Normative Shortcuts and the Hermeneutic Singularity

Photo of tools

(Originally posted on the COHUBICOL research blog.)

Legal normativity is an important theme for COHUBICOL, particularly how its nature might change when the medium that embodies it moves from text to code- and data-driven systems. Normativity is a useful concept in thinking about the role of law and of legal systems; it refers to the purposive force of (textual) legal instruments and rulings that, subject to their interpretation and potential contestation, require citizens to act (or not act) in certain ways.

How we act and behave are impacted by many other forces which might be purposive, coercive, or even unknown to us.

Importantly, however, normativity arises not just from legal instruments and rulings. How we act and behave are impacted by many other forces which might be purposive, coercive, or even unknown to us. Although it is not without certain limitations, Lessig’s influential analysis of four ‘regulatory modalities’ – social norms, the market, architecture/software code, and law (Lessig 2006) – gives a basic sense of how our actions are impacted by multiple external forces, sometimes acting concurrently and/or in concert with one another. This existence of alternative sources of normativity is important to bear in mind, especially where they lack the public and democratic character of law (Hildebrandt 2008). In this blog post I want to propose the concept of the ‘normative shortcut’ as a means to better understand the consequences of software code that lacks these characteristics. I’ll then speculate on the idea of the ‘hermeneutic singularity’ as one such potential consequence in the field of legal analytics.

Narrow and broad normativity

Piekarski and Wachowski (2018) usefully distinguish between two forms of normativity, narrow and broad. The narrow form is exhibited in specific rules that instruct us how (not) to act. These rules are derived from pre-existing conventions, customs, or social use. An example might be ‘thou shalt not kill’, a rule that might be seen as a reflection of our (contemporary) sense that killing is wrong, rather than a command that is at odds with our intuitive sense of how to behave. With narrow normativity, the convention or social practice is ‘converted’ into a specific rule.

By contrast, broad normativity is embodied in less precise standards that guide behaviour without specifying too exactly how we should or should not act. Piekarski and Wachowski give examples of dinner party etiquette or the design patterns that structure how users interact with software (consider, for example, the concepts of ‘windows’ and ‘swiping’ in graphical user interfaces). In each case there may be an ‘ideal’ behaviour – how one is generally expected to act – but the precise contours of that behaviour are not specified absolutely in advance.

In either case, one can see an element of flexibility, and an element of time. With narrow normativity, the rule will only be set down once the relevant group has accepted and internalised its requirements – the rule is thus flexible in the sense that its contours reflect existing social practice, while it takes time to develop and evolve (a corollary being that when social views and practices change, so do the rules – consider, for example, the contemporary legal enshrinement of gay marriage). With broad normativity, the ‘proper’ way to act is defined sufficiently widely that, at the moment of action, the individual is to an extent free to act contrary to what was expected – the individual has flexibility to interpret the rule and how she will act in response to it, which in turn requires time. I have previously (Diver 2018) referred to the ‘space’ between the rule’s text on the page and its instantiation in action as the ‘hermeneutic gap’; in order to retain interpretive flexibility the individual must be afforded the opportunity to comprehend the rule, without which the gap is collapsed and interpretation precluded.

The normative shortcut

This is where the question of a normative shortcut arises. When designers create the code of software systems, they create constitutive normativity that to an extent defines, from the outset, how the users of those systems may act (Hildebrandt 2008; Searle 1995) by means of what actions the system affords and disaffords them (Diver 2018). The system’s code thus constitutes and delimits behaviour according to the decisions made by its designer as to what (dis)affordances should be included. This has the potential to remove the elements of flexibility and time described above: the code dictates what the user may do, without necessarily being a reflection of existing social practice, and it does so in a potentially very precise and immediate manner, removing the time for reflection that in turn enables flexibility of interpretation and action.

When designers create the code of software systems, they create constitutive normativity that to an extent defines, from the outset, how the users of those systems may act

This is the normative shortcut that the designer takes – she can jump very quickly from identifying the behaviour she would like users to engage in (likely defined according to commercial imperative) to ensuring that that is what they in fact do, writing code that creates the architectures that ensure few or no other possibilities. In this way, there is no need to reflect pre-existing socially-sanctioned behaviours à la narrow normativity, and simultaneously the hermeneutic gap that allows flexibility in interpretation and behavioural responses is collapsed. The designer moves around these obstacles by means of the normative shortcut, imposing on users whatever normativity she wishes.

A ‘hermeneutic singularity’?

From the perspective of COHUBICOL, the normative shortcut is problematic in that designers can implement code-driven systems whose normativity has significant implications (e.g. through fiscal or property transfers) but that neither reflects existing social practices (particularly those embodied in the legal system as we know it, with all its familiar features for stabilising expectations and mediating conflict) nor allows space for interpretation, since the code’s execution is pre-determined and its execution immediate, regardless of any intervening circumstances that might merit a different approach.

From a data-driven perspective, the normative shortcut is perhaps harder to identify, but is no less important (indeed, its subtler effects in this context might merit even greater concern). The design of data-driven systems (e.g. machine learning (ML) models) that process legal text in order to predict outcomes may – in a less immediate sense than user-facing code architectures – adapt legal practice and expectations in ways that are not compatible with the ideal of a flexible and adaptable legal system. The effect of the outputs of such systems, based as they invariably are on data from past decisions, might be that certain lines of reasoning become entrenched because predictive analytics promotes certain forms of argumentation over others. The risk here is that the resulting case law becomes brittle and less capable of adapting to social change. The normativity of judicial rulings – both narrow and broad – might then be ‘short-cut’ by the forms of litigation that are promoted by such systems, because the arguments led in court are limited to what their predictions suggest is likely to be successful. This might ultimately lead to a homogenisation of legal argumentation that reflects the interests and biases of those whose legal actions are reflected in the training data.

…the reliance on ML analysis of legal texts creates path dependencies for future legal actions

Without being too alarmist, we might then envisage a moment of hermeneutic singularity, after which the reliance on ML analysis of legal texts creates path dependencies for future legal actions, the seeds of those dependencies having been planted by the normative shortcut of an ML system’s design compounded by its reliance on the unchanging text of past judgments. The space for interpretation might thus be constricted, however imperceptibly, each time an ML-predicated legal argument becomes enshrined in a judicial decision. This incremental predetermination is presumably not what we expect or wish from a legal system in a society that is rapidly changing on numerous fronts; in order that the courts retain their role as arbiter of the rule of law, it may be that we ought actively to ensure that certain normative shortcuts cannot be taken.

Image by Todd Quackenbush.

Selected references

  • Diver L., “Law as a User: Design, Affordance, and the Technological Mediation of Norms” (2018) 15 SCRIPTed 4
  • Hildebrandt M., “Legal and Technological Normativity: More (and Less) than Twin Sisters” (2008) 12 Techné: Research in Philosophy and Technology 169
  • Lessig L., Code: Version 2.0 (New York: Basic Books, 2006)
  • Piekarski M. and W. Wachowski, “Artefacts as Social Things: Design-Based Approach to Normativity” (2018) 22 Techné: Research in Philosophy and Technology 400
  • Searle J.R., The Construction of Social Reality (New York: Free Press, 1995)