It’s a well-known yuppie hack. Four friends call an Uber for the bar; the one with the lowest fare hails the ride. That they are going from the same place to the same stop simultaneously doesn’t faze the pricing algorithm. I once had a friend go to such lengths as to search for rides to destinations to which he planned to drive, only to exit the app before getting in his car. He wanted to trick the algorithm into believing he was a highly price-sensitive customer. He had a streak of winning the hailing wars until the rideshare apps seemed to catch on to his game. The anecdote is funny; the implication that apps are tracking you to set a price is not.
Using algorithms to set prices, including personal prices, has innovated some industries, but may also destabilize traditional consumer protection and antitrust policy. In markets like rideshare where personal data usage is unavoidable, policymakers can use machine learning techniques to prevent algorithmic abuse, but idolizing customer privacy comes with anticompetitive risks of its own.
The Mixed Bag of Algorithmic Pricing
Algorithms make prices both more dynamic, in that they can respond in real time to market conditions, and more personalized, in that price tags can be based on individual consumers’ behavior and traits. Personalized pricing becomes surveillance pricing when companies obfuscate what personal details they use to tabulate a price, leaving users to guess how best to play the game. It also gets creepy fast. If your search history reveals you are travelling to attend your father’s funeral, JetBlue can charge much more for a trip home. If a property manager’s pricing algorithm connects some data dots to conclude you are going through a divorce and want to live in the same apartment building as your kids, you could pay an additional premium in rent.
The economics are not straightforward. Sometimes more information about consumers actually improves competition and lowers prices. Other times, it allows dominant firms to usurp consumer surplus and act as de facto monopolies. Regardless of market effects, sellers having near-perfect information changes the power dynamic of economic transactions, leaving consumers to feel cheated when companies use their immutable characteristics or personal desperation against them.
A Rideshare Case Study
Rideshare apps are unique in the personalized pricing discourse in that they function as a clearing house for buyers and sellers of driving. Unlike other marketplaces that compete in the markets they serve—for example, Amazon selling its Basics products alongside traditional producers—rideshare apps don’t have company drivers. Unlike other industries employing algorithms simply to boost margins, a rideshare company’s success hinges on cracking a highly complicated optimization problem. Long before Uber, taxis priced dynamically based on traffic density; the rideshare innovation is minimizing driver downtime by pairing nearby drivers and riders with mutually acceptable rates; hyper-segmentation is intrinsic to the market.
This renders useless the common policy prescriptions to deal with algorithmic pricing, such as mandating explainable AI (whose outputs are at least in theory comprehensible) or warnings that prices are personalized. Such transparency works in other markets, but when the pricing mechanism is necessarily proprietary and necessarily personal, a lack of transparency easily becomes a license for surveillance. Rideshare is the example par excellence of how algorithmic pricing can move from a dynamic model to a surveillance one, as each customer becomes a market of one.
The rideshare industry also has many of the conditions that tilt algorithmic pricing towards monopolistic and unfair practices. Strong network effects mean the market is naturally divided among only a handful of competitors, especially outside city centers. Those few competitors have similar “weak” segments of the market, meaning they charge discounts and surges to the same types of customers as their rivals. Rides cannot be resold, and customers have limited knowledge of what similar customers are charged unless, like my friends, they are huddled together. As a result, platforms require lots of insight about riders to offer their service, while riders have little ability to judge its value or fairness.
This can turn mutually beneficial economic transactions into one-sided ones with little consumer surplus. Unsurprisingly, consumer protection watchdogs and policymakers have accused rideshare apps of using phone battery percentage or the number of competitors’ apps on your phone to approximate desperation. Although Uber and Lyft deny allegations of using data this way, both acknowledge opaque company assessments of “trip purpose” affect prices.
The industry illuminates how greater pricing transparency affects key features of modern digital markets. While technological innovations can provide meaningful checks on murky pricing mechanisms, some may protect consumers’ anonymity at the expense of their welfare.
Zero Knowledge Proof-Based Audits
Traditional audits that require snapshots of company data are fruitless when government agencies cannot verify the data and models continuously update. Demanding outright access to evaluate the model won’t work either: without any firm malfeasance, models may learn to maximize profit by feeding regulatory bodies an unobjectionable version of the algorithm while directing actual customers toward a nefarious black box version. But audits are not a dead end entirely. Zero knowledge proofs (ZKP) are increasingly capable of verifying machine learning, and companies will probably soon routinely require them in business transactions. This makes them a solid foundation for accountability with little regulatory lift.
ZKPs confirm a model’s inputs without a company revealing private data or model weights. A company generates a mathematical proof that its algorithm contains only specified factors. Verifying that proof requires little computation and results in mere milliseconds of latency. Instead of algorithms re-routing audit spot checks, ZKP can operate on customer devices to verify every price is the result of an algorithm that excludes predatory considerations. The same ZKP circuit architecture can verify a model even as it continuously adjusts the weights of millions of parameters.
Pricing models can still play tricks by, for example, compressing a number of predatory considerations into a single “desperation vector” before feeding it to the model. The zero knowledge proof can only verify that the vector was a factor, not that the vector’s components essentially measure customer vulnerability. And algorithms will naturally look for proxy variables that capture a desperate customer’s high willingness to pay. Outlaw low phone battery as a consideration, and the algorithm will approximate the same through screen brightness, scroll speed, and degraded signal strength. If regulators ban any factor that has plausibly nothing to do with supply and demand fluctuations, the algorithm can reverse engineer results similar to those that use banned parameters by tweaking the weights of acceptable ones.
Still, well-architected ZKP verification is a useful check on an algorithm’s ability to minimize consumer welfare. When the law enumerates the most socially unacceptable factors and their most direct, market-unrelated proxies, and users can mathematically verify models do not use them, it hems in algorithms and re-establishes trust with customers. When algorithms rely on coarsely correlated factors or tweak weights, this steers the market back to segmentation rather than personal targeting. Desperation vectors still pose a serious threat, but courts may be able to sniff them out with clever data analysis.
The Limits Of Privacy Preservation
Other technological bounds can augment ZKP to protect users from being individually profiled, but even the most privacy-conscious regulators would do well to refrain from single-mindedly pursuing customer anonymity. Privacy guarantees sometimes inhibit competition and can theoretically facilitate algorithms colluding to set supracompetitive prices without intent or evidence. In game theory, colluding rivals have an incentive to cheat because each can temporarily win more customers by undercutting a competitor’s prices. But when algorithms can instantaneously match a rival’s lower prices, the lack of lag eliminates the incentive to undercut. Instead of a race toward lower prices, algorithms can tend toward testing consumer tolerance for ever-higher prices. Secrecy is the only way out of the upward spiral. The less transparent the pricing, the more unstable the collusion, as rivals regain the incentive to surreptitiously undercut one another. Eliminate secrecy for the sake of transparency, and the regulatory scheme can end up facilitating tacit collusion.
Here, again, rideshare provides a window into how the dynamic plays out. For years, Uber has implemented and innovated a statistical method known as differential privacy. This is a way of inserting the minimal amount of statistical noise necessary such that a random person’s information can be removed from the dataset without affecting analytic conclusions. Differential privacy prevents Uber from charging a loyal customer’s exact willingness to pay based on his rider history, but it also prevents Lyft from poaching that customer with a discount seen only on his screen. Private history on the app is the only price consideration that rivals cannot see and the only way a platform can offer a lower price without competitors knowing. Eliminate this, and the overlap of competitors’ “weak” market segments grows, as does the incentive to collude.
Mandating differential privacy also raises the already high barrier to entry for newcomers. With large data pools, a firm needs minimal statistical noise to remove a single person without consequence to the dataset. Uber’s dataset remains extremely useful while that of a newcomer held to the same privacy standards is minimal.
The Only Way Out Is Through
No solution will perfectly preserve the market expansion algorithmic pricing enables while completely preventing customers from being personally targeted. Algorithmic pricing may complicate antitrust and consumer protection to the point where we all long for the days of taxi cabs and being stranded at the airport. But for better or worse, algorithmic pricing is here to stay: attempts to eliminate price discrimination outright can be as arbitrary and distortionary as the practice itself. That doesn’t mean we need to resign ourselves to nebulous prices dictated by unaccountable algorithms. The journey to ensure fair pricing may be long, but at least we can hitch a ride.