Pages

Thursday, April 07, 2022

A Systems Approach to Cheap Speech: Flash Trades, Engagement Levers, and Destabilization Attacks

For the Balkinization symposium on Richard L. Hasen, Cheap Speech: How Disinformation Poisons Our Politics-and How to Cure It (Yale University Press, 2022).

Julie E. Cohen

Rick Hasen’s timely and important book links disinformation-based strategies for election manipulation to the platform-based, massively intermediated information infrastructures that enable them. This essential contribution comes at a time when policymakers are, finally, paying systematic attention to platforms as sources of democratic vulnerability. They are not, however, paying attention in quite the right way, and Hasen’s exposition suggests some important policy shifts. In particular, as Hasen recognizes, strategies that regulate audience targeting, drawn from the privacy governance toolkit, can (and should) supplement the traditional election law toolkit.

Mechanisms for audience targeting are not the only platform feature of concern, however. Platform-based, massively-intermediated information systems are continually, iteratively optimized to amplify content based on its ability to drive user engagement—and, therefore, to privilege outrage and volatility over deliberation, reasoned contestation, and truth production. Although these systems were not designed for the principal purpose of undermining democratic governance, their affordances invite and amplify disinformation-based destabilization attacks to which democratic political systems are particularly ill-equipped to respond. Systems thinking about disinformation-based strategies for election manipulation requires attention not only to tools for audience design but also to the affordances that amplify destabilization attacks.

Systems for “Content Moderation”

For many, debates about content governance within platform-based speech environments are first and foremost about “content moderation”—an activity revolving around post hoc review of content posted by users. The content moderation frame encourages systems thinking of a sort, but its insistent focus on post hoc intervention systematically forecloses attention to vitally important pieces of the disinformation puzzle.

Broadly speaking, content moderation involves flagging objectionable content for review and possible removal. In operation, content moderation systems involve considerable variation and complexity. The content in question may be flagged by users or identified by automated means; it may be flagged after posting or identified and quarantined mid-upload; it may be removed or subjected to a lesser sanction (such as downranking, shadowbanning, or demonetization); and the user who attempted to post it may or may not be given notice and/or the opportunity to seek review by a human moderator. In the case of Facebook/Meta, a very few high-profile cases are taken up by the Facebook Oversight Board, which may make a recommendation to Facebook on how to treat similar items.

One important shortcoming of the content moderation frame is that it simply doesn’t fit the problem. As evelyn douek explains, its post hoc, atomistic orientation suggests a comparison to courts—and, from that standpoint, today’s privatized and largely opaque processes fall far short of those that users might have a right to expect. Reforms designed to make content moderation processes more recognizably quasi-judicial also are doomed to fail, however, because of the sheer scale of content moderation operations and the automated tools they require. For douek, those conclusions point toward designing administrative content removal processes capable of enacting regularized systemic interventions.

The more important objection to the content moderation frame, however, involves what lies outside it. Systems for post hoc content moderation that leave underlying ex ante mechanisms for content immoderation undisturbed don’t and can’t significantly diminish flows of disinformation because they ignore the features of the platform environment that optimize it for disinformation to begin with. An important strength of Cheap Speech, which sets it above many other discussions of content moderation’s defects, is its willingness to confront the content immoderation problem.

Systems for Audience Targeting

As Hasen recognizes, a more serious discussion about content governance requires consideration of mechanisms for content provision as well as those for content takedown—and that requires attention to the ad-based digital business models on which (many) platform-based information systems rely. Ad-based business models in turn rely heavily on mechanisms for audience targeting and microtargeting, so regulatory attention to those mechanisms seems only  logical. But proposals to regulate audience targeting mechanisms, which blend elements from election law and privacy law frames, also leave a vitally important piece of the disinformation puzzle unacknowledged and unaddressed.

Broadly speaking, tools for audience targeting offered within platform-based, massively-intermediated information systems allow third-party advertisers to specify the types of audiences they want to reach. Here too, there is considerable room for variation. Large national and regional businesses seeking to position their offerings for maximum appeal use filters designed to match particular demographic groups with ads designed for them. Local businesses rely heavily on geographic filters and may also use other demographic filters to the extent feasible and permissible—so for example, sporting equipment stores can target suburban parents and nail salons can target women, but landlords cannot target based on race, ethnicity, or religion. Political campaigns and interest groups attempt to target candidate appeals and issue ads for maximum uptake by favorably predisposed voters. Digital adtech companies, for their part, allow advertisers to select demographic parameters and/or to target ads to various predefined groups. More sophisticated advertisers, including some political campaigns, can use information they have collected using various methods—website registration, bespoke apps, requests for event tickets, and so on—to specify target audience parameters more precisely.

Crucially, all of this activity involves not only conventional strategies for demographic segmentation but also tools for behavioral profiling designed by both platforms and the digital marketing consultants who use their services. So, for example, a cosmetics company might target ads for a hair loss remedy to men aged 30 to 50 who buy high-fashion brands—or it might microtarget based on other behavioral data, such as time spent using dating apps or browsing tips on how to minimize the appearance of aging. A political campaign might use its database of voters who lean conservative on education policy to target ads to demographically similar audiences—or it might microtarget appeals using particularly hot-button language to those whose browsing behavior reveals particularly high engagement with content advocating strong parental control of education. Actors wishing to conduct disinformation-based destabilization attacks can rely on capabilities for behavioral profiling to microtarget audiences whose emotional buttons they think they can push. For example, they can microtarget ads for YouTube videos opposing vaccination to those whose browsing behavior reflects engagement with conspiracy-themed content more generally.

The theoretical linkages between capabilities for microtargeting and disinformation uptake seem straightforward enough. Political scientists who study voting mechanisms have shown that single-party legislative districts (whether created deliberately or via self-sorting) tend to entrench differences of political opinion. And there is some evidence that ranked-choice voting mechanisms may increase civility and reduce polarization. By analogy, it makes sense to think that legislation banning or limiting targeting and microtargeting might make it harder for disinformation campaigns to flourish because it would effectively require lumping together audiences of different political persuasions.

In reality, however, proposals to restrict or ban political targeting and microtargeting are unlikely to counter disinformation-based destabilization attacks effectively, for two principal reasons. First, the election law and privacy law frames on which such proposals rely tend to exclude from coverage the very types of communications and the very types of targeting that represent disinformation operators’ stock in trade. Begin with election law. As Hasen recognizes, many communications that would not qualify as covered political advertisements under current law, because they do not advocate for or against a specific candidate, are nonetheless crafted for politically polarizing effect. By tweaking definitions, one might expand current coverage to include certain types of issue ads pertaining to matters or candidates currently on the ballot, but the communications used in destabilization attacks often are not so easy to characterize. Election laws also typically balance speech and anti-corruption values by excluding small-dollar expenditures from reporting and disclosure requirements. Disinformation campaigns, however, do no need to spend large sums to produce large impacts via user-driven uptake and social circulation, and some such campaigns rely on posts that are not ads at all.

Privacy governance, for its part, tends to be conceptualized through the lens of privacy self-management, or privacy as control over one’s own data. (I’ve discussed the inadequacies of that paradigm in more detail elsewhere.)  Privacy regulations drafted to facilitate privacy as control allow consensual targeting, and that makes them particularly ill-suited to combating disinformation-based attacks. Today’s digital political campaigns are chiefly crafted to exploit the consent of the willing, beginning with contact information supplied by interested voters and relying on processes of social circulation to spread their messages more widely. When a political campaign organized around a candidate or a ballot issue chooses to spread disinformation, its message can spread readily among those willing to be targeted. Privacy statutes crafted around notions of privacy as control also tend to exclude practices of so-called contextual advertising, in which advertisers bid to target their ads next to particular types of content rather than targeting consumers directly, but contextual advertising is an important part of the disinformation playbook. Contextual advertising exceptions in proposed laws to restrict or ban political targeting and microtargeting create generous loopholes for disinformation-based destabilization attacks to persist and thrive.

The more significant problem with attempting to fight disinformation by regulating tools for audience targeting and microtargeting, however, is that such efforts ignore other, less visible aspects of the platform business model that also play a major role in driving disinformation’s spread and uptake.

Systems for Content Amplification

Tackling content governance within platform-based, massively-intermediated information environments requires consideration of all of the ex ante mechanisms for content immoderation that platforms employ, including not only tools for audience targeting and microtargeting but also tools for content amplification. Selective, strategic amplification for user engagement underwrites all parts of the platform business model, including both algorithm design and ad pricing.

Consider, once again, the platform advertising dashboard. As we’ve seen, advertisers can use the dashboard to communicate their wishes, selecting demographic parameters for their target audiences or supplying more detailed and data-driven behavioral profiles. But the dashboard is a tool for two-way communication. Through it, platform operators conduct automated, real-time auctions that perform two sets of simultaneous functions. They pit would-be advertisers against one another to secure desired placements, and at the same time they train the machine learning processes underlying the auctions to reward more effective ads—specifically, ads that produce greater user engagement and social circulation—with better placements. Functionally, then, the dashboard is also an engine for flash trading in economies of user attention and engagement that platforms themselves work to produce, and it rewards disinformation-based destabilization attacks for their efficacy at generating engagement.

As that description is intended to suggest, moreover, the advertising dashboard and the machine learning engine behind it represent only parts of a larger whole that is oriented first and foremost toward keeping eyeballs on the platform. And, crucially, mechanisms for controlling audience design via targeting and microtargeting are not the only platform features that work to circulate content to users. Equally as important, though far less visible to the external eye, are the internal engagement levers that amplify certain types of content. Platforms continually optimize and reoptimize for user engagement, routing, suggesting, and upranking items that, based on past data, are likeliest to generate interaction and social recirculation. Content that generates outrage, including especially outrage-generating content that plays to partisan extremes, does well on those metrics.

Whether this business model “works” in the conventional sense—i.e., whether it produces the sorts of conversion ratios that commercial advertisers care about—is beside the point. Commercial advertisers understand one thing very well: The dominant platforms, particularly those operated by Facebook/Meta and Google, are where the eyeballs are. Adversarial attackers, meanwhile, do not care about efficacy in the way commercial advertisers (maybe) do. The question is not whether every attack works but whether some (enough) can leverage platform-provided engagement levers to achieve maximum uptake. For any particular attack, the measure of success is its decontextualized and apparently uncontrolled circulation—from users to other users and user groups, from the originating platform to other platforms, and in a few lucky cases, to coverage by mainstream broadcast and print media. And because the platform business model mandates optimization for engagement, platforms have little incentive to institute more global measures designed to undercut destabilization attacks.

It is clear, however, that platforms have the tools to dampen viral circulation, should they choose or be required to do so. Processes of social circulation are sometimes described as “organic”—and, to be fair, they are designed to exploit irrational tendencies of human social groups—but there is nothing natural about them. Technical and organizational processes can be reeingineered. Internal processes now trained singlemindedly on optimization for engagement could be retrained to respond differently to rapid spread. Practices like Facebook’s/Meta’s Xcheck program, which gives certain high-profile accounts leeway to violate the company’s own content policies based on the engagement they generate, could and should be discontinued. Flash trading dashboards could be constrained to offer transparent, fair, and equal pricing. Disclosures could reach beyond “transparency theater” to operational reality, shedding meaningful light on how both internal engagement levers and outward-facing flash trading dashboards work. These are solvable problems.

Putting All the Systems Together

A signal virtue of Cheap Speech is its willingness to reach beyond the conventional tools of election law to address election-related problems. To respond effectively to platform-facilitated destabilization attacks, however, it is necessary to acknowledge all of the mechanisms in play. Disinformation-based destabilization attacks thrive within platform-based, massively-intermediated information environments constructed and iteratively fine-tuned both to enable audience targeting and to amplify the content that generates the most engagement. An effective response to the pathologies of cheap speech must address both systems. 

Julie E. Cohen is the Mark Claster Mamolen Professor of Law and Technology at Georgetown Law. You can reach her by e-mail at jec@law.georgetown.edu.