Section 1 - Introduction and Problem Statement
Against the backdrop of Article 34 and the wider Digital Services Act (DSA), I believe that algorithmic decisions, engineered by the systems that generate them, threaten democratic life. I show how that harm is caused, why it sits at the threshold where democracy's thin minimum conditions cease to hold and how that threshold can be measured to assess how close to collapse democracy is.
Digital media platforms, news aggregators and administrative systems have manoeuvred the democratic state into a position where it is almost blind: barely able to recognise whether citizens are exercising agency in the space of reasons. Platforms provide a semblance of interaction. The space of reasons within which political content is selected is reduced to the point where our capacity to contest it becomes restricted. Algorithmic curation, in exercising power, has not closed the space of reasons outright, yet. For digital platforms to survive and grow the traffic passing through their systems, they must provide a minimal illusion of choice. In 2026, platforms edge us towards closure of the space of reasons while a thin, phenomenal residue of participation remains, present in the feed and in the like, but absent from the reasoning that put us there.
There is a need to make a distinction between the harm courts address and the harm I expose. Platforms continue to be subjected to legal pressures. These address almost entirely the individual harms such as addiction and impact on mental health caused by the operation of the platforms. The argument runs ‘a particular law or statute was not complied with which has resulted in measurable harms’. Currently, Meta are facing civil charges focused on privacy violations under COPPA (Children’s Online Privacy Protection Act), design and deployment of harmful platforms and deceptive statements (essentially claiming that there was no harm when internal communications flagged this on numerous occasions) (In re Social Media Adolescent Addiction/Personal Injury Products Liability, 2026, pp. 3–5) [hereafter In re Social Media Litigation]. Such accountability works at the phenomenal level. There is a named harm, attached to an established law with identifiable claimants. In the Meta case thousands of individual claims are collapsed into one case. This form of accountability is not trivial. Financial penalties can reach hundreds of millions of euros: in 2024 the European Commission fined Meta €797.72 million for abusive practices benefitting Facebook Marketplace (European Commission, 2024b). Where such penalties are capped at a fraction of turnover, they can be viewed as a function of the cost of doing business rather than a constraint on it, what Pasquale terms treating “the resulting fines as a (trivial) cost of doing business” (Pasquale, 2015, p. 144). Alongside them, court documents expose internal methodology that would otherwise stay proprietary.
Litigation does not reach the harm I am concerned with. What it exposes, through discovery, is evidence of it. Meta's own internal documents describe its headline harm-reduction features, Quiet Mode and Take a Break, as tools that "served as proof points but were not effective at reducing behaviors" (In re Social Media Litigation, 2026, p. 17), while a separate internal document frames teenage sensitivity to dopamine not as a risk to be managed but as something that "presents opportunities for us" (In re Social Media Litigation, 2026, p. 18). The intention recorded in Meta's own documents is not ignorance of harm managed poorly. It is engagement pursued deliberately, including a documented decision to reduce advertising load specifically for teenage users on Instagram Reels in order to increase their engagement and "maximize long-term revenue growth" (In re Social Media Litigation, 2026, p. 21). The end point of that intention is a visible tool of reduction sitting alongside a design practice moving in the opposite direction: toward more engagement, not less. That gap between the phenomenal gesture of harm reduction and the substance behind it is not evidence of negligence, an act of omission. It is evidence of a choice, an act of commission, sustained without justification to those it affected. This is a defining feature of domination, not non-compliance. Democratic harm, in my focus, does not register as an injury to identifiable individuals. The harm operates on the space of reasons itself, settling what counts as participation before a post is written or a feed is scrolled.
My starting point is Diamond and Morlino's definition of vertical accountability. This provides the lens for where and how the harms occur. In their account it is "the obligation of elected political leaders to answer for their political decisions when asked by voters or constitutional bodies" (Diamond and Morlino, 2004, p. 25). Whilst the original framing runs in one direction, "from citizens to leaders" (Diamond and Morlino, 2004, p. 25), the communication is better considered as bidirectional, between citizen and state. The two directions of this channel correspond to what Pettit calls being authorised "as a voice worth hearing and an ear worth addressing" (Pettit, 2002, p. 350). We are heard when raising a demand and addressed when power is justified. Platforms can be interpreted as threatening both. This bidirectional channel is now compromised by algorithmic interference in two broad areas of life. First, the increasing application of algorithmic decisions within civic systems such as welfare, health and policing. Second, the redirection of answerability towards curated platform content, displacing the citizen as the party owed a response. Platforms thereby come to occupy part of the position Diamond and Morlino assign to those who govern, exercising power over citizens' political judgements while owing them no account of it. The first of these areas is one for future development, adapting the model I have created to platform behaviour in this paper.
In this I build on Klonick's account of platforms as the "New Governors" of online speech, private systems of governance that write and enforce their own rules and lack direct accountability to their users (2018, pp. 1599, 1603). I depart from her account on two points. First, Klonick finds platforms motivated to reflect their users' expectations because their "economic viability depends on meeting users' speech and community norms" (pp. 1603, 1625). Responsiveness of that kind is not answerability. Where platforms curate content and order feeds, they do not only reflect users' expectations; they help create them and then respond to what they have created. That circularity goes to the heart of domination in the space of reasons and it is the managed performance of responsiveness I describe below. Second, her remedy lies in changes to the platforms' own architecture and governance (p. 1603). The DSA audit record examined in Sections 3 to 5 shows what follows when the governed party also authors the benchmarks it is measured against. The Harm Index therefore tests for what internal reform cannot supply: an addressee outside the platform and standing to contest its rules.
To be a democracy, even at its thinnest point, we must establish that there is an identifiable space and route providing the phenomenal measure point, capable of establishing that there is communication. In keeping with my thin concept, this provides a simple but powerful test: either there is communication or there is not. Where there is no communication, we cease to be acting within a democracy and leave the assessment of how weak democracy is behind. Being a phenomenal measure, Diamond and Morlino's framework does not examine whether the conditions that make this capacity possible in the first place are themselves intact. That gap can appear even where communication looks intact, the online reaction to a newsfeed, for example, regardless of the substance or agency behind that reaction. This is the almost blind, the space between what appears to be a democratic exchange and the reasoning process that produced it. In 2004, that gap was latent rather than active. No agent had the capacity to exploit it at a national or transnational democratic scale.
Applying Forst's concept of noumenal power allows this gap to be treated as more than a technical defect in communication caused by algorithmic process. If it were just a technical issue, a solution could be engineered, closing the gap and solving the problem. The 2026 Meta proceedings, referenced above, show that an engineering solution exists for the phenomenal harms the case targets. What we encounter alongside, however, is a problem within the space of reasons itself, one no equivalent technical fix reaches. Noumenal power operates through justifications that move agents to think or act in ways they otherwise would not (Forst, 2017, pp. 39, 49). Platforms structure what is experienced, when it is experienced and how frequently the experience arrives. Forst is explicit that this power is not itself a normative verdict: it "can be the power of a good or a bad teacher... of a liberating revolutionary or a dictator" and "the concept of power itself determines neither the evaluation nor the means used to move someone to think or to do something" (Forst, 2017, p. 49). Platform engagement sits closer to the dictator end of that spectrum than the good teacher end. ‘Teacher’ in that we are provided with content that informs. ‘Dictator’ in that the logic determining what we receive is set by weightings we cannot see, did not consent to and have no standing to contest, regardless of whether that logic happens to serve our interests. Power can also be exercised through physical force, a limiting case that in its extreme form lies outside the realm of power altogether and reflects instead the lack of it (Forst, 2017, p. 50). Physical domination and noumenal domination are not the same operation. Physical domination bypasses the space of reasons. It requires no belief or consent, nothing but bodily compliance, which is why Forst locates it outside the realm of power altogether. Noumenal domination works the opposite way. It does not need to compel the body. It captures the space of reasons itself, producing belief or acceptance without ever needing to force it, so that what looks like reasoning from the inside was never open to genuine contest. What distinguishes legitimate exercises of noumenal power from illegitimate ones is not its presence, no society is free of it, but whether the justifications it rests on remain open to contest and revision. Power that persists through comprehensive, structuring justifications Forst calls rule, and rule "can be well justified or it can rest on bad justifications". Where those justifications close off contestation entirely, presenting an asymmetrical relation as natural, unalterable or simply given, and leave those subjected no real alternative, the result is domination (Forst, 2017, p. 49). What I expose in applying my Harm Index, in Sections 3 and 4, is how platforms move towards dominating the space of reasons.
This positions the democratic danger along a continuum rather than at a single, absolute break, even though at the very end of that continuum a break does exist, the move from noumenal to dominating. At the reciprocal end, communication is genuinely two-way: citizens can contest reasons, question justifications and have what Forst calls "standing as equal normative authorities", the co-authors of the order that governs them (2017, p. 43). As democracy slips, communication remains formally intact but becomes increasingly managed and asymmetrical, moving toward other-defined rather than co-determined rules of justification. ‘Other-defined’ refers to the weightings and rules which determine what I see and do. At the thin end of the continuum, democracy is not simply weak and asymmetrical. It is threatened. As the asymmetrical nature of the space increases, this threat moves closer to the point of domination. The communicative appearance of accountability persists while the reciprocal conditions that make accountability democratic have been hollowed out.
This is the lens through which the gap I have identified should be read. The issue is not whether communication occurs, in its broadest sense it does, but whether it remains reciprocal, preserving citizens as equal normative authorities. A citizen may subscribe, like or post content but if the algorithmic structure has already determined what will count as meaningful input, the exchange is communicative in form while structurally non-reciprocal in substance. In Forst's terms, that is the beginning of domination: a justificatory order in which the subject is still addressed, but not as an equal participant in the practice of justification (Forst, 2017, pp. 43, 49). The democratic form remains visible but its noumenal basis has been thinned to the point where communication becomes less a relation of co-authorship than a managed performance of responsiveness. Structures, as well as people, sustain this thinning without needing to argue for it afresh each time. They reproduce themselves by “suggesting that how they function is ‘natural’”, so that "a 'second nature' can develop" and "the normative power of the factual is reproduced" (Forst, 2017, p. 46). Seen this way, the problem is not that democracy disappears all at once, as in a more noticeable event such as a coup, but that it is gradually hollowed out as communicative channels persist while the justificatory space becomes increasingly asymmetrical, closing off the possibility of reciprocal critique and meaningful contestation.
Algorithmically driven systems now exercise exactly this kind of noumenal power in two directions: an upward harm, in which commercial recommender architectures pre-structure the preferences and beliefs citizens bring to political life before any institutional accountability process engages them, and a downward harm, in which the state's exercise of power is driven by shielded proprietary architecture, disclosure exemptions and outright obfuscation. All are corruptions of vertical accountability and are not currently held to account. Existing measurement work maps algorithmic threats to democracy across three democratic domains at once (Sandri and Novelli, 2026, p. 7). To my knowledge, no framework yet tests, against live records, whether the minimum conditions for contesting algorithmic power hold. What is required is a Harm Index. My paper builds one, deliberately constrained to three indices rather than many more. A model with too many indices invites disagreement over definitional detail rather than over the substance of the harm itself. A thicker model, built for the full range of democratic typologies, can follow, but only once a thin model has proven it can measure that edge before it erodes further. My three indices are Purpose, Petition and Peer: whether a system's governing logic can be identified, whether there is a working route to challenge it and whether those subject to it have any standing in writing its rules. Using the European Union's Digital Services Act (DSA) as a live case, I show that current verification is not yet the foil to platform self-assessment it claims to be.
The European Union's DSA (Regulation (EU) 2022/2065) is a highly developed, real-world attempt to operationalise an approach resembling the Purpose index my framework specifies. Very large online platforms (VLOPs) and very large online search engines (VLOSEs) must conduct annual systemic risk assessments under Article 34, adopt mitigation measures under Article 35 and submit to independent audits under Article 37, with the resulting risk assessment reports, audit reports and audit implementation reports made public under Article 42. Since the first publications in 2024, this has produced two full rounds of published reporting across nineteen platforms (Albert, 2025), a body of dated material against which any proposed measurement instrument can be tested.
Using the model, my paper answers two questions. First, can a Harm Index built to operationalise Purpose be applied to real reporting, rather than remaining a set of illustrative indicators? Second, does the DSA's current audit architecture, in which each platform selects and pays its own auditor, achieve anything resembling the cross-platform concurrence an external verifier is supposed to provide, or does it only look like it does, providing phenomenal outputs that lack substance? My first question is a methods check. The second is my contribution to the problem of measuring democratic harm at the edge. My answer is that a Harm Index can be applied. In answering the second question, it does not achieve concurrence. Verification under the DSA is inconsistent from platform to platform, but the deeper problem is what I call coordinated restraint rather than the absence of common standards: one auditor who broke from that coordinated approach was replaced in the following reporting round (Albert, 2025). Sections 3 and 4 examine how aspects of the DSA's architecture function as work-arounds that auditors and platforms have used to their advantage.
I proceed as follows. In Section 2 I create and explain the Harm Index applying three indices, Purpose, Petition and Peer. Section 3 sets out the DSA's own audit architecture: how a finding is reached, what a disclaimer means and why my model scores it as it does. Each of the twelve indicators is mapped onto specific Articles that generate evidence for it, providing the evidentiary foundation the test cases in Section 4 applies. In Section 4 I apply the Harm Index to published risk assessment and audit reports from three VLOPs across both DSA reporting rounds. Section 5 analyses what the test case shows about the audit architecture as a whole, beyond the three sampled platforms. Section 6 states the limits of the exercise. Section 7 concludes.
Section 2 - The Harm Index and Digital Services Act
Following Goertz's (2006, pp. 35 - 39) account of concept structure, the Harm Index adopts a necessary-and-sufficient design. The concept is democratic standing, in Forst's sense of standing, as equal normative authorities within an order (2017, p. 43), with the indicators designed to capture the status of that standing. I ask whether a system meets the minimum conditions of a contestable space of reasons for those subject to it. The Harm Index measures that standing, with harm being the concept’s failure. It has three parts, each measured by an index: Purpose (architectural transparency), Petition (justificatory addressability) and Peer (minimal co-authorship). Each is necessary and none can substitute or add weight for another. Together, for the minimal concept the Index measures, they are sufficient. In Goertz's terms, the indices are joined by logical AND rather than OR, with AND taking the minimum (p. 40). The Index does not follow what Goertz calls an ‘m-of-n’ rule, under which a case qualifies by meeting enough conditions, any of which may be missing (p. 36). All three indices must be met for standing to hold.
The score is read through two tests, each dichotomous; together they yield a three-valued score. At the threshold, the question is whether standing exists at all: if any index fails, the system scores 0. Above the threshold, the question is how securely standing holds: precariously, at the edge of domination (1) or securely (2). The same rule governs both. Each indicator is coded 0, 1 or 2. Each index score is the minimum of its indicators, with the headline score being the minimum across the three indices, not their sum or average. Figure 1 sets out this structure, adapting Goertz's notation (2006, p. 61). Figure 2 shows how the score is read, as two successive tests, each an AND across the three indices. The first is the threshold: standing holds only if all three indices score at least 1; a system failing it scores 0 (Figure 2a). The second applies only where standing holds and asks whether it is secure: if all three indices score 2, the system scores 2; otherwise, 1 (Figure 2b). The two tests together are equivalent to taking the minimum. No step involves substitution, so no indicator can compensate for another (p. 46): at every level, a part either meets the condition or the whole fails it. In operational terms, the trigger for further attention is always the lowest score. A strong score elsewhere does not offset a weak one; it is the weakest indicator that sets the system's overall result. Table 2.1 shows the threshold test and Table 2.2 the combined outcomes across the graded scale.
Figure 2a. Does standing hold? Figure 2b. How secure is standing?
Table 2.1. Threshold test: does standing hold?
| Purpose Holds | Petition Holds | Peer Holds | Standing Holds |
|---|---|---|---|
| F | F | F | F |
| T | F | F | F |
| F | T | F | F |
| F | F | T | F |
| T | T | F | F |
| T | F | T | F |
| F | T | T | F |
| T | T | T | T |
Note: T = the index holds (scores 1 or 2); F = it fails (scores 0). Standing holds only when all three indices hold (logical AND); otherwise D = 0. All eight combinations are shown.
Table 2.2. Combined outcomes across the graded scale
| Purpose | Petition | Peer | D |
|---|---|---|---|
| 0 | 0 | 0 | 0 |
| 0 | 2 | 2 | 0 |
| 2 | 0 | 2 | 0 |
| 2 | 2 | 0 | 0 |
| 1 | 1 | 1 | 1 |
| 2 | 1 | 2 | 1 |
| 2 | 2 | 1 | 1 |
| 2 | 2 | 2 | 2 |
Note: D is the minimum of the three index scores: 0 = domination, 1 = precarious, 2 = secure. Rows are selected to show the rule: a single failing index sets D to 0 and below the top band the weakest index sets the result.
Because this Index is calibrated to work at the edge of democracy, that one weak indicator is not simply a minor technical shortfall. It is exactly the kind of gap through which democracy's thin minimum conditions begin to erode. A supplementary breakdown, showing every indicator within each index, may be reported alongside the headline score, but only as a diagnostic map showing where that trigger sits, not as a substitute for the headline figure itself: a system scoring well across Purpose and Petition while scoring 0 throughout Peer has not achieved the reasonable overall position a summed or averaged score would suggest, it has failed categorically on one index and the headline score must show that failure rather than dilute it. This distinction matters practically as well as conceptually. Locating exactly where the trigger sits, whether Purpose is low while Petition and Peer are high, or the reverse, allows repair to be targeted at the specific indicator, the index, or the system as a whole, rather than treated as a single undifferentiated deficiency. As I show in Sections 3 and 4, this matters directly to how exclusions within an audit are scored. Where an area is audited but a decision is made to exclude it from the scope of what the audit examines, for instance, an area under active EU investigation that auditors have been advised to leave untested, that exclusion is not an absence of information. It is a choice made within a process that otherwise occurred. Recording it as null would treat it as data that does not exist, letting it fall outside the minimum-aggregation logic I apply and disappear from the headline score exactly where the trigger is most needed. Recording it as 0 keeps it inside the calculation, consistent with the criteria already set out: no independent verification occurred for that area, which is the condition a 0 is defined to mark, regardless of whether the absence resulted from the area never being scoped in or from a decision to scope it out mid-process.
My Harm Index is applied at the level of the individual system. In the DSA context, the individual platform's audited service is the unit of analysis, rather than the company that may control several platforms. A single company can operate one well-governed system and one badly governed system in parallel, in the same way a state can hold free elections while operating a captured judiciary. Collapsing multiple platforms or apps into a single company-level or country-level score would lose the information my Index is designed to expose. System-level scores can be aggregated, by company or by jurisdiction, but the base unit of measurement remains the system, scored on three indices.
Each index asks one question. Purpose (architectural transparency) asks whether a system's governing logic can be identified with sufficient clarity, by those affected by it, to be challenged. Petition (justificatory addressability) asks whether there exists an external party, beyond the system's own operator, powerful enough to receive and act on an individual's demand for justification. Peer (minimal co-authorship) asks whether those subject to a system have any structural standing to contest the rules that govern it, as distinct from contesting individual outputs the rules produce. Each formal term and its shorthand are used interchangeably throughout what follows.
Keeping with the thin end operationalisation, three tests govern selection for the indicators. First, an indicator must correspond to a failure mode already documented in practice rather than just being conceivable. This keeps the Index focused on what issues exist, rather than what might happen. As a developmental model, failure modes not yet documented can be folded into future iterations of the Index as they emerge and as the DSA or other system applies. Second, an indicator must be scoreable by a party outside the system under scrutiny, using evidence that party can realistically obtain. My Harm Index uses published, dated, citable material, transparency reports, audit reports and regulatory submissions, rather than internal documentation. Third, a score of 0 must mark a state meaningfully closer to the unaccountable exercise of power, not merely to suboptimal or unpolished practice, so that the Index registers categorical gaps rather than differences of degree. A system that discloses its logic in slightly less detail than a competitor is not, on that basis alone, closer to democratic harm, but a system with no independent verification of its declared logic at all is. For the same reason, each indicator is scored on a three-point scale rather than a finer one: the three points mark three qualitatively distinct institutional states, absent, reactive and proactive, rather than gradations of quality an external verifier cannot reliably distinguish. The evidence for each indicator, and why it earned its place against these three tests, is given where that indicator is introduced below.
My indicator-based approach adapts the method proposed by Sandri and Novelli (2026), who frame AI's threat to democracy as a principal–agent delegation problem and propose institutional assessability as the central condition for democratic control: ex-ante documentation of how a system works, ex-post reconstruction and verification by independent experts, and contestability (p. 10), operationalised through observable, largely publicly available indicators (pp. 7, 13). As I show, self-reporting in respect of the DSA creates opportunities for reinterpretation of what evidence counts. The Harm Index shares a commitment to externally verifiable, dated evidence. It differs from Sandri and Novelli in two respects that matter for what follows: it specifies an aggregation rule, taking the minimum (see Figures 1 and 2), where Sandri and Novelli leave aggregation unspecified, so that strength on one dimension cannot compensate for failure on another. This is deliberately narrower in scope, three indices rather than a cross-domain framework spanning information ecosystems, elections and public administration, in order to test whether a thin model can be applied to a single, active regulatory case rather than illustrated across several at once.
Table 2.3 tests the indicators for Purpose. It asks five questions:
- Disclosure: has detail been provided on how the system works and was it offered freely or only when someone asked?
- Independent verification: has anyone checked the disclosure is accurate and does that happen routinely or only after something's gone wrong?
- Public availability: how easily can results of verification be accessed?
- Methodological concurrence: are all systems measured against the same, externally set rules, or does each get judged by its own?
- Depth of verification: was the system's logic checked, or simply confirmed to exist?
Table 2.3. Purpose (Architectural Transparency) - whether a system’s governing logic is disclosed, verified, visible, consistently benchmarked and genuinely examined
| Indicator | Scoring Guidance (0 / 1 / 2) |
|---|---|
| Disclosure of optimisation objective or declared decision logic | 0: Not disclosed even on request 1: Disclosed reactively, on request or under legal challenge 2: Disclosed proactively on a continuous reporting cycle |
| Independent verification of declared logic against operation | 0: No independent verification exists 1: Verification occurs only after a specific complaint or incident 2: Routine, continuous verification independent of complaint |
| Public availability | 0: Findings not published 1: Findings available only via request (e.g. Freedom of Information) 2: Findings proactively published on a fixed cycle |
| Methodological concurrence across subjects | 0: No common standard exists; each subject’s verification is conducted against criteria that subject, or its own hired verifier, was free to define 1: A common procedural framework applies (e.g. a shared assurance standard) but substantive benchmarks remain subject-defined 2: Verification is conducted against a common, externally specified substantive standard applied consistently across subjects |
| Depth of verification | 0: Verification confirms only that a stated control or piece of documentation exists 1: Verification confirms the control produces some observable effect, without characterising the underlying logic or measuring the effect’s size or direction 2: Verification characterises the actual ranking, weighting or optimisation logic, or draws on the system’s own internal experimental evidence (e.g. holdout-group testing) |
Methodological concurrence tests something the other indicators do not directly capture. Transparency is usually understood as a property of the relationship between a single system and a single observer: is enough disclosed, verified and published for that observer to form a judgement. Where many structurally similar systems are subject to the same disclosure regime, a further question arises: are those systems being measured against a shared standard, such that a given score for one system means the same thing as an equal score for another, or is each system’s verification conducted against a benchmark that system, or its own hired verifier, was free to define for itself.
A system can score well on disclosure, verification and public availability for every subject individually, while still failing to produce comparable information across subjects, if the standard against which each subject is verified is not itself held constant. This is a practical concern. Four platforms in the DSA's first reporting round, AliExpress, Facebook, Instagram and TikTok (Albert, 2025), received a Disclaimer of Opinion on their systemic-risk obligations (Articles 34 and 35), meaning their auditor declined to state any conclusion at all, positive or negative, because a European Commission investigation was open at the time. This was not four auditors independently reaching the same judgement. The European Contact Group, the industry body representing the major firms conducting DSA audit work, has set out the position that auditors should in many cases refrain from concluding where the Commission is investigating, an approach already evident in the first cycle and restated in guidance updated in 2025 (Albert, 2025). A coordinated industry position on when not to exercise the judgement Article 37 exists to require, not an absence of standards but an agreed one. The DSA's audit regime, structured to permit this, even while formally requiring disclosure, verification and publication from every subject, has not delivered architectural transparency in the sense my thin standard requires; it has delivered nineteen separate bilateral relationships between platforms and their chosen auditors, applying the common vocabulary of a single regulatory framework. Nineteen private conversations do not add up to one public safeguard.
Depth of verification tests something distinct from methodological concurrence. Both can fail independently of one another. Concurrence asks whether verification is conducted against a shared standard across subjects. Depth asks a prior question of any verification exercise considered on its own: whether it examines what a system’s governing logic actually does, or only whether a stated control exists and produces some observable effect. This focuses on an issue arising from the early stages of the DSA audits. Recommender-system audits under the DSA’s first reporting round were assessed independently by several researchers and the finding recurs across sources that did not coordinate with one another. Bouchaud and Ramaciotti (2024, p. 14), evaluating methodologies used across social media recommender-system audits more broadly, reached a closely related conclusion by a separate route. Reviewing the first round of recommender-system audit findings across platforms, Chapman (2025) found that because platforms and auditors developed their operative definitions independently of one another, comparing the results was, in his words, like "comparing apples to artichokes." What the audits verified, on inspection, tended to be narrower still than inconsistent definitions alone would suggest. YouTube’s auditors reported that they had inspected changes to recommender outputs before and after a user modified a setting and confirmed that the setting influenced the system’s main parameters; Meta’s auditors, examining Facebook, confirmed only that a preference control was accessible from the interface (Solarova et al., 2026, p. 6). Neither exercise characterised the ranking or amplification logic the control was supposed to govern or measured the size or direction of the effect the control produced. This matters. Platforms already hold better evidence than this. Product teams run large numbers of internal experiments against holdout groups every year specifically to measure what changes to recommender design actually do (DSA Observatory, 2025) and this experimental evidence does not appear to be what audits examine. The gap, in other words, is not that the relevant information about algorithmic behaviour does not exist. It is that verification does not reach it. Meßmer and Degeling (2023, p. 2) warned, before the first audits, that the DSA's lack of concrete audit guidelines could foster "audit-washing", that is, "platforms exploiting audits to legitimise their practices and neglect responsibility". The first audit cycle suggests that risk has materialised, and my paper adopts the term for it.
Together, concurrence and depth complete the picture disclosure, verification and public availability leave open. A system can score well on those three, poorly on concurrence because its benchmark is self-defined and poorly again on depth because even its self-defined benchmark was never tested against the system’s actual behaviour. These are three separate ways for a transparency regime to fail while still producing what looks, from the outside, like extensive documentation. As Pasquale warns, transparency may simply provoke complexity as effective at defeating understanding as secrecy itself; transparency is an interim step towards intelligibility, not a substitute for it (Pasquale, 2015, p. 8).
Petition, the second index, asks whether an external party, beyond the system’s own operator, is empowered to receive an individual’s demand for justification and act on it. The DSA captures this indicator. Article 20 establishes an internal complaint-handling mechanism, through which a platform reviews its own moderation decisions; Article 21 goes further, requiring platforms to submit to review by externally certified out-of-court dispute settlement bodies, independent of the platform and accredited by a national Digital Services Coordinator. Initially, this looks like exactly the external addressee my Harm Index requires. It is qualified in one respect that the indicators need to register precisely, Article 21 bodies are expressly barred from imposing a binding outcome. The text states that a certified body "shall not have the power to impose a binding settlement of the dispute on the parties". Platforms must engage with the process in good faith and must report on outcomes as part of their mandatory transparency reporting under Article 24. However, nothing in the Article requires them to implement what the body decides. Pettit draws the same line. Standing requires being able to command attention and respect and "it is not enough that others condescend and confer such attention and respect by way of responding to a supplication or petition" (2002, p. 350). A route of petition that the platform may answer or decline at will does not yet confer standing. Article 21 therefore satisfies the external-addressee half of Petition while falling short, by design rather than by accident, of the enforceability half.
Table 2.4 tests the indicators for Petition. It asks four questions:
- Existence of an external addressee: is there someone to go to who is not the platform itself?
- Enforceability: how binding are remediation requirements?
- Accessibility: is the route free to a genuine complainant, or does cost or complexity stand in the way?
- Outcome reporting: are results published in enough detail to show whether the process actually works?
Table 2.4. Petition (Justificatory Addressability) - whether a working, independent, accessible and accountable route to challenge a system exists
| Indicator | Scoring Guidance (0 / 1 / 2) |
|---|---|
| Existence of an external addressee for individual grievances | 0: No channel external to the system’s own operator 1: Internal complaints process only (DSA Article 20) 2: External, independently certified body empowered to review and rule (DSA Article 21) |
| Enforceability of findings | 0: Advisory only, no obligation on the operator to respond 1: Operator must engage in good faith and report the outcome, but is not required to implement the finding (DSA Article 21’s non-binding standard) 2: Binding finding with a remediation obligation |
| Accessibility of the channel to an individual citizen | 0: Realistically usable only by well-resourced actors 1: Formally open to individuals but carries cost or procedural burden for the claimant 2: Free or nominal-cost, low-barrier, direct individual access |
| Outcome reporting | 0: Outcome statistics not reported 1: Aggregate statistics reported but not broken out by claim type or reversal rate 2: Outcome statistics reported with sufficient granularity to assess whether the channel functions (e.g. reversal rate, response time), consistent with DSA Article 24 |
The fourth indicator, outcome reporting, is included based on the Article 24 requirement that platforms report exactly the data needed to score this indicator as part of their routine transparency reporting. Scoring it requires no separate data-collection exercise. Appeals Centre Europe, one of the certified Article 21 bodies, identifies why this indicator matters in practice. Of 1,565 decisions taken between January and August 2025, 77 percent overturned the platform's original decision, but around two-thirds of those were default decisions, issued because the platform had not supplied the content needed to review the dispute. Where the Centre could examine the content, it overturned the platform in roughly half of cases (Appeals Centre Europe, 2025, pp. 15–16). Where platforms responded, they implemented only around half of the decisions against them (pp. 35, 42). The mechanism can therefore be stalled by withholding content and, when it rules against a platform, declined. A regime can score close to full marks on disclosure of outcomes and still leave the underlying grievance unresolved. Reporting that a decision was overturned is not the same thing as requiring the platform to act on it.
A working comparator is useful to illustrate that the enforceability gap in Article 21 is a design choice rather than an unavoidable feature of external redress mechanisms. The United Kingdom's Freedom of Information Act 2000 gives individuals a structurally similar right: where a public authority refuses or mishandles a request for information, the requester may apply to the Information Commissioner's Office for a decision under section 50. The Commissioner's decision notice is, in the ICO's own description, legally binding (Information Commissioner's Office, n.d.); if a public authority fails to comply, the Commissioner may certify that failure to the High Court, where it may be treated as contempt of court, and either party may appeal to the First-tier Tribunal. The mechanism is not immune to challenge, but the baseline is compliance, not invitation. The contrast with Appeals Centre Europe lies at both ends of the process. Where a public authority withholds the information the Commissioner needs, the Commissioner can compel it by information notice (Freedom of Information Act 2000, s. 51); where a platform withholds the content a dispute body needs, the body can only issue a default decision. Where the Commissioner finds against an authority, the authority must comply; where the Appeals Centre finds against a platform, compliance is the platform's choice. In the cases reported only around half of such decisions were implemented (Appeals Centre Europe, 2025, pp. 35, 42). On the two indicators Article 21 does not fully satisfy, the FOIA/ICO regime scores at the top: an external body independent of the public authority and a finding the authority is legally required to implement. The comparison also bears on the concurrence problem identified in Purpose. FOIA addressability runs through a single statutory regulator applying one consistent standard to every public authority. DSA addressability runs through a range of independently certified ODS bodies, each setting its own procedural rules, which, as one legal commentary reviewing the emerging practice has already noted, "tend to differ between bodies" (Freshfields, 2025). The DSA’s redress architecture is not simply weaker than the ICO model on enforceability. It reproduces, within addressability itself, the same fragmentation the concurrence indicator was built to catch in the audit regime.
Peer, the third index, asks whether those subject to a system have any structural standing to contest the rules that govern it, rather than only the individual outputs those rules produce. This is the index the DSA supports least directly. Co-authorship is the one condition the Act does almost nothing to secure. Two mechanisms bear on it. Article 40 creates two research access routes. Under Article 40(4), on a reasoned request from the Digital Services Coordinator of establishment, platforms must give vetted researchers access to data "for the sole purpose of conducting research that contributes to the detection, identification and understanding of systemic risks" and to assess risk mitigation measures (Regulation (EU) 2022/2065, Art. 40(4)), with researcher status granted against the conditions in Article 40(8). Article 40(12) adds a narrower route, limited to data "publicly accessible in their online interface" (Art. 40(12)). Neither route reaches system logic itself: explanations of the design, logic and functioning of algorithmic systems can be required only by the Digital Services Coordinator or the Commission (Art. 40(3)). Article 48 provides for voluntary crisis protocols, strictly limited to "extraordinary circumstances affecting public security or public health" (Art. 48(1)). The Commission encourages platforms to participate in drawing them up (Art. 48(2)) and "may, where necessary and appropriate", involve civil society organisations (Art. 48(3)), and the protocols' measures are tied to activation in a crisis (Art. 48(4)(b)–(c)). It is a discretionary channel that switches on in extraordinary circumstances rather than a permanent one. Neither mechanism gives an ordinary citizen a route into contesting governing logic directly; both are, at best, proxy channels through which researchers or civil society bodies might do so on citizens' behalf.
Understanding of system logic under the DSA is therefore not absent but tiered. Depth of access increases as distance from the public increases. Individuals receive outputs and summaries through statements of reasons (Regulation (EU) 2022/2065, Art. 17) and the main parameters of recommender systems (Art. 27). The public receives aggregates and audit conclusions (Arts 15, 24(5), 42(4)). Researchers receive data (Art. 40(4), (12)). Only the auditor (Commission Delegated Regulation (EU) 2024/436, Art. 5(2)) and regulators (Regulation (EU) 2022/2065, Art. 40(3)) reach the systems themselves, a version of what Pasquale calls “qualified transparency” (Pasquale, 2015, p. 142). From them, the public receives a conclusion and, where an audit is disclaimed, nothing. The wider public is the DSA's addressee, yet it stands furthest from understanding what it is meant to hold to account.
Table 2.5 tests the indicators for Peer. It asks three questions:
- Contesting channel: is there a route or method to challenge the system logic?
- Independence: is the body undertaking the challenging free of the platform's funding or control, or is it effectively marking its own homework?
- System access: can the system logic be verifiably challenged?
Table 2.5. Peer (Minimal Co-authorship) - whether those subject to a system have any real, independent standing to contest its governing logic
| Indicator | Scoring Guidance (0 / 1 / 2) |
|---|---|
| Permanent channel to contest the governing logic itself, not just individual outputs | 0: No such channel exists 1: Channel exists but is narrow or conditional (e.g. activated only during a declared crisis, as under DSA Article 48) 2: Permanent, general-purpose channel available at all times |
| Independence of any structural-input mechanism from the entity whose systems are contested | 0: Mechanism is internal to, or controlled by, the regulated entity, including where the entity has replaced the mechanism following adverse findings 1: Mechanism is nominally external but selected or funded by the regulated entity 2: Mechanism is fully independent, governed by parties other than the regulated entity |
| Functioning access to system-level data enabling external contestation of the governing logic | 0: No external access to system-level data exists 1: Access exists (e.g. DSA Article 40 vetted-researcher access) but is documented as incomplete or unreliable 2: Access is complete, reliable and enables independent reconstruction or verification of the governing logic |
The evidence available for the third indicator, functioning access, illustrates why access alone does not satisfy the index. An academic audit of the data researchers actually receive through Article 40(12) access found it to be "substantially incomplete and systematically biased" (Bekavac and Mayer, 2026, p. 3), meaning the one channel built to approximate structural co-authorship does not, on inspection, reliably work. A mechanism can exist on paper, in the sense the first indicator asks about, while failing the third indicator because the access it provides cannot support the kind of independent reconstruction of governing logic that would make contestation meaningful.
Together, the three tables are not three independent checklists sitting side by side. They describe a single relationship that properly functions as a whole. Purpose identifies what a system does and, where verification is meaningful, who is answerable for it. Without an identifiable author, or a proxy standing in for one, there is no one to direct a demand for justification toward and Petition has nothing to attach to. Petition, in turn, gives an individual complaint somewhere to go, but a channel that only ever resolves single disputes, one at a time, after the fact, leaves the governing logic that produced the dispute untouched. It is Peer that gives affected parties any standing to contest that logic directly, rather than relitigating its outputs indefinitely. Peer, of this kind, depends on the same identifiability that Purpose was designed to establish in the first place: a party cannot meaningfully contest rules whose author or logic it cannot see. The three indices are, in this sense, mutually constituting rather than cumulative. A regime that scores unevenly across them, strong on Purpose and weak on the other two, has not partially met the standard. It has built one working part of a mechanism that does not function without the other two. Sandri and Novelli reach a similar conclusion about the interlocking of delegation failures, in which opacity prevents contestation and accountability is hollowed out "even while formal institutions remain outwardly intact" (2026, p. 13). The Harm Index formalises that interdependence as a necessary-and-sufficient structure, so that the failure of any one condition registers in the headline score.
Some candidate indicators that appear in governance literature were considered and deliberately excluded. Toledo's (2026) systematic review of algorithmic accountability in public administration identifies five recurring governance dimensions, among them ethical governance and public trust (pp. 13, 16). Neither survives the three tests set out above. Ethical governance, typically operationalised, asks whether a system’s values-based commitments are well drafted, which is a matter of degree rather than a threshold a system either crosses or does not and scoring it would reintroduce exactly the comparative, quality-of-polish judgement the third test excludes. Public trust fails a different test: it is an outcome of the conditions my Harm Index measures, not one of those conditions itself. Toledo's own framework situates it later in the same way, as a product of democratic legitimacy rather than one of its conditions (Toledo, 2026, p. 24). Including it would risk treating a symptom of unaccountable power as though it were evidence of the power’s presence or absence, when the causal relationship runs the other way. A further candidate, a direct measure of algorithmic fairness or bias in outcomes, was excluded specifically to follow my paper’s thin commitment: a fairness indicator would require the Index to adjudicate whether a system’s outputs are substantively just, a thick, contested normative question my paper has deliberately declined to tackle, rather than the narrower, thin question of whether the system’s governing logic is identifiable, challengeable and open to contestation regardless of what that logic turns out to produce. My Harm Index measures whether accountability's preconditions, identifiability, challengeability, standing to contest, are met. It does not, by design, measure the fairness of any particular decision those preconditions allow. A model built to do both would no longer be thin enough to sit at the edge where my paper needs it to operate.
Petition and Peer do not carry equal empirical weight in the DSA test cases that follow. Article 24's outcome-reporting requirement means the data needed to score most of Table 2.4's indicators already sits inside the same transparency reports this paper draws on for Purpose, resulting in Petition being provided for. Peer is limited, reflecting the DSA's structure. Article 40's access regime is the only mechanism bearing on it directly, the evidence available about that mechanism is critical rather than merely thin. No comparable volume of published, cross-platform data exists to test it the way Purpose and Petition can be tested. Each indicator is given formal scoring guidance and a brief account of why it is included in the Harm Index, shaped in part by recent DSA-specific commentary examining recommender-system audit practice, beyond the sources cited directly above (DSA Observatory, 2025). What this Section has not done, deliberately, is show what any of these scores looks like against a real system. That is Section 4's task.
References, Data and Footnotes
References, data and footnotes for the full paper are on the Eroding the Edge of Democracy page.