Law As Computation in the Era of Artificial Legal Intelligence

tags

Mireille Hildebrandt

(Hildebrandt, M.: Law As Computation in the Era of Artificial Legal Intelligence. Speaking Law to the Power of Statistics)

Law as information: Law as a coherent web of speech acts informed by tenets of legal certainty, justice, and instrumentality. Balancing these tenets requires judgment defined by its anticipated performative effect, a judgment neither mathematically certain nor based on subjectivist psychology. Concerns expressed about undetectable infringements on legal protections and about judgment atrophy.

Notes

Conceptual foundations of the law

NOTER_PAGE: (2 . 0.4011887072808321)

next to impossible to map the multilevel framework of partially contradictory legal requirements

NOTER_PAGE: (2 . 0.5765230312035661)

regulatory system that stays ahead of the myriad of legal constructions developed by corporate law firms

NOTER_PAGE: (2 . 0.6062407132243685)

access to legal counsel for those not extremely rich has become highly problematic

NOTER_PAGE: (3 . 0.09361069836552749)

economic understanding of law has implied a turning point for the idea of lawyers as dignified stewards of individual justice and societal order

NOTER_PAGE: (3 . 0.2436849925705795)

‘the law’ as equivalent with ‘the legal services market’

NOTER_PAGE: (3 . 0.3231797919762258)

not sure we should take the assumptions underlying the narrative of commodified legal services for granted

NOTER_PAGE: (3 . 0.38930163447251115)

preliminary question of how law does and should constitute and regulate such economic markets

NOTER_PAGE: (3 . 0.4353640416047548)

Instead of looking at law as a product or service, or as something that regulates a given population, I investigate law as constitutive of societal defaults

NOTER_PAGE: (4 . 0.12481426448736999)

distinguishing the mathematical theory of information that grounds both computer science and artificial intelligence from the notion of meaningful information that is aligned with human language

NOTER_PAGE: (4 . 0.30460624071322434)

law in terms of speech act theory, highlighting the performativity of positive law

NOTER_PAGE: (4 . 0.37221396731054973)

Law as information

NOTER_PAGE: (4 . 0.5081723625557207)

information either refers to ‘formativeness’ or to ‘aboutness’

NOTER_PAGE: (4 . 0.6835066864784547)

Data that is stored in a computing system without being used can be seen as information-about-something that – until it is employed in some way or another – has no formative impact as such

NOTER_PAGE: (4 . 0.6983655274888558)

Often, the formative aspect of the relationship is based on the performative nature of specific types of language usage

NOTER_PAGE: (5 . 0.15676077265973254)

speech or text does not merely describe but often does what it describes

NOTER_PAGE: (5 . 0.32243684992570576)

depend on shared habits, patterns of behaviour and on the anticipation that practices such as drinking from a cup and enrolling for an education will make sense to others, who live in the same shared world

NOTER_PAGE: (6 . 0.1575037147102526)

this particular type of formativeness (constituting meaning) ties in with human language use; we cannot assume that other organisms or data-driven artificial agents are capable of generating meaning in this particular sense

NOTER_PAGE: (6 . 0.3707280832095096)

force of law depends on the interlocking of performative speech acts

NOTER_PAGE: (6 . 0.6508172362555721)

If law were only dependent on enforcement it would not be law but administration, discipline or simply a matter of violence

NOTER_PAGE: (6 . 0.7013372956909361)

sources of law, in that sense, do not merely contain information about the law but institute the law as such

NOTER_PAGE: (7 . 0.14041604754829123)

Legal certainty, predictability and openness

NOTER_PAGE: (7 . 0.25408618127786037)

law as ‘the prophecies of what the courts will do

NOTER_PAGE: (7 . 0.30757800891530457)

links law with experience and prediction

NOTER_PAGE: (7 . 0.3231797919762258)

opposition between logic (inexorable certainty) and statistics (calculated uncertainty)

NOTER_PAGE: (7 . 0.40936106983655274)

Uncertainty is inherent in what Herbert Hart famously described as the open texture of fundamental legal concepts, 28 safeguarding the creative and highly artificial nature of human intercourse

NOTER_PAGE: (7 . 0.5690936106983655)

relationship between performativity and legal certainty

NOTER_PAGE: (8 . 0.10772659732540861)

emphasized the potential of machine learning where it comes to ensuring legal certainty

NOTER_PAGE: (8 . 0.16047548291233282)

Law as information must, therefore, be understood as referring to law as a coherent web of speech acts that inform the consequences of our actions, itself informed by the triple tenets of legal certainty, justice and instrumentality that hold together jurisdiction (the force of law), community (even if between strangers) and instrumentality (the policy objectives of the democratic legislator)

NOTER_PAGE: (9 . 0.14710252600297177)

antinomian tension between these three tenets

NOTER_PAGE: (9 . 0.22585438335809804)

judgment that is neither mathematically consistent nor based on subjectivist psychology; the judgement is defined by the performative effect it anticipates

NOTER_PAGE: (9 . 0.2303120356612184)

mistake the mathematical simulation of legal judgment for legal judgment itself

NOTER_PAGE: (9 . 0.31054977711738485)

judgment itself is predicated on the contestability of any specific interpretation

NOTER_PAGE: (9 . 0.32689450222882616)

very fact that computability implies translation and simulation also implies that judgments are not computable if computability is understood as a perfect translation

NOTER_PAGE: (9 . 0.4739970282317979)

Law as computation

NOTER_PAGE: (9 . 0.5884101040118871)

computer science is built on a separation between signs and their meaning that is radically different from the semantics of human language

NOTER_PAGE: (10 . 0.27488855869242196)

feedback is information about the past and the present that supposedly informs the future

NOTER_PAGE: (10 . 0.4190193164933135)

Much depends on who decide whether the system ‘gets it right’, which depends on the definition of task T (in machine readable terms) and performance metric P (which determines progress)

NOTER_PAGE: (11 . 0.19316493313521543)

no relation with that of the illocutionary force or performativity of speech acts

NOTER_PAGE: (11 . 0.22659732540861813)

data does not speak for itself; it requires the attribution of meaning by those who use the data

NOTER_PAGE: (11 . 0.3402674591381872)

suggested that information is data plus meaning.

NOTER_PAGE: (11 . 0.34323922734026746)

Machines work with signs, they do not speak human language

NOTER_PAGE: (11 . 0.4576523031203566)

cognition is based on computation and manipulation of signs, not on generating meaning

NOTER_PAGE: (11 . 0.49108469539375926)

we may attribute meaning

NOTER_PAGE: (11 . 0.5059435364041605)

Meaning differs from information in its dependence on consciousness, self-consciousness and unconsciousness

NOTER_PAGE: (11 . 0.5549777117384844)

entanglement of self-reflection, rational discourse and emotional awareness

NOTER_PAGE: (11 . 0.5891530460624071)

differentiate their normative force from the force of law

NOTER_PAGE: (12 . 0.23848439821693906)

The opacity of ML

NOTER_PAGE: (13 . 0.17310549777117384)

deliberate hiding of the source code and the relevant training and test data

NOTER_PAGE: (13 . 0.17607726597325407)

issues with the contestability of the output

NOTER_PAGE: (13 . 0.24071322436849923)

warrants a legal obligation to work with open source software and to provide access

NOTER_PAGE: (13 . 0.40564635958395245)

neither domain experts in law nor those subject to law have developed any skills to scrutinize these systems

NOTER_PAGE: (13 . 0.4866270430906389)

we may simply not detect incorrect interpretations

NOTER_PAGE: (13 . 0.650074294205052)

shift from reason to statistics

NOTER_PAGE: (13 . 0.6827637444279346)

processes signs (data) not meaning (content)

NOTER_PAGE: (14 . 0.11069836552748885)

different schools of statistics generate different outcomes that will make a real difference

NOTER_PAGE: (14 . 0.30609212481426445)

bias inherent in ML applications becomes relevant in terms of the law

NOTER_PAGE: (14 . 0.45096582466567603)

Quasi invisible infringements

NOTER_PAGE: (14 . 0.5668647845468053)

taking for granted that current legal practice entails no such infringements

NOTER_PAGE: (14 . 0.5995542347696879)

visualize previously invisible discrimination, unsubstantiated conclusions, missing facts, hidden arguments or arbitrary decision-making

NOTER_PAGE: (14 . 0.6344725111441307)

judgmental atrophy

NOTER_PAGE: (15 . 0.25854383358098065)

while smart machines get better in simulating the skills of their human trainers, domain experts may fall prey to gradual deskilling

NOTER_PAGE: (15 . 0.34323922734026746)

lack consciousness and self-consciousness and even if these were to be epiphenomena of biological evolution, 67 they make ‘a difference that makes a difference’

NOTER_PAGE: (15 . 0.4063893016344725)

recognize when the automation goes awry, misrepresents relevant cases or misinterprets relevant causation

NOTER_PAGE: (16 . 0.7213967310549777)