For all his genius as a musician and performer, James Brown never got his head around taxes. His run-ins with the IRS were well documented, at one point leading to a jail sentence. Between recordings and countless live performances, he clearly struggled to keep abreast of his fiscal responsibilities, which begs the question: might a computational system of tax law have helped the Godfather of Soul stay in the IRS's good books?
Computational law concerns the ‘mechanization of legal analysis’,Footnote 1 as illustrated by Bob Kowalski's attempt to translate the British Nationality Act 1981 into predicate logic.Footnote 2 This is a form of ‘commoditisation’ of legal knowledge,Footnote 3 of which a core ambition is to provide individuals with cheap, efficient legal advice, thereby improving access to justice (‘A2J’). A fully-worked logical representation of the 1981 Act could take user responses to questions (e.g. ‘were you born in the UK?’) and return legal advice on citizenship status.
But why stop at legal advice? If cost-minimisation and time-saving are A2J priorities, then a machine statement of advice (‘You are likely to be a British citizen’) is useful, but a definitive machine statement of law (‘You are a British citizen’) is better. Machine statements of advice (‘MSAs’) cut the time and costs of engaging a lawyer; machine statements of law (‘MSLs’) go further, cutting the time and costs of engaging a courtroom, making final legal conclusions free and instantaneous.
Of course, we might be uncomfortable with the idea of MSLs. If so, it's time we started articulating why; not only because A2J compels us to, but also because MSLs are closer than you might think.
In English law, a worker must ascertain whether she is an employee or self-employed for income tax purposes. A test has emerged piecemeal through the common law, and it includes considerations like the engager's control over the worker,Footnote 4 whether the contract is personal to the worker,Footnote 5 and whether the worker bears any financial risk,Footnote 6 among other things. These factors are not weighted,Footnote 7 making the law in this area ‘uncertain in terms of administration and compliance’.Footnote 8
HMRC responded to this uncertainty in 2016 by providing an online test, entitled ‘Check employment status for tax’ (CEST).Footnote 9 Users answer multiple choice questions, including, for instance, whether the engager can move the worker between different tasks. The design of the classifier has been criticised for failing to include a test known as ‘mutuality of obligation’,Footnote 10 though this test may be of limited relevance.Footnote 11 Strikingly, the CEST page indicates that, should a taxpayer use CEST to determine her employment status, ‘HMRC will stand by the result given’ (as long as the information supplied is correct, and the relevant arrangements are not ‘contrived’).
HMRC has thus effectively guaranteed the validity of determinations produced by a computational legal algorithm. Although it would take primary legislation to confer legally binding status on CEST classifications,Footnote 12 HMRC's guarantee has meant that CEST produces classifications of law binding on the only branch of government poised to litigate. Moreover, any court doubting the validity of a CEST determination might nonetheless find in favour of the taxpayer on the basis that the determination combined with the guarantee constitute a ‘specific undertaking’ by HMRC, protected under the legitimate expectations doctrine.Footnote 13
CEST thus goes quite far beyond issuing MSAs; in fact, it's inches away from issuing fully-fledged MSLs. If it wished, Parliament could take the final step of making CEST rulings legally binding. Should it?
Consider three counterarguments. The first is that a system comprising a finite set of decision rules might ‘run out of rules’ without reaching the required confidence threshold to classify a worker's employment status.Footnote 14 This raises controversial questions of jurisprudence: Philip Leith argues that legal expert systems presuppose a Hartian view that ‘most law is well agreed’ and thus determinable, contrary to contemporary consensus for a ‘dynamic’ understanding of the law.Footnote 15 But computational legal models fit into dynamic jurisprudence also; in Dworkin's model for instance, Hercules as a ‘judge of superhuman intellectual power and patience’ (processing all pre-interpretive materials and assessing ‘fit’ and ‘justification’) cannot be replicated by actual judges,Footnote 16 but appears almost synonymous with machine learning.Footnote 17 In cases of real interpretive indeterminacy, Dworkin invokes decision-making on the basis of ‘substantive political convictions’; a residual judicial (or indeed legislative) competence behind the algorithm would be necessary and desirable for such rarities, but for all other cases MSLs could be available.Footnote 18
The second argument concerns the open texture of the law.Footnote 19 Hart illustrates this semantic challenge with reference to the word ‘vehicle’;Footnote 20 the concept of a ‘vehicle’ has fuzzy boundaries (e.g. does it include bicycles?), such that computational systems may have difficulty using it. In response, note that open texture troubles human judges also, who are unlikely to agree on whether a bicycle is a ‘vehicle’. Furthermore, computational systems can leverage comprehensive ontologies for use in extensional (i.e. enumerative) definitions, or lexicographic resources for intensional (i.e. feature-based) definitions, supplemented by ‘common sense’ knowledge bases like Cyc.
Thirdly, MSLs might raise transparency concerns. CEST currently fails this test badly, withholding from the user the reasoning behind decision-making (it could easily publish this). A greater concern is Benjamin Alarie's claim that complete legal certainty will come at the price of vast complexity – think thousands of factual features each having minute impacts on the final case decision, obscuring why the case was decided that way.Footnote 21 This is unappealing; opacity of legal principles would entail stagnation of the law or worse. To prevent this, the classification task could be made more coarse-grained, sacrificing some certainty for greater transparency.
There are no apparent slam-dunk arguments against adoption of legislation giving effect to MSLs, particularly in the tax context, where interpretation is largely literalist, statute-based, and complicated for human judges to apply. The A2J imperative therefore compels the elevation of CEST classifications to MSLs (subject to legislative competence to amend the algorithm). If there are other counterarguments, say them loud: otherwise, it's time for this tax machine to get on the scene.