describes how “one or more sub-regions can be determined to be under-supplied by the service providers as compared to one or more other sub-regions by forecasting or making a prediction for a future instance in time (e.g., a time ahead of the current time) that a sub-region will be under-supplied at the future instance in time… In this manner, a service provider [such as a driver] can, for example, recognize which sub-regions will be under-supplied in the future instance in time (e.g., ten minutes from the current time), and move to a particular location to increase business.”

    The same patent application says that in some implementations, Uber’s system uses factors including historical data, weather, and calendar information (like events) to determine when an area will be undersupplied.

    Drivers across different markets often find surge notifications are not a reliable indicator of demand, however, so they react to the surge in other ways. Some drivers follow the messaging instructions and respond positively to surge pricing. Others engage with surge based on how they believe other drivers respond to Uber’s incentives: they might avoid the areas Uber directs them to, or drive at times that Uber doesn’t explicitly highlight so that there’s less competition.

    Others have noticed that the surge pricing will disappear if drivers flock to the area, so they consider waiting just outside the edge of a surge area to help sustain its rise.

    Many drivers check the passenger app to see what kind of competition they’re up against. When surge is at 1.5x the usual rates, some drivers strategically check the passenger app to assess whether they should accept rides at that rate or wait for surge to increase. If there is a lot of visible competition, the strategic move is to accept ride requests, but if there are only a few other sedans nearby, they may wait for surge to go up.

    A text Uber sent to a driver, who posted it in a forum with the comment:

    “No thanks, uber. Having surveyed the sitch via the passenger app, my market is fairly saturated with upwards of 10 drivers at the moment.” - Peter M., Uber driver


    Jake W., an Uber driver, posted this screenshot of surge pricing to a forum with the comment: “Headed down.” Surge pricing motivated him to head to the area in red.


    A notification from Uber to a driver who tried to log off: “Demand is very high in your area. Make more money, don’t stop now!”

    Uber’s ability to reduce wait times for passengers and increase rides for drivers through an algorithmic analysis of where everyone is can be a huge benefit to the app’s users, and a boon to cities seeking to reduce traffic and pollution.

    A screenshot from the “God-view” Uber has of passenger and driver activity. Blue lines are where the cars have gone, green lines are people that are going to get picked up, and eyeballs are people who have the app open, but haven’t yet ordered a ride. This screenshot and explanation was featured in a 2011 presentation at a Node.js meet-up by Curtis Chambers, an Uber engineering manager, titled "Distributed Web Architectures."

    But Uber doesn’t tell drivers how long a surge is expected to last or how many drivers have been alerted to a surging area—information that could help drivers better plan their shifts and their next direction, be it towards or away from a surge zone.

    Ok been sitting in this surge for half an hour and not one ping!!!” – Dave H., Uber driver, posted to a forum

    In forums and in interviews, many veteran drivers advise against “chasing the surge” because they find it isn’t a reliable source of fares. One study on driver behavior found that “more than half of interviewed drivers, however, were not influenced by surge pricing.” Some only pursue surge zones if they’re nearby; if they have to drive 20 minutes to get to one, it might disappear before they arrive. Others find themselves sitting in a surge zone waiting for 15 to 20 minutes without getting a ride request.

    Drivers are not told how many passengers are waiting in that area, or how many other drivers are alerted to the same zone. While Uber engineers have a "God-view" of passenger and driver activity, the service's drivers lack that data.

    The passenger play

    Passengers learn to game the system, too. Some drivers report that passengers will place their pick-up location outside of a surge zone, and then call drivers to direct them to where they really are. Some drivers also express frustration when they receive ride requests from outside the surge areas when they’re waiting in a surge because it means they have to work at lower rates. They risk a hit to their cancellation rating, which Uber monitors, if they decline rides. In the San Francisco market, for example, drivers are expected to have a cancellation rate of less than 5 percent, or they risk losing access (being fired) to the system:

    Screenshot from Uber’s partner website in San Francisco.

    Drivers, and users more broadly, lack an understanding of how algorithmically-curated information flows are filtered and manipulated for easier (and sometimes, misleading) consumption. They learn how to evaluate algorithmic incentives against their own experience, and the experiences of other drivers.

    Uber’s access to real-time information about where passengers and drivers are has helped make it one of the most efficient and useful apps produced by Silicon Valley in recent years. But if you open the app assuming you’ll get the same insight, think again: drivers and passengers are only getting part of the picture.

    Alex Rosenblat is a researcher and technical writer at Data & Society, a New York research institute focused on social, cultural and ethical issues arising from data-centric technological development.

    *Names have been changed or modified to preserve driver anonymity.