Autonomous Agents and Extension of Law: Policymakers Should be Aware of Technical Nuances
This post expands upon a theme from Samir Chopra and Lawrence White’s excellent and thought-provoking book – A Legal Theory for Autonomous Artificial Agents. One question pervading the text: to what extent should lawmakers import or extend existing legal frameworks to cover the activities of autonomous (or partially autonomous) computer systems and machines? These are legal frameworks that were originally created to regulate human actors. For example, the authors query whether the doctrines and principles of agency law can be mapped onto actions carried out by automated systems on behalf of their users? As the book notes, autonomous systems are already an integral part of existing commercial areas (e.g. finance) and may be poised to emerge in others over the next few decades (e.g. autonomous, self-driving automobiles). However, it is helpful to further expand upon one dimension raised by the text: the relationship between the technology underlying autonomous agents, and the activity or results produced by the technology.
Two Views of Artificial Intelligence
The emergence of partially autonomous systems – computer programs (or machines) carrying out activities at least partially in a self-directed way, on behalf of their users, is closely aligned with the field of Artificial Intelligence (AI) and developments therein. (AI is a sub-discipline of computer science.) What is the goal of AI research? There is probably no universally agreed upon answer to this question – as there have been a range of approaches and criteria for systems considered to be successful advances in the field. However, some AI researchers have helpfully clarified two dimensions along which we can think about AI developments. Consider a spectrum of possible criteria under which one might label a system to be a “successful” AI product:
View 1) We might consider a system to be artificially intelligent only if it produces “intelligent” results based upon processes that model, approach or replicate the high-level cognitive abilities or abstract reasoning skills of humans ;or
View 2) We might most evaluate a system primarily based upon the quality of the output it produces – if it produces results that humans would consider accurate and helpful – even if the results or output came about through processes that do not necessarily model , approach, or resemble actual human cognition, understanding, or reasoning.
We can understand the first view as being concerned with creating systems that replicate to some degree something approaching human thinking and understanding, whereas the second is more concerned with producing results or output from computer agents that would be considered “intelligent” and useful, even if produced from systems which likely do not approach human cognitive processes. (Russell and Norvig, Artificial Intelligence: A Modern Approach, 3 Ed, 2009, 1-5). These views represent poles on a spectrum, and many actual positions fall in between. However, this distinction is more than philosophical. It has implications on the sensibility of extending existing legal doctrines to cover the activities of artificial agents. Let us consider each view briefly in turn, and some possible implications upon law.
View 1 – Artificial Intelligence as Replicating Some or All Human Cognition
The first characterization – that computer systems will be successful within AI when they produce activities resulting from processes approaching the high-level cognitive abilities of humans, is considered an expansive and perhaps more ambitious characterization of the goals of AI. It also seems to be the one most closely associated with the view of AI research in the public imagination. In popular culture, artificially intelligent systems replicate and instantiate – to varying degrees – the thinking facilities of humans (e.g. the ability to engage in abstract thought, carry on an intelligent conversation, or to understand or philosophize concerning concepts at a depth associated with intelligence). I raise this variant primarily to note that despite (what I believe is a) common lay view of the state of the research- this “strong” vision of AI is not something that has been realized (or is necessarily near realization) within the existing state-of-the art systems that are considered successful products of AI research. As I will suggest shortly, this nuance may not be something within the awareness of lawmakers and judges who will be the arbiters of such decisions concerning systems that are labeled artificially intelligent. Although AI research has not yet produced artificial human-level cognition, that does not mean that AI research has been unsuccessful. Quite to the contrary – over the last 20 years AI research has produced a series of more limited, but spectacularly successful systems as judged by the second view.
View 2 – “Intelligent” Results (Even if Produced by Non-Cognitive Processes)
The second characterization of AI is perhaps more modest, and can be considered more “results oriented.” This view considers a computer system (or machine) to be a success within artificial intelligence based upon whether it produces output or activities that people would agree (colloquially speaking) are “good” and “accurate” and “look intelligent.” In other words, a useful AI system in this view is characterized by results or output are likely to approach or exceed that which would have been produced by a human performing the same task. Under this view, if the system or machine produces useful, human-like results, this is a successful AI machine – irrespective as to whether these results were produced from a computer-based process instantiating or resembling human cognition, intelligence or abstract reasoning.
In this second view, AI “success” is measured based upon whether the autonomous system produces “intelligent” (or useful) output or results. We can use what would be considered “intelligent” conduct of a similarly situated human as a comparator. If a modern auto-pilot system is capable of landing airplanes in difficult conditions (such as thick fog) at a success rate that meets or exceeds human pilots under similar conditions, we might label it a successful AI system under this second approach. This would be the case even if we all agreed that the autonomous autopilot system did not have a meaningful understanding of the concepts of “airplanes”, “runways”, or “airports.” Similarly, we might label IBM’s Jeopardy playing “Watson” computer system to be a successful AI system since it was capable of producing highly accurate answers, to a surprisingly wide and difficult range of questions – the same answers that a strong, human Jeopardy champions would have produced. However, there is no suggestion that Watson’s results were the result of the same high-level cognitive understanding and processes that likely animated the result of the human champions like Ken Jennings. Rather, Watson’s accurate output came from techniques such as highly sophisticated statistical machine-learning algorithms that were able to quickly rank possible candidate answers through immense parallel processing of large amounts of existing written documents that happened to contain a great deal knowledge about the world.
Machine-Translation: Automated Translation as an Example
To understand this distinction between AI views rooted in computer-based cognition and those in “intelligent” or accurate results, it is helpful to examine the history of computer-based language translation (e.g. English to French). Translation (at least superficially) appears to be a task deeply connected to the human understanding of the meaning of language, and the conscious replication of that meaning in the target language. Early approaches to machine translation followed this cue, and sought to convey aspects to computer system – like the rules of grammar in both languages, and the pairing of words with the same meanings in both language – that might mimic the internal structures undergirding human cognition and translation. However, this meaning and rules-based approach to translation proved limited and surprised researchers by producing somewhat poor results based upon the rules of matching and syntactical construction. Such system had difficulty in determining whether the word “plant” in English should be translated to the equivalent of “houseplant” or “manufacturing plant” in French. Further efforts attempted to “teach” the computer rules about how to understand and make more accurate distinctions for ambiguously situated words but still did not produce marked improvements in translation quality.
Machine Learning Algorithms: Using Statistics to Produce Surprisingly Good Translations
However, over the last 10-15 years, a markedly different approach to computer translation occurred – made famous by Google and others. This approach was not primarily based upon top-down communication of the basics of constructing and conveying knowledge to a computer system (e.g. language pairing and rules of meaning). Rather, many of the successful translation techniques developed were largely statistical in nature, relying on machine-learning algorithms to scour large amounts of data and create a complex representation of correlations between languages. Google translate – and other similar statistical approaches – work in part by leveraging vast amounts of data that has previously been translated by humans. For example, the United Nations and the European Union frequently translate official documents into multiple languages using professional translators. This “corpus” of millions of paired and translated documents became publicly available electronically over the last 20 years to researchers. Systems such as Google Translate are able to process vast numbers of documents and leverage these paired, translated translation to create statistical models which are able to produce surprisingly accurate translation results using probabilities – for arbitrary new texts.
Machine-Learning Models: Producing “intelligent”, highly useful results
The important point is that these statistical and probability-based machine-learning models (often combined with logical-knowledge based rules about the world) often produce high-quality and effective results (not quite up to the par of nuanced human translators at this point), without any assertion that the computers are engaging in profound understanding with the underlying “meaning” of the translated sentences or employing processes whose analytical abilities approach human-level cognition (e.g. view 1). (It is important to note that the machine-learning translation approach does not achieve translation on its own but “leverages: previous human cognition through the efforts of the original UN translators that made the paired translations.) Thus, for certain, limited tasks, these systems have shown that it is possible for contemporary autonomous agents to produce “intelligent” results without relying upon what we would consider processes approaching human-level cognition.
Distinguishing “intelligent results” and actions produced via cognitive intelligence
The reason to flag this distinction, is that such successful AI systems (as judged by their results), will pose a challenge to the task of importing and extending of existing legal doctrinal frameworks – (which were mostly designed to regulate people) into the domain of autonomous computer agents. Existing “type 2” systems that produce surprisingly sophisticated, useful, and accurate results without approaching human cognition are the basis of many products now emerging from earlier AI research and are becoming integrated (or are poised to become ) integrated into life. These include IBM’s Watson, Apple’s SIRI, Google Search – and in perhaps the next decade or two – Stanford’s/Google’s Autonomous self-driving cars, and autonomous music composing software. These systems often use statistics to leverage existing, implicit human knowledge. Since these systems produce output or activities that in some cases appear to approach or exceed humans in particular tasks, and the results that are autonomously produced are often surprisingly sophisticated, and seemingly intelligent – such “results-oriented”, task specific (e.g. driving, answering questions, landing planes) systems seem to be the near path of much AI research.
However, the fact that these intelligent-seeming results do not result from systems approaching human-cognition is a nuance that should not be lost on policymakers (and judges) seeking to develop doctrine in the area of autonomous agents. Much – perhaps most of law – is designed and intended to regulate the behavior of humans (or organizations run by humans). Thus embedded in many existing legal doctrines are underlying assumptions about cognition and intentionality that are implicit and are so basic that they are often not articulated. The implicitness of such assumptions may make these assumptions easy to overlook.
Given current trends, many contemporary (and likely future) AI systems that will be integrated into society (and therefore more likely the subject of legal regulation) will use algorithmic techniques focused upon producing “useful results” (view 2), rather than focusing on systems aimed at replicating human-level cognition, self-reflection, and abstraction (view 1). If lawmakers merely follow the verbiage (e.g. a system that has been labeled “artificially intelligent” did X or resulted in Y) and employ only a superficial understanding of AI research, without more closely understanding these technical nuances, there is the possibility of conflation in extending existing legal doctrines to circumstances based upon “intelligent seeming” autonomous results. For example, the book authors explore the concept of requiring fiduciary duties on the part of autonomous systems in some circumstances. But it will take a careful judge or lawmaker to distinguish existing fiduciary/agency doctrines with embedded (and often unarticulated) assumptions of human-level intentionality among agents (e.g. self-dealing) from those that may be more functional in nature (e.g. duties to invest trust funds). In other words, an in-depth understanding of the technology underlying particular autonomous agents should not be viewed as a technical issue. Rather it is a serious consideration which should be understood in some detail by lawmakers in any decisions to extend or create new legal doctrine from our existing framework to cover situations involving autonomous agents.