By Matthew J. Donigian*
The foreperson is considered an essential component of the American jury. After being selected foreperson, a juror is delegated the responsibility of harnessing the fact-finding power of the jury to reach an efficient and just verdict. Although jurisdictions employ different methods when selecting a foreperson, the foreperson almost always functions as the jury’s leader and exercises her power by organizing discussion as she sees fit. In recent years various jurisdictional foreperson selection processes have been scrutinized by empirical studies that aim to discover if the foreperson’s role gives her unequal influence in a jury’s verdict. These studies indicate that the foreperson does possess considerable power and that certain methods of selecting a foreperson may be injecting inequity into the system.
Recently, amazing progress has been made in the fields of computing and artificial intelligence (AI), which may allow for an artificial intelligence to take over many of the administrative tasks that are currently taken care of by the foreperson. Since the advent of the first digital computer—ENIAC—in 1947, the ratio of cost to processing power in computers has increased exponentially. This reality has produced stunning advances in computing power over the last half century. In 1986, the Cray X-MP supercomputer was capable of .220 gigaflops of processing power at the cost of about $15 million. Compare this to the HPU4Science cluster, which has 6,000 gigaflops (6 teraflops) of processing power and cost about $30,000 to build in 2011 using commercially-available gaming computer hardware. In addition, groundbreaking advancements in machine audio-visual recognition capabilities, natural language processing, and general increases in the sophistication of artificial intelligence are making technological feats once relegated to science fiction novels not only possible, but affordable; two stunning examples of this technology are the Forum and Blog Threaded Content Analysis project that is capable of analyzing “large volumes [of] social media content” and providing summaries useful to intelligence operations and Google’s self-driving car.
Exponential increases in computing power will inevitably allow an artificial intelligence to perform the responsibilities of a jury foreperson. Handing over the responsibility of controlling the logistical affairs of the jury to an artificial intelligence may prevent the inequities created by unequal delegation of power to the jury foreperson.
II. Problems with the Human Foreperson
As expected, jury forepersons have more influence than other jurors during
deliberations. Forepersons also participate much more than other jurors, and “account for about 25%-31% of speaking during deliberations . . . .” In addition, the foreperson makes “twice as many novel statements of facts or opinions” when compared to other jurors.
Jury forepersons are selected differently depending on the jurisdiction. In most states the jurors themselves are free to choose the foreperson at the beginning of the trial, while in some states the judge selects the foreperson before deliberations begin. Depending on the selection method employed, different sources of bias may have an effect on which juror is ultimately selected. For example, when a judge selects a foreperson they may give preference to a juror “who nods at the right times and seems to interpret the case in the same way that judge does.” In states where jurors select their own foreperson there is evidence that education, occupation, and expertise influence the jurors’ choices.
Does the bias injected by the selection of a foreperson detract from the ideal the jury is meant to live up to? This question largely depends on what ideal we have in mind. In the not-so-distant past juries were almost exclusively composed of the wealthier elements of society, and the Supreme Court’s recently adopted ideal of the cross-sectional jury is in stark contrast to the composition of juries at the founding. In Thiel v. Southern Pacific Co., the Court explains why the cross-sectional jury upholds democratic ideals: “[j]ury competence is an individual rather than group or class matter. That fact lies at the very heart of the jury system. To disregard it is to open the door to class distinctions and discriminations which are abhorrent to the democratic ideals of trial by jury.”
III. Technology Required to Create an AI Foreperson
A foreperson requires three major abilities in order to perform her function. First, a foreperson must be able to hear jury deliberations. Second, a foreperson must be able to think about what is said by other jurors and process that information. Third, (ideally) a foreperson should be able to use this information to direct the jury’s discussion. If a computer is to play the role of foreperson then it must be able to perform all of these functions as well as a human.
A. Speech Recognition
Jury deliberations present several challenges that voice recognition software must overcome if an artificial intelligence is to take over the role of foreperson. The three most puzzling obstacles are the format of jury deliberations, which may often involve speakers talking over one another in varying volumes and voice types, the speed at which deliberations may take place, and the constant turnover of new jurors each time a trial takes place.
These are major challenges for modern speech recognition software because in order to increase accuracy, current software must be sensitive to a user’s voice while remaining extremely insensitive to outside noise. This makes the best speech recognition software on the market very good at recognizing the speech of a long-time user with a noise-cancelling microphone but inept at accurately recognizing speech in a multi-party conversation.
The most popular speech recognition software on the market today is Dragon NaturallySpeaking (Dragon). Although it is amazingly accurate, Dragon is inadequate for application in the jury room for two reasons: first, Dragon owes much of its accuracy to its ability to be “trained” by a user by analyzing the user’s voice when reading preset written materials; and second, Dragon recommends completing training exercises in a location with similar ambient noise to the location where a user will be dictating so that Dragon can better isolate the user’s speech from other noises.
Recent developments in speech recognition algorithms have led to the creation of speech recognition software that—while less accurately than trained software like Dragon—can analyze speech without being trained by a user and can be used in everyday life. The most popular of this new wave of software are Apple’s Siri and Google Voice Search. And while these technologies have much of the speech recognition capabilities that would be needed by an AI foreperson, they have been criticized as inaccurate, with one review claiming that Siri processes speech accurately in noisy conditions only eighty-three percent of the time. Since many queries that are processed by Google Voice Search and Siri are short and to the point, these numbers would likely be far lower when presented with the unique challenges of the jury room.
B. Critical “Thinking” and Natural Language Processing
In order for an artificial intelligence to perform the duties of foreperson, recognizing the speech of jurors is not enough. The artificial intelligence must also be able to accomplish the more complex task of understanding the context and nuances of natural language. Without this ability, the machine is simply replacing the court reporter instead of becoming an element of the deliberating body. While this capability is still outside the reach of current technology, there are several projects that show promise that this accomplishment will become feasible in the near future.
IBM stunned the world on February 14, 2011, when “Watson,” a supercomputer, beat Jeopardy! Champion Ken Jennings in the first and second rounds of a three-round tournament. Accomplishing this feat required Watson to hear and answer a broad array of questions, many of which included slang terms or cultural references. Watson was able to win at Jeopardy! largely because of IBM’s natural language processing technology that allowed the supercomputer to “understand” the context of the question being asked and provided a degree of representation and reasoning based on the processing of the information.
Natural language processing capability is essential to any artificial intelligence foreperson. A skilled foreperson should also be able to sense consensus and call for straw-poll votes when the debate has coalesced. Although Watson’s capabilities suggest that this technology will be available in the future, it is not yet adequate for this application. Watson’s current capabilities are focused on answering specific and structured questions like those in Jeopardy!, and although jStart (Watson’s commercialized version) has been used in more practical applications like processing patient discharge reports in plain-text and generating follow-up alerts for healthcare providers, it is not built to process the massive amount of data and ambiguity present in live-action juror deliberations.
Perhaps the most important part of an artificial intelligence is its ability to communicate useful information to humans. If all Watson did was “think” about Jeopardy! questions it may produce interesting academic data, but its astounding capabilities would be lost on much of the population. What makes Watson truly amazing is its ability to synthesize the previous two capabilities—speech recognition and natural language processing—and then render the correct answer. IBM allowed Watson to “think” by giving it a plethora of academic resources, the entirety of Wikipedia, and other reference materials.
However, having access to a world of knowledge is not enough. Watson also needed to be able to sift through this information to identify the knowledge that was pertinent to each question. IBM’s DeepQA technology gave Watson this ability by allowing Watson to consider the different ways that a question could be interpreted and then use probabilities and a complex search algorithm to hone in on the most likely answers.
Technology like Watson is currently inaccessible to state and federal courts due to the extremely large investment required for Watson to perform at the level it did during Jeopardy!. This investment is computing power, and Watson had an unbelievable amount at its disposal: 90 clustered IBM Power 750 servers (each retailing at over $85,000) churning out over 80 teraflops of computing power. Even if Watson-like technology was more cheaply available, it still would not be able to perform the functions of a competent jury foreperson.
IV. Future Feasibility as Evidenced by Exponential Increases in Computing Power
Although computer hardware and artificial intelligence software are currently incapable of performing the functions of the jury foreperson, this capability may not be as far off as many think. According to Ray Kurzweil, the Engineering Director of Google, computer speed is not only increasing at an exponential rate, but the rate at which computer speed is increasing is also increasing exponentially. At this rate, supercomputer power will increase to around 500 times faster than the world’s fastest supercomputer, the Titan CrayXK47, by the year 2020.
Kurzweil not only predicts exponential increases in computing power, but also exponential increases in computing price-performance. According to Kurzweil, this means that “later in this century . . . a thousand dollars of computation will be trillions of times more powerful than the human brain.” This amount of computing power would be more than capable of performing the functions required of a jury foreperson, and affordable computers capable of performing this feat will most likely be available far before the type of system mentioned by Kurzweil.
The constitution guarantees a criminal defendant the right to a trial by a jury of her peers. This right places limitations on the role that an artificial intelligence can and should play in the jury room. However, by relinquishing the leadership functions of the foreperson to an artificial intelligence, the jury can avoid any corruption associated with the foreperson’s selection and the power afforded her.
Eventually, artificial intelligence will be able to perform managerial and organizational tasks far better than any human foreperson. An artificial intelligence could provide jurors with an accurate breakdown of how much each juror has contributed, a catalog of those contributions, and could ensure that jurors regard jury instructions when delivering a verdict.
In the near future, increases in technology will allow for an artificial intelligence to enter the jury room as a powerful tool. In order to prepare for this reality, ethical concerns regarding the extent to which an artificial intelligence should participate in deliberations must be considered, and any implementation must carefully balance the defendant’s constitutional right to a trial by a jury of her peers and the interests of fairness and efficiency.
* J.D., University of Illinois College of Law. B.A., Political Science, University of Massachusetts at Amherst.
 The foreperson is usually elected at the beginning of deliberation and is charged with leading the jury’s discussion. The foreperson will often call for straw-poll votes to check the jury’s progress toward a verdict.
 CRAY X-MP/48: 1986–1990, SCD Supercomputer Gallery, http://www.cisl.ucar.edu/computers/gallery/cray/xmp.jsp (last visited May 8, 2013).
 Adam Stevensen, High Performance Computing on Gamer PCs, Part 1: Hardware, Ars Technica (Mar. 30 2011, 11:30 PM), http://arstechnica.com/science/2011/03/high-performance-computing-on-gamer-pcs-part-1-hardware/2/.
 Noah Shactman, Air Force’s Top Brain Wants a ‘Social Radar’ to ‘See Into Hearts and Minds,’ Wired (Jan. 19, 2012, 6:30 AM), http://www.wired.com/dangerroom/2012/01/social-radar-sees-minds/.
 See Ashlee Vance, Google’s Self-Driving Robot Cars Are Ruining My Commute, Bloomberg Businessweek (Mar. 28, 2013), http://www.businessweek.com/articles/2013-03-28/googles-self-driving-robot-cars-are-ruining-my-commute (explaining the effects of Google’s self-driving cars on driving).
 Traci Feller, What the Literature Tells us About the Jury Foreperson, 22 Jury Expert 42, 42 (2010), available at http://www.thejuryexpert.com/wp-content/uploads/FellerNov2010Vol22Num6.pdf.
 Thiel v. Southern Pacific Co., 328 U.S. 217, 220 (1946).
 See Learn Some Quick and Easy Tricks to Using Dragon NaturallySpeaking, Nuance, http://www.nuance.com/naturallyspeaking/customer-portal/tips-tricks.asp (last visited May 8, 2013) (explaining how to “train” Dragon to recognize your speech).
 Philip Elmer-DeWitt, Minneapolis Street Test: Google Gets a B+, Apple’s Siri Gets a D, CNNMoney (Jun. 29, 2012, 6:42 AM), http://tech.fortune.cnn.com/2012/06/29/minneapolis-street-test-google-gets-a-b-apples-siri-gets-a-d/.
 David R. Martin & Jim Fitzgerald, IBM’s Watson Beats ‘Jeopardy!’ Champs Ken Jennings and Brad Rutter in First Public Test, Mass Live (Jan. 13, 2011, 10:04 PM), http://www.masslive.com/news/index.ssf/2011/01/ibm_watson_beats_jeopardy_champs_ken_jennings_brad_rutter.html.
 UNC Healthcare: How Big Data Was Leveraged to Reduce Medicaid Re-Admissions, IBM Software: jStart Portfolio, http://www-01.ibm.com/software/ebusiness/jstart/portfolio/uncMedicaid.html (last visited May 8, 2013).
 Ray Kurzweil, How My Predictions Are Faring 135–36 (2011), available at http://www.kurzweilai.net/images/How-My-Predictions-Are-Faring.pdf.
 See Introducing Titan, http://www.olcf.ornl.gov/titan (last visited May 8, 2013) (describing Titan’s theoretical peak performance as 20 petaflops, or 2×1016 flops, compared to Kurzweil’s projection of around 1×1019 flops in 2020).
 Ray Kurzweil, Kurzweil Responds: Don’t Underestimate the Singularity, MIT Tech. Rev. (Oct. 19, 2011), http://www.technologyreview.com/view/425818/kurzweil-responds-dont-underestimate-the-singularity/.