On the Investigation of Black Boxes
* Report: Nicholas Diakopoulos. Algorithmic Accountability: On the Investigation of Black Boxes. Knight Foundation and the Tow Center on Digital Journalism at Columbia Journalism School.
"The past three years have seen a small profusion of websites, perhaps as many as 80, spring up to capitalize on the high interest that mug shot photos generate online. Mug shots are public record, artifacts of an arrest, and these websites collect, organize, and optimize the photos so that they’re found more easily online. Proponents of such sites argue that the public has a right to know if their neighbor, romantic date, or colleague has an arrest record. Still, mug shots are not proof of conviction; they don’t signal guilt. Having one online is likely to result in a reputational blemish; having that photo ranked as the first result when someone searches for your name on Google turns that blemish into a garish reputational wound, festering in facile accessibility. Some of these websites are exploiting this, charging people to remove their photo from the site so that it doesn’t appear in online searches.
It’s reputational blackmail. And remember, these people aren’t necessarily guilty of anything. To crack down on the practice, states like Oregon, Georgia, and Utah have passed laws requiring these sites to take down the photos if the person’s record has been cleared. Some credit card companies have stopped processing payments for the seediest of the sites. Clearly both legal and market forces can help curtail this activity, but there’s another way to deal with the issue too: algorithms. Indeed, Google recently launched updates to its ranking algorithm that down-weight results from mug shot websites, basically treating them more as spam than as legitimate information sources. With a single knock of the algorithmic gavel, Google declared such sites illegitimate. At the turn of the millennium, 14 years ago, Lawrence Lessig taught us that “code is law”—that the architecture of systems, and the code and algorithms that run them, can be powerful influences on liberty. We’re living in a world now where algorithms adjudicate more and more consequential decisions in our lives. It’s not just search engines either; it’s everything from online review systems to educational evaluations, the operation of markets to how political campaigns are run, and even how social services like welfare and public safety are managed. Algorithms, driven by vast troves of data, are the new power brokers in society. As the mug shots example suggests, algorithmic power isn’t necessarily detrimental to people; it can also act as a positive force. The intent here is not to demonize algorithms, but to recognize that they operate with biases like the rest of us. And they can make mistakes. What we generally lack as a public is clarity about how algorithms exercise their power over us. With that clarity comes an increased ability to publicly debate and dialogue the merits of any particular algorithmic power. While legal codes are available for us to read, algorithmic codes are more opaque, hidden behind layers of technical complexity. How can we characterize the power that various algorithms may exert on us? And how can we better understand when algorithms might be wronging us? What should be the role of journalists in holding that power to account? In the next section I discuss what algorithms are and how they encode power. I then describe the idea of algorithmic accountability, first examining how algorithms problematize and sometimes stand in tension with transparency. Next, I describe how reverse engineering can provide an alternative way to characterize algorithmic power by delineating a conceptual model that captures different investigative scenarios based on reverse engineering algorithms’ input-output relationships. I then provide a number of illustrative cases and methodological details on how algorithmic accountability reporting might be realized in practice. I conclude with a discussion about broader issues of human resources, legality, ethics, and transparency."
"An algorithm can be defined as a series of steps undertaken in order to solve a particular problem or accomplish a defined outcome. Algorithms can be carried out by people, by nature, or by machines. The way you learned to do long division in grade school or the recipe you followed last night to cook dinner are examples of people executing algorithms. You might also say that biologically governed algorithms describe how cells transcribe DNA to RNA and then produce proteins—it’s an information transformation process. While algorithms are everywhere around us, the focus of this paper are those algorithms that run on digital computers, since they have the most potential to scale and affect large swaths of people. Autonomous decision-making is the crux of algorithmic power. Algorithmic decisions can be based on rules about what should happen next in a process, given what’s already happened, or on calculations over massive amounts of data. The rules themselves can be articulated directly by programmers, or be dynamic and flexible based on the data. For instance, machine-learning algorithms enable other algorithms to make smarter decisions based on learned patterns in data. Sometimes, though, the outcomes are important (or messy and uncertain) enough that a human operator makes the final decision in a process. But even in this case the algorithm is biasing the operator, by directing his or her attention to a subset of information or recommended decision. Not all of these decisions are significant of course, but some of them certainly can be. We can start to assess algorithmic power by thinking about the atomic decisions that algorithms make, including prioritization, classification, association, and filtering.
Sometimes these decisions are chained in order to form higher-level decisions and information transformations. For instance, some set of objects might be classified and then subsequently ranked based on their classifications. Or, certain associations to an object could help classify it: Two eyes and a nose associated with a circular blob might help you determine the blob is actually a face. Another composite decision is summarization, which uses prioritization and then filtering operations to consolidate information while maintaining the interpretability of that information. Understanding the elemental decisions that algorithms make, including the compositions of those decisions, can help identify why a particular algorithm might warrant further investigation." (http://towcenter.org/research/algorithmic-accountability-on-the-investigation-of-black-boxes-2/)
Typology of Countermeasures
"There are a number of human influences embedded into algorithms, such as criteria choices, training data, semantics, and interpretation. Any investigation must therefore consider algorithms as objects of human creation and take into account intent, including that of any group or institutional processes that may have influenced their design. It’s with this concept in mind that I transition into devising a strategy to characterize the power exerted by an algorithm. I’ll start first with an examination of transparency, and how it may or may not be useful in characterizing algorithms. Then I’ll move into how you might employ reverse engineering in the investigation of algorithms, including both theoretical thinking and practical use cases that illustrate the technique. I conclude the section with certain methodological details that might inform future practice in developing an investigative reporting “beat” on algorithms, including issues of how to identify algorithms for investigation, sample them, and find stories.
Transparency, as it relates to algorithmic power, is useful to consider as long as we are mindful of its bounds and limitations. The objective of any transparency policy is to clearly disclose information related to a consequence or decision made by the public—so that whether voting, buying a product, or using a particular algorithm, people are making more informed decisions. Sometimes corporations and governments are voluntarily transparent. For instance, the executive memo from President Obama in 2009 launched his administration into a big transparency-in-government push. Google publishes a biannual transparency report showing how often it removes or discloses information to governments. Public relations concerns or competitive dynamics can incentivize the release of information to the public. In other cases, the incentive isn’t there to self-disclose so the government sometimes intervenes with targeted transparency policies that compel disclosure. These often prompt the disclosure of missing information that might have bearing on public safety, the quality of services provided to the public, or issues of discrimination or corruption that might persist if the information weren’t available. Transparency policies like restaurant inspection scores or automobile safety tests have been quite effective, while nutrition labeling, for instance, has had limited impact on issues of health or obesity. Moreover, when the government compels transparency on itself, the results can be lacking. Consider the Federal Agency Data Mining Reporting Act of 2007,19which requires the federal government to be transparent about everything from the goals of data mining, to the technology and data sources used, to the efficacy or likely efficacy of the data mining activity and an assessment on privacy and the civil liberties it impacts. The 2012 report from the Office of the Director of National Intelligence (ODNI) reads, “ODNI did not engage in any activities to use or develop data mining functionality during the reporting period.”20Meanwhile, Edward Snowden’s leaked documents reveal a different and conflicting story about data mining at the NSA. Even when laws exist compelling government transparency, the lack of enforcement is an issue. Watchdogging from third parties is as important as ever. Oftentimes corporations limit how transparent they are, since exposing too many details of their proprietary systems (trade secrets) may undermine their competitive advantage, hurt their reputation and ability to do business, or leave the system open to gaming and manipulation. Trade secrets are a core impediment to understanding automated authority like algorithms since they, by definition, seek to hide information for competitive advantage.21Moreover, corporations are unlikely to be transparent about their systems if that information hurts their ability to sell a service or product, or otherwise tarnishes their reputation. And finally, gaming and manipulation are real issues that can undermine the efficacy of a system. Goodhart’s law, named after the banker Charles Goodhart who originated it, reminds us that once people come to know and focus on a particular metric it becomes ineffective: “When a measure becomes a target, it ceases to be a good measure.”22 In the case of government, the federal Freedom of Information Act (FOIA) facilitates the public’s right to relevant government data and documents. While in theory FOIA also applies to source code for algorithms, investigators may run into the trade secret issue here as well. Exemption 4 to FOIA covers trade secrets and allows the federal government to deny requests for transparency concerning any third-party software integrated into its systems. Government systems may also be running legacy code from 10, 20, or 30-plus years ago. So even if you get the code, it might not be possible to reconstitute it without some ancient piece of enterprise hardware. That’s not to say, however, that more journalistic pressure to convince governments to open up about their code, algorithms, and systems isn’t warranted. Another challenge to using transparency to elucidate algorithmic power is the cognitive overhead required when trying to explicate such potentially complex processes. Whereas data transparency can be achieved by publishing a spreadsheet or database with an explanatory document of the scheme, transparency of an algorithm can be much more complicated, resulting in additional labor costs both in the creation of that information as well as in its consumption. Methods for usable transparency need to be developed so that the relevant aspects of an algorithm can be presented in an understandable and plain-language way, perhaps with multiple levels of detail that integrate into the decisions that end-users face as a result of that information. When corporations or governments are not legally or otherwise incentivized to disclose information about their algorithms, we might consider a different, more adversarial approach.
While transparency faces a number of challenges as an effective check on algorithmic power, an alternative and complementary approach is emerging based around the idea of reverse engineering how algorithms are built. Reverse engineering is the process of articulating the specifications of a system through a rigorous examination drawing on domain knowledge, observation, and deduction to unearth a model of how that system works. It’s “the process of extracting the knowledge or design blueprints from anything man-made.”23 Some algorithmic power may be exerted intentionally, while other aspects might be incidental. The inadvertent variety will benefit from reverse engineering’s ability to help characterize unintended side effects. Because the process focuses on the system’s performance in-use it can tease out consequences that might not be apparent even if you spoke directly to the designers of the algorithm. On the other hand, talking to a system’s designers can also uncover useful information: design decisions, descriptions of the objectives, constraints, and business rules embedded in the system, major changes that have happened over time, as well as implementation details that might be relevant.24,25 For this reason, I would advocate that journalists engage in algorithmic accountability not just through reverse engineering but also by using reporting techniques, such as interviews or document reviews, and digging deep into the motives and design intentions behind algorithms. Algorithms are often described as black boxes, their complexity and technical opacity hiding and obfuscating their inner workings. At the same time, algorithms must always have an input and output; the black box actually has two little openings. We can take advantage of those inputs and outputs to reverse engineer what’s going on inside. If you vary the inputs in enough ways and pay close attention to the outputs, you can start piecing together a theory, or at least a story, of how the algorithm works, including how it transforms each input into an output, and what kinds of inputs it’s using. We don’t necessarily need to understand the code of the algorithm to start surmising something about how the algorithm works in practice. Inputs Outputs Inputs Outputs (A) I/O Relationship Fully Observable (B) Only Output Observable Inputs Outputs Inputs Outputs (A) I/O Relationship Fully Observable (B) Only Output Observable Figure 1. Two black box scenarios with varying levels of observability. Figure 1 depicts two different black-box scenarios of interest to journalists reverse engineering algorithms by looking at the input-output relationship. The first scenario, in Figure 1(A), corresponds to an ability to fully observe all of an algorithm’s inputs and outputs. This is the case for algorithms accessible via an online API, which facilitates sending different inputs to the algorithm and directly recording the output. Figure 1(B) depicts a scenario in which only the outputs of the algorithm are visible. The value-added model used in educational rankings of teachers is an example of this case. The teacher rankings themselves became available via a FOIA request, but the inputs to the algorithm used to rank teachers were still not observable. This is the most common case that data journalists encounter: A large dataset is available but there is limited (or no) information about how that data was transformed algorithmically. Interviews and document investigation are especially important here in order to understand what was fed into the algorithm, in terms of data, parameters, and ways in which the algorithm is used. It could be an interesting test of existing FOIA laws to examine the extent to which unobservable algorithmic inputs can be made visible through document or data requests for transparency. Sometimes inputs can be partially observable but not controllable; for instance, when an algorithm is being driven off public data but it’s not clear exactly what aspect of that data serves as inputs into the algorithm. In general, the observability of the inputs and outputs is a limitation and challenge to the use of reverse engineering in practice. There are many algorithms that are not public facing, used behind an organizational barrier that makes them difficult to prod. In such cases, partial observability (e.g., of outputs) through FOIA, Web-scraping, or something like crowdsourcing can still lead to some interesting results." (http://towcenter.org/research/algorithmic-accountability-on-the-investigation-of-black-boxes-2/)