Social Networks and Mediation
The Missing Element in Social Media Governance
In early October, Michael and Susan Dell sat in the Oval Office to report on one of the largest philanthropic commitments in recent history, intended to seed investment accounts for twenty-five million American children.[1] While one may reasonably ask whether the very wealthy use donations of this scale to gain economic or political advantage, the donation was undoubtedly a major benefit for children. Social networks, however, fixed on something else: the face of a woman. Mockery of her appearance, speculation about her health and plain cruelty were multiplied and served to millions of users by systems designed to reward exactly that kind of content.
"New forms of violence are evolving out of our growing reliance and use of online and digital communications" says Amnesty International, and added: "People spread hate, fear, and misinformation on social media in ways we’ve never seen before since they are rarely held accountable for their actions."
These episodes reveal what happens when public and private conversations possess every element capable of accelerating social unrest, and no element capable of processing it.
The failures: dyadic clashes
I remember the period of greatest centralization in content moderation, the so-called “woke era”, roughly between 2018 and 2022, when platforms like Facebook built internal quasi-judicial processes to evaluate speech, complete with panels, appeals and published decisions.[2] “I disapprove of what you say, but I will defend to the death your right to say it,” runs the phrase attributed to Voltaire, which I deeply stand by,[3] and it made me question those procedures from the start.[4]
Despite their institutional and procedural facade, these mechanisms frequently operated under opaque dynamics. There was no genuine principle of bilateralism (a symmetrical right of defense), which turned disputes into dyadic clashes in which the power asymmetry built into the platform’s infrastructure ended in unilateral decisions. Critics from very different positions found the results wanting: a Cato Institute analysis concluded that the Oversight Board did not protect expression consistently,[5] while others noted that a body receiving appeals by the hundreds of thousands issued only a few dozen full decisions a year.[6]
Then the pendulum swung. After the acquisition of Twitter in 2022 and the change in the US administration, platforms dismantled much of that apparatus: third-party fact-checking ended in the United States, restrictions on political topics were lifted, and crowd-sourced Community Notes became the main corrective.[7]
In a few years we moved to the opposite pole: very little process, and recommendation systems that intensify whatever generates the most engagement, misinformation and harassment included.[8]
These are not symmetrical situations. In the first case, the problem was the excessive or arbitrary exercise of moderation power. In the second, it is a combination of a lack of appropriate procedures and the algorithmic amplification of certain content. Both problems share a common institutional origin: the absence of a legitimate mediation architecture.
The tension between freedom of expression and the protection of the community is dyadic when framed as a confrontation between two positions attempting to impose themselves upon one another. The typical response is to choose one of the extremes: more moderation or less moderation.
What is needed is an institutional, architectural mediator that transforms the raw, binary conflict (the dyadic tension between toxic expression and free speech) into a stable dynamic system.
The missing element
In my upcoming book The Constitutive Third, I argue that every persisting dynamic order requires a third element that articulates its opposing poles.[9] The book does not address social networks directly, but its framework applies to any order whose stability depends on a mediating third. The productive question concerns where the articulating function actually sits, and in what condition, more than whether such a third exists. When a proper third degrades, its function migrates: some other element of the system absorbs it and performs it, well or badly. Read this way, the two phases of platform governance look less like opposite policies and more like two successive failures of the same function.
The past era, that of centralized moderation, displays what the book calls parasitic mediation: the persistence of the form of a mediating third after its function has decayed. The Meta Oversight Board had the outward marks of a tribunal (cases, reasoning, something like precedent), yet it was funded through a trust endowed by the company whose decisions it reviewed, and its jurisdiction was defined by that same company. Equality of the parties, consistency of the rule and impartiality are the invariants that identify a genuine third, and an articulator that is also a party to the conflict fails the last of them so completely that the book treats it as a contradiction in terms. Formal validity without functional mediation is precisely what users experienced as opacity.
The present era seems marked by the absence of any mediating institution. The structural analysis suggests something more uncomfortable: the articulating function has migrated into the recommendation system. Today the ranking algorithm decides which disagreements escalate and which fade, which voices reach millions and which reach no one. That system is, in effect, the third of the digital public square, and it is captured by one of the interests it should mediate: the commercial imperative of engagement, which treats hostility as a resource. A captured third does more damage than an absent one, because it still occupies the place where a genuine mediator could stand. Functional migration does not imply that the algorithm acts as a genuine mediator: it is an element that occupies and partially performs an articulating function, yet it does so as a captured element, serving one of the poles it should mediate.
This is why the problem is architectural. More moderation and less moderation both act on content, and both leave the distributing function in the hands of a party.
Mediation: an articulated alternative
An articulated solution would not abolish the tension between free speech and moderation; it would organize that tension into a structure capable of moving forward without the system flying apart.
An individual may fully retain their right to express an opinion, even one that is extreme, unpopular, or profoundly misguided. However, it does not follow that a platform is obligated to use its recommendation systems to exponentially amplify the reach of that opinion.
This distinction allows us to move beyond a series of false dichotomies. Moderating expression is not the same as moderating its distribution. Allowing someone to speak does not entail guaranteeing them amplification. And reducing the spread of harmful behavior does not necessarily require removing the content.
The insight has circulated for some years under the formula “free speech is not the same as free reach,” coined by Renée DiResta and later adopted, at least nominally, by X itself.[10] What the formula lacks is a foundation, and the legal perspective can supply one. Freedom of expression protects an individual prerogative, whereas algorithmic distribution is a matter of institutional architecture. They belong to different categories, and so the two need not be treated in the same way. Speech is something a person does; reach is something a system allocates. And allocation by a system is exactly the kind of power that, in every other domain, we subject to rules of independence, transparency and review.
The third in the distribution rule
If the articulating function now sits in the ranking system, that is where a genuine third has to be built. Panels, juries and appeals matter, but they come afterwards, as escalation pathways. A third located only in review bodies would leave the decisive function (who sees what, and how fast) untouched.
Embryonic examples already exist. Community Notes, usually presented simply as the deregulatory alternative to fact-checking, rests on an interesting mechanism: a note is displayed only when it is rated helpful by users who usually disagree with one another.[11] Researchers have generalized the idea as bridging-based ranking, in which content gains reach according to the agreement it produces across divides.[12] In the terms of the book, this is an attempt to preserve asymmetry: the poles remain distinct, and the system rewards what can travel between them. The mechanism is still slow, limited in coverage and vulnerable to coordination, and it governs annotation more than distribution as a whole. Even so, it shows that a third can be designed into the architecture itself.
When high-velocity, high-toxicity content surges, the same logic can guide the response: temporary friction proportional to viral velocity, context tags from cross-perspective reviewer pools, or routing disputed content into deliberative community spaces. The speech remains, while its explosive energy is absorbed and converted into contextual conversation. The book’s term for this is “plasticity”: the capacity of a third to deform under load, absorb tension without rupturing, and recover its shape when the load passes.
Nested mediation
A plastic mediating architecture cannot operate at a single level, detached from legal systems, democratic norms and human rights standards. The book describes complex stable systems as “nested”, because each level has its own third, and the levels are connected by formal mechanisms of escalation. Applied to social networks, the three levels could look like this:
Micro level: user flags, immediate cooling-off tools, and friction applied by the distribution rule itself during anomalous spikes.
Meso level: decentralized community juries (randomized, representative panels of users, advised by context models) reviewing disputed cases.
Macro level: escalation to independent, cross-platform bodies, accountable to public law and international human rights frameworks.
The macro level is less utopian than it sounds. The European Digital Services Act already contains several of its pieces: certified out-of-court dispute settlement bodies independent of the platforms, systemic risk assessments that must address the design of recommender systems, a crisis response mechanism, and the obligation for very large platforms to offer at least one recommendation option not based on profiling.[13] Whatever one thinks of European regulation in general, these are institutional thirds with legal existence, and they deserve to be discussed on their structural merits.
Caveats and clarifications
This appears to be a suitable institutional possibility, yet there is a conceptual risk: that this third ends up being nothing more than another moderation body, albeit one with more members and more sophisticated procedures.
That would not be sufficient. The constitutive third is not merely an arbiter positioned between two sides; rather, it is an entity that alters the conditions under which those sides interact, enabling a dynamic that could not be sustained through binary confrontation alone.
The book offers a simple test for the difference. Remove the element and observe what happens. If the relation between the parties continues as before, the element was a facilitator, useful and dispensable. If the relation loses its stability, the element was constitutive. A review panel placed on top of an engagement-maximizing feed would pass this test as a facilitator. A distribution rule that rewards cross-perspective agreement, connected to panels and independent bodies through review and appeal, would be constitutive, because removing it would change how the conversation itself unfolds.
For the same reason, the three levels must be connected through review and appeal procedures. Otherwise we risk creating three independent structures instead of a polycentric architecture, a failure the book calls “fragmentation”.
The mechanisms described above also need refinement:
Friction and delay should apply during extraordinary spikes in spread, and never automatically to all controversial material.
Context tags should distinguish between verified information, contested claims, opinion, satire, and content taken out of context. A user panel cannot become a universal arbiter of truth.
Deliberative spaces should be forums of genuine participation, where those expressing minority opinions keep a real voice and visibility.
Human-AI panels require clear rules on selection, independence, rotation, transparency and appeal, and protection against political capture and coordinated manipulation.
The decisive requirement is impartiality. Whoever operates the third cannot be a party to the conflict it mediates. A mediating layer owned, funded and calibrated by the platform alone would reproduce the parasitic form of the previous era. Independence of funding, of appointment and of rule-setting is a structural condition of the whole proposal. The third element does not require unrestricted state regulation of recommendation algorithms; the editorial freedom of the platforms and the rights of users must be taken into account in the design. Platforms may claim curation as their own editorial expression;[14] if so, their editorial freedom belongs to one of the poles, which is precisely why it cannot also serve as the third.
Finally, any intervention must be proportional to the risk and reversible. The system must be able to act in dangerous situations without turning an exceptional measure into a permanent restriction.
The scapegoat as a false third
The Dell episode illustrates one more point. René Girard described how communities in conflict restore a fragile unity by converging on a single victim: the scapegoat unites the group through its exclusion.[15] Online mobs reproduce the mechanism at scale. For a few days, millions of users who agree on nothing else agree on a target, and the resulting sense of shared purpose feels, from inside, like community. Structurally it is a counterfeit of mediation: a unity produced by exclusion, which leaves the underlying conflicts intact and simply relocates them onto someone who cannot answer. An architecture that rewards engagement will tend to manufacture these counterfeit mediations, because nothing produces engagement more reliably than a shared target.[16]
A control architecture vs. an architecture of dynamic stability
It is essential to point out the difference between a control architecture and an architecture of dynamic stability.
In a control architecture, such as the hybrid systems Meta and Twitter built between 2018 and 2022, the goal was to reduce deviation from specific parameters. In an architecture of constitutive mediation, the aim is to preserve the system’s capacity to process its own internal tensions.
This completely shifts the perspective on social media moderation. A social network should not aspire to eliminate disagreement, not even intense disagreement, because disagreement is part of its democratic function. The problem arises when distribution mechanisms transform disagreement into a dynamic of hostility that destroys the conditions necessary for deliberation.
In this theory, therefore, stability does not mean the absence of conflict; it means the existence of an institutional capacity to process that conflict.
Final thoughts
Why should we remain unable to imagine new institutional forms while this situation persists? We can begin to talk about new forms of coexistence on social networks and stop being paralyzed by the current mess.
Free speech and social responsibility must coexist in tension; one should not destroy the other. The mediating architecture should enforce what we can call a principle of "Asymmetric Visibility and Deliberation": protecting the right of speakers to voice extreme or unpopular opinions (preserving free speech), while protecting the community from the algorithmic amplification of targeted harassment and systemic misinformation (preserving social protection). It does so by separating the right to speak from the entitlement to algorithmic distribution.
The principle rests on five structural conditions for a legitimate architecture of constitutive mediation, drawn from the argument above:
Separation of speech and reach: expression is protected as an individual prerogative; distribution is governed as institutional architecture.
Impartiality of the mediator: whoever operates the third cannot be a party to the conflict it mediates.
Cross-perspective validation: amplification and context are earned by what travels between positions, and minority views keep a real place in deliberation.
Proportionality and reversibility: interventions scale with risk and expire when the risk passes.
Nested accountability: each level answers to the one above through transparent procedures of review and appeal, up to public law.
By anchoring social network governance in these five conditions, we could move away from both corporate censorship and algorithmic chaos, toward a resilient, self-correcting digital public square.
The purpose of mediation is to give disagreement a structure within which it can remain a condition of coexistence.
[1] The Oval Office appearance took place on 7 October 2026. The $6.25 billion pledge, announced in December 2025, commits $250 to each of 25 million children. See Hello! Magazine, “Billionaire Michael Dell breaks his silence after wife Susan’s White House appearance” (October 2026), https://www.hellomagazine.com/us/928613/billionaire-michael-dell-breaks-his-silence-after-wife-susans-white-house-appearance/; Proto Thema, “Billionaire Michael Dell defends wife after cruel comments about her appearance at White House” (13 October 2026).
[2] During the period of greatest centralization in content moderation (approximately between 2018 and 2022), companies designed hybrid systems combining artificial intelligence, internal Trust & Safety teams and external entities, such as the Meta Oversight Board (Facebook) or the Trust and Safety Council (Twitter).
[3] The phrase was not written by Voltaire. It is a synthesis of his thought by his British biographer Evelyn Beatrice Hall, in The Friends of Voltaire (1906).
[4] Gastón Rey, “Freedom of speech and social networks immunity,” Medium (2020), https://medium.com/@gastonrey76/freedom-to-speech-in-the-case-of-social-networks-inmunity-bfc49d3b717f
[5] The report found that when the Board reviewed content categorized as left-leaning, it was allowed to remain online 90% of the time, against 41% for content identified with the political right. David Inserra, “New Report Analyzing Meta’s Oversight Board Finds Mixed Impact Advancing Free Expression,” Cato Institute (18 June 2025), https://www.cato.org/blog/new-report-analyzing-metas-oversight-board-finds-mixed-impact-advancing-free-expression
[6] Engadget, “Meta’s Oversight Board made just 53 decisions in 2023” (2024). These entities had structural limitations regarding their independence, concerning the origin of their funds, the delineation of their jurisdiction, and their reliance on the platform’s infrastructure.
[7] The internal regulatory landscape of these platforms changed radically following Elon Musk’s acquisition of Twitter in 2022 (renamed X) and became further entrenched after the change in the US administration. In its corporate policy updates on Meta Newsroom (January 2025), Meta announced the end of its third-party fact-checking program in the United States, the adoption of a Community Notes model, and the lifting of previous restrictions on political topics such as gender identity and immigration.
[8] El País, “Ending fact-checking on social media fuels hate speech and harassment, experts warn” (10 January 2025).
[9] Gastón Rey, The Constitutive Third: The Architecture of Stable Dynamic Systems (2026, under editorial review, College Publications), chapters 15 to 17.
[10] Renée DiResta, “Free Speech Is Not the Same As Free Reach,” Wired (30 August 2018); X, “Freedom of Speech, Not Reach: An update on our enforcement philosophy” (April 2023).
[11] Stefan Wojcik et al., “Birdwatch: Crowd Wisdom and Bridging Algorithms can Inform Understanding and Reduce the Spread of Misinformation” (2022).
[12] Aviv Ovadya and Luke Thorburn, “Bridging Systems: Open Problems for Countering Destructive Divisiveness across Ranking, Recommenders, and Governance,” Knight First Amendment Institute (2023).
[13] Regulation (EU) 2022/2065 (Digital Services Act), articles 21 (out-of-court dispute settlement), 34 and 35 (systemic risk assessment and mitigation), 36 (crisis response mechanism) and 38 (recommender systems not based on profiling).
[14] In Moody v. NetChoice, LLC, 603 U.S. 707 (2024), the US Supreme Court treated platforms’ curation of their feeds as editorial activity protected by the First Amendment.
[15] René Girard, Le Bouc émissaire (Paris: Grasset, 1982); English translation The Scapegoat (Johns Hopkins University Press, 1986).
[16] This is not merely a form of false mediation; it is also a real mechanism for generating social cohesion, albeit a cohesion achieved through violence and exclusion. For this reason, one might also speak of “counterfeit constitutive mediation” or “violent pseudo-mediation.” The Girardian mechanism does indeed produce a form of unity, but it does so by sacrificing a third party, rather than by introducing an entity that might legitimately articulate the poles. This distinction is highly significant for the theory, as not every third party produces a legitimate constitutive mediation: some generate stability by eliminating an element of the system.
