Il 'nodo borromeo' dell'intelligenza artificiale e la sua regolazione
Andrea Ottolia. Professore associato di diritto commerciale nell’Università di Genova
Report held at Workshop: “Important Issues of Business Law” (Company Law, Artificial Intelligence & Intellectual Property Law) for Business People - Shinawatra University - 16 March 2018”.
Il presente lavoro identifica i tre formanti dell’innovazione svolta attraverso l’intelligenza artificiale (i) nei dati, (ii) negli algoritmi e (iii) nei modelli derivati e dimostra come l’intervento regolatore in questa materia debba tener conto di ciascuno di essi. Ponendo l’attenzione in particolare al problema dell’utilizzo di sistemi di intelligenza artificiale c.d. “black box” si esplorano le diverse opzioni regolatorie “ex post” ed “ex ante” indicando per queste ultime a quali condizioni sia auspicabile un regime di trasparenza orizzontale (ovvero inter partes) oppure verticale (ovvero nei confronti dell’Autorità competente).
This work identifies the three elements of the innovation achieved through artificial intelligence in (i) data, (ii) algorithms and (iii) models and demonstrates how the regulatory intervention in this field should take account of each of them. With particular attention paid to black box A.I. the paper analyses different regulatory options both “ex post” and “ex ante”, describing for the latter under which conditions we should apply “horizontal transparency” (i.e. inter partes) v. “vertical transparency” (i.e. to the competent Authority).
Artificial Intelligence – black box – Intellectual Property – Big Data – Internet of Things – GDPR
Artificial intelligence (AI) has recently gained wide attention in the legal discourse . While the topic is not new (its origins go back to the Fifties ), it has become of paramount importance for the ubiquitous advantages provided by AI in several innovation processes . The present work identifies which are the different regulatory levers in this field and draws some ideas that Regulators may take into consideration when intervening in this particular area of law and technology.
2. Methods: the Borromean knot of data, algorithms and models
The first observation takes a broad view on the topic and suggests that any legal intervention in this field should carefully consider all the different components of the innovation process generated by AI. Such process (to which I conventionally refer to also as “computational innovation” ) can be described by the image of the Borromean knot  where each ring is equally important and removing any ring results in two unlinked rings.
The three rings of computational innovation are data, algorithms and models: (i) data are the basic sources of any inductive process of knowledge production, (ii) algorithms are (at this stage of technology) software capable to learn, (iii) models are the results generated by a mixture of the formers: they consist of software that is capable to take certain decisions based on rules that are independently developed by algorithms once they have been “educated” by a particular set of data. Private and public regulation should never underestimate the inter-relations between these three “rings”: limiting its action just to one or two of them may result in unwanted chilling effects.
2.1. Only looking at data
We may consider a project set by a public Institution to build a digital commons starting from a collection of analog data. By limiting the attention solely at the data, the Institution may end up with a contractual agreement with a private AI service provider, resulting in a free digitalization of the analog collection, capable of facilitating its availability in a human readable form: this outcome may be “traded off” by the private party (that digitizes, structures and analyzes the data) by obtaining an exclusive right on the final models “extracted” from the data (or “educated” through the Institution’s data). Without going into the niceties of such right, the ideal candidate for such legal protection would be trade secrecy law .
This kind of choice may be legitimate (even if it somehow resembles the “proto-copyright”, i.e. the “privilegia” provided in the Fifteen century by the Republic of Venice to certain cultural enterprises to ensure exclusivity on publishing businesses). However, from the standpoint of computational innovation, this kind of agreement would betray the purpose of creating a digital commons: indeed, by creating an intellectual property on the models, it would facilitate the “enclosure” of the “computational public domain”, i.e. the part of public domain consisting of the inner correlations between data.
2.2. Only looking at algorithms
We may then consider contractual or regulatory schemes meant to make algorithms freely available through free software regimes , e.g. by contractually imposing the publication of the source code. The real impact of such solutions in terms of circulation of knowledge may be significantly frustrated in cases where data collections are kept secret or are difficult to rebuild due to network effects: here, data controller may educate algorithms through such secret collections and generate proprietary and secret models .
2.3. Only looking at data and algorithms
Other examples of the different risks generated by partial approaches in this field may emerge in a B2B contractual relation (e.g. between a private entity “A” providing services of data analysis to a big data holder “B”) where the parties miss to regulate the ownership regime for the models generated during contractual execution. This will generate a great level of uncertainty and conflict. In particular, where intellectual property on the data belongs to B, this will argue that A cannot retain any ownership over the derived model due to an interference with its intellectual property rights. On the other side, A will have a great incentive to ..