I wanted to elaborate a bit more on some thinking that has been developing over the past few years. I posted this on a social channel.
It is well established that common law versus civil law based systems have some key structural differences. Most obviously: common law involves customs, conventions and, in the case of the UK an unwritten constitution. Generally it has always been my perception that common law based legal systems can be more agile or adaptive. This may be a desirable trait especially in times of rapid change. But it also comes with problems and, in the case of AI, with possibly quite unique vulnerabilities. This is what this post tried to make salient.
Common law systems evolve via precedent and (re) interpretation of past cases. As such they may generate legal outcomes that imply unequal treatment or could be interpreted as “privileges” that, with the aid of AI traversing a societies rules engine may become ever more sharply exposed.
Until the advent of large context window generative AI systems, much of a legal system’s complexity was protected by search costs. It was expensive to determine all the rules, precedents, exceptions, administrative practices, conventions and interpretations bearing on a particular situation, still more expensive to compare thousands of similarly situated people or firms.
AI radically reduces that cost as “inference meets retrieval”.
A relatively codified system such as many civil law based systems has an explicit rules graph: provisions have identifiable sources, hierarchies, amendments, exceptions and definitions. Contradictions can in principle be detected as contradictions and turning legislation into logical representations so that it can be audited, tested and sometimes executed computationally.
In an AI-mediated society, that architecture potentially makes the legal system easier not merely for machines to apply, but for citizens to interrogate: “Why does A receive treatment X while observationally similar B receives Y?”. Such interrogation may result in a political economy that exposes the case-specific “flexibility” for what it may well be: privileges.
Common-law systems contain a larger amount of law whose practical content can be distributed over precedent, doctrines, customary practice, administrative discretion, interpretive conventions and institutional norms. Effectively, what AI can do is to generate a giant function that spits out the rules engine that is applicable for any vector of attributes x:
It then naturally raises the question why feature vectors and that are very similar produce different outcomes. The ability to reverse engineer a societies’ rules engine through AI may raise broad questions on provenance and in the process, it may expose heterogeneity in the applied set of rules. Much of that heterogeneity will have legitimate explanations, but some may look remarkably like rents, exemptions, grandfathering, access privileges or arbitrary historical accidents.
This is naturally a problem in societies that have grandfathered in their respective feudal origins and the associated privileges into the rules engine.
This has the possibility to throw up many dimensions of social conflict at a time when the social contract and key founding pillars of capitalist organization of Western societies is under increasing stress. I think that this is simply due to the fact that capitalism was not designed for the type of network-based economic platforms that have emerged with the digital age.
Relative to civil law-based societies, where technically the codified law should already be easily accessible, AI may make a common-law system more accessible. Yet, there remains ambiguity in translating law into computational logic as this still represents an act of interpretation: concepts such as reasonableness or proportionality may not be convertible into deterministic thresholds without making substantive choices.
AI lowers the cost of reconstructing a society’s effective rules engine. Legal systems whose legitimacy depends heavily on rules being dispersed across precedent, convention, discretion and historically accumulated exceptions may consequently face greater political pressure than systems in which the operative rule set is comparatively explicit, hierarchical and codified.
Common law based societies and vulnerability to adversarial AI
One of the mental models that I have entertained is that some societies are uniquely vulnerable to adversarial and forms of hybrid warfare. I see here in particular foreign influence and money post 2008 global financial crisis as an important vector. I have tried to highlight this to some (former) national security professionals as a distinct case. In the UK, much of this unique vulnerability may be derived from the deep rooted rentierism that may be adjacent to the remnants of feudalism in particular around the role of real estate in economy and politics and in its unique position to shape and direct human flourishing.
In this reply I laid out the argument of why I think common law societies may be uniquely vulnerable. The argument here in short in this thread reply is that adversarial strategic players, which could be domestic or international, may deploy AI to identify ways of strategically directing the societal updating of the rules engine through shaping and/or creating new case law. A formalization of this argument may be as follows.
Common law is a dynamically updating rules engine where the legal state at a point in time , is not determined solely by legislative processes but it can be updated, and directed through adjudication
where is the case brought before the court, the arguments advanced, and the jurisdiction. Which cases get brought to the attention of teh courts, hence, can have dramatic influence over the direction in which the legal system updates.
Adversarial AI deployment may sharpen this process of legislative interference in a way that can drastically pivot away power from democratically elected representatives. An adequately resourced actor can search the state space of potential plaintiffs, factual configurations, jurisdictions, judges, etc to ask something like:
Here is the value of the resulting legal system to actor A with AI representing the cost of a case brought to the courts. An adversarial player that aims to maximize the dynamic payoff for itself may simply need to identify the sequence of cases that moves toward its preferred state.
Security vulnerability of adversarial precedent formation
Since the legal system is basically code, an adversarial interest group may effectively identify the set of cases that, taken together with popular beliefs can be maximally destabilizing. Just like hackers find vulnerabilities in code, there may be adversarial precedent formation or precedent poisoning.
A machine-learning analogy could be that an adversarial actor may pollute the training data resulting in machine learned algorithms to potentially reinforce bias that is desirable for the attacker. Marketing is a simple case of such preference manipulation and of course, the recent case of the sting against Reform UK highlights that this machinery is well at work in politics in the UK whereby in-transparency in the financial system creates a manifold of backdoors for dark money to effectively reshape societal discourse and/or create pressure.
The legal system may function well, judges need not be corrupt, captured, or even mistaken. The vulnerability of a society sits upstream in the selection into the adjudicative process. An adversarial actor may disproportionately determine which legally ambiguous boundary cases reach authoritative courts, thereby controlling the training data from which the common-law rules engine evolves. This fundamentally can undermine conceptionalizations of sovereignty and democratic agency fundamentally.
In democratic societies, the sovereign chooses something approximating
That is, the legal system or the rules engine is chosen in a way that maximzes the (perceived) sovereigns welfare. This happens first and foremost through elections in democratic societies.
An adversarial actor , that could be a large multinational corporation, or a wealthy network of elites, an ideological organisation, or a foreign agent may be able to move the actual legal trajectory
toward their preferred legal system without ever winning a legislative contest simply through the court system. This amounts to a hacking of societies rules engine and de-facto reduces sovereignty.
The sovereign technically retains the ability to legislate around an undesirable precedent, and higher courts can reverse it. Yet, correction may become increasingly costly.
It should be clear that this vulnerability may be unique to common law systems as precedent can be an authoritative source of law: courts may be required to follow an earlier decision. In traditional civil-law systems an individual judgment generally has a weaker law-producing role.
As a result, an adversarial actor may wish to identify the sequence of plausible disputes that has the highest probability of moving the body of precedent from its present state towards the legal regime preferred by the adversarial actor.
That is a reinforcement-learning problem and strategic litigation that aims to use legal cases to achieve broader political, social or economic objectives already exists. AI potentially changes its scale, cost and precision and as a result, its mere existence requires much greater needs for transparency around the actors that are directly or indirectly involved. Ultimately, the bounding case may well be that there needs to be full transparency over the population of claims and economic interests.
Is the UK’s uniquely vulnerable?
I feel that the UK is uniquely vulnerable on so many dimensions. First, with London as a financial hub it naturally is particularly open to (dark) money flowing into the UK and thereby making it possible for adversarial players to hack the legal system to shape law in a direction beneficially to the interest group represented by.
Second, since such attacks tap in and require legal representation owing to the licensing regimes, it effectively may create a distinct conflict of interest whereby the dark money effectively cross subsidises occupational groups or channels steady income to individual privileges of occupational licensed the country has very strong licensing regimes that effectively create large conflicts of interests
Third, as a country, the UK that does not have a written constitution and it further lacks a fail safe as it is relatively easy to change fundamentals of British law up to the point of, de-facto, undermining commitments of the British state to fundamental human rights.
Fourth, its first-past-the-post electoral system makes it uniquely vulnerable, in particular in light of the ever shrinking party membership and in particular in contexts with a media system that has effectively been hollowed out, allowing politicians to effectively campaign on national platforms and policy issues while, de-facto, neglecting their constituencies in a way that may fail to produce the relevant disciplining negative backlash locally.