Archive for category: artificial intelligence

I arrive at the last text of this series with a sense (rare, and useful) of having found a problem whose conceptual architecture is not yet consolidated in the literature, and to which my training allows a singular entry. I want, in this closing, to do three things: make explicit the synthesis that connects the four layers I developed; articulate the relationship between this program of reflection and my doctoral thesis, which has a different object but compatible toolkit; and map four research questions that this synthesis suggests for the medium-term horizon.

The unifying thesis

I will state the thesis as synthetically as possible, and then unpack it.

Vitrinization, the burnout society, the degraded equilibrium of platforms, the transformation of public deliberation by generative AI – all these phenomena are contemporary instances of a single object: the political economy of cognition under personalized choice architectures at scale. The transversal operator that connects all layers is framing.

The intuition is the following. Framing, as Tversky and Kahneman demonstrated in 1981, is the elementary cognitive gesture by which the presentation of a problem produces the preference about the problem. That gesture can operate:

• At an individual scale – a human interlocutor, a therapist, a laboratory experiment.
• At an algorithmic scale – the TikTok feed, the Google ranking, the order of search results.
• At a strategic scale – behavioral advertising that orbits political conflict, electoral microtargeting.
• At an epistemic scale – the LLM response that summarizes a complex debate, the AI that assists the decision-maker.

At each scale, framing operates with the same formal structure but at radically different magnitudes. A Tversky-Kahneman experiment affects dozens of subjects. A recommendation algorithm operates on billions. An LLM response cuts across millions of queries per day. The magnitude is new; the mechanism, not.

The consequence is that problems that previously seemed to belong to separate domains – individual cognitive biases, institutional design, platform regulation, AI governance – are, in fact, instances of the same problem. And they should be analyzable with a common vocabulary.

How the four layers stitch together

To briefly recapitulate:

The cognitive layer (click to see) establishes the microfoundation. Framing acts on the individual, producing preferences contingent on the mode of presentation. Tversky-Kahneman, and the work I did in my master’s thesis on framing in GMCR applied to the Cocó viaducts conflict, provide the starting point – with the important methodological caveat that elicitation of these preferences is itself subject to framing, configuring a recursion that is part of the method and cannot be eliminated from it.

The strategic layer (click to see) shows how individuals whose preferences depend on presentation produce, in aggregate, collective equilibria. Le Bon, Freud, and Türcke supply the historical lineage of the problem of crowds; Schelling, Festinger, and Akerlof-Kranton supply the modern game-theoretic vocabulary. The result, under the current architecture of platforms, is a Nash equilibrium of mutual exhaustion – vitrinization as the dilemma of the attention commons.

The institutional layer (click to see) shows that the architecture under which these equilibria form is not given by natural law. Castells provides the broad sociological frame, with the network society as a historical construction in dispute; Acemoglu and Johnson, in Power and Progress, show the generic thesis of technological underdetermination; Sunstein, with the notion of sludge, provides the operational instrument. The European DSA, the Australian law of 2024, and the Brazilian Bill 2630/2020 are institutional experiments in progress, whose empirical effects are beginning to be measured.

The algorithmic layer (click to see) shows that the entry of generative AIs adds a new layer of cognitive mediation that acts recursively on the three earlier ones. It amplifies or substitutes for human System 2 (Stiegler as pharmakon); it redistributes productivity unequally (Brynjolfsson, Li, and Raymond); it closes the reflexive training-production-training circuit (Rahwan and team; emerging literature on model collapse); and it calls into question Lévy’s bet on collective intelligence – when the collective comes to include non-human agents, the concept needs to be rediscussed.

The stitching: the four layers form a stack. Each layer constrains the layer above and modifies the layer below. Layer 4 produces content that affects the framing of Layer 1; Layer 1 produces preferences that aggregate into the equilibrium of Layer 2; Layer 2 produces phenomena that motivate interventions in Layer 3; Layer 3 designs the environment in which Layer 4 is trained. The system is closed, and therefore requires systemic analysis.

Why this synthesis is a contribution

There is a piece of literature for each layer. There is little literature – almost none in the lusophone space – that integrates the four. There are books in the philosophy of technique (Stiegler, Han, Türcke, Sibilia), in the political economy of technology (Acemoglu-Johnson, Zuboff), in the sociology of the network society (Castells), in behavioral economics (Thaler-Sunstein, Kahneman, Sunstein-Sibony-Kahneman in Noise), and in applied game theory (Schelling, and more recently the work of Acemoglu-Ozdaglar-Siderius). There is also the technical literature on LLMs and machine behavior (Rahwan and team, Brynjolfsson, the alignment community).

Little of this material talks to itself. Philosophers do not model; modelers do not philosophize. Behavioral economists usually stop at Layer 1; game theorists usually stop at Layer 2; political scientists go to Layer 3 but rarely connect it to the quantitative toolkit of the previous ones. The AI community has its own journals, conferences, and vocabulary, with little interface with the political economy of information.

My position combines a toolkit from game theory (GMCR), behavioral economics (framing effects), agent-based modeling (current focus of the doctorate), concrete institutional experience (Government of Ceará, Mayor’s Office of Fortaleza, UNICEF projects), and active interest in generative AI. It is a rare combination (I think), and I have discovered, writing this series, to be an operative combination. It allows formulating questions that no discipline, in isolation, formulates.

Articulation with the doctoral thesis

I want to be clear about how this series relates to the work I have been developing in the doctorate, because clarity matters.

My thesis, at the School of Economics of the University of Porto, investigates the conditions under which cooperation emerges among cognitively heterogeneous agents in agent-based models. It is a theoretical-methodological problem, not a substantive problem about social networks. It has a defined scope, supervision, schedule, and does not overlap with the object of this series.

But the relationship between the two is precise and fertile. ABM with cognitively heterogeneous agents is exactly the natural toolkit to quantitatively investigate the substantive phenomena this series maps. Schelling, who sustains much of the argument of Layer 2, is the grandfather of the methodology. Vitrinization, seen through the lens of Layer 2, is a problem of emergent cooperation (or its failure) among cognitively heterogeneous agents under a specific incentive architecture – exactly the kind of problem ABM is built to investigate.

The best formulation of the relationship is therefore: the thesis develops the toolkit; this program of reflection maps the substantive domain in which that toolkit can be, in parallel and future research, applied. The two works feed each other – the conceptual clarity gained here orients the methodological choices of the thesis; the technical rigor of the thesis provides the instrument to return to these questions with greater robustness in parallel papers and on the post-doctoral horizon.

Four research questions for the horizon

The four questions I list below are, in my assessment, the most promising for a medium-term research program – not for the current thesis, which has its own scope, but for what comes after and for parallel papers that naturally connect to the toolkit I am consolidating.

Question 1: ABM modeling of the dilemma of the attention commons with cognitively heterogeneous agents.

How does one model, in ABM, the emergence (or non-emergence) of cooperation in environments where agents have different susceptibilities to framing, different weights for social comparison, different bounded-rationality frontiers? Applied to the problem of vitrinization: under what parametric conditions (cognitive heterogeneity, intensity of algorithmic framing, exit cost) does an equilibrium of mutual exhaustion emerge? Which perturbations in the architecture displace the system into other basins of attraction? It is a question that naturally connects the thesis toolkit to the substantive domain of this series, and that could generate a sequence of articles on the post-thesis horizon.

Question 2: Preference elicitation under framing recursion – honest protocols.

Recognizing that the instrument of preference measurement is subject to the same bias it studies, is it possible to design elicitation protocols that internalize this recursion as part of the method? A convincing answer would involve triangulation among direct elicitation, revealed preferences, and experiments with varied framings – and, ideally, formal models of the distortion that the protocol itself introduces. It would be a methodological contribution of interest to all applied work in GMCR and in modeled conflict in general. It has direct affinity with what I learned in the master’s thesis.

Question 3: Algorithmic framing as treatment variable in electoral behavior.

Is it possible to design quasi-experimental studies that isolate the effect of algorithmic framing on electoral behavior? I have a descriptive pilot on the blog itself – the analysis of media sentiment in the 2024 Fortaleza municipal elections. What is missing is moving from description to causal identification. There is starting literature, dominated by studies with privileged access to platform data; the interesting question is what can be done with public Brazilian and Portuguese data, mobilizing observational causal inference (synthetic control, discontinuity in algorithmic changes, instruments through exogenous exposure shocks).

Question 4: Human-AI coupling in political decision-making – experiment.

How does human-AI coupling change the quality of collective decision in conflict problems? More specifically: when a decision-maker consults an LLM to analyze a conflict, does the framing the model produces in its response alter the equilibrium of the conflict? And in which direction – convergence or divergence? Increase or reduction of epistemic diversity?

It is, in my reading, the most original and potentially impactful question. Possible design: an experiment in which real decision-makers solve conflicts modeled in GMCR – or in ABM with humans and AI agents interacting in the same environment – half with LLM assistance and half without, measuring the stability of the equilibrium reached, the quality of deliberation, and the polarization of preferences before and after. It connects Layers 1, 2, and 4 directly, and opens exactly the technical space my thesis is consolidating: ABM with cognitively heterogeneous agents, now extended to include machine agents alongside human agents.

Closing

I began this series trying to name a diffuse unease – a structural fatigue, a vitrinization that seems to define the present. I end it with a conceptual architecture and a parallel research program.

What happened in the middle was the realization, which I consider useful, that we are not facing separate problems – privacy, polarization, digital mental health, automation, AI governance. We are facing manifestations of a unified problem: how societies organize, govern, distribute, and dispute the collective cognitive resource, given the mediation technologies currently available.

This problem is old at heart. It is in Plato on writing, in Le Bon on crowds, in Marx on alienation, in Adorno on the culture industry. But it takes a specific form in this decade, with specific instruments and at a specific scale. To think it well requires a vocabulary still being built, and quantitative tools still being developed – including the toolkit I am consolidating, in another front, in my thesis.

I hope this series has been a small contribution to that construction. And that it orients, in the coming years, both the doctoral work in progress and the parallel and subsequent agenda this reflection has made visible to me. I found the axis. Now it is work.

Deoclécio Paiva de Castro

Ph.D. Candidate in Economics, Faculty of Economics, University of Porto
M.Sc. in Mathematical Modeling and Quantitative Methods, UFC
B.Sc. in Industrial / Production Engineering

References

ACEMOGLU, D.; JOHNSON, S. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. New York: Public Affairs, 2023.

ACEMOGLU, D.; OZDAGLAR, A.; SIDERIUS, J. “A Model of Online Misinformation”. Review of Economic Studies, 2024.

AKERLOF, G. A.; KRANTON, R. E. Identity Economics: How Our Identities Shape Our Work, Wages, and Well-Being. Princeton: Princeton University Press, 2010.

BRYNJOLFSSON, E.; LI, D.; RAYMOND, L. R. “Generative AI at Work”. Quarterly Journal of Economics, vol. 140, no. 2, 2025.

CASTELLS, M. The Rise of the Network Society. Oxford: Wiley-Blackwell, 2010. [First edition: 1996.]

CASTRO, D. P. Efeito de Enquadramento no Modelo de Grafos para Resolução de Conflitos com uma Aplicação ao Conflito das Obras de Construção dos Viadutos do Cocó [Framing Effect in the Graph Model for Conflict Resolution with an Application to the Cocó Viaducts Construction Works]. Master’s thesis – Federal University of Ceará, Fortaleza, 2022.

CASTRO, D. P. “Cinturão Digital do Ceará e as mídias digitais” [The Ceará Digital Belt and Digital Media]. Proceedings of Social Media Brasil 2011, Fecomercio, São Paulo, 2011.

CASTRO, D. P. Posts on ChatGPT o1, AI, and electoral analysis. deocleciocastro.com, 2024.

FREUD, S. Civilization and Its Discontents. New York: W. W. Norton, 2010. [Original German: 1930.]

HAN, B.-C. The Burnout Society. Stanford: Stanford University Press, 2015.

HIPEL, K. W.; KILGOUR, D. M.; FANG, L. Conflict Resolution Using the Graph Model: Strategic Interactions in Competition and Cooperation. Cham: Springer, 2018.

KAHNEMAN, D. Thinking, Fast and Slow. New York: Farrar, Straus and Giroux, 2011.

LE BON, G. The Crowd: A Study of the Popular Mind. Mineola: Dover, 2002. [Original French: 1895.]

LÉVY, P. Collective Intelligence: Mankind’s Emerging World in Cyberspace. New York: Plenum, 1997. [Original French: 1994.]

RAHWAN, I. et al. “Machine Behaviour”. Nature, vol. 568, pp. 477-486, 2019.

SCHELLING, T. C. Micromotives and Macrobehavior. New York: W. W. Norton, 1978.

SHUMAILOV, I. et al. “AI Models Collapse When Trained on Recursively Generated Data”. Nature, vol. 631, pp. 755-759, 2024.

SIBILIA, P. O Show do Eu: a intimidade como espetáculo [The Spectacle of the Self: intimacy as spectacle]. Rio de Janeiro: Nova Fronteira, 2008.

STIEGLER, B. Symbolic Misery, Volume 1: The Hyperindustrial Epoch. Cambridge: Polity Press, 2014.

SUNSTEIN, C. R.; SIBONY, O.; KAHNEMAN, D. Noise: A Flaw in Human Judgment. New York: Little, Brown, 2021.

TÜRCKE, C. Erregte Gesellschaft: Philosophie der Sensation. Munich: C. H. Beck, 2002.

TVERSKY, A.; KAHNEMAN, D. “The Framing of Decisions and the Psychology of Choice”. Science, vol. 211, no. 4481, pp. 453-458, 1981.

ZUBOFF, S. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. New York: Public Affairs, 2019.

There is a moment, in any serious reflection on social networks, when the technological temptation reappears. It is the moment when one says: “the networks are like that, it is the nature of the thing”. Or its symmetrical reverse: “these technologies are inherently exploitative… extracting attention, data, and behavior is what they were built to do”. In either case, technology is ascribed a fixed nature that determines the social outcome.

I want to argue, in this third text of the series, that this ascription is wrong – and that it conceals the political operation that produced the current state of affairs. Social network technologies are not intrinsically anything. What we see today is the result of specific institutional choices, made by specific actors, in specific regulatory contexts. And what has been done can be undone or redone.

I have personal empirical evidence of this. In 2011, I was responsible for the creation and coordination of the social networks of the Government of the State of Ceará. I later worked for years in the Mayor’s Office of Fortaleza, with activity on projects such as the UNICEF Urban Centers Platform and the Secretariat for Conservation and Public Services. I was also responsible for the creation and coordination of FB Podcast, hosted at Colégio Farias Brito – a private secondary-school institution – which became notable as the first podcast studio in Brazil with free public access, dedicated to discussions in education, innovation, and entrepreneurship. I lived from the inside the phase in which it was possible to imagine social networks and digital infrastructure as instruments for participation, transparency, and dialogue between State, society, and citizen.

That imagination was not naïve; it was plausible and theorized. The networks could have become what they promised. They did not because specific choices were made elsewhere. It is this argument I want to develop, anchored in a literature that combines network sociology, political economy, and institutional design.

Castells: the network society as social fact, not technological destiny

It is not possible to discuss what the networks “could have been” without going through Manuel Castells. The trilogy The Information Age – The Rise of the Network Society (1996), The Power of Identity (1997), End of Millennium (1998) – is the most influential sociological reference of the last quarter-century on what the network has structurally changed in contemporary societies.

The central thesis is twofold. First, there is a change in the mode of development: industrial capitalism gave way to informational capitalism, in which productivity depends crucially on the capacity to generate, process, and apply information. Second, there is a change in the social form: the dominant social structure ceases to be hierarchical and becomes network-shaped – composed of nodes and flows, with an inclusion/exclusion logic that operates by being in or out of the network, no longer only by static class.

The political consequence, which Castells developed in later works – Communication Power (2009) and Networks of Outrage and Hope (2012) – is that the network is simultaneously a terrain of new forms of domination and of new forms of resistance. The movements of the Arab Spring, the Spanish 15M, Occupy, the June 2013 protests in Brazil, and more recently the 2019 Chilean uprising and the articulations of Brazilian indigenous movements are all comprehensible within Castells’s frame. The network was not destined for extraction; neither is it destined for liberation. It is contested terrain.

Castells’s argument lends sociological density to what I am saying: the current configuration of commercial social networks is not “what technology produces”. It is what a specific allocation of power produced over that technology, at a specific historical moment, with little organized contestation. And organized contests can change the outcome – as the very network-based social movements Castells documents show.

Power and Progress: technology is underdetermined

Daron Acemoglu and Simon Johnson published in 2023 Power and Progress, a book that may be the most important recent economic contribution on the topic. The central thesis can be summarized in one sentence: the path a technology takes is not determined by the technology itself; it is determined by the allocation of power around it.

The authors make a brilliant historical reading. The same mechanical technology that could have benefited weavers in early-nineteenth-century England was used, given the prevailing relations of power, to displace them. The same electrification that could have shortened the working day was used, given the prevailing relations of power, to intensify it. The story is not “technology liberates” nor “technology enslaves”. It is: technology is malleable; organized power decides the form of the malleability.

Applied to social networks, the argument is disconcerting. The same network communication technology that enables global access to knowledge, horizontal political articulation, unprecedented cultural expression – and that produced real episodes of these, as Castells documented – was configured, given the property regime and regulatory environment of the early 2000s in the United States, as infrastructure for the extraction of behavioral surplus. In the precise sense Shoshana Zuboff gives in The Age of Surveillance Capitalism (2019): our attention and our behavioral traces are the raw material of prediction products sold in secondary markets.

There was no inevitability in this. There was decision.

Who designs the sludge

Cass Sunstein, in the book that bears its name (Sludge, 2022), introduced a notion complementary to his famous “nudge”. If a nudge is institutional friction designed to facilitate desired behavior, sludge is friction designed to hinder it.

The important observation is that digital platforms are specialists in asymmetric sludge. It is trivially easy to post and like; it is deliberately hard to delete the account, disable notifications, opt out of the algorithmic feed, export your data. The interface is designed for low friction toward engagement and high friction toward exit. This asymmetry of frictions is not incidental. It is architecture – and architecture is choice.

If the equilibrium of mutual exhaustion I described in the previous text is sustained by this asymmetry of frictions, then redesigning the frictions is a legitimate instrument of public policy. The point is not to ban the networks; it is to invert the sludge – to require that exit be as easy as entry, that control be default, that data be portable, that opting out of personalization be one click.

This is part of what the Digital Services Act in Europe is trying to do, especially by requiring more transparency, expanding user control, and allowing users of large platforms to opt for non-personalized recommendations. Obligations of algorithmic transparency, restrictions on targeted advertising to minors, and the requirement of systemic-risk reports from very large platforms all point in that direction. It is one of the first serious attempts, on a democratic scale, to redesign the sludge of platforms.

Other experiments are underway. Australia approved in November 2024 a law that establishes a minimum age of 16 for accounts on certain social platforms, imposing on the companies the obligation to take reasonable steps to prevent under-16 accounts. Brazil has discussed for years the Fake News Bill (PL 2630/2020) in variants ranging from promising to problematic, depending on the wording. The United States debates Section 230, though with legislative paralysis that reflects the veto power of the platforms themselves. China, with its autocratic regulatory model, imposes screen-time restrictions on minors that would be inadmissible in democracies but that show, once again, that technology is malleable when regulatory power wants it to be.

The relevant academic question – and one that enters the research agenda this series sketches – is: which of these redesigns produces, empirically, a displacement of the equilibrium toward less exhausting states? Theory says it should work. Empirical evidence is being built. There is here a clear research opening in comparative policy evaluation with quantitative method.

Benkler, Schneider, and the other possible architecture

Yochai Benkler, in The Wealth of Networks (2006), had already described a possible alternative architecture: commons-based peer production. Wikipedia is the classic example: a non-market cooperative infrastructure produced in network, governed by collectively agreed rules, and in operation for more than twenty years.

More recently, Nathan Schneider in Governable Spaces (2024) and the literature on “platform cooperativism” explore models of democratic governance of platforms – digital cooperatives, collectively owned platforms, decentralized networks such as Mastodon (on the ActivityPub protocol) and Bluesky (on the AT Protocol). These are not utopian proposals; they are experiments in operation, with documentable successes and failures.

The point is not that these alternatives are fully realizable at global scale. It is that their existence proves that the current configuration is not the only possible one, and therefore the current configuration is a choice.

Back to Ceará in 2011

I rediscover here, in theorized form, what I lived as practice. In 2011, in Ceará, we were still using the networks in a “pre-engagement-optimized-by-algorithm phase”. The feed was chronological, incentives for the user were different, targeted advertising was at another stage of development, and, fundamentally, the marginal profit of the network did not yet depend so centrally on prolonged capture of attention. It was possible, in that design, to do public communication that answered questions, expanded transparency, and established reasonable dialogue with citizens. It was the environment in which the Cinturão Digital do Ceará – 3,000 km of fiber optic, 92% urban population coverage – made sense as a democratic project.

I am not nostalgic for the tools of 2011. They were more primitive in many respects. But the institutional configuration and business model of the platforms at that moment allowed a public use that became progressively more difficult. What changed was not the network technology in any fundamental sense… it was the choices of the platforms (and some squared head persons in my way) and the regulatory environment that permitted them.

This observation has particular weight for any public manager who lived through that transition: it is lived empirical proof of Acemoglu and Johnson’s argument, and direct resonance with Castells’s thesis. Technology is underdetermined. Its destiny is power allocation.

The political conclusion

There is a difficult but inescapable conclusion. If vitrinization is the result of architecture, and if architecture is choice, then vitrinization is a problem of public policy, not an individual or moral problem. To ask people to be less vain, more conscious, more resilient, is to ask them to assume individual costs to solve a systemic problem. It works as poorly as asking individuals to solve global warming by changing light bulbs.

Intervention has to be structural. And structural means institutional. It means redesigning incentives, frictions, metrics, data-ownership rules, governance models.

But there is an additional problem, which opens the topic of the next text. The very tools of analysis – including behavioral economics, game theory, and network sociology – are being transformed in real time by the emergence of generative AIs. AIs are not just one more object on the regulatory agenda. They are a new layer in the cognitive architecture we are trying to analyze – and one that changes what it means to analyze.

That is the topic of the next text (CLICK HERE FOR PART 4).

References

ACEMOGLU, D.; JOHNSON, S. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. New York: Public Affairs, 2023.

BENKLER, Y. The Wealth of Networks: How Social Production Transforms Markets and Freedom. New Haven: Yale University Press, 2006.

CASTELLS, M. The Rise of the Network Society. Oxford: Wiley-Blackwell, 2010. [The Information Age, vol. 1. First edition: 1996.]

CASTELLS, M. The Power of Identity. Oxford: Wiley-Blackwell, 2010. [The Information Age, vol. 2. First edition: 1997.]

CASTELLS, M. End of Millennium. Oxford: Wiley-Blackwell, 2010. [The Information Age, vol. 3. First edition: 1998.]

CASTELLS, M. Communication Power. Oxford: Oxford University Press, 2009.

CASTELLS, M. Networks of Outrage and Hope: Social Movements in the Internet Age. Cambridge: Polity Press, 2012.

CASTRO, D. P. “Cinturão Digital do Ceará e as mídias digitais” [The Ceará Digital Belt and Digital Media]. Proceedings of Social Media Brasil 2011, Fecomercio, São Paulo, 2011.

SCHNEIDER, N. Governable Spaces: Democratic Design for Online Life. Berkeley: University of California Press, 2024.

SUNSTEIN, C. R. Sludge: What Stops Us from Getting Things Done and What to Do about It. Cambridge, MA: MIT Press, 2022.

ZUBOFF, S. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. New York: Public Affairs, 2019.

There is something in the present that leaves me intellectually uneasy, and I have been trying to name this discomfort. It is not a private matter: it has surfaced in conversations with colleagues in Portugal, with friends in Brazil, with people I knew from my earlier public-sector work and who today hold a range of positions. There is a diffuse exhaustion, the sense that we are all carrying out a task whose terms we never formally accepted, a permanent demand to appear, to perform, to remain relevant, and an exhaustion that does not yield to eight hours of sleep, a vacation, or therapy. The phenomenon is collective enough not to be an individual symptom, and structural enough not to be a matter of willpower. It is something to be thought through.

I bring to this thinking a specific trajectory. In 2011, I worked as a strategic-planning officer for digital media in the Government of the State of Ceará. That year I presented, at the Social Media Brasil conference held at Fecomercio in São Paulo, a case on the use of digital media articulated to the Cinturão Digital do Ceará project, a state-owned fiber-optic infrastructure of 3,000 km, with planned coverage of 92% of the state’s urban population, designed to support distance education, telemedicine, public transparency, and direct dialogue channels with citizens. The text I wrote argued, citing explicitly Pierre Lévy on cyberspace, Foucault on power as productive network, Santaella and Lemos on connective cognition, and Rosenau on governance, that the combination of the infrastructure plus the qualified presence of public power on social networks could generate “a democratic and participatory environment, [with] broad-ranging discussions for the emergence of a network of collaborative solutions”. The normative horizon, stated explicitly in the conclusion of that text, was Lévy’s collective intelligence at the service of what Liberation Theology has called a just and supportive society.

Almost fifteen years later, it is difficult to reread that text without discomfort. Not because the premises were wrong… the theoretical toolkit was robust, and several of the references are the same ones I encounter today in serious discussions about networks, but because the institutional configuration that followed made that aspiration almost unreachable. Not through failure of the technology, but through capture of its governance. The same platforms that then seemed infrastructure for participation became infrastructure for the extraction of behavioral surplus. Lévy’s collective intelligence was replaced by what I have been calling, in conversations and drafts, vitrinization (from the Portuguese ‘vitirine’ – ‘shop window’, in English): people constructing appearances, lives, happinesses, and opinions that exist only in the digital field, always artificially shaped to fit patterns that exist precisely to compare us to one another. A world of digital ideas, in an almost Platonic sense, where lived life becomes a paler shadow of the posted life.

I am, at the moment of writing, reading Freud’s Civilization and Its Discontents. I had already read Guy Debord’s The Society of the Spectacle, Byung-Chul Han’s The Burnout Society, and Paula Sibilia’s O Show do Eu [The Spectacle of the Self]. I recently began Gustave Le Bon’s The Crowd: A Study of the Popular Mind, which Freud, incidentally, discusses explicitly in Group Psychology and the Analysis of the Ego, from 1921. There is a clear lineage here on collective behavior: Le Bon, Freud, Türcke, Han. The diagnoses are phenomenologically precise. I miss, however, something my training allows me to look for: a theory that explains not only what we feel, but why this equilibrium is stable. Why people who know that the networks leave them worse off continue using them. Why “the machine that cannot stop” is, at once, individually rational choice and collective trap.

This question is close kin to those I have been working on in my doctorate. I am a doctoral student in Economics at the School of Economics of the University of Porto, with current research centered on the emergence of cooperation in agent-based models (ABM) with cognitively heterogeneous agents. Earlier, at the Federal University of Ceará, I defended a master’s thesis on framing effects in the Graph Model for Conflict Resolution, applied to the conflict of the Cocó viaducts in Fortaleza. The tools (game theory, behavioral economics, multi-agent modeling) converge on a common object: how individual decisions, under specific choice architectures, produce collective outcomes that no one asked for but everyone sustains.

I intend, in this series of five texts, to organize a conceptual architecture that connects the humanist diagnosis of contemporary exhaustion (Lévy critically revisited, Castells, Han, Türcke, Sibilia, Debord, Le Bon, Freud) with the quantitative toolkit of behavioral economics and game theory. The background thesis is simple and perhaps still underexplored: vitrinization is a contemporary instance of a broader problem – the political economy of cognition under personalized choice architectures at scale.

This problem operates in four layers, and I will dedicate one text to each:

1. The cognitive layer, where framing acts on the individual (this text).
2. The strategic layer, where framing produces non-cooperative collective equilibria.
3. The institutional layer, where the architecture of the game is a political choice, not a technological destiny.
4. The algorithmic layer, where generative AI enters as a new layer – simultaneously amplifier and attenuator.

A fifth text will close the series with the synthesis and a research agenda that articulates this program with my doctoral thesis. Let us proceed to the first layer.

Framing as cognitive technology

In 1981, Amos Tversky and Daniel Kahneman published in Science an experiment that became classic. They presented subjects with a public-health problem (a disease that could kill 600 people) and two treatment options. The difference between the two experimental groups did not lie in the options: it lay in the framing. For one group, outcomes were described in terms of lives saved; for the other, in terms of lives lost. The choices reversed. Preferences, supposedly stable according to standard economic theory, turned out to be sensitive to the frame in which the problem was presented.

This finding opened decades of research in behavioral economics and established an epistemologically fundamental point: framing is not noise over preference; it is constitutive of it. In many decision contexts, the observed preference is not independent of the framing that produces it. It emerges from the interaction between subject, context, and form of presentation.

When I worked on my master’s thesis applying the framing effect to the Graph Model for Conflict Resolution – in the concrete case of the Cocó viaducts conflict in Fortaleza – I was operationalizing this intuition in a multilateral conflict setting. Decisions on major urban works involve actors with apparently fixed preferences: environmental movements, government, contractors, residents. What I showed is that these preferences, and therefore the stable equilibria of the conflict, change radically depending on the framing of the options.

It is necessary, however, to be methodologically honest about a central difficulty in the use of GMCR and of conflict models in general. The elicitation of decision-makers’ preferences is, by far, the hardest problem of the method. It is not just one step among others: it is the step at which everything can collapse. Decision-makers rarely manage to articulate complete preferences over all possible states of a conflict. When they do articulate, they do so incompletely, inconsistently, context-dependently, and (here a disconcerting circle closes) sensitively to the framing of the interview itself. The instrument we use to measure preferences in conflict contexts is itself subject to the same mechanism we are trying to model. This is a genuine recursion, not a methodological flaw to be circumvented. To recognize it openly is, in my assessment, part of what may become an original contribution of the research program this series sketches.

From experiment to algorithm: framing industrialized

The point I want to defend now is that something profound has changed between the Tversky-Kahneman experiment and the present. In 1981, framing was artisanal: a researcher designing two versions of a problem, a doctor speaking with a patient, a lawyer preparing an argument. It was a manual instrument.

In the 21st century, framing has become computational infrastructure. The Instagram feed algorithm, TikTok, X, YouTube execute, thousands of times per second, the same gesture that Tversky and Kahneman executed one experiment at a time. Each user, in each session, receives a customized framing: the same content, ranked in different order, juxtaposed with different other content, in different emotional moments, generates different decisions. Cass Sunstein, in works after Nudge, called this “personalized choice architecture”.

The difference in scale is not trivial; it is qualitative. A framing experiment affects dozens of subjects. A recommendation algorithm operates over billions of users simultaneously, with real-time feedback and adaptive optimization of the framing that maximizes engagement. Daron Acemoglu and Simon Johnson, in Power and Progress (2023), describe what one can call, with them, a preferential manipulation technology at industrial scale.

The empirical evidence anchoring this argument is strong, though a critical reading of the most well-known study of the topic requires care. In 2015, Eytan Bakshy and a Meta team published in Science a seminal paper documenting that algorithmic ranking contributes materially to the composition of what each user sees — although the authors themselves, funded by Meta, attributed greater weight to individual choices than to the algorithm, a reading that received substantial methodological criticism. Subsequent studies, with expanded access to Meta data, indicated that changes in the algorithmic feed can alter patterns of exposure and use, though effects on political attitudes and polarization are more difficult to identify. For the argument of this text, the central point is more restricted: the algorithm participates in the composition of the decision environment. The algorithm, in other words, is industrial framing.

Lévy revisited: from aspirational cyberspace to instrumental architecture

It is worth, at this point, returning to Pierre Lévy, an author whose presence in my 2011 repertoire I already noted in the opening of this series. Lévy described cyberspace, in Cyberculture (1997), as “the new means of communication that arises from the worldwide interconnection of computers”, including in this description not only the infrastructure but also “the human beings who navigate and nourish this universe”. The formulation is hospitable: the network includes, by construction, the human agents who inhabit it and constitute it as co-authors.

What happened between 1997 and 2026 is, in part, an inversion of that hospitality. The network still includes humans, but no longer as co-creators of a space of collective intelligence rather as input to a behavioral-extraction operation that frames their choices to optimize engagement. The difference is not in the Lévy of 1997; it is in the institutional configuration that took over the infrastructure he was describing. This will be the central argument of the third text in this series.

The epistemological consequence

There is a consequence here that seems to me underexplored in Brazilian, Lusophone, and even international academic public debate. If we accept the behavioral evidence (that framing constitutes preference) and if we accept the empirical evidence (that the algorithm frames at scale) then preferences expressed in digital space should not be read as simple pre-existing preferences of users. They are co-produced preferences: part comes from the subject, part comes from the computational architecture.

This destabilizes both the liberal discourse about “freedom of choice” in media consumption, and the reactive discourse about “total manipulation”. The truth is more uncomfortable: we are subjects partly authors, partly co-authored by an infrastructure we do not control and whose design we never voted on.

The practical question that opens is: if framing is a technology, and if this technology operates on our cognition at scale, who has the right to use it? Under what rules? With what responsibilities? These questions lead directly to the next layer — what happens when billions of individuals co-produced by algorithms interact with one another?

That is the topic of the next text. (CLICK HERE FOR PART 2)

References

ACEMOGLU, D.; JOHNSON, S. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. New York: Public Affairs, 2023.

BAKSHY, E.; MESSING, S.; ADAMIC, L. “Exposure to ideologically diverse news and opinion on Facebook”. Science, vol. 348, no. 6239, pp. 1130-1132, 2015.

CASTRO, D. P. Efeito de Enquadramento no Modelo de Grafos para Resolução de Conflitos com uma Aplicação ao Conflito das Obras de Construção dos Viadutos do Cocó [Framing Effect in the Graph Model for Conflict Resolution with an Application to the Conflict of the Cocó Viaducts Construction Works]. Master’s thesis — Federal University of Ceará, Fortaleza, 2022.

CASTRO, D. P. “Cinturão Digital do Ceará e as mídias digitais” [The Ceará Digital Belt and Digital Media]. Proceedings of Social Media Brasil 2011, Fecomercio, São Paulo, 2011.

DEBORD, G. The Society of the Spectacle. New York: Zone Books, 1995. [Original French: 1967.]

FOUCAULT, M. Power/Knowledge: Selected Interviews and Other Writings 1972-1977. New York: Pantheon, 1980.

FREUD, S. Civilization and Its Discontents. New York: W. W. Norton, 2010. [Original German: 1930.]

FREUD, S. Group Psychology and the Analysis of the Ego. New York: W. W. Norton, 1990. [Original German: 1921.]

HAN, B.-C. The Burnout Society. Stanford: Stanford University Press, 2015.

KAHNEMAN, D. Thinking, Fast and Slow. New York: Farrar, Straus and Giroux, 2011.

LE BON, G. The Crowd: A Study of the Popular Mind. Mineola: Dover, 2002. [Original French: 1895.]

LÉVY, P. Cyberculture. Minneapolis: University of Minnesota Press, 2001. [Original French: 1997.]

SIBILIA, P. O Show do Eu: a intimidade como espetáculo [The Spectacle of the Self: intimacy as spectacle]. Rio de Janeiro: Nova Fronteira, 2008.

SUNSTEIN, C. R. Sludge: What Stops Us from Getting Things Done and What to Do about It. Cambridge, MA: MIT Press, 2022.

THALER, R.; SUNSTEIN, C. R. Nudge: Improving Decisions about Health, Wealth, and Happiness. New Haven: Yale University Press, 2008.

TÜRCKE, C. Erregte Gesellschaft: Philosophie der Sensation. Munich: C. H. Beck, 2002.

TVERSKY, A.; KAHNEMAN, D. “The Framing of Decisions and the Psychology of Choice”. Science, vol. 211, no. 4481, pp. 453-458, 1981.