Working Paper
Stage 2 established the theoretical foundation for this investigation. Drawing on Forst’s concept of noumenal power, I argue that algorithmic systems exercise a form of power over individuals, as political agents, that shapes observable democratic activity. This is a denial of agency which occurs prior to actions. The obligation of justification that this generates cannot be discharged through horizontal accountability mechanisms alone. Stage 3 applies the argument against Diamond and Morlino’s eight dimensions. I test whether the dimensions are capable of detecting and measuring the condition Stage 2 identifies. The answer is that they are not, but the reason is instructive. The dimensions do not fall short. They continue to function as normative and phenomenal measures of democratic quality, and the framework they provide remains as analytically strong as it was designed to be in 2004. What the dimensions cannot do is reach the architectural level at which the conditions of democratic agency are formed. That level was not part of the environment the framework was designed to address. The gap lies not within any single dimension but in what all of them presuppose: the conditions that make democratic activity possible before measurement begins. It is here that the framework’s reach ends and the case for an additional dimension begins. The dimensions provide the necessary foundation to incorporate a ninth dimension.
Stage 3 examines three dimensions: participation, responsiveness and the rule of law. They represent, respectively, inputs, outputs and legal foundations of democracy. Applying the architectural argument from Stage 2, algorithmic effects should be measurable across all three. What my analysis reveals is not three separate problems but a single limitation expressing itself in three different registers.
Diamond and Morlino’s participation dimension measures the formal capacity of individuals to engage in democratic life: voting, organising, petitioning and making demands of power.1 A good democracy, they argue, must ensure that citizens are in fact able to make use of these formal rights to influence the decision-making process (Diamond and Morlino, 2004, p. 23). These are observable activities, measurable through voting data returns and more independent indices such as V-Dem. The dimension works on the basis that the conditions under which those formal rights are exercised remain broadly intact. If citizens are prevented from voting through identification requirements, the dimension can register and account for that constraint. Where there is a one party state with nominal elections, this can be factored into a judgement on how participatory the process genuinely is. The architectural argument puts the dimension under pressure in scenarios where there is apparent freedom to vote, a range of parties and no outward coercion. On its face, this looks like a neutral environment. In 2026, however, there is an unseen determinant impacting participation prior to citizens making informed choices. This is the condition Habermas identifies when he argues that algorithm-controlled platforms are ‘far from neutral’ in their operation, distorting the communicative conditions on which political agency depends (Habermas, 2022, p. 163). What platforms appear to offer as open participation diverges, internally, from what they deliver. The divergence operates before any act of participation takes place.
The divergence Habermas identifies highlights an aspect of participation the dimension cannot resolve on its own. An examination of a counterfactual scenario highlights this limitation: if the shaping of participatory agency were just an additional form of observable constraint, not present in 2004, such as voter identification requirements or one party competition, it could in principle be factored into an assessment of participatory quality. It would form one of the variables applicable to measuring the quality of participation specifically and democracy more generally. There would be no need for a ninth dimension. My argument is that the architectural condition is categorically different from those the dimension was designed to capture. It does not constrain participation from outside. It fabricates the conditions of participation from within, before any observable activity takes place. That is a different kind of problem, requiring a different kind of analytical approach. High voter turnout, for example, registers as positive engagement within Diamond and Morlino’s framework. Where that turnout has been shaped by algorithmically curated information, however, the distinction between acting and being moved to act becomes structurally obscured. While the observable trajectory remains identical to an external observer, the underlying political agency is fundamentally distinct. The direction is the same, the agency is not. In a simplified model, the dimension records the output as a percentage of outcomes. It cannot, on its own, assess the conditions that produced the outcomes. As Alnemr (2025, pp. 2–3) demonstrates, algorithmic systems promote political polarisation and compromise the conditions of political agency before citizens enter the public sphere at all. Dean’s concept of the ‘fantasy of participation’ explains why this is so difficult to detect. Citizens believe they are genuinely active. The experience of using the digital environment is indistinguishable, from within, from meaningful democratic participation. What the dimension cannot assess is the degree to which the architecture has pre-structured what citizens encounter and how their demands are formed. Dean describes this as a structural condition of depoliticisation in which the performance of participation substitutes for its substance (Dean, 2005, pp. 60–61). Participation may be formally robust while its conditions are being structurally hollowed out and the dimension has no means for detecting the difference.
My argument does not depend on endorsing any particular theory of what participation should look like. My own position sits toward the thin end, where a minimum of features would be in place. That leaves open further discussion as to whether participation is of a participatory, deliberative or liberal democratic nature. The architectural problem here cuts across theories of democratic theory. This matters for the strength of my claim. Deliberative democracy requires the conditions for reasoned public exchange. As established, those conditions are structurally compromised at the architectural level rather than simply degraded at the output stage. Representative democracy requires that citizens form preferences independently and hold governments to account through the mechanisms of election and representation. Where algorithmic systems pre-structure those preferences before any institutional engagement occurs, the representational chain is mediated at its source: the government that is elected, and the policies it pursues, reflect an architecturally shaped input rather than an independently constructed one. Participatory democracy, even in its thinner forms, requires only that citizens can meaningfully engage in political life through the available channels. Where the conditions of engagement are structurally invisible, participation is hollowed out regardless of how many channels formally exist. Any empirical data the dimension generates about participation loses democratic meaning in this condition, rendering discussion about which type of democracy is in operation premature. The architectural problem is prior to any dispute about which theory of democracy is correct. It operates at the level of the conditions that concepts of democracy presuppose without examining.
Applying a deliberative framework, Castelló et al. (2025) attempt to address this problem. They reveal where that approach reaches its limit, arguing that social media platforms have evolved into commodified spaces whose business models incentivise hate speech, misinformation and political fragmentation. Their response is to enhance platforms’ deliberative capacity across four dimensions: transparency and accountability, openness and inclusiveness, conduciveness to argumentation and consequentiality (Castelló et al., 2025, pp. 12–15). Their four ‘capacities’ describe, in deliberative terms, what a democratically legitimate platform environment must provide. Each, however, presupposes what the architectural argument identifies as structurally absent. Conduciveness to argumentation requires that users engage in reasoned exchange rather than merely circulate content, their “essential dialectics” (Castelló et al., 2025, p. 14). However, the architecture determines what counts as engagement before any dialectic begins. Consequentiality requires that deliberation connects to collective decision-making in ways that citizens can recognise and contest. Problematically, where the architecture shapes both what citizens understand themselves to want, our 2026 problem, and what they understand themselves to be entitled to demand, the chain between deliberation and consequence is broken. Their institutional recommendation, that platforms be reconceptualised through a social contract between platform firms and stakeholders (Castelló et al., 2025, p. 16), underestimates the structural asymmetry at stake. As established in Stage 2, platforms are not parties to a contract in any conventional sense. They, increasingly, exercise quasi-sovereign power while simultaneously contesting the authority of democratic institutions to govern them. A social contract presupposes roughly equal parties capable of negotiating terms. That presupposition does not currently hold, and until it does, the deliberative capacities Castelló et al. identify remain aspirational rather than operational. What their framework cannot provide is a standard for the architectural conditions that must be in place before any of those capacities can function. That is the gap the ninth dimension is designed to address.
The shaping of participatory conditions is not confined to social media, but it is most often the aspect which is highlighted. It operates across the full range of digital systems through which political life is now mediated. Google’s search ranking determines how political information is ranked and delivered. YouTube’s recommendation system shapes what political content is consumed and in what sequence. News aggregators such as Apple News and Google News2 curate what counts as relevant journalism, algorithmically determining which stories reach which citizens and in what order. Each produces a personalised information environment that fragments the shared epistemic foundation on which democratic participation depends. What one citizen receives as politically true, urgent or legitimate may bear little resemblance to what another encounters, not because of differences in their reasoning, but because of differences in what their algorithmic environment has made available to them. Pariser’s filter bubble (2011, pp. 9–10) and Sunstein’s echo chamber (2017, pp. 9–11) diagnose this condition at the level of content. While the proliferation of harmful content remains a pressing empirical concern, the architectural argument advanced here operates at a conceptually prior level. This argument operates at the prior level. By the time the content is delivered it is too late. The architecture shapes all individuals who do the seeing before any content arrives.
The upward direction is not, however, the only way in which the participation dimension is placed under architectural pressure. Participation operates in a second direction, beyond the public sphere. Where the first direction concerns the upward flow of citizen agency, the second concerns the exercise of state power through algorithmic systems that participation itself does not reach. Vertical accountability requires that the state’s exercise of power over individuals remains visible, attributable and challengeable. Where it does not, citizens cannot contest decisions that affect them, representatives cannot examine the actions of government and the relationship between state and citizen is hollowed out from within. As Stage 1 of this investigation established, both the Palantir contracts and the DWP’s Universal Credit Advances Model illustrate this condition. In each case, the systems that govern citizens’ relationship with the state operate under Freedom of Information exemptions and proprietary protections that place their logic beyond public reach. The Public Law Project’s TAG register identifies 55 active algorithmic systems across UK public administration, the majority operating with limited transparency (Public Law Project, 2024). This condition is not exceptional. It is the way in which public administration now routinely operates. The consequence for participation is structural, removing an avenue through which participation and accountability are traditionally linked. A citizen may petition against the deployment of machine learning in welfare decisions, as over 300,000 did in relation to Palantir contracts in April 2026, while the systems that govern their relationship with the state continue to operate unchanged. The formal channels of participation function. The substantive connection to the addressee that matters is removed. Petitioning against a system whose logic cannot be inspected is participation directed at a locked digital cage.3 The participation dimension has no mechanism for detecting this condition. It registers the petition. It cannot register the architectural opacity against which the petition is directed. In Forst’s terms, the individual retains the phenomenal status of a participant while being noumenally reduced to a data point: present in the count, absent from the conversation (Forst, 2017, p. 42).
Responsiveness asks whether governments respond to the wants of citizens. Diamond and Morlino, using Powell, identify “three links in the chain of democratic responsiveness”: preferences are formed and structured by competing political parties, aggregated through electoral processes into a government of policy makers and translated by elected officials into policy outcomes (Diamond and Morlino, 2004, pp. 27–28). As a measure of democratic quality, it works well across a range of democratic styles. For responsiveness to function as a democratic measure, two conditions must hold. First, the preferences and demands to which governments respond must be formed independently, by citizens exercising genuine deliberative agency. Second, those preferences must be directed toward an identifiable addressee, a government or institutional arm, capable of receiving and acting on them. Where algorithmic systems intervene, shaping both what individuals take themselves to want and curating what they understand themselves to be entitled to demand, both conditions are compromised. Powell’s first link in the chain, preference formation, is compromised before the chain begins. The commercial logic of engagement maximisation results in the information environment, through which preferences are formed, being optimised not for deliberative quality but for retention and reaction. An environment optimised for retention and reaction rather than considered exchange is not one in which independent preferences can reliably form. These are not typical features of a responsive process. As Castelló et al. (2025, pp. 1–2) demonstrate, platform architectures systematically undermine information quality, creating structural power asymmetries that shape what citizens understand themselves to want before any formal political demand is articulated. The architecture registers both as engagement data, weighting content by its capacity to retain attention rather than its democratic quality. A government responding to the preferences this environment produces can score highly on Diamond and Morlino’s responsiveness dimension while responding to something that is, at least in part, an architecturally pre-formed output rather than an independently constituted political demand. Responding to algorithmically amplified pressure instead of considered preference removes the space for longer term deliberation, further eroding the conditions under which genuine democratic responsiveness can operate.
Where algorithms systematically prioritise content that provokes strong emotional responses, the democratic consequence is that responsiveness risks becoming structurally biased toward populist outputs. This is not a contingent risk erasable through governance. It is a structural tendency built into the logic of the platforms through which political preferences are increasingly formed. Maly identifies this condition as ‘algorithmic populism’: the process through which algorithms function not as neutral distributors but as active co-constructors of political identity, producing ‘the people’ as an engagement-optimised construct rather than an independently formed democratic constituency (Maly, 2018, pp. 8, 13–14). The architecture rewards engagement over deliberation, with engagement metrics substituting for genuine public support (Maly, 2018, p. 9; Rădulescu, 2025, p. 9). Empirical evidence confirms the directional character of this tendency and its political direction. Huszár et al. (2022, p. 2), in a large-scale randomised experiment involving two million daily active Twitter accounts across seven countries, found a consistent trend: in six out of seven countries studied, the mainstream political right enjoys higher algorithmic amplification than the mainstream political left. The study is notable because it was conducted in part by Twitter’s own Machine Learning Ethics, Transparency and Accountability Team, making it a case of a platform’s own researchers confirming the amplification bias built into its architecture (Huszár et al., 2022, p. 1). More recently, Ye, Luceri and Ferrara’s (2025, pp. 1, 5) audit of X’s recommendation system during the 2024 US Presidential Election found that neutral new accounts show a default right-leaning bias in content exposure and that both left and right-leaning accounts encounter amplified exposure to content aligned with their own political stance and reduced exposure to opposing viewpoints.4 A government responsive to this environment is not responding to its citizens. It is responding to an algorithmically curated and directionally biased approximation of them. Where that approximation systematically amplifies populist and emotionally charged content, responsiveness becomes a transmission mechanism for architecturally produced political pressure rather than a measure of democratic quality.
UK immigration policy illustrates the dangers of influencing responsiveness. Increased government responsiveness to anti-immigration sentiment registers positively within Diamond and Morlino’s framework. The current Labour government has it as a major agenda point, as did previous governments. What the dimension cannot capture is the extent to which that sentiment has been amplified and channelled by recommender systems optimised for engagement rather than reflective judgement. The 2024 summer unrest provides a concrete illustration of the mechanism. Following the Southport murders of July 2024, the recommendation algorithms of social media companies amplified misleading and hateful messaging, driving protests that turned violent and targeted Muslim and migrant communities. The House of Commons Science, Innovation and Technology Committee found that social media business models incentivised the spread of misinformation and that algorithmic recommendations played a role in driving divisive narratives during the crisis period (House of Commons Science, Innovation and Technology Committee, 2025). Amnesty International’s subsequent analysis of X’s open-source recommender algorithm identified systemic design choices that favour contentious engagement over safety, concluding that the platform’s algorithmic design and policy choices contributed to heightened risks during the wave of anti-Muslim and anti-migrant violence (Amnesty International, 2025). The amplification was not incidental. It was architectural. The longer-term shaping of anti-immigration sentiment through platform dynamics is documented in the trajectory of Reform UK and Restore, whose rise has been partly attributed to algorithmic amplification on Meta and other platforms, functioning as accelerants for populist sentiment that pre-exists the architecture but is normalised and scaled through it. The government’s subsequent policy responses, including accelerated deportation measures and a hardened public tone on asylum, represent responsiveness to a political environment that has itself been architecturally shaped. The policy response may be entirely genuine. The preferences it responds to are not entirely independent. Responsiveness, in this condition, risks becoming a measure of governmental sensitivity to algorithmically curated public pressure rather than to the considered preferences of democratic citizens.
If, as Diamond and Morlino argue, the rule of law is “the base upon which every other dimension of democratic quality rests” (Diamond and Morlino, 2004, p. 21), limitations identified here carry greater consequences. A structural inadequacy in the rule of law dimension does not affect only that dimension. It undermines the normative foundation on which the remaining seven rely. My argument in this section is that algorithmic systems introduce amplified inadequacy, not by violating the rule of law on its own terms, where existing statutes and guardrails provide partial mitigation, but by operating at a level that the law has not yet been designed to reach.
In 2025, 4.2 million people’s faces were scanned by UK police using live facial recognition technology, more than in any other European capital or Western democracy (Big Brother Watch Team, 2026). The de jure rule of law remained intact. No explicit statutory provision was violated. The de facto reality, however, preceded the legal framework: individuals were subject to forms of algorithmic identification they were unable to inspect, challenge or refuse prior to their operation. The High Court challenge brought by Shaun Thompson and Silkie Carlo in January 2026 confirms the point. Legal challenge was available, but only after the architectural condition had already taken effect. The issue is not illegality in the conventional sense. It is a structural inadequacy: the framework of law exists, but it does not yet reach the space in which power is being exercised.
This structural inadequacy was confirmed, paradoxically, by the court’s own verdict. The High Court dismissed the claim, ruling the Metropolitan Police’s updated deployment policy to be in accordance with the law and prescribed by law under Articles 8, 10 and 11 of the European Convention on Human Rights (R (Thompson and Carlo) v Commissioner of Police of the Metropolis [2026] EWHC 915 (Admin), para 229). The court found that internal safeguards, including mathematically mapped crime hotspots and mandated administrative authorisation processes, satisfied the qualitative requirements of foreseeability under the Convention (paras 95–98, 127, 229). Crucially, however, the court reaffirmed that its supervisory jurisdiction in judicial review is strictly restricted to checking whether a public body acts within its legal powers, not to examining the underlying operational merits of the technology itself (para 5). By ruling that a rigorous internal procedural trail is sufficient to satisfy human rights standards (paras 217–227), the court reduces the rule of law to procedural compliance. The inner mathematical logic of the algorithm, the training datasets and the corporate operational experience driving deployment selections remain a closed domain insulated from public scrutiny or judicial oversight (paras 194–199, 184). The High Court’s verdict structurally instantiates the very noumenal power imbalance established in Stage 2, transforming the rule of law into an administrative shield for automated state action rather than a mechanism for democratic contestation. Further, the Metropolitan Police settled a damages claim brought by Shaun Thompson arising from his wrongful identification, even as the court upheld the policy’s legality (para 27; Big Brother Watch Team, 2026). The architecture passed the legal test while producing a compensable wrong. De jure rule of law was satisfied. The de facto harm had already occurred.5
When evaluated through Diamond and Morlino’s framework, this verdict exposes the vulnerability of the rule of law when confronted with automated power. For Diamond and Morlino, a robust rule of law requires transparency, equality and predictability so that citizens can understand and legally check how state power is being exercised. The court’s reasoning satisfies the formal conditions of that standard while leaving its substantive requirements unmet. The claimants have announced their intention to appeal, confirming that the architectural problem identified here remains live and unresolved.
Pasquale’s analysis of legal automation sharpens the structural inadequacy the Thompson and Carlo verdict exposes. He argues that the removal of identifiable persons from decision-making processes breaks the accountability chain that democratic legitimacy requires (Pasquale, 2019, pp. 5–6). The rule of law, on this account, is not simply a set of substantive constraints on state power. It is a relational structure in which power is exercised by persons who can be identified, questioned and held responsible. Where algorithmic systems replace that relational structure with automated processes, the law does not fail on its own terms. It is circumvented at a level outside the reach of its own mechanisms. The Metropolitan Police’s live facial recognition deployment illustrates the condition. No identified officer decided to scan Shaun Thompson’s face. A system did. The legal challenge that followed had to be directed at the policy authorising the system rather than at any person who had exercised judgment about him. The accountability chain was not broken. It never existed. The architecture replaced the person before the law had occasion to reach either.
Erickson and Gregory’s analysis of algorithmic specification identifies how deep the problem is. Where Pasquale identifies the displacement of persons, Erickson and Gregory identify the displacement of legal concepts themselves. When legal ideas are translated into computational form, the translation does not simply automate the application of existing meaning. It transforms the meaning itself. Terms that were designed to remain interpretively open, accommodating moral development and contextual judgment, are converted into fixed parameters, closing down the interpretive space that legal language was designed to preserve and producing a conceptual rigidity that the law’s normative function cannot survive (Erickson and Gregory, 2025, p. 5). The authority to determine what legal concepts mean passes from courts and legislatures to the engineers and training datasets that define the system’s inputs, an unauthorised reassignment of interpretive authority from legal institutions to technical systems (Erickson and Gregory, 2025, p. 8). The problem does not stop there. As algorithmic systems are deployed repeatedly, their outputs begin to shape institutional expectations and narrow what counts as a legitimate interpretation, feeding back into the legal system and influencing future practice (Erickson and Gregory, 2025, p. 5). The algorithm becomes, as Erickson and Gregory put it, an “unacknowledged legislator, shaping legal meaning under the guise of technical necessity” (Erickson and Gregory, 2025, p. 4). At that point the rule of law is not merely formally present while substantively hollow. It is being gradually displaced by a parallel architecture of meaning, one that runs alongside the human system while quietly supplanting it. In the live facial recognition case study, the concept of lawful identification is no longer interpreted by an officer exercising judgment in relation to a person. It is computed by a system whose classificatory logic is neither publicly articulated nor judicially reviewable and whose repeated deployment is quietly generating the normative categories by which future deployments will be assessed. The rule of law is formally present. Its interpretive substance has been evacuated.
The three dimensions examined in Stage 3 share a single structural problem. Each continues to function as a normative and empirical measure of democratic quality. What none of them does is reach the level at which the conditions of democratic agency are formed before observable activity begins. The participation dimension cannot detect the architectural fabrication of participatory conditions. Responsiveness cannot distinguish independently formed preferences from algorithmically curated approximations of them. Rule of law cannot reach the level at which legal concepts are transformed by their encoding into systems whose inner logic remains closed to judicial and democratic scrutiny. These are not failures of the dimensions. They are the boundary of the framework’s reach. It marks the space a ninth dimension must occupy, in 2026 and beyond: one capable of measuring the architectural conditions that determine the quality of democratic agency before any of the existing eight dimensions begin their work.