Machine learning is becoming ever more common. Artificial intelligence programs like IBM's Watson can gobble up huge corpuses of knowledge, attempting to recreate human cognitive processes. These self-teaching machines can learn how to create novel recipes, master Jeopardy, or answer legal questions.
The technology poses major questions, not just for lawyers but for the law itself. Take, for example, automatic sales pricing algorithms, the type that are already common in online trading and travel booking. A machine learning program, applied to such algorithms, could possibly learn the benefits of price fixing, according to a recent scholarly article from the Oxford Centre for Competition Law and Policy, raising significant antitrust concerns.
Will Tech "Disrupt" Antitrust Enforcement?
Machine learning, or "cognitive computing" systems, attempt to recreate the human learning process in electronic form. Machine learning applies the process of observing, evaluating and deciding on an iterative basis, allowing programs to become more "intelligent" over time. Already, rudimentary machine learning is common in eDiscovery programs, which learn to detect responsive documents based on user inputs. As the technology increases, so too do its possible applications. For example, one legal start-up, ROSS Intelligence, is looking to apply machine intelligence to answering legal questions.
The fast pace of technological development, particularly around self-learning machines, raises many legal concerns, particularly around antitrust enforcement. Legal scholars Ariel Ezrachi, of Oxford, and Maurice Stucke, of the University of Tennessee, recently surveyed some of these concerns. The pair surveys the many ways artificial intelligence may collude to distort markets.
The Autonomous, Collusive Machine
An industry's automatic pricing algorithms don't need to get together in a back room for collusion to occur. Instead, to several similar computer programs could "promote a stable market environment in which they predict each other's reaction and dominant strategy," reducing competition and consumer choice, according to Ezrachi and Stucke.
The pair identified four major, non-exclusive categories of potential machine learning based collusion. Most of these are simply computerized versions of existing anticompetitive practices. For example, algorithms can be designed with collusion in mind (category I), a single algorithm could set prices for multiple actors (category II), or programs could learn to predict anothers' behavior, creating tacit collusion (category III).
The most legally challenging is the "Autonomous Machine," however. This fourth category of collusion occurs when "machines, through self-learning and experiment, determine independently the means to optimize profit." Essentially, machine intelligence can learn that collusion is its optimal strategy. Traditional concepts of intent and agreement don't apply and legal liability is unclear. Here, "the lack of evidence of any anticompetitive agreement or intent," the pair write, "may result in AI self learning escaping legal scrutiny."
That level of machine intelligence based collusion remains largely theoretical at this point, however. That leaves legislators and enforcement agencies plenty of time to address such risks. In the meantime, the Department of Justice has already indicated that it's willing to pursue tech-based antitrust. In April, the DOJ announced its first e-commerce antitrust action for price collusion, pursing online sellers on Amazon's marketplace who used shared, collusive algorithms to set prices.