dans le cade de la Digital service act Week le 1 juillet 2020
mercredi 1 juillet 2020
Ceci est la traduction en anglais par Valeriya Tsekhanska d'un article : Nos réseaux sociaux, notre régulation par Serge Abiteboul et Jean Cattan, paru dans Grand Continent
The networks we love to hate
The networks we love to hate
We love to communicate. We love to share. We love to debate, exchange, sometimes vigorously. The Internet has multiplied our natural abilities to do so in countless forms. To respond to this desire to communicate, a myriad of services has developed over the years. Among them are what are called digital social networks – services that allow us to expose aspects of our personality, a profile of ourselves, and to be constantly in contact with everyone else. Once upon a time, it was Myspace. Nowadays, a wide variety of services have emerged in this double dimension of mediation and communication of individuals, ranging from Facebook, Snap, or Twitter to WT Social and Mastodon. Although not their primary role, other services such as Wikipedia, Google Maps, YouTube or Jeuxvideo.com also allow such exchanges.
Little by little, we have acquired various profiles and identities. Every day, we build more relationships - whether friendly or not -, out of shared interest, if not out of love. We stay in touch, we post, we follow each other, we like, we comment, we react, until we cannot take it anymore, and want to stop this altogether. But, more often than not, we delve deeply into this exuberance of social ties.
These services attract more and more people by utilizing the network effect. Something interesting happens in our social sphere, in a much wider circle, our “contacts” join it, the information spreads virally, we feel the need to be there too. We want to be there. We enrich our traditional social ties with these services, we forge new ties as well. This is where we often learn what makes our community, our neighbourhood, our country react. The information provided by the users themselves opens up new horizons, from the most local to the most general, horizons that are broader than those of traditional media.
The social networks become an essential exchange facility, a new frontier of our public space. And like every architecture, social networks are becoming normative. From Baron Haussmann to Lawrence Lessig, whether real or virtual, whether boulevards or computer code, the architecture is law, - code is law. Social networks participate in defining the content that we exchange. It becomes obvious when we have to limit what we write to 280 characters, or when our messages almost instantly disappear. We do not express ourselves in the same way. Some art forms are born of these constraints, but not only.
Gide said, “Art is born of constraints, lives off struggles and dies of freedom”. Social networks, as we know them, give rise to conflicts between different freedoms. They draw their essence from the freedom of expression and the right to be heard for many people who have had no say before: the activist of a suburban association, the adolescent of a lost territory... They can die of conflicts with other fundamental rights: how to reconcile freedom of expression with the right to be protected against misinformation as well as the right not to be defamed, insulted or harassed on these same networks?
Beyond the parameters that directly define the content and formation of our messages, most social networks also define what is massively read, seen or doomed to fall into oblivion of cyberspace. They choose the content they will show to you in order to maximise their profits. Someone would say “It’s the business model, stupid!”. Social networks are not all dedicated to the good of humanity, it is not their (main) objective. Inevitably, companies always seek profitability, and social networks often aren’t exception. It is not judgment, but only an observation.
The social media’s business model is essentially based on advertising. Whether it is justified or not, such model encourages a heavy exploitation of our personal data and an increasing profiling. The more the social networks know about us, the more profitable its advertising space becomes.
However, to be more profitable, advertisements should not only be targeted, but also must be seen. So that we stay longer online, social networks promote content that raises user engagement: makes us react, attracts us, and makes us stay. As a result, we are more exposed to advertising, sponsored content and the content that sponsors the platform. Our attention has become a vast market. As Reed Hastings, co-founder and CEO of Netflix, once said: “ we actually compete with sleep (…) and we're winning!”, - a contemporary version of the “available brain time” that once prevailed on television. What is true for audio-visual services, is also true for the networks: we continue to spend more and more time in front of our screens, and such commitment quickly turns into an addiction.
Like any “addict”, we always need more. Whatever makes us react can be very extreme: starting from a tiny kitten, and ending with something quite outrageous. Often this is still not enough, and we are forced to invent stories, gossips, and bring out conspiracies. We would get bored otherwise.
Fortunately, commitment to noble causes, true feelings, humour, poetry and art on social media saves us from all this. Some beautiful ideas and intellectual innovations circulate there as well. They remind us how infinite human creativity is. We love to see a crowd mobilise for just causes. We admire the youth that appropriates struggles through new means of communication, the ones that their ancestors failed to grasp. We take advantage of this every day, and we are pleased. In essence, every day we prove that social networks can be an instrument at the service of our development, if not the liberation of individuals and people, because, after all, isn’t it all about connecting people and creating society?
However, there is a risk: if we sit back and allow social networks to evolve on its own, and if their dark side prevails, one day we would have to pay the price for not being able to master the unbridled expression of our own desires and impulses, abandoning the public sphere to a business model that commodifies people.
The toxic content that misuses the social networks comes in different forms: terrorism and child pornography, threat messages and fake news, misinformation, cyberbullying, and violation of privacy. Society is concerned and demands to take measures to fight them. The Cambridge Analytica case was the first revealer of such phenomenon: how the use of social networks, and in particular of political advertising on them, can influence a major collective decision. Our exposure to hatred continues to grow, along with the fear that it may turn into violence, which is unfortunately very tangible.
If we do not want to scorn social networks, we must show that there is another way. This requires education and engagement of citizen Internet users. Some organisations offer alternative models. This is the case, for example, of WT Social, - a social network that focuses on information, developed by Jimmy Wales in 2019. The message is to draw inspiration from the methods of funding, presentation and moderation of Wikipedia to avoid the spread of fake news.
If such initiatives are good to take, will they manage to preserve this energy in social networks that makes us addicted? Perhaps, but it is very unlikely that they alone will succeed in replacing the major platforms in foreseeable future, as some of those platforms already have billions of users and know perfectly well how to profit from the network effect and the unlimited financial resources. It is safe to say that we will have to rely on some form of state intervention to force these network giants, that we have nurtured by our membership, to work for the good. Even if the market were to provide solutions to this end, the spread of fake news or disseminating hatred are not problems that can be left to the wisdom of the crowd and the supply and demand relationship.
In our system of representative democracy, institutions with a democratic legitimacy have a say. However, the task of guaranteeing the freedom of expression for everyone, while protecting people from excesses of social networks is exceptionally difficult. There is a narrow margin, for example, between thwarting disinformation and instituting state information. Institutions must intervene in an open framework under the watchful eye and control of a responsible society. In order to achieve the necessary balance in every area: regional, national, cultural and linguistic, it is essential to ensure the participation of all interested communities, researchers, civil society, users and representatives of the State. Everyone must be able to participate in identifying the problems encountered by the platforms, in defining solutions and in the monitoring the actions taken.
To say that nothing is done with the announced propagation of fake news, hatred, and personal data commodification, is wrong. Society and politicians have taken up the subjects. An emergency has arisen as if these subjects were new. Suggestions are made, decisions are taken on both sides: the side of social networks themselves, and the side of public authorities. States react and parliaments legislate, sometimes even under the pressure of current events and emotions. This is done out of sequence, with no overview, while trying to find solutions, often separately, to different facets of the same problem. Indeed, the responses remain inadequate.
In this ambient chaos, main lines can still be identified. To describe the situation during the 2018 edition of the Internet Governance Forum, - the UN summit that was held that year in Paris, - Emmanuel Macron suggested three possible paths, each of them corresponding to a model for regulating social networks:
- The first way is that of regulation by market players themselves, and self-regulation with inductive intervention of state authorities. If this is the path that allowed the creativity of the web, that allowed it to flourish, it has shown its limits in terms of regulating social networks.
- The second way is that of a direct control of content by the State and the imposition of obligation to withdraw harmful content. This is the path chosen by various legislative initiatives, including in France. In addition to serving as a tool for undemocratic regimes, it has so far been ineffective.
- The third way is the one advocated for Europe by Emmanuel Macron at the Internet Governance Forum, a path yet to be developed: “We must build a new path through regulation, whereby States, Internet users, civil societies, and all stakeholders are able to effectively regulate”. Concretely, this path of regulatory supervision that we want to be open and democratic.
The third path is more complex, but potentially more effective and more respectful of everyone's rights. We believe that this will be the way to combine many initiatives emerging from society as a whole, to generate new ones. We also believe that it can be designed in Europe, and serve as a model on a global scale.
The 1st way: self-regulation and inductive intervention
What about published content that can have a harmful effect? Both in the United States and in Europe, it is subject to the social network’s limited liability regime. It is this regime that has enabled the success of services whose content is fed by users. For its opponents, this is a flawed regime to which we owe the appearance of the worst excesses encountered on the web.
In 1996, such regime was instituted in the United States by Section 230 of the Communications Decency Act: “No provider or user of an interactive computer service` shall be treated as the publisher or speaker of any information provided by another information content provider”. Due to the foregoing reasons, this measure is dedicated to preserve the Internet expansion and market freedom, as well as to enhance user control over their own online activities. This includes imposing responsibility on parents to protect their children against harmful content. The accountability of service providers went hand in hand with the accountability of adults.
Considering that the most powerful of the social networks that we use are based in the United States, it is important to have in mind this limited liability regime that informs the legal culture with which they are imbued: the principle is that social networks are not responsible for content propagated on their platforms if they do not have editorial control over it.
In 2001, the European Union adopted eCommerce Directive 2000/31 and its Article 14. This Article, which was added to the American regime, very relatively limits the scope. In its transposition in France, Article 6 of the Law for Confidence in the Digital Economy of 21 June 2004 as interpreted by the Conseil constitutionnel, establishes the first bricks of the law on responsibility of the “hosts”. The host - being a social network, - is responsible if it actually had knowledge of the illegal nature of the content, if it did not act promptly to remove such data, or did not prohibit access to such content once it came into its knowledge. That is to say that, concretely and very summarily, a social network, wherever it is established in the world, is obliged to remove manifestly illegal content of which it has become aware, either because such content has been reported, or because it was characterised as such by a judge.
Note that these American and European texts were developed before social networks got to be as we know them now; for example, Facebook came into existence in 2004, YouTube - in 2005, Twitter - in 2006. Tensions arise from differences in appreciation of the principle of freedom of expression on both sides of the Atlantic: quasi-religious reading across the Atlantic of the 1st Amendment to the American Constitution, more flexible interpretation in Europe of Article 10 of the European Convention on Human Rights, which may have limitations that are "necessary in a democratic society".
Furthermore, in Europe, and especially in France, the intervention of the judge remains important. The Law For Confidence in the Digital Economy has opened the way to a whole series of summary proceedings, by which a judge can urgently order platforms to remove content. While deadlines are often considered too long by complainants, the procedures and case law exist.
In this context, the first question arises: what is illegal content? Between clearly illegal and legal content there is a range of “grey” data that must be assessed by the platform. The qualification of this content is not simple for a social network’s moderator, and can even divide magistrates.
To “frame” the practices, a social network engages in the development of a code that defines if the contents is acceptable by “community standards”. This is not without raising a question of its legitimacy or authority as a private actor to define what can or cannot not be accepted by the services, admittedly private, but also defined as public spaces. The social network then processes the content and decides if it corresponds to its community standards. Different means are implemented in order to do this.
Human resources - which are extremely variable depending on platforms - are allocated to moderate the content reported as potentially harmful by users of the platform, or automatically detected as such.
Among other questions, mass human moderation on social networks raises a problem of the wellbeing of moderators. More and more studies, reports and articles highlight the suffering of moderators, as they have to view violent or pornographic content all the time, spending hours in the trash of the web. People with few qualifications and often in distant countries have to asses sometimes very complex legal questions, that may also require in-depth knowledge of the culture or current affairs in the country where the relevant events are taking place.
In fact, people are fairly mediocre in reporting illegal or contrary to the network’s standards content. Majority of reports are inaccurate and are made under the influence of emotions. Despite their lack of knowledge and skills, moderators do better, because they are trained and less directly involved. Nevertheless, they also have their cultural and personal biases. Their task is not easy especially for hate messages, and perhaps even harder for fake news.
An alternative is to delegate to algorithms this work of moderation that is difficult for humans to sustain. How a software can detect harmful content? Typically, machine learning techniques are used. The software is first trained on a sample of content annotated by humans who have qualified them as acceptable, hate message, violence, etc. Confronted with new content, the software will seek to determine the qualification of messages which this content resembles the most.
Algorithms are already widely used in the detection of illegal content, in particular with regard to terrorism and child pornography. They would already be better than human moderators, even if they still are far from perfect: the problem is complex with the ambiguity of languages, irony ... And they especially lack the context. Because of this, despite the progress made in this direction, the ultimate decision to withdraw, or not, content is taken by a human (with some limited exceptions like certain terrorist content or child pornography). One day the question will arise whether we can accept to delegate such decisions to a software. This question deserves to be debated.
Social networks therefore mobilise both human and technological resources to moderate content. In addition to the encountered practical difficulties, the choice of fully internalising the problems posed by social networks is questioned in terms of the following principles:
i. Is the relative absence of recourse to a judge in their procedure acceptable?
ii. Is it desirable that they alone decide their content withdrawal policy?
iii. As the efforts vary considerably from one social network to another, at what point a social network can be considered sufficiently diligent to identify and remove content? What should it measure? Against which standard to judge?
How to answer these questions staying within the fundamentals of EU law?
Pending a real substantive debate on these questions, the platforms remain in a flexible legal environment and have a very wide margin of self-regulation. This is what allows them to arrogate the vocabulary of the State to set up, for example, “supreme courts” which, like that envisaged by Facebook, would challenge the choices made by the platform as to whether or not to remove content. The public space is so remote on the platforms, and the space left vacant by the state is so large, that the platforms can afford such regal behaviour.
Whether you are a defender of freedom of expression or a representative of public order, it does not work. The self-regulatory regime leaves far too much space for platforms. Depending on their goodwill and their means, they will be either excessively or insufficiently diligent. Moreover, what is the legitimacy of these private platforms to solve alone this variety of problems with only a posteriori intervention of the authorities acting on behalf of the people? What is their legitimacy to decide what can and cannot be said in a public space, to define this public space?
In fine, it is a question of defining a point of balance and of resolving questions that have major societal importance: questions of memories, of relationship with the body, with discrimination, between citizens and their representatives etc. It cannot be up to platforms alone to define this point of balance. It is up to them to participate in its implementation, yes, but it is up to the community to define it. This is why, at some point, an intervention by the state and society is required. The question is how and to what end?
The 2nd way: direct control of content by the State
In addition to the system established at the European level, different levels of response have been put forward in the Member States to address the problems posed by the publication and dissemination of content, whether it is manifestly illegal or located in a grey area.
The first case concerns the fight against disinformation. Briefly, the French law of 22 December 2018 relating to the manipulation of information opened the possibility, during electoral periods, to seize the judge of summary proceedings to decide, within forty-eight hours, on the withdrawal of content that constitutes “inaccurate or misleading allegations or accusations of a fact with the aim to alter the sincerity of the upcoming election that are disseminated in a deliberate, artificial or automated and massive manner”. In addition, the platforms must report on the measures adopted to stem the spread of fake news, or provide transparent information on the remuneration “received in return for the promotion of [...] information content [ relating to a debate of general interest]”.
Another case concerns the treatment of hate speech. The initiative comes from Germany. Since 2018, the so-called NetzDG law notably requires social networks to remove manifestly hateful content in twenty-four hours, effectively referring the platform to assess what it is. The law provides for a penalty of up to 50 million euros for the platform’s failure to comply with this obligation. According to the latest information, the German government wanted to further reinforce its system.
It is difficult to say at present what will be the regime of the regulation of hate content in France. The anti-hate law bill proposed by Laetitia Avia in March 2019 gave rise to strong reactions. As it stands, with large lines, it imposes an obligation to remove illegal content of different kinds within 24 hours. In the law passed by the Assembly at second reading, a time limit of one hour was also established for content notified by the authorities to any platform as terrorist in nature. The sanction amounts could reach 4% of the company turnover. A blocking measure targeting mirror content is also planned.
Among the most resounding criticisms are those of the European Commission, recorded in a notice of 22 November 2019. According to the Commission, the measures disproportionately restrict the freedom to provide services; the reach of the aim pursued - human dignity - on the identification of the targeted content remains unclear; the proposed measures are not targeted and relate to excessively broad range of online platforms; the notification conditions are not precise enough; there is a risk of excessive deletion of content given the time limit and the amount of the penalty; there is no guarantee against general monitoring of content by platforms; and finally France should not legislate independently when the Commission could legislate on this issue within the framework of the Digital Services Act expected for 2020. It should be noted, however, that despite the Commissions’ announcement that it wanted to act, no specific plan hitherto has been published.
More generally, even though we fully understand the need for public protection, we may question whether the legislative initiatives like those pushed in France and Germany do not approach the problem from the wrong end.
Paradoxically, by increasing the amount of sanctions and strengthening the mechanisms of repression to which social networks could be subject, governments risk accentuating the power of these platforms. If no control is exercised over the methods of propagation and withdrawal of content beyond the question of whether content remains one or twenty-four hours on the platform, the State will remain blind to the choices made by the platform, both in the way it marks certain content and in its opt-out policy. Even more so as the small platforms - which do not have the means to implement a device of suppression of the contents to comply with the law - are condemned to completely depend on the large platforms for moderation, to considerably limit the communication of the Net surfers, or to be sanctioned.
To underline the difficulties encountered, mention should be made of the case where content notified by the judge and withdrawn reappears elsewhere in the same form or a slightly altered form. This problem was first found in a judgment of 3 October 2019 by the Court of Justice of the European Union (aff. C-18/18) in a defamation case on Facebook. It follows from this judgment that a court is entitled to order a social network like Facebook to remove information identical to that already withdrawn, but also information deemed “equivalent”. Content is deemed equivalent by the Court if it remains “in substance unchanged from that which gave rise to the finding of illegality, and comprising the elements specified in the injunction, and that the differences in the wording of this content equivalent to that characterizing the information previously declared unlawful are not such as to compel the host to carry out an independent assessment of this content ”.
The situation in which the content deemed illegal reappears almost instantly elsewhere after a potentially lengthy procedure was clearly problematic. However, as this withdrawal order can be made “globally, within the framework of relevant international law”, the room for manoeuvre and the impact of the decision on social networks are potentially considerable. Does all content needs to be compared with the removed content to determine if they are equivalent? Who can judge this equivalence?
Also, by adopting specific rules, the various European States introduce legal differences to the detriment of the consistency of the internal market. In order to homogenise everything, and even before the NetzDG law, the European Union adopted in 2016 a code of conduct, and in 2018 - a series of recommendations for platforms. These recommendations define certain methods for the notification of illegal content, to what extent hosts can take proactive measures or use automated detection methods, what remedies are open to content publishers, how to deal with the specific case of terrorist content, etc. But all this remains very general and unrestrictive: Member States need only to take due account of it.
By entrusting an administrative authority with the task of assessing the diligence of a platform in the removal of content, the State finally pushes aside all the rest of the institutions, starting with justice and society, from defining what can be said in a public space. Ultimately, the administration finds itself in a role of direct control of the content by the sanction of the absence of withdrawal, and thus our democracies promote (if it were necessary) a model which can only satisfy much less democratic countries.
The phenomena we face are frightfully complex. They play on our psychology, our individual and collective dynamics in digital social networks, but also in real life. We must open up the regulation of social networks to society as much as possible. The supervision of social networks can open us up to this alternative, not so much in that we would give specific missions to an administrative authority, but in that it will allow society to seize the problem and provide real and human solutions.
The 3rd way: agile supervision regulation
During his speech at the Internet Governance Forum in November 2018 where he announced the need to follow the third path, Emmanuel Macron also announced a partnership between the French State and Facebook, whose vocation was precisely to fuel reflection on this third way.
For six months, ten members of the French administration and three rapporteurs had access to Facebook premises and staff to observe, exchange and question the modalities of moderation of hate content by the social network. As part of this mission, judges, experts in computer science, telecoms and regulation, representatives of the police in charge of cybercrime and people responsible for the fight against racism and discrimination rubbed shoulders. The question put to them was the regulation of hate messages on Facebook in France. The goal was to come up with solutions that can extend to (i) other platforms, (ii) other types of content such as fake news, and (iii) the European framework.
During this mission, Facebook thus allowed French civil servants to study the functioning of its moderation, and in particular, to visit some of its moderation centres in Europe. Due to the time constraint, during the three months of in situ observation, the members of the group did not have access to the computer code of the social network. On the other hand, they had the opportunity to be exposed to main algorithms underlying this code. If they were able to speak with engineers and moderators of the platform, this was done in the presence of company officials. If the discussions were direct and if the controversial subjects were not avoided, the members of the group were exposed to the reality that Facebook wanted to present them, which is of course only a part of reality.
From this mission, a report was produced outlining several elements of findings and proposals.
Firstly, the report indicates that Facebook is determined to seriously tackle the problems encountered by social networks, and that it invests in this direction significant human and technical capital. This is, of course, to be contrasted with the smaller efforts of other platforms that have neither the means, nor perhaps the urge to confront these real problems. The report also notes that, even with good will and resources, the social network is struggling to solve the problems its services pose to society.
A crucial finding is that social networks play a role of “de facto editorialisation”. By structuring the content published by their users, by choosing which to “promote”, they go beyond the role of mere information host. The algorithms defined by social network designers accelerate the spread of certain content, feature it, and make it visible to a massive number of people. The exposure of the public to problematic content is rooted in these choices. The non-exposure (or a lesser exposure) must therefore also be the result of the social network’s editorial choices. This observation is the starting point of any approach to regulating social networks that aims to be effective.
As a result, the authors of the report suggest that the social network should “internalise” the objectives assigned to it by the public authorities in the fight against, for example, hateful content. The adopted angle is different from that adopted by the legislator until now. It would consist not of deciding what is to be deleted, but of ensuring that social networks put in place necessary measures for the non-propagation of hate content according to criteria set not only internally by the social network.
These objectives and criteria vary with the cultures, languages and societal themes playing out in each State, and must take due account of the diversity of European States and peoples. The implementation of regulation can, therefore, only be done effectively at a national level. To this end, the report advocates replacing the logic of the “destination” State (the State of victim’s residence) for that of the State of the “installation” of a platform (often Ireland). That said, according to this model, the regulators must coordinate through a European regulator, which would have sufficient power to dialogue from a position of strength with major social networks, would be able to define the rules of the game and curtail excesses of national regulators under pressure of national events.
The regulation should then be built around the empowerment of social networks, which are themselves held to a “duty of diligence” vis-à-vis their users: in this way “social networks would undertake to assume a responsibility vis-à-vis their users regarding abuses by other members, and attempts to manipulate the platform by third parties.”
An important aspect of the report is that not all platforms are subject to the same regime. Only the most important - the “systemic platforms” - are subject to such ex ante regulation, to avoid the regulation limiting the emergence of new innovative companies. As proposed by the report, it could be conceived that platforms whose number of users reaches 10-20% of the population of a Member State would be considered as such.
Medium-sized platforms are not a priori affected. If one of them appears to pose particular problems or to have particularly negative effects, it could be promoted in the big league and fall under the influence of the regulator. It must be emphasized that small and medium-sized platforms that are generally not subject to this regime are not exempt from all moderation: they are obviously still bound to apply the laws.
At the national level, the regulator must have the means to assess the results of the measures taken by the social networks it regulates. To this end, social networks are subject to very high transparency standard. This transparency is essential, because it alone allows a serious and meaningful evaluation of the moderation work. As such, platforms must, for example, provide information on their moderation modalities, their statistics related to moderation, including false positives and false negatives, - content unjustly withdrawn or wrongly authorised. Transparency also concerns the content scheduling functions: what content is promoted, why, with or without remuneration?
Platforms should also facilitate the reporting of problematic content. Internet users concerned by this content, whether they have published or reported it, are notified of the results of the procedures, and may appeal the decision.
In short, an “agile” national regulator is interested in global dynamics without focusing on specific cases, which can be escalated to be resolved by a judiciary. The focus is shifted from an obligation of result to an obligation of means, even if, as a last resort, the obligation of result persists within the framework of the law. To ensure compliance with the public objectives set for social networks, the regulator may impose heavy fines if they do not fulfil the obligation to take measures imposed on them.
The most serious challenge that any moderation faces is its acceptance by society: if it does not do enough - it is an accomplice of Satan, if it does too much - it is censorship. This is a pitfall on which, we believe, the self-regulation and direct moderation of the state have failed. Supervision regulation consisting of a tête-à-tête between the state and the social network would not be immune to rejection by society. The report suggests social networks enter into an informed political dialogue with all stakeholders: the regulator, of course, but also the government and its services, the legislator, the justice system, and civil society (in particular, associations and research labs). All of them participate in the definition of objectives, in evaluation, in monitoring appeals, in the construction of learning databases.
The national regulator in charge of supervising social networks would have the task of ensuring the accessibility to the outside world, of organising debates around the definition of objectives, and of involving the whole of society in the supervision process. It plays a central role in sharing information that describes the services of social networks, in particular those that explain their algorithmic choices. Finally, based on the objectives set by the political power, the regulator would be responsible for the resolution of both general and specific problems, and for the diligent settlement of disputes.
The whole structure remains in a delicate balance. Only judges can decide on the legality of the content. The regulator should oversee the operation of systemic platforms. The system derives its effectiveness from the complementarity of their roles.
Conclusion: Supervision, an open door to society
Society is already aware of the problems posed by social networks. This led social networks to take action to attempt “calming down”, while staying within the framework of the first path described. These issues have also taken up political agendas. Measures taken within the second path, such as the NetzDG law, helped systemic platforms realise that their business could not continue without profound changes.
The appetite of Internet users for social networks shows that these networks deserve to be saved. In this article, we have insisted on supervision regulation, as a way to address the problem without harming the essential contributions of social networks to society, to allow everyone to express themselves, to stay informed, to share and communicate with others, -in short, the 3rd path. However, this is only one side of the problem.
Education. We also need to learn to use social networks, to treat each other well and to learn to respect others. This places education at the heart of the system. Today, Article L. 312-15 of the French Education Code provides, for example, that “moral and civic education aims in particular to encourage students to become responsible and free citizens, to develop critical thinking and to adopt a thoughtful behaviour, including in their use of the Internet and online public communication services.” Critical thinking can be developed, for example, by encouraging the editing and moderation of online content, such as on Wikipedia, for example. Social networks are the fruit of scientific and technical revolutions. Such teaching must therefore also take this dimension into account. In particular, computer education is essential to understand how these networks operate. If we are to be masters of our environment, we must understand what it is made of. This education in critical and algorithmic thinking in the age of social media cannot be limited to a young audience. It must cover all age groups. All of us are concerned!
Data and algorithms. The algorithms for detecting content that needs to be discarded are based on annotated data compilation. The algorithm “learns” to separate the wheat from the chaff using this data. This data compilation allows to distinguish between a true information and misinformation, a hate message or just a little caustic text, etc. They must, therefore, be placed at the service of the whole society, and not only be used by large platforms that have the means to build such programmes. Small businesses also must have access to it, otherwise there is a risk of strengthening oligopolies. This calls for these compilations to be considered a common good, - “data of general interest”. Obviously, data sharing must be carried out with respect for the protection of privacy and business secrecy, possibly after anonymisation and/or consolidation. The world of research and civil society have their place in the constitution of these data sets, which must be produced in a permanent agreement between all the parties concerned. Beyond the question of data, researchers should be encouraged to explore new ways for algorithmically detecting harmful content. For example, work on cyberbullying has shown that these situations can sometimes be used more effectively by observing the graphs specific to attacks (clusters going towards a person), rather than by relying on the analysis of the actual words.
Citizen engagement. Violent extremism of all kinds has invaded social networks. Organisations have chosen to fight them on the same ground. This is the work carried out by Moonshot CVE which manages to establish a demographic and geographic analysis of the audience using data from social networks. By identifying that in such a region, such a public is more prone to utter hate speech online, social action becomes easier. In the United States, together with the Anti-Defamation League and the GenNext Foundation, Moonshot CVE has launched a program called Redirect method to combat propaganda by ISIS and the White supremacists. People doing specific research denoting one of these two trends are redirected from Google Ads or YouTube videos to organisations and content that are able to deconstruct propaganda discourse. Because, as associations note, once the content has been removed, the vulnerable person who looks for such content will always be left in its position. Allowing them to get in touch with certain people or see certain content can cause them to transform.
In such approaches, engagement goes beyond just the social network alone and involves third-party organisations. The steps and knowledge required are far too numerous to be internalised by a social network. These tasks must be a responsibility of specialised organisations, actors on the ground like Life After Hate, an organisation of former hate-mongers at the service of the fight against hate.
The social network regulator. Supervision is based on transparency, and therefore on understanding how social networks work. This makes strong participation of the whole society possible, that can induce the economic actor to open the door to “regulation by society” (Paula Forteza). Supervision becomes not only the responsibility of an authority, but of society as a whole, which can take action to deploy the most appropriate remedies. A role of the regulator is to mobilise and empower the whole of society. To this end, the regulator must be a tool for dialogue between government and its departments, judiciary, researchers from all disciplines, associations and Internet users.
To transform social networks, the regulator must adapt to its purpose and rely on the forces of social networks in order to ultimately become a social network itself.
 This paper does not represent an official view of any institution or person other than its authors.
 On sociology, history and typology of social networks, see in particular the works of Pierre Mercklé, notably La sociologie des réseaux sociaux, La Découverte, 2016 and La découverte des réseaux sociaux. A propos de John A. Barnes et d’une expérience de traduction collaborative ouverte en sciences sociales, in Réseaux, 2013/6, n ° 182, pp. 187 and s.
 The social media timeline dating back to the 1970s is available on Wikipedia: https://en.wikipedia.org/wiki/Timeline_of_social_media . See also V. Schafer, Les réseaux sociaux numériques d’avant, in Le temps des médias, 2018/2, n ° 31, pp. 121 and s.
 On social networks, their architecture and their influence on democracy, see also Amaelle Guiton, Réseaux sociaux : ont-ils enterré le débat public?, in Revue Projet, 2019/4, n ° 371, pp. 26 and s.
 On the parallel between computer code and Parisian architecture under Napoleon III, see Lawrence Lessig, Code. Version 2.0, Basic Books, 2006, p. 127.
 A. Gide, L’évolution du théâtre, Nouveaux prétextes, Mercure de France, 2014.
 The expression is borrowed from the title of a workshop entitled It's the business model, stupid! Targeted advertising and human rights organised at Rightscon 2019 in Tunis on June 13, 2019.
 The fact of requiring a lot of data to better target the public receiving an advertisement is not obvious and must in any case register in accordance with the principle of data minimisation established by Article 5 of the General Data Protection Regulation (2016/679).
 The quote of Netflix’ president, Reed Hastings, taken from Rina Raphaël, Netflix CEO Reed Hastings: Sleep Is Our Competition, Fastcompany.com, June 11, 2017. The available brain time is an expression of Patrick Le Lay, former CEO of TF1 back in 2004.
 The Statista.com site reports that a world average time spent on social networks increased from 90 minutes in 2012 to 136 minutes in 2018.
 The data used for the Cambridge Analytica case is taken from Alex Hern, Cambridge Analytica: how did it turn clicks into votes? Theguardian.com, May 6, 2018. For examples of advertisements developed according to the types of personalities defined see Jeremy B. Merrill and Olivia Goldhill, These are the political ads Cambridge Analytica designed for you, Qz.com, January 10, 2020.
 The speech by Emmanuel Macron at the Internet Governance Forum is available in full via Elysee.fr.
 US Code, chapter 47, section 230.
 For the decision of the Conseil constitutionnel on the LCEN, see Recital 9 of the decision n°2004-496 of June 10, 2004.
 For an illustration of a judgment of the European Court of Human Rights on freedom of expression, see ECHR, Handyside v. the United Kingdom , judgment of December 7, 1976.
 On the case of moderators, in addition to the numerous articles and dedicated reports, see the work of Sarah T. Roberts, Behind the screen, Yale University Press, 2019 and articles by Casey Newton for The Verge, including Bodies in seats of June 19, 2019 available on Theverge.com.
 On Facebook's “cour suprême ”, see among others Le Monde and AFP, « Bientôt une « cour suprême » de Facebook, pour statuer sur les publications supprimées », Lemonde.fr, January 25, 2020; and Brent Harris, Preparing the Way Forward for Facebook's Oversight Board, about.fb.com, January 28, 2020 for all information related to its operation.
 On the power and difficulties of moderation with software, see, for example, Ex Machina: Personal Attacks Seen at Scale, Ellery Wulczyn, Nithum Thain, and Lucas Dixon, WWW '17: Proceedings of the 26th International Conference on World Wide Web.
 On France's position, see the tribune of seven French Ministers: « Mettre fin à l’impunité » on the Web : sept ministres s’engagent à lutter contre la haine en ligne », Lemonde.fr, June 18, 2019.
 The European Commission’s code of conduct of May 2016 on illegal online hate speech is available on the ec.europa.eu website. See also the recommendation of the Commission of March 1, 2018 on measures to effectively fight against illegal online content, C (2018) 1177 final.
 The report of the mission “Regulation of social networks – Facebook experiment” published in May 2019 is available via numerique.gouv.fr.
 On a similar approach that consists in considering that the main problem lies in the organization of online content, see H. Murphy and M. Murgia, Can Facebook really rely on artificial intelligence to spot abuse , Ft.com, November 8, 2019 and, in particular, the conclusion by Ms. Sasha Havlicek of the Institute for Strategic Dialogue: “If you don't address the underlying tech architecture that amplifies extremism through the algorithmic design, then there is no way to outcompete this”.
 S. Abiteboul, G. Dowek, The Age of Algorithms, Cambridge University Press (translated from Le temps des algorithmes, Le Pommier), 2020.
 On the use of digital data, see S. Abiteboul and V. Peugeot, Terra Data, Qu'allons-nous faire des données numériques?, Le Pommier, 2017.
 The Moonshotcve.com site presents the organisation’s work on combating online violence, including all of the work of mapping extremist speech.
 Information relating to the redirect method is available in particular on redirectmethod.org and on the ADL website.
 Just like Life After Hate, other organizations also rely on connecting with former « extrémistes violents », in particular to define the most suitable responses to an extreme online speech. As such, see the Against Violent Extremism program of the Institute for Strategic Dialogue.
 On the regulation by society, see Claire Legros, Paula Forteza: « Les citoyens doivent participer à la régulation des plates-formes numériques », Lemonde.fr, November 19, 2018.