Working Paper
In April 2026, over 300,000 individuals signed two petitions calling on the UK government to terminate its contracts with Palantir Technologies (38 Degrees, 2026a; 38 Degrees, 2026b). These contracts, covering the NHS, MoD and multiple police forces, were awarded through established parliamentary and administrative processes. With a combined value exceeding £600 million and strong indications of further expansion (NHS England, 2023a; Ministry of Defence, 2025; HM Government, 2025), such contract awarding would ordinarily pass as routine business, allowing governance to proceed without democratic challenge. In this case, however, the contracts and the backlash they have generated provide a useful starting point for examining how democratic accountability operates in relation to technologically complex and morally contested systems.
What distinguishes these contracts is not simply their scale, they have the potential to impact all individuals who have data recorded in the UK regardless of nationality. The openly stated views of Palantir’s founders, the company’s deep involvement in global security and military logistics and ongoing concerns around data sovereignty, particularly in relation to the US CLOUD Act, combine to produce a form of governance that sits uneasily within democratic expectations. This tension grew in April 2026. Palantir published a twenty-two point manifesto on X, drawn from Karp and Zamiska (2025), setting out a clear, technology driven political position (Palantir Technologies, 2026). Its dismissal of concerns around AI weaponry as “theatrical” (point 5), reframes issues of accountability as secondary to technological inevitability. Similarly, classifying equality as “dogma” (point 21), positions cultural identity within a grading system, rejecting some cultures as less valuable than others.
Palantir represents a particularly visible instance of this transformation. It is not, however, exceptional. In the UK, the Public Law Project’s TAG register identifies 55 active algorithmic systems, most of which operate with limited transparency and restrict meaningful scrutiny (Public Law Project, 2024b). The Department for Work and Pensions’ Universal Credit Advances Model illustrates the problem well. As a machine learning system used to assess fraud risk in advance payment requests, it operates under substantial opacity. Its architecture, specifications and operational data are withheld under Freedom of Information exemptions. Claimants subject to its decisions are therefore unable to assess the basis on which those decisions are made (Department for Work and Pensions, 2025b).
Where scrutiny emerges, it is typically reactive and partial. The DWP’s fairness analysis, published in July 2025 following external pressure and covering April 2024 to March 2025, found that older claimants and non-UK nationals were disproportionately over-referred for review. These findings were later confirmed by the National Audit Office (Department for Work and Pensions, 2025a; National Audit Office, 2025). Similar dynamics appear beyond the UK. In the Netherlands, the SyRI welfare fraud detection system was ruled unlawful by the District Court of The Hague in 2020 for violating the right to privacy under the European Convention on Human Rights. This was the first case in which a state algorithmic system was invalidated on such grounds (van Bekkum and Zuiderveen Borgesius, 2021, p. 324).
Together, these cases point to a transformation in how political power operates. Existing measures of democratic quality do not adequately capture this shift because they assume forms of decision-making that remain visible, attributable and challengeable. Where governance depends on opaque and privately developed systems, these assumptions no longer hold. I will argue that an additional element is required, one capable of holding to account the implications of algorithmic dependence, opacity and the erosion of contestability within decision-making itself.
Diamond and Morlino’s eight dimensions of democracy (Diamond and Morlino, 2004) frame how we can judge the strength of a democracy. The dimensions work well, serving to frame what are normative aspects of democratic life. They are as equally valid in 2026 as they were in 2004. However, what they do not do is account for the presence of technology, particularly the use of algorithmic decision making systems. They embed assumptions, prioritise particular forms of knowledge and, as in the case of Palantir, are increasingly accompanied by explicit political visions that sit outside democratic scrutiny.
The framework’s limitations in this respect are not a flaw in its design. They reflect the conditions under which it was constructed. Diamond and Morlino’s dimensions rest on an assumption that was reasonable in 2004: that the environment within which democratic activity takes place is broadly neutral. Citizens are treated as political subjects who form preferences, exchange reasons and exercise their right to hold their government accountable. The eight dimensions measure how well democracy performs within that environment. What they do not measure is whether the environment is now capable of sustaining the conditions those dimensions require, it could not. Our 2026 digital environment plays a constitutive role capable of shaping what political information we engage with, how it is weighted and delivered, how engagement is made available and how state power is increasingly exercised. This happens before individuals deploy any of the capacities the dimensions are designed to reflect. This constitutive role is most precisely understood through Forst’s concept of noumenal power “to have and to exercise power means to be able—in different degrees—to influence, use, determine, occupy, or even seal off the space of reasons for others” (Forst, 2015, p. 116). Algorithmic systems exercise precisely this kind of power, shaping the normative space of democratic agency before it is deployed.
Vertical accountability, as Diamond and Morlino conceive it, describes the relationship between citizens and the state, the ways in which citizens form political judgements, make demands of power, hold government to account and through which the state exercises power over individuals in ways that remain visible, attributable and challengeable. Algorithmic systems have corrupted this accountability process at both ends. They degrade the upward flow of citizen agency and the downward exercise of state power in ways that the existing dimensions of democratic quality cannot detect.
Understanding how the degradation occurs requires examining each direction in turn. The first direction is upward. Vertical accountability depends on citizens being able to form coherent political judgements, exchange reasons and exercise pressure on those who govern them. The algorithmic architectures of online platforms, search engines, content aggregators, recommendation systems and social media, shape the informational environment through which this upward flow of accountability operates. This is not just a social media problem. It is a platform architecture problem. Google’s search ranking determines what political information is findable. YouTube’s recommendation system shapes what political content is consumed and in what order. News aggregators algorithmically curate what counts as relevant journalism. Critically, these systems do not deliver the same environment to all citizens in the way that, for example, newspapers or magazines do. Decisions made by and about individuals include their search history, location, demographic characteristics and inferred preferences. These are some of the factors which determine what each person sees. This produces divergent and personalised information that fragments our shared knowledge on which democratic activity depends. What one citizen understands to be 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 identifies this dynamic as the filter bubble, the personalised informational enclosure produced by algorithmic curation (Pariser, 2011, pp. 9–10), while Sunstein demonstrates that such enclosures produce echo chambers that reinforce existing beliefs, systematically excluding the kind of exposure to alternative perspectives that democratic deliberation requires (Sunstein, 2017, pp. 9–11). Across all of these systems, what is visible, credible and shareable in the public sphere is both shaped and individualised before citizens exercise any of the capacities on which vertical accountability depends. As Alnemr demonstrates, these systems promote political polarisation, undermine the epistemic quality of public deliberation and compromise the conditions of rational political agency before citizens enter the public sphere at all (Alnemr, 2025, pp. 4–5). Habermas argues that the platform revolution represents not an expansion of existing media but a structural caesura (break) comparable in significance to the introduction of printing, one in which algorithm-steered platforms fragment the shared epistemic basis of the public sphere into self-enclosed echo chambers that acquire the status of competing publics. This fundamentally alters the conditions under which democratic opinion and will formation can operate (Habermas, 2022, pp. 158, 162). The result is that vertical accountability is weakened at its foundation. Citizens hold governments to account on the basis of preferences and beliefs that have been algorithmically shaped and individually differentiated before the accountability process begins.
The second direction is downward and more opaque. 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 democratic relationship between state and citizen is hollowed out from within. Algorithmic systems embedded within state institutions have created this condition. Ulbricht and Egbert demonstrate the problem clearly in their study of Palantir’s deployment in German federal policing. The modular and proprietary character of such platforms makes it exceptionally difficult for parliamentary opposition, civil society and affected citizens to assess either the scope of data use or the reasoning that produces decisions affecting them: the ‘what’ and the ‘why’. They conclude that platform regulation in the public sector “is not only about technology-specific regulation but also about general mechanisms of democratic control, such as the separation of power, public transparency, and civil rights” (Ulbricht and Egbert, 2024, p. 12). The MoD’s own procurement documents acknowledge that its data analytics capabilities have become so dependent on Palantir architecture that changing supplier would cause disproportionate technical difficulties (Ministry of Defence, 2025). This is institutional lock-in stated openly in an official government document, creating a form of algorithmic dependence that places the exercise of state power beyond the reach of democratic accountability by design rather than by accident.
The DWP’s Universal Credit Advances Model illustrates the same condition at the level of individual citizens. The system withholds its model architecture, specifications and operational data under Freedom of Information exemptions, meaning that claimants subject to its decisions cannot scrutinise the basis on which those decisions are made (Department for Work and Pensions, 2025b). When the DWP’s own fairness analysis found that older claimants and non-UK nationals were being disproportionately over-referred for review, the finding was not proactively published — it emerged only after sustained external pressure (Department for Work and Pensions, 2025a; National Audit Office, 2025). The system exercises power downward over individuals while remaining opaque to those individuals, to civil society and to the elected representatives who are formally responsible for it. Beyond the UK, the District Court of The Hague’s ruling against the SyRI system established that this is not merely a failure of administrative practice but a violation of fundamental rights, the first judicial recognition that state algorithmic systems can breach the right to privacy under the European Convention on Human Rights (van Bekkum and Zuiderveen Borgesius, 2021, p. 324).
What these cases share is not a problem of transparency. It is a specific form of opacity, one that is structural, legally protected and increasingly treated as a normal feature of public administration rather than an exception to it. The normalisation has embedded itself into our democracy and is being used to justify its own use. The individual subject to these systems encounters state power removed from accountability processes. They cannot identify the assumptions encoded in the system, contest the reasoning that produced the decision affecting them, and, as the Palantir case illustrates, may find that the company whose system governs their relationship with the state holds an explicit political vision that has never been subject to democratic deliberation. Forst’s concept of noumenal power captures this condition precisely. These are systems that seal off the space of reasons for those subject to them, structuring what citizens can demand, contest and justify in their relationship with the state before any formal accountability mechanism is engaged (Forst, 2015, p. 116).
Both directions corrupt the dimension of participation in ways that are formally invisible. Participation can register as robust, citizens vote, petition, campaign and engage, while the conditions of participatory agency have been hollowed out simultaneously at both levels. On the upward direction, political preferences are shaped before citizens engage with one another in the public sphere. The platform environment pre-structures what citizens understand themselves to want, believe and oppose before they participate in any accountability process. On the downward direction, the state’s exercise of power over citizens is increasingly mediated by algorithmic systems that participation itself cannot reach. A citizen may petition against the use of machine learning in welfare decisions, as 300,000+ have done in relation to Palantir, while the systems that govern their relationship with the state continue to operate behind FoI exemptions and proprietary architectures that place them beyond public reach.
This is not a failure of participation in the formal sense. Citizens retain the right to vote, to organise, to petition and to challenge. What has changed is the environment in which those rights are exercised. On one side, the informational conditions of participation have been algorithmically pre-structured and individually differentiated before citizens engage. On the other, the decisions that participation is supposed to influence are increasingly produced by systems whose logic participation cannot enter. The Diamond and Morlino framework provides guidance for measuring the activity of participation. It has no mechanism for assessing whether the conditions that make participation meaningful, shared information, contestable reasoning, visible and attributable state power, remain intact. The eight dimensions remain capable of measuring the outputs of democratic life. This is tested through a range of empirical rubrics such as V-Dem, Freedom House’s Freedom in the World index and the Economist Intelligence Unit’s Democracy Index. What they cannot assess is whether the prior conditions that give those outputs democratic substance are being systematically eroded. A democracy can score well across all eight dimensions while the architectural foundations of meaningful agency are quietly hollowed out. Citizens can participate formally while being unable to contest the systems that shape what they know, what they believe and what the state does to them. Accountability appears to function while power is being exercised by the algorithmic layer.
This is the gap. It is structural rather than empirical, and it cannot be addressed by refining the existing dimensions. The dimensions ask how well democracy performs. The gap concerns whether the conditions under which democratic performance is possible remain in place. Closing it requires not a revision of Diamond and Morlino’s framework, but an addition to it, a ninth dimension directed at the conditions under which democratic agency becomes possible: the visibility, contestability and democratic answerability of the algorithmic systems. The following sections develop the theoretical foundations for that dimension, drawing on Forst’s account of noumenal power and the distinction between the upward and downward directions of vertical accountability, before setting out its content, indicators and regulatory implications.