Cross-eyed glances of a lawyer and a mathematician.
This dossier, proposed by Laurence Pécaut-Rivolier, advisor to the Court of Cassation, and Stéphane Robin, research director at INRA, was separated into three episodes.
Link to the first episode, published on April 14.
Link to thesecond episode, published on April 15.
3 – A concrete example: the digital labour code
31 – The concept
Announced by Article 1 of Ordinance No. 2017-1387 of September 22, 2017 on predictability and securing relations, the purpose of the digital labour code, according to the law, is to allow,”in en response to a request from an employer or employee about its legal situation, access to legislative and regulatory provisions as well as to conventional stipulations, in particular of branch, business and establishment, subject to publication, applicable to it.”
It is actually a tool, both a search engine that allows, from keywords and some basic information (company name, etc.), to find the article of the code that is intended to s apply, as well as the text of the applicable industry or inter professional collective agreement (at least for the fifty most important of them that are so far in the database),and toolkit for its users (simulators, act models). ).
Revolution for some, basic tool and deceptive27 for others,the digital work code is a use of AI with the aim of facilitating access to information for all. It allows any user, using common words,to obtain an accessible and understandable answer by all based on the legal and conventional text that applies s’applique to his situation. It thus aims to make labour law “accessible and effective to all” in the words of the Minister of Labour28, since,through the use of keywords derived from common language, one can find the right text, which, with regard to the labour code whose thickness has often been mentioned, is a considerable step forward, but even more so when it comes to collective agreements, which are for the employee or even the employer. .
In practical terms, the Digital Labour Code is a free en online platform that currently offers three services:
A database incorporating the texts of the labour code, that of the main collective agreements and explanatory sheets, allowing from common French keywords and not necessarily legal (holidays,disputes) to obtain the return to the documents appearing most suitable;
A database of collective agreements to verify, this time by a system of legal keywords, a navigation in their provisions by theme and/or keywords;
A licenciement toolkit, including simulators that allow the calculation of frequent compensation that can be easily calculated from the response to some factual data (separationpay, notice of resignation, gross/net salary, precarious pay and notice of termination),and document models.
The concept is all the more in line with the idea of a public service because it provides that “the employer or employee who avails himself of the information obtained through the “digital labour code” [i.e.], in the event of a dispute, presumed in good faith”2929 . Therefore,the consultation of this database offers a rare legal certainty to the person who uses it.. The stakes are high: the labour law intelligence services (in the Direccte) are currently responding to more than 900,000 labour law requests per year, from the simplest to the most complex cases.30. In three months,the digital labour code has already attracted nearly 420,000 visits.31.
This tool is therefore an intelligent database, but is not intended to perform a legal analysis. It may give a textual correspondence to a keyword, but not help to legally qualify the facts submitted to it. It may give the applicable amount of compensation if the dismissal is without real and serious cause, but it cannot say whether,in the circumstances,the dismissal is without real and serious cause. . This is, obviously, a considerable innovative help for the user, whether he is an employer, an employee, or even a lawyer. To a question expressed in “everyday” words, it links with the legal text that resolves the situation – if it exists – and if necessary with official websites (e.g. public service sites) providing useful explanations, and with calculators when the request concerns a salary or compensation amount dependent on objective data (seniority,applicable agreement, working time, etc.). But it does not analyze the situation and does not give a legal answer to particular cases, if only because it does not incorporate, except at the margin in the calculators,the case law.
32 – Achievement32
To carry out the digital labour code, the Directorate General of Labour went into start-up mode 33, bringing together a team composed of lawyers from the Directorate General of Labour or recruited specifically for this purpose and specially dedicated computer scientists, who worked together, in an incubator, for nearly two years. Many tests have been organized thanks in particular to a beta version accessible very early online and whose codes are publicly accessible on a Github repository, including the uses of the machine learning community. The site of the digital labour code offers the user various tools that illustrate the distinction between mechanistic models and empirical models made above.
The former (“toolbox”) propose, for example, to calculate the duration of a notice to be respected in the event of resignation or to estimate a severance pay. The words “calculation” and “estimate” should not be understood in the statistical sense or in the sense of AI. On the contrary, these calculations and estimates are based on a perfectly mechanistic approach: the user is invited to answer a series of closed questions that iteratively guide the system.. The algorithm thus progresses in a deterministic way within the legal texts as the situation submitted to it is gradually clarified. The conduct of this process therefore mimics the legal reasoning to arrive at the determination of the rule of law applicable in this case and the establishment of a (rangegamme of) duration or compensation prescribed by law. If the situation is particularly atypical,the system may not provide an answer and offer to contact a legal service directly. .
This work is not an AI in the usual sense, but it is remarkable proof that it is possible to encode the sens decision-making process in a formal way and thus offer an opinion whose motivation is provided by the actual course of the algorithm. Its implementation, however, is based on two conditions rarely met in the legal field (and in most of the problems that AI tackles). First of all, it deals with a section of labour law that lends itself particularly well to formalization and automation.. But above all, its implementation en required ant work to the department’s services to perform the encoding of all these rules in a form suitable for automatic processing, and will ask to continue this same painstaking work as part of the permanent updates..
But the site of the digital labour code also proposes a free method of questioning (inen “easy French”) texts relating to labour law. This is not to provide an opinion (let alone a decision)on a particular situation or dispute, but,more modestly,to guide the user (employer or employee) to the relevant texts or regulations. .
The Mechanisms AI Are Complementary Mechanisms research Traditional and Answers Preset by the experts In labour law. The service is based In Effect GreatlyFor This That Concerns its mode not Forced, on Tools From AI And Particular Tools Treatment Automatic Of Language since he It’s about interpret A Query Formulated In Language free (unlike questions) Closed). For Determine The Texts And Regulations the most Probably Relevant. The System Used Here relies on a Algorithm analogous to That Used by the Engine Google search Based Mainly on associations Statistics between words Observed In A Great document base. This Algorithm Has Obviously, Had Be Led On A base of Data that the Texts Official Could Hardly Be − So It Is True that the Language Legal Is, Sometimes, Enough Away ” French easy.” The System Proposed Rests Particular on one Algorithm Representation Of Text From Laboratories Google and Led for the French Via a body of documents Available on the internet and Particular the whole pages of the site Wikipedia. The System Is So Initially Based on one Algorithm Learning Not Supervised The Sense Where It proposes Texts « Close “The Query Formulated. His Use a, for example, Example, Permitted Put In link the Terms “Navigo Pass” and ” Refund transport costs per Employer ». The System Put In Place Improves Same Gradually The relevance of Its Answers In asking Each User If The Answer Proposed Him Has Been useful, Way similar to the Process Described (1323). A Algorithm Supervised Allows Then Discriminate The Better between the Answers Useful And Those who Are Least.
This second tool, which is part of the AI, is obviously not immune to the biases inherent in any learning system (dependence on the learning database, the effect of choosing a particular adjustment criterion, etc.). but the non-decision-making nature of the response it produces obviously reduces the consequences of these biases..
33 – The outlook
Should we go further in the development of the digital labour code, as some would like? If textes documentary improvements are obviously possible, including the increase in the number of standard acts of branch conventions available,the question can be asked of the ability to integrate the case law into the tool, i.e. the interpretation given to the texts by the Court of Cassation, or even to answer, thanks to this database of case law, more targeted questions of users (“Can PuisI be dismissed for a delay at work?”
On the first point, and subject to being certain of having a mechanism to update the jurisprudential basis in real time, the service would obviously be tenfold,without changing the very design of the tool..
On the second point, on the other en hand, the approach would be totally different,which would be to give the digital code the ability to provide a legal analysis and response to a situation of fact mentioned by an individual. The step to take would be dangerous,especially from a series of hypothesis questions too elementary to ensure the reliability of the diagnosis.
The digital labour code tool is valuable and unique in terms of access to information.. If it is not intended to propose a decision, it helps to guide a non-expert user in an area of which he is not necessarily familiar, which is already a lot. Of course, it deliberately limits its “predictive” ambitions to a particularly marked and quantifiable field of labour law and is based, for this reason, on a resolutely mechanistic approach, thus not blindly trusting the algorithms derived from AI.
To want to broaden its predictive ambition would distort it and present much less guarantees.
In design, there is no neutrality, there are only choices, if only by default in implementation, that require questioning the balance line that one wants to establish between secure response all digital and truly adapted response of which the actor is still the best guarantor. The concept of loyalty, although differently understood by mathematicians and lawyers, can help, even compel, to draw this line..
We must admit that in the area of justice, AI can greatly help in accessing information but cannot replace human intervention. And that, in this area more than any other, only a true cross-understanding of data and issues through a strong dialogue between those who can conceive and those who can use is fundamental. and possible. This interaction will obviously be facilitated if professionals in both fields are warned of these issues from their initial training, or even if a fraction of them specialize in them..
Predicting is not an end in itself – otherwise the conclusion would be simple: algorithms can obviously predict,and better than any human, but why? – since we do not know how the prediction is made and why we want it. The motivation for the decision is part of the decision itself and underpins its legitimacy..
Failing to place the judge in his decision-making, it might be tempting to give algorithmic predictions the same value as expert advice. But it would be, on the one hand,to underestimate the effect of intimidation produced by a “mathematical” result that makes it difficult not to submit to it most often. It would be, on the other hand, to forget that an expert can be asked about the reasoning or methodology that led him to his opinion,questions to which an algorithm will be well in trouble to answer..
Conversely,for the tasks that are decided to entrust to it, the AI requires a complete mastery of its design. No space can be left to default design choices, for ease,by habit for the computer scientist. It will be necessary for the lawyers to be aware of this and to seize the tool so that the AI obeys them and not the other way around..
To do this, immersing yourself in understanding how it works is a responsibility of lawyers,and decision-makers in the judicial world. Under this condition, the contribution of AI to justice can be a considerable step forward and a progress that will be no less.
- P. Januel, Codification, review of the work in progress and prospects, Dalloz news, June 26, 2018: “For the higher coding commission, the code name to what is only an interactive service is inappropriate. She is concerned about the”overdoing”of the code label.”
- Press conference on 16 Jan2020 for the launch of the digital labour code.
- Ord. 1387-2017, 22 Sept.2017, art. 1er,al. Two of them.
- Source: Ministry of Labour.
- Source: Ministry of Labour.
Mme 32. The authors thank the Director General of Labour, Mr. Yves Struillou,the Deputy Director, Mr. Laurent Villboeuf,and the Project Manager at the Directorate General of Labour, Catherine Lissarague,for directeur leur allowing them to observe in situ the design of the digital labour code during the year 2019 and to have given them access to all the documents on the design of the tool..
- J.-F. Kerléo, “Public Service in Start-up Mode,” AJDA 2020. 83 .