“Old Crime, New Code” – DOJ Outlines its Views on When Software and AI Can Facilitate Collusion

Last week at a conference chaired by the head of Axinn’s West Coast antitrust practice, Antitrust Division Criminal Deputy Daniel Glad outlined the distinction between benign pricing software and “algorithmic collusion,” and reiterated that policing algorithmic collusion is a continued enforcement priority. He emphasized, however, that while algorithms represent a potentially new mechanism through which collusion can occur, the “mechanics of collusion” via algorithms “are more conventional than they look on first inspection” and are subject to the same legal analysis and evidentiary burdens as traditional cartels.

Glad argued that algorithmic conduct is not “beyond the reach of criminal antitrust enforcement” when prosecutors can prove beyond a reasonable doubt that competitors agreed to fix prices by knowingly substituting “shared non-public competitive information for the independent decision-making that the antitrust laws require of competitors,” whether “through architecture, through information sharing, or through follow-the-algorithm understandings.” And he emphasized that the ability of law enforcement to detect collusion “does not diminish,” but “grows,” as the use of software tools leaves behind an increasingly longer digital trail of logs, records, queries, artifacts, and timestamps. 

Glad noted that DOJ thinks critically about the line between “ordinary vertical software contracts” and “hub-and-spoke algorithmic arrangements” that amount to an agreement among competitors “to eliminate competition among themselves.” Whether DOJ can prove a rim, meaning an agreement among competitors, “is the element that determines whether [there] is a vertical arrangement subject to the rule of reason, or a horizontal conspiracy subject to the per se rule.” 

Glad’s Guidance

Glad pointed to several potential indications that – on “the right record” – could indicate “anticompetitive coordination” among competitors and thus open the door to criminal enforcement:

Proving an agreement. Glad first argued that when competitors “understood that their sensitive non-public data will be used to set prices for competitors” and participated in the exchange, revenue management platform, or large language model “on that understanding,” such evidence could support allegations that use of a common algorithm amounted to an agreement to fix prices. 

Applied specifically to large language models or any other artificial intelligence, he argued that “if your pricing system depends on your competitors’ confidential inputs to function, you should expect us to ask why that is not anticompetitive coordination.” 

Glad’s assessment can be broken down into several steps: 

• Adopt a common pricing algorithm: He began with the predicate that competitors adopt the same pricing model. 

• Competing users knowingly rely on each other’s non-public data: Beyond knowing and collective adoption, DOJ will assess users’ collective knowledge of the model’s inputs and outputs. Glad asked whether “each firm understood” – for instance, through “publicized terms of service” – “that its data would shape” model inputs or outputs for its competitors. In other words, the concern is not models whose runtime function (the model’s execution process that generates a pricing recommendation) uses only data from a single competitor. Nor, as is permitted in RealPage’s consent decree and noted in his remarks, is the concern about all forms of model training (the process of analyzing data to define and refine its rules) that combine competitor data: DOJ’s civil settlement with RealPage permits training its model using combined competitor data where the data are sufficiently aged, aggregated, and anonymized. In that case, RealPage can continue to train its model using combined competitor data, so long as it uses inactive lease data aged at least 12 months with a geographic variable no narrower than the state level.

• Users know the algorithm produces pricing recommendations competitors will collectively receive and rely on: When users know they and their competitors are using a model that combines competitors’ “confidential economic data,” that tees up the ultimate question of what the model did with competitors’ non-public data. Glad explained the potential for concern when there is evidence that the model produces recommended prices that competitors each “understood” their fellow competitors would not only receive, but “rely on” or “follow.”   

Proving intent. Beyond proof of an agreement among competitors, antitrust crimes include an intent requirement. There must be proof beyond a reasonable doubt that conspirators knowingly agreed to fix prices. Here, Glad argued that “intent travels with the human decision to contribute to and rely upon the system.” In other words, a compliance assessment cannot start and end with an analysis of the data in and data out but should account for how the model is marketed and how and why it is used. 

Compliance is King

For AI companies and users alike, Glad emphasized the importance of robust compliance and training efforts geared toward sales, pricing, and procurement teams that use artificial intelligence and algorithmic tools. 

That means educating executives that antitrust risks don’t only come from smoke-filled rooms but can result from use of commercially available “pricing software, revenue management platforms, and AI-driven decision tools.” 

• That begins by knowing which algorithmic tools your company uses as part of its competitive decision-making, whether for pricing or procurement, but also bidding or salary-setting. 

• From there, the tools’ inputs are important to understand. That means asking what data your company and other users provide the tool, and whether your confidential data trains the tool or feeds its outputs. 

• The next step is examining the tool’s outputs, with Glad urging that particular attention is warranted when competitors receive and follow strategic recommendations they know are fed by their competitors’ non-public data. 

"Software cannot launder collusion. When competitors exchange competitive intentions in a hotel suite or through a trade association, it is well settled that that raises antitrust concerns. So too with a text thread or a common algorithm." - Acting Deputy Assistant Attorney General Daniel Glad, Antitrust Division, U.S. Department of Justice

www.justice.gov/...