Information, increasingly, is everywhere. Machines gather information, process it, and automatically communicate it, often in terms humans understand. Bots tweet on Twitter; Fitbits communicate a user’s activity record; Project Tango devices render 3D maps; and IBM’s Watson can now argue. With algorithms increasingly writing, drawing, and even debating, a central question for regulators, courts, and scholars is to what extent the First Amendment protects speech generated by algorithms. If algorithmic communication falls within First Amendment coverage, regulators will have a more difficult time governing it. But if it does not, courts will need to explain how the exclusion can sit comfortably with First Amendment theory and current doctrine.
Stuart Minor Benjamin positions the puzzle of algorithmic speech as part of a larger project in understanding First Amendment jurisprudence and its expansion and contraction. In previous work, Benjamin has asked how hard it would be to expand First Amendment coverage; in Algorithms and Speech, he asks how hard it would be to narrow the existing jurisprudence to exclude a practice that would otherwise be covered. Benjamin recognizes the potential regulatory consequences of First Amendment coverage of algorithmic speech. But he surveys Supreme Court caselaw and concludes that there is no principled way to exclude many algorithmic communications from speech protection without excluding much other communication that we deem squarely within the First Amendment’s coverage.
Algorithms and Speech does valuable work in laying out the current state of expansive First Amendment doctrine, and in identifying the Supreme Court’s reluctance to create new exceptions. Benjamin also clarifies that the coverage of algorithmic speech is not just a matter of making analogies to earlier media. Search engine results may be like editorial decisions, the claim of Eugene Volokh and Donald Falk’s 2012 white paper, but Benjamin is intent on finding an underlying reason why both are covered that goes beyond structural similarities.
The touchstone of First Amendment coverage, according to Benjamin, is the sending of a substantive message. Benjamin points out that when a sendable and receivable message has actually been sent, the Supreme Court has never found that message to be outside First Amendment coverage (a point that is historically untrue, but correct within current jurisprudence). Excluding algorithmically generated speech would upend existing First Amendment caselaw, requiring either the drawing of arbitrary lines or the exclusion of much of what is currently considered to be speech. Benjamin explains that the arguments against covering algorithmic speech, such as distinguishing it as corporate speech or commercial speech, would also leave core First Amendment institutions such as newspapers unprotected. What is most important to Benjamin is consistency, and the article is admirable in trying to craft rules that apply equally to all.
Benjamin carves out several important limitations. First, an algorithm that does not communicate a substantive message will not be protected. Second, because Benjamin hinges First Amendment protection on the communication of a message (but interestingly, not its receipt), some companies may have to indicate that they are editors. Third, as is the case with newspapers, laws of general applicability such as labor laws, tax laws, and most antitrust laws can apply to algorithmic speakers with no First Amendment ramifications. The government just can’t ban or compel substantive communication.
Benjamin makes a convincing case for the protection of search engine results under current First Amendment doctrine. The recent SDNY decision in Zhang v. Baidu, where a district court judge found First Amendment protection for Baidu’s search results, shows that judges are likely to agree. The article is also painstakingly honest in trying to maintain the cohesiveness of the Court’s First Amendment reasoning. But by positioning the question of algorithmic speech within current jurisprudence and around the model of search engine results, Benjamin limits the scope of the article in several ways. Algorithms and Speech does important work and will likely be a foundation that others will build on, but it leaves several more difficult questions for another day.
Benjamin steers away from reasoning from First Amendment theory. He instead navigates the “guideposts” of Supreme Court jurisprudence, accepting Cass Sunstein’s assertion that the First Amendment in practice is incompletely theorized. This leads to an unstated bias towards the jurisprudential status quo. It’s not clear precisely why preserving the guideposts of current jurisprudence is the right approach. The primary explanation offered is that disturbing the status quo of jurisprudence will threaten other media that more clearly rest at the heart of First Amendment interests.
We may soon be at a stage, however, where upending existing First Amendment caselaw is to some extent inevitable. If algorithmic speech does get full First Amendment protection, how will the intent requirements in many of the categories of unprotected speech get implemented? And if algorithmic speech does not get protection, how will we distinguish that content from human speech without threatening many of the values underlying the First Amendment? Algorithms and Speech takes a fascinating first step, while in its caselaw-driven approach leaving a number of important questions on the near horizon.
The second way in which the article is less daring that it could be stems from the model Benjamin chooses for algorithmic speech. The search engine model—of an algorithm running according to its programmers’ general intent—runs the article into limitations right where the questions get most interesting. Benjamin’s touchstone for First Amendment coverage, based on the Spence test, is the intentional communication of a substantive message by a human being. When the algorithm is no longer a tool for its user, but an artificial intelligence, Benjamin suggests the connection to a human speaker might become sufficiently attenuated that First Amendment coverage might no longer be appropriate. The problem is, as Bruce Boyden has pointed out, that the line between tool and independent message generator is exceedingly difficult to draw.
The flurry of scholarship around algorithmic speech shows the variety of ways a First Amendment problem can be framed: with a focus on the speaker, the message, the medium, or the listener. Benjamin’s focus is in large part on an intentional speaker. The recipient of a message matters to Benjamin to the extent that the recipient can identify the speaker as communicating a message, but the intentional speaker is for Benjamin the core of what turns information into First-Amendment-relevant speech.
Robert Post has taken a different approach to First Amendment coverage, explaining that First Amendment protection extends to a medium of communication because of its status as a social practice (both Tim Wu and Andrew Tutt have applied this to search engines). The question is whether certain kinds of algorithmic speech are sufficiently like protected media or have acquired enough of their own cultural meaning to deserve protection. This is not a question courts usually want to explicitly ponder.
A third way to frame First Amendment interests is to talk about the message itself, and whether its contents reflect high or low value speech. But as Benjamin points out, the Supreme Court has repeatedly rejected this message-oriented approach, at least with respect to speech made by an intentional speaker.
A fourth way to frame First Amendment interests is to look at the broader communications environment, including the listeners, readers, and receivers of communication. And this approach highlights a problem with both Benjamin’s and Boyden’s potential exclusion of algorithmic speech once a human is no longer involved in the selection of a message. Even an autonomous Tolstoy Bot would still be creating works that appear to readers to be identical to speech by humans. At that future phase of algorithmic creativity, censorship of Tolstoy Bot would affect readers the same way as censorship of Tolstoy: they would have access to less information, and would perceive the government’s actions as censorship. Under multiple theories of the First Amendment, this kind of censorship would raise problems—even if there is no human meaningfully involved in the crafting of the substantive message listeners could receive.
The article’s focus on a speaker’s agency as the touchstone of First Amendment protection thus may in practice prove to be both too inclusive—including actions as speech merely because a speaker claims to have a substantive message—and too underinclusive—rejecting speech that clearly looks like speech to a reader. The speaker’s agency approach may be most consistent with current jurisprudence, but the more difficult question for a future article is whether it is the right approach to future technologies, and why.