Artificial Intelligence And Big Data: Good For Innovation?
Source – forbes.com
Artificial intelligence is firmly embedded throughout the economy. Financial services firms use it to provide investment advice to customers, automakers are using it in vehicle autopilot systems, technology companies are using it to create virtual assistants like Alexa and Siri, and retailers are using artificial intelligence (AI) together with customers’ prior sales histories, to predict potential purchases in the future, to name but a few examples. The potential of AI to boost economic growth has been discussed in numerous forums, including by Accenture, the Council on Foreign Relations, the McKinsey Global Institute, the World Economic Forum, and President Obama’s Council of Economic Advisers, among others.
The most dramatic advances in AI are coming from a data-intensive technique known as machine learning. Machine learning requires lots of data to create, test and “train” the AI. Thus, as AI is becoming more important to the economy, so too is data. The Economist highlighted the important role of data in a recent cover story in which it stated “the world’s most valuable resource is no longer oil, but data.” In this sense, both the ability to obtain data about customers, together with the ability to program AI to analyze the data, have become important tools businesses use to compete against each other, and against potential entrants.
A potential entrant that lacks access to good data faces substantial hurdles, and this has led some regulators to question the extent to which control over data creates barriers to entry. For example, in December 2015 FTC Commissioner Terrell McSweeney asked: “Can one company controlling vast amounts of data possess a kind of market power that creates a barrier to entry?” This is a worry, because if barriers to entry are too high, entrants will not enter, established firms will not feel competitive pressures, and innovation may suffer. Thus, in March 2017 CFPB Director Richard Cordray noted: “We recognize that data access makes it possible to realize the many benefits of competition and innovation.”
The hurdles faced by entrants may have implications beyond competition and innovation. A common technique that entrants currently use to overcome the lack of customer data is to train their AI on publicly available datasets. But, if these datasets are biased in some way, then the resulting AI will reflect the bias. The worry is that if many entrants use similarly biased datasets, then bias quickly propagates, as argued by Amanda Levendowski.
Access To Data Helps Firms Move Down A Learning Curve
An established firm’s access to data may allow it to take advantage of a learning curve, which may exacerbate barriers to entry for other firms (Michael Spence’s 1981 article in Journal of Political Economy is a classic on this topic). An established firm’s access to data allows it to refine its AI over time, allowing it to get better at offering its AI-enabled product over time. Due to this learning-curve effect, there are increasing returns to scale in data control. As Daron Acemoglu and Simon Johnson put it, “the real power—for good and ill—is in software and increasing returns to data. If one self-driving car company does well initially, it will be able to collect more data—and then further improve its algorithm. Other companies will not be able to catch up.”
The learning-curve effect is not just about learning how to use your data better, it is about how to better organize your business around the new technology. Take the case of Netflix and Blockbuster in the early 2000s. At the time, Blockbuster was the dominant firm in the video rental business and Netflix was the technology-oriented niche player. The two firms had very different business models: Blockbuster targeted impulse renters; Netflix targeted technologically sophisticated movie buffs via DVDs by mail. Netflix relied on sophisticated logistics software to track its rentals and manage its DVD inventory. Blockbuster eventually realized the threat posed by Netflix and attempted to replicate Netflix’s DVD by mail business, but was unable to do so successfully. By that time, Netflix had benefited from shipping and tracking millions of DVDs and had moved so far down the learning curve, and refined its logistics so well, that Blockbuster was unable to replicate it effectively.
Difficulty accessing data can create a barrier to entry, which has two implications. First, consumers don’t benefit from the innovative new products that a potential entrant might offer. Second, established firms, knowing that there is little chance of entry from a potential rival, may not see the need to innovate and improve on their existing products. Without the threat of entry, then, there may be fewer incentives to innovate.
Is Innovation Suffering? The Evidence Is Mixed
Despite the potential creation of entry barriers, there seems to have been lots of entry by innovative firms, at least in the recent past. Venture capital investment in AI startups grew by 40% between 2013-2016, according to McKinsey Global Institute. Anecdotally, companies such as Facebook, Google, Amazon and Apple have bought up hundreds of innovative startups over the past decade. A recent article in the New York Times suggested these acquisitions were evidence of anticompetitive conduct. But it is not that simple; we do not have the counterfactual of how many of these startups would have been created but for the opportunity of being bought out by a larger incumbent. Put differently, the very fact that firms like Facebook, Google, Amazon and Apple buy up startups may itself cause more firms to startup in the first place.
Likewise, there seems to be lots of investment by established firms. A recent report by McKinsey Global Institute estimates that established firms spent $2 to $3 billion on AI-related acquisitions in 2016, and between $18 and $27 billion on internal corporate investment in AI-related projects in 2016. Thus, it does not seem like entry and innovation have suffered in the recent past, but there are still a couple reasons to worry about the future.
We have been enjoying a long economic expansion (the third longest in U.S. history). Perhaps there has been entry by so many startups because, thanks to the expansion, there is lots of access to venture capital. When and if venture capital starts to dry up—for example if there is a recession—barriers to entry could become more important. And there are some signs that venture capital is tailing off; seed-stage financing is apparently down 40% from a peak in mid-2015, and some of the decline may be due to fear of established firms, particularly large, platform-oriented technology firms.
Academic research has highlighted, theoretically and empirically, that economic incentives push large platform firms to extract more and more value from the firms that rely on them. Theoretical work in this area includes papers by Joseph Farrell and Michael Katz, and more recently by Geoffrey Parker and Marshall Van Alstyne. Recent empirical work by Wen Wen and Feng Zhu document that when a platform starts to appropriate features of its developers’ applications, the application developers cut back on innovations to those applications. If at some point there is evidence that AI and big data has created an unfair advantage, and innovation is suffering, then what are the remedies?
Data Portability As A Potential Policy Solution
If customer data is deemed to be a critical resource, some form of customer data portability may be a solution. Under such a model, a customer would maintain possession of some core data that he or she could then take from one company to a rival, much in the way that a phone customer can take his or her phone number from one provider to another (which was not always the case). In principle, this should help reduce barriers to entry, because any potential customer of a new entrant could easily shift her data from the established firm to the entrant. Moreover, the ability of the customer to do this creates an incentive for the established firm to innovate and improve upon its existing services for the customer.
In practice, the effectiveness of customer data portability will likely vary across sectors of the economy. The value a consumer gets from a social media platform derives from a network effect—the connection to all the others on the platform. Even if a customer could “port” her own data to a rival social media platform, she would not be able to port her friends’ data, and so customer data portability would have little effect on competition and innovation in this sector. In contrast, customer data portability would have a larger effect on competition and innovation in other sectors such as online shopping or financial services, where a customer could port her prior purchase or transaction history.
But how would this remedy affect the established firm, which initially attracted the customer? One argument is that the originator would have less incentive to invest in the relationship between it and the customer, because ultimately rivals may benefit from that investment, not the originator. But a different argument is that the company may have more incentive to invest in the relationship between it and the customer, so as to create as much value from the relationship so as to not lose the customer to a rival.
Critics of the idea also highlight the important issues of privacy and security. For example, if data can be easily moved from one company to another, might some of the data “leak” to bad actors? Yet, these are issues that companies increasingly need to address anyway. Starting in May 2018, U.S. companies doing business in the EU will need to comply with a comprehensive suite of privacy rules spelled out under the EU’s General Data Protection Regulation (GDPR). Moreover, consumers do not seem to value privacy very highly. A recent NBER study by Susan Athey, Christian Catalini and Catherine Tucker found that most MIT undergraduate students would be willing to trade personal information in exchange for a free pizza. Ultimately, there may be technological solutions that help maintain customer privacy while allowing for easy customer data portability, as outlined by Ryan Calo in a recent AI policy paper.
In the meantime, policy makers will need to weigh potential risks to consumers from privacy breaches against the potential benefits to consumers from increased competition, and ultimately increased innovation, that may result from customer data portability. Indeed, a year ago the Obama Administration conducted a Request for Information on data portability, and earlier this year the European Commission conducted hearings on customer data portability.
In summary, AI has the potential to provide many benefits to our economy and society. However, startups and established firms that are just beginning to use AI need access to data in order to train their AI systems. Difficulty in accessing the necessary data can create a barrier to entry, potentially reducing competition and innovation. Data portability is not a panacea, but provides many benefits and will likely be part of a suite of solutions ultimately embraced by regulators.