Participation asymmetry in EU AI governance: a longitudinal analysis of consultation demographics and regulatory outcomes

As AI systems increasingly mediate access to employment, credit, education and public services, who shapes the rules governing these technologies is a question of democratic power. The EU AI Act, adopted in May 2024 as one of the first binding comprehensive horizontal regulatory frameworks for artificial intelligence, was designed to demonstrate that democratic governance could shape – not merely react to – AI development through a “human-centric” approach balancing innovation with fundamental rights protection.

The legitimacy of this regulatory project rested on democratic inputs. Under the EU’s Better Regulation framework, stakeholder consultation is the primary mechanism through which citizens and civil society can shape policy in expert-driven domains. For the AI Act, this translated into three major consultations between 2020 and 2024. Yet barely 16 months after adoption, and before most provisions entered into force, the European Commission proposed comprehensive weakening through the “Digital Omnibus” (COM [2025] 836 final). The proposal would make enforcement conditional on industry readiness, create exemptions for processing sensitive personal data and dilute accountability requirements. This rapid retreat raises a diagnostic question:

This paper argues that part of the answer lies in systematic participation asymmetry during the Act’s development – asymmetry that conditioned not only policy details but also the architecture of accountability. The argument concerns the specific mechanism that Better Regulation places at the centre of citizen access to expert-driven regulation; alternative legitimacy channels, including parliamentary deliberation and electoral accountability, are acknowledged as partially operative.

Through mixed-method analysis of European Commission consultations (2020–2024), we document a pronounced transformation in participation demographics. Total responses collapsed 75% between Phase 1 (White Paper, n = 1,215) and Phase 2 (Legislative Proposal, n = 303). Citizen participation fell from 30.8% to 3.96% during the legislative consultation – the moment of highest political stakes – while industry actors consolidated at 47%–54% of submissions in Phases 2 and 3. The 2024 implementation consultation maintained this asymmetry (47.2% industry and 5.74% citizens), indicating that exclusion was structural rather than contingent.

Three mechanisms explain how asymmetric participation conditioned regulatory outcomes, each operating at a distinct analytical level. Resource disparities filtered citizens out as consultation shifted from accessible policy questions to complex legal texts requiring professional expertise. The resulting concentration of participation among institutionally positioned actors did not merely reduce citizen numbers; demographic thinness and structural power imbalance reinforced one another, as participation barriers and institutional access co-occurred among the same set of actors. Institutional power imbalances then conditioned whose input counted: industry actors secured preferential access to decision-making venues, while civil society was relegated to advisory roles without voting authority. Definitional flexibility extended asymmetry beyond the consultation period itself, enabling subsequent narrowing of ambiguous provisions through interpretation processes dominated by the same actors who shaped them.

We distinguish between influence – where consultation outcomes align with industry preferences without constraining the regulatory instrument’s broader public functions – and structural capture – where participation asymmetries become embedded in regulatory architecture, constraining future governance possibilities. This distinction is formalised in the Theoretical Framework below.

Empirically, this study provides the first cross-phase longitudinal comparison of AI Act consultation demographics spanning 2020–2024, including original quantitative analysis of the 2024 implementation consultation. We document that civil society and industry identify prohibited AI practices at systematically divergent rates (mean 5.1x difference across six Article 5 categories). Normatively, we assess whether procedural consultation, as currently designed, can provide the democratic authorisation that durable AI governance requires.

Two research questions structure the analysis: how did participation demographics shift across the AI Act’s consultation phases and what mechanisms explain the pattern; and how was asymmetric participation associated with the Act’s regulatory provisions and what does this imply for democratic accountability in AI governance?

This paper proceeds as follows: Related Work reviews the relevant scholarship; Theoretical Framework develops the analytical concepts; Research Design and Methodology describes the study design; the Consultation Deficit documents the empirical patterns, addressing RQ1 through quantitative analysis of participation demographics across all three phases; From Consultation Asymmetry to Regulatory Architecture traces asymmetry to provisions; Governance Implications and Democratic Assessment addresses RQ2 through provision-tracing and public value assessment; and Conclusion follows. The analysis operates at the level of institutional conditioning and patterned association rather than deterministic causation, with the distinction between influence and structural capture operationalised through the four-indicator rubric in the Theoretical Framework, disciplining all inferential moves throughout.

EU consultation scholarship consistently documents organised-interest dominance in European Commission regulatory processes. Rasmussen and Carroll (2014) demonstrate that business and occupational associations represent the dominant category in Commission online consultations, finding that their participation is approximately 11 percentage points higher in these exercises than in the broader population of registered interest groups. Quittkat (2011) finds that open consultations remain dominated by business representatives and Northern member states, raising questions about whether the regime achieves its stated inclusive goals. Bobbio (2019) identifies an intrinsic contradiction in participatory arrangements: they simultaneously give voice to citizens and deploy that participation to legitimise decisions already framed, placing participants in the position of deliberating within pre-defined agendas rather than genuinely shaping them.

Independent institutional and empirical evidence confirms these structural patterns. The European Court of Auditors (ECA), 2019 evaluation of 26 Commission consultations, found no dedicated quality indicators, limited transparency about how input is weighted and inadequate citizen outreach (ECA, 2019, paras. 30–32, 44–48) – with only 33% of surveyed participants agreeing that the Commission takes account of citizens’ opinions (ECA, 2019, Annex III). Neumann et al. (2025), analysing 220 consultations from 2014 to 2021, found median citizen participation of 49 responses per consultation against 136 organisational responses, with for-profit organisations generating four times more input than nonprofits. These patterns establish the structural baseline against which the AI Act analysis proceeds: what we find is not an anomaly but a characteristic.

Critical analysis of the AI Act has identified significant regulatory limitations. Paul (2024) shows that the risk-based framework’s categorisation encodes contestable political judgments about which AI applications warrant intervention, while Molavi Vasse’i (2024) argues that the framework’s structural effect is to concentrate regulatory obligation on a narrow band of use cases, leaving the majority of AI applications under minimal oversight, and to relegate civil society to a consultative Advisory Forum without voting rights in the Act’s central decision-making body – replicating at the governance architecture level the participation asymmetry this paper documents at the consultation stage. Bakiner’s (2023) content analysis of 302 responses to the consultation on the Commission’s proposed Regulation Laying Down Harmonised Rules on Artificial Intelligence (COM [2021] 206 final) adds a constitutive dimension: he finds that the Commission’s regulatory text stabilised a particular sociotechnical imaginary – treating AI as a fully controllable technology – before stakeholder input could contest it, with businesses prioritising legal certainty and narrow definitions, non-governmental organisations foregrounding harm and structural inequality and almost no topical overlap between business and academic respondents.

What none of these studies has systematically examined is the consultation process itself as a site where democratic authorisation was – or was not – achieved. No study has documented participation demographics across the complete policy cycle, traced stakeholder feedback to final provisions or assessed consultation as a constitutive rather than merely input-gathering process.

The regulatory capture literature provides the analytical resource for this gap. Carpenter and Moss (2013) distinguish graduated levels from weak influence to strong capture, where regulatory action is “consistently or repeatedly directed away from the public interest” toward the regulated industry. Yackee and Yackee (2006) demonstrate that systematic participation asymmetry shapes not only policy details but also fundamental problem framing. This paper uses structural capture – a condition where participation asymmetries become embedded in regulatory architecture, constraining the instrument’s capacity to serve functions beyond industry interests – as the operative analytical concept, distinguished from influence, where outcomes align with industry preferences without foreclosing alternative governance possibilities.

Sheila Jasanoff’s (2004) concept of co-production holds that ways of knowing and ways of organising society are produced together. Applied to AI governance, regulatory consultation does not merely gather input on pre-existing technology – it actively constitutes what counts as “AI,” what problems require governance and who possesses authority to decide. Three constitutive pathways make this dynamic visible.

Problem constitution: consultation responses define what counts as AI risk warranting intervention, whether framed as a technical bias problem amenable to auditing or as a structural power asymmetry requiring institutional redesign. Actor authorisation: the demographic composition of participants determines which forms of knowledge are treated as credible, privileging technical and commercial expertise when industry actors predominate. Solution delimitation: the range of policy options narrows to those articulable within the dominant consultation discourse, such that proposals requiring capacities absent from the participant pool become structurally implausible. These pathways explain how resource disparities, institutional power imbalances and definitional flexibility map from distributional patterns into governance consequences: asymmetric participation does not merely skew preferences but also conditions the categories through which regulation is conceived.

Demonstrating that consultation shapes regulatory meaning; however, it does not by itself establish whether the resulting provisions lock asymmetry into governance architecture. A second analytical concept is needed. This paper distinguishes between influence, where consultation outcomes align with industry preferences without constraining the regulatory instrument’s broader public functions, and structural capture, where participation asymmetries become embedded in regulatory design, constraining the instrument’s capacity to serve functions beyond industry interests. The distinction is consequential: influence leaves alternative governance possibilities open; structural capture forecloses them.

The co-production theory explains how asymmetric consultation constitutes regulatory meaning, tracing the mechanism through which participation patterns shape problem definitions, authorise particular actors and narrow the range of feasible solutions. The structural capture rubric operates at a different level: it assesses whether the provisions that result from this process embed asymmetry into institutional design. The analytical sequence reflects this division of labour: the co-production analysis establishes how consultation shaped the Act’s regulatory categories, and the rubric then determines whether the resulting provisions constrain future governance possibilities in ways consistent with capture rather than mere influence.

This distinction is operationalised through four indicators (Table 1) ^[1]. Path dependency (does the provision structure future decisions in ways that systematically advantage industry?); implementation asymmetry (do implementing mechanisms privilege industry while marginalising public oversight?); contestability foreclosure (does the provision foreclose public-interest reinterpretation?); and limited reversibility (would correcting the provision require legislative amendment?). A finding of structural capture requires that all four indicators be satisfied at sufficient depth.

The framework is structural rather than intentionalist. Structural capture describes an emergent condition that need not have been planned to be consequential. A related but distinct term completes the analytical vocabulary. Where structural capture describes the embedded condition, structural vulnerability describes its downstream consequence: the susceptibility of capture-consistent provisions to implementation pressure, as documented through the Digital Omnibus case. Article 5 (prohibited practices) serves as a negative case: meaningful prohibitions survived the legislative process despite industry’s participation advantage, confirming that capture is concentrated in the classification and assessment architecture rather than evenly distributed across the Act. Article 27 (fundamental rights impact assessments) is a partial case where capture-consistent dynamics are present but incompletely embedded.

Mark Moore’s (1995) public value framework holds that legitimate public action requires substantive value, fair procedures and political legitimacy through democratic authorisation. Crucially, authorisation cannot be replaced by procedural compliance alone. Bozeman’s (2007) concept of public values failure identifies a characteristic risk: when market-oriented logics are imported into domains where different normative commitments should govern, democratic authorisation is displaced. Political theorists of European governance distinguish three legitimacy dimensions: input legitimacy (governing “by” the people), output legitimacy (governing “for” the people) and throughput legitimacy – the procedural quality, transparency and accountability of governance processes (Scharpf, 1999; Schmidt, 2013).

The two frameworks are analytically interdependent. Co-production explains the mechanism through which asymmetric participation shapes regulatory meaning, while public value theory provides the evaluative criteria for assessing outcomes. Together, they yield the bridging proposition that organises the empirical analysis: asymmetric co-production initially creates the appearance of throughput and output legitimacy while systematically hollowing out input legitimacy. As structural vulnerabilities accumulate and translate into implementation pressure, all three legitimacy dimensions are eroded. Input legitimacy is undermined through the exclusion of affected publics from constitutive consultation stages. Throughput legitimacy is compromised by opacity in how inputs are weighted and by the rejection of coordinated civil society demands without transparent justification. Output legitimacy is weakened through regulatory provisions conditioned by the participation profile rather than by a democratically representative range of risk assessments.

This framework positions consultation as one constitutive mechanism within a multi-pressure system. Member State preferences channelled through Council positions, informal lobbying, Commission deliberations and Parliament’s policy priorities also shaped the Act’s final architecture – pressures that fall outside what consultation data can observe. The analysis maintains that participation asymmetries structured the conditions within which those additional pressures operated, without claiming that consultation inputs were the sole or primary determinant of regulatory outcomes.

This study uses an explanatory sequential mixed-methods design (Creswell and Plano Clark, 2017). The Better Regulation framework, operationalising Article 11 of the Treaty on European Union, conducts consultations through the “Have Your Say” portal, open to any individual or organisation. The ECA (2019) found systematic shortcomings in this process, including unclear objectives and limited transparency about how input is weighted – structural deficiencies directly relevant to AI Act consultation dynamics. Quantitative analysis documents demographic composition across three strategic consultation phases to identify patterns of inclusion and exclusion; qualitative provision-tracing examines mechanisms through which participation asymmetries were associated with regulatory outcomes. Integration occurs through the co-production framework, which treats participation patterns as constitutive of regulatory character rather than as contextual background.

Integration between the two stages was theory-driven. The quantitative analysis established two findings that informed the qualitative case selection. First, in Phase 2 – the consultation context for the legislative proposal itself – industry held 53.5% of submissions against 3.96% citizen representation, establishing that the legislative architecture was shaped under conditions of pronounced industry numerical dominance. Second, the Phase 3 data revealed a mean 5.1 × divergence between civil society and industry in identification of prohibited AI systems across the six Article 5 categories (Table 5), establishing that the dominant constituency systematically under-identified the harms that prohibition provisions are designed to prevent. Together, these findings generated the expectation that provisions delegating interpretive authority to the dominant stakeholder group would exhibit capture-consistent characteristics. They also pointed to a specific class of provision: those at which the regulated entity self-determines whether its own systems pose the very risks it has been documented to under-identify. Article 6(3) was selected as the qualitative trace case because it occupies precisely this site – the point where Annex III classification architecture meets provider self-assessment. On this logic, it is the theoretically critical site for assessing whether demographic asymmetry was associated with capture-consistent regulatory design. This selection logic ensured that the qualitative stage tested the co-production mechanism where demographic asymmetry would be most analytically consequential, if it operated as theorised.

The AI Act’s development involved three distinct strategic phases (Table 2). The White Paper consultation (February–June 2020, 116 days, n = 1,215) posed high-level questions accessible to non-specialists. The legislative proposal consultation (April–August 2021, 108 days, n = 303) required engagement with 108 pages of legal text. The implementation guidance consultation (November–December 2024, 29 days, n = 383) focused on operationalising prohibitions and AI system definitions. Our argument concerns the magnitude and systematic direction of demographic asymmetry across these phases, which exceeds what functional differences in consultation type alone plausibly explain.

To facilitate analysis of 1,902 submissions totalling approximately 4,000 pages, we developed a local Retrieval-Augmented Generation pipeline for automated retrieval and targeted thematic search. The tool functioned exclusively as a retrieval and flagging mechanism; all interpretive and analytical decisions remained with the author. The tool is publicly available at https://github.com/thodoris/ec-consultation-analysis. Robustness was addressed through triangulation across independently produced data sources: Commission synopsis reports, the CEPS analytical report commissioned separately by the AI Office and original analysis of raw CSV response data. Two concrete validation steps complemented this triangulation. First, all quoted material from consultation responses used in the Article 6(3) analysis was manually verified against the original responses on the Have Your Say portal. The Siemens Energy, Siemens AG, EDRi and vzbv passages cited below were all checked for wording, attribution and contextual representativeness. Second, Retrieval-Augmented Generation outputs characterising industry preferences on self-assessment, classification thresholds and definitional scope were checked for convergence with the independently produced characterisations in the CEPS (2025) analytical report and the Commission synopsis reports. This convergence is what the triangulation claim above operationalises. These steps do not substitute for inter-coder reliability procedures, which were not feasible within the analytical workflow used here, but they document that interpretive decisions were anchored to source material rather than treated as self-validating outputs of the retrieval pipeline.

Two primary inferential constraints bound the analysis: the counterfactual problem (no alternative consultation process exists against which to compare) and the single-provision limitation (Article 6(3) was selected theoretically, not through systematic provision-level sampling and should not be extrapolated as representative of the Act’s architecture as a whole). Regulatory outcomes may also have been primarily determined by Member State preferences, informal lobbying or Parliament and Commission priorities that would have produced similar results regardless of consultation input. These limitations discipline all inferential claims: the analysis establishes structural association and conditioning, not determination.

The White Paper consultation (Phase 1) represents the most democratically diverse participation profile in the AI Act’s development. EU citizens constituted 30.8% of 1,215 valid responses – near-parity with business and industry at 29.0% (ratio 1.06:1). Non-market actors combined represented 44.0% in Phase 1 (citizens 30.8% and civil society organisations 13.2%) ^[2]. The accessible framing mattered: high-level questions about AI’s societal implications did not require technical expertise to engage. Citizens and non-governmental organisations strongly favoured new binding legislation establishing “red lines” prohibiting certain AI applications, while business representatives frequently argued that existing frameworks were sufficient and cautioned against over-regulation. This normative divergence anticipated the ideological asymmetries that would characterise subsequent phases.

The legislative consultation (Phase 2) reveals a pronounced shift. As format shifted from accessible policy questions to 108 pages of legal text requiring specialised regulatory vocabulary, the participant pool narrowed to those with institutional resources to engage. Total submissions fell 75% to 303, with demographic shifts substantially exceeding the absolute decline. Industry achieved majority control at 53.5%, while EU citizens declined to 3.96% – a 96.8% reduction in respondent count (from 374 to 12) and a 14-fold industry advantage in the citizen-to-industry ratio (0.07:1, from 1.06:1) ^[3]. Eight EU Member States had zero representation, indicating that nearly a third of the Union was entirely absent from the legislative consultation. Organised civil society increased proportional representation (13.2%–24.4%), suggesting responsive mobilisation as regulatory stakes materialised, but this was insufficient to compensate for citizen withdrawal. The consultation format operated as an exclusionary mechanism – not through intentional gatekeeping but through design choices that predictably filtered out non-specialist participants without a corresponding outreach strategy.

The implementation consultation (Phase 3) confirms that demographic asymmetry persists across the full regulatory cycle. Industry maintained 47.2%, while citizens remained at 5.74% – comparable to Phase 2’s 4.0%. The analytical report commissioned from CEPS acknowledged this explicitly: “The stakeholder representation in this survey appears significantly skewed and not representative of the general population… This suggests a heavy over-representation of industry and technical stakeholders, and an under-representation of the public” (CEPS, 2025, p. 13). The skew was visible, entered into the official record and nonetheless reproduced.

Table 3 summarises stakeholder composition across all three phases; Figure 1 visualises the longitudinal trajectory.

Taken together, the three phases document a longitudinal arc that cannot be described by any single phase in isolation. The citizen-to-industry ratio moved from near-parity in Phase 1 (1.06:1) to near-total inversion in Phase 2 (0.07:1), recovering only marginally in Phase 3 (approximately 0.12:1). By November 2024, conditions that might plausibly have generated demographic correction existed – a new institutional actor, a formally adopted regulation, heightened public awareness – yet none produced it. Citizens remained at 5.74%, virtually unchanged from Phase 2’s 3.96%. This non-occurrence of demographic correction under conditions structurally available for recovery constitutes stronger evidence of structural embedding than the Phase 2 collapse itself, which could be attributed, however partially, to the novelty of the format shift.

The 29-day Phase 3 window, substantially compressed relative to Phases 1 and 2 (108–116 days), compounded this pattern. Rapid-turnaround windows disproportionately advantage organisations with dedicated regulatory affairs capacity. That the AI Office selected a window less than a third the length of the preceding phases – despite access to Phase 2’s documented underrepresentation and the ECA (2019) audit findings – is itself analytically part of the pattern, consistent with resource disparities operating through temporal design as well as content.

Geographic concentration reinforced these dynamics. Germany, Belgium and France accounted for 44%–56% of submissions throughout the cycle. Belgium’s share doubled to 27.4% in Phase 2 as Brussels-based institutional actors dominated. US participation rose from 4.9% to 8.6% as binding obligations materialised. Within industry, large companies constituted 62.6% of company respondents in Phase 1 and approximately 65% in Phase 3: the dominant stakeholder group was itself internally unrepresentative of the interests it nominally contained.

A contemporaneous benchmark reinforces the structural interpretation. The DSA pre-legislative consultation (June–September 2020, n = 2,863) achieved 66% EU citizen participation under the same institutional framework and in the same year as the AI Act White Paper. That higher citizen participation was demonstrably attainable under comparable institutional conditions is the most direct evidence that the AI Act’s consultation architecture, not the regulatory domain, accounts for the exclusion documented here.

Demographic asymmetry is associated with ideological asymmetry. Phase 3’s structured consultation data enables quantitative assessment. Respondents rated the importance (1–10 scale) of clarifying seven elements in the AI system definition; cross-tabulation by stakeholder type reveals significant divergence in what different actors wanted regulation to accomplish (Table 4).

Civil society prioritised clarifying “inference” mechanisms (8.36) – seeking broad coverage of complex algorithmic decision-making – while rating “adaptiveness” low (4.20). Industry prioritised “adaptiveness” (7.78), seeking narrow definitions that exclude systems not learning after deployment, thereby potentially exempting large categories of algorithmic systems from oversight. The alignment between industry organisations (7.78) and AI providers (7.58) on “adaptiveness”, contrasted with civil society’s low rating (4.20), is consistent with coordinated boundary-work seeking exclusionary definitions. Citizens exhibited distinct priorities – high ratings for “autonomy” (8.15) and “adaptiveness” (7.45), reflecting concern about systems operating independently of human control, and a low rating for “inference” (4.20) – reflecting a lay framing of AI harm that the consultation’s technical vocabulary did not translate into regulatory influence.

The systematic direction of these divergences – industry consistently prioritising narrower regulatory scope – is consistent with the co-production dynamic identified in the Theoretical Framework: when industry-dominated consultation shapes which definitional elements receive priority, the operational meaning of “AI system” is constituted through a process that privileges market-oriented interpretations over rights-protective ones. The same actors who prioritised narrow “adaptiveness” definitions are, under the adopted regulatory architecture, the actors who make the initial self-determination of whether their own system poses “significant risk” under Annex III.

A related asymmetry emerges in how stakeholders perceive whether AI systems fulfil prohibition criteria. Phase 3 data reveal epistemic divergence across all six Article 5 categories (Table 5).

Civil society identified concrete examples of prohibited systems at a mean rate of 54.0%; combined industry stakeholders did so at 10.6% – a mean ratio of 5.1×. The gap is widest for crime risk prediction (9.3×) and facial image scraping (9.1×) and narrowest for social scoring (2.5×). Three explanations are plausible and likely operate simultaneously. Civil society’s mission – serving marginalised communities, documenting discrimination complaints – may enable perception of harms that industry’s structural position obscures. Industry respondents may strategically minimise identification to reduce perceived regulatory burden, consistent with the high industry N/A rate (56.9% versus 33.0% for civil society). Or civil society may apply broader definitions, with divergence widest precisely where prohibition criteria are least precisely specified. What the data establish unambiguously is that the consultation’s participant pool is systematically associated with low rates of prohibited-system identification. When industry dominates at 47.2% in Phase 3, the Commission receives signals disproportionately suggesting that Article 5 prohibitions either do not apply to current systems or that respondents lack the information to assess applicability. When regulatory guidance is produced on the basis of this asymmetric epistemic input, it institutionalises the understanding generated by the participant pool’s dominant constituency – consistent with what Jasanoff (2004) terms problem constitution through co-production.

The asymmetries documented above replicate structural features of EU Better Regulation consultations identified by independent institutional evaluations. The ECA’s audit found no dedicated quality indicators, consultation strategies published in only 12 of 22 examined cases and only 7 of 23 synopsis reports detailing analytical methods (ECA, 2019). Neumann et al. (2025), analysing 220 consultations from 2014 to 2021, found median citizen participation of 49 responses versus 136 organisational responses, with citizen engagement correlating with nonprofit mobilisation (β = 1.46, p < 0.001) but not with for-profit engagement – indicating that citizens participate primarily when civil society intermediaries mobilise them, consistent with the resource disparities mechanism. Because the patterns observed in the AI Act replicate those documented across dozens of unrelated consultations, explanations focused solely on AI Act-specific political dynamics cannot account for the asymmetry. The structural location of the problem – embedded in the Better Regulation architecture itself – is what requires explanation and reform.

The consultation patterns documented above raise a central question: was asymmetric participation associated with the character of the regulatory provisions that emerged? This section traces the single critical case identified through the case selection logic set out in the methodology – the Article 6(3) self-assessment architecture. Article 6(3) satisfies all four structural capture indicators, and the three co-production mechanisms are jointly visible in its development. Findings from this case should not be extrapolated as representative of the Act’s regulatory architecture as a whole; they demonstrate that capture-consistent dynamics operated in at least one consequential provision, consistent with the institutional conditioning logic developed above. Trilogue outcomes reflect multiple institutional pressures beyond formal consultation, including bargaining between Parliament, Council and Commission and lobbying through informal channels; the analysis asks whether the systematic direction of provision-level outcomes is consistent with the asymmetric participation profile documented above.

From consultation preferences to regulatory architecture: Industry respondents in Phase 2 expressed a strong preference for provider self-assessment over mandatory third-party conformity evaluation. Siemens Energy argued that “in-house conformity assessment bodies must be allowed to assess the conformity of AI systems/applications” (Siemens Energy AG, 2021, p. 3). Siemens AG added that mandatory registration was “not necessary or proportionate” (Siemens AG, 2021, p. 3). What emerged in the final Act reflects this orientation closely. Most high-risk AI systems listed in Annex III of the AI Act (points 2–8) are subject to provider self-assessment rather than independent evaluation. Article 6(3) introduces a discretionary filter allowing providers to determine that their system, despite Annex III listing, does not pose “significant risk of harm” if it satisfies one of four conditions: performing a “narrow procedural task,” being intended to “improve the result of a previously completed human activity,” “detecting decision-making patterns or deviations” without replacing prior human assessment, or performing a “preparatory task.” A provider developing a hiring AI system can self-determine that it “improves the result of” human recruiters’ screening, classifying its own system as non-high-risk. No external body reviews the assessment before market entry. A regulatory design placing interpretive authority outside the regulated entity would retain Annex III automatic classification with third-party conformity assessment as the default, inverting the burden from provider self-exemption to provider-initiated, externally verified exception. That the adopted architecture placed interpretive authority with the regulated entity – over coordinated civil society opposition – is the outcome this case documents.

Civil society opposition and its fate: Civil society organisations mobilised against the adopted architecture with coordination. EDRi, backed by 118 civil society signatories, demanded closure of what they characterised as the Article 6(3) loophole, arguing that it “risks undermining the entire AI Act, shifting to self-regulation,” and called for restoration of the Commission’s original automatic risk-classification approach (European Digital Rights [EDRi], 2023). The Federation of German Consumer Organisations (vzbv) grounded the democratic illegitimacy of self-assessment in independently commissioned survey evidence. Drawing on TÜV-Verband’s consumer research, vzbv noted that a large majority of German consumers favoured independent pre-market safety assessment over provider self-certification (vzbv, 2021). Despite this mobilisation, the discretionary filter survived trilogue intact.

Four-indicator assessment: Article 6(3) satisfies all four structural capture indicators cumulatively, with each indicator’s decision rule (Table 1, footnote) met by specific features of the provision and its trilogue history:

1. Path dependency is satisfied because reform exceeds delegated administrative correction: full revision requires legislative amendment under the ordinary legislative procedure. The formal corrective capacity that Article 43(6) confers on the Commission, to adopt delegated acts requiring third-party assessment, depends on institutional will not yet demonstrated.

2. Implementation asymmetry is satisfied because no mandatory pre-market external review applies: providers self-interpret “narrow procedural task” and “significant risk,” and the ex-post review mechanism under Article 80 operates reactively, after market deployment.

3. Contestability foreclosure is satisfied because the four conditions are sufficiently abstract to sustain a wide range of provider interpretations and because no mandatory public-interest review mechanism is attached to the self-determination.

4. Limited reversibility is satisfied because coordinated civil society demands – backed by 118 signatories and surveyed consumer evidence – were raised and rejected during trilogue and because correction now depends on co-legislative procedure rather than administrative adjustment. Considered cumulatively, the four indicators are satisfied at sufficient depth to support the structural capture characterisation. This distinguishes Article 6(3) from the Article 5 case, where meaningful prohibitions survived, and from Article 27, where capture-consistent dynamics are present but incompletely embedded because reversibility remains more accessible through delegated authority. All three co-production mechanisms are present in this case. Resource disparities meant civil society engaged less effectively with the technical complexity of the Annex III/Article 6(3)/Article 43 interaction. Institutional power asymmetries meant that the coordinated civil society opposition documented above was rejected, while industry preferences were incorporated. Definitional flexibility enabled providers to self-interpret key terms, extending interpretive autonomy from consultation into ongoing implementation.

Assessed through the public value framework, the regulatory architecture documented above constitutes deficits across all three legitimacy dimensions. Input legitimacy requires that those subject to regulation could meaningfully shape decisions. The Act exhibits a significant gap: groups structurally more exposed to fundamental rights risks from AI – communities subject to AI-mediated discrimination, workers under algorithmic management – were substantially underrepresented in shaping it (Scharpf, 1999). No mandatory participatory design requirements govern high-risk AI deployment. The fundamental rights impact assessment mechanism (Article 27) exemplifies this gap: Parliament’s original mandatory language was weakened to discretionary guidance in Recital 96, leaving deployers free to determine whether to consult affected stakeholders. With trade unions at only 4.95% of Phase 2 responses, the regulatory framework addresses whether hiring algorithms are discriminatory, but not whether hiring decisions should be algorithmic. When structurally exposed groups are underrepresented in consultation, the problems the regulation addresses risk being defined by those least exposed to them.

Throughput legitimacy concerns the quality and inclusiveness of governance processes (Schmidt, 2013). The Article 6(3) self-assessment architecture delegates ongoing governance authority to regulated actors without pre-market external review. Institutional arrangements reinforce this pattern: industry secured AI Board seats with decision-making authority, while civil society was relegated to the Advisory Forum in a consultative role – replicating the participation asymmetry documented during consultation at the governance level. The ECA’s finding that only 7 of 23 synopsis reports detailed analytical methods for processing data (ECA, 2019) indicates that participants submitted feedback without knowing how it would be evaluated relative to other inputs. The same asymmetry is visible at the standards-setting stage that operationalises the Act’s requirements, corroborating the implementation asymmetry indicator beyond the consultation phase: Kilian, Jäck and Ebel (2025) find that CEN-CENELEC JTC 21 standardisation committees are predominantly composed of large enterprises, with SMEs, civil society organisations and academia structurally under-represented because of the same resource constraints documented above.

Output legitimacy requires that governance effectively addresses public needs. The Act centres individual rights – transparency, non-discrimination and data protection – while provisions for collective contestation, participatory oversight and democratic control over whether AI should be deployed for particular purposes remain weak. The solutions that emerged – market-compatible rights frameworks, provider self-assessment and technical standardisation – correspond to the constrained range of options available to the consultation’s participant pool. By constructing trustworthiness primarily as auditability, the regulatory discourse channelled questions of democratic authorisation – whether particular AI deployments should occur at all – into questions of technical mitigation and procedural verification.

The Digital Omnibus proposal (COM [2025] 836 final) provides a provisional test of this dynamic. Barely 16 months after the AI Act’s adoption, the Commission proposed amendments that would make high-risk system enforcement conditional on industry-defined “adequate support measures,” create a GDPR carve-out enabling processing of special category personal data for “bias detection and correction” (Article 4a) and eliminate registration for self-exempted systems. The Omnibus’s origins are multi-causal, including the 2024 electoral shift toward competitiveness priorities and broader deregulatory pressures. The co-production analysis adds a specific dimension: the consultation process was the site at which the operative meanings of “risk,” “high-risk” and “significant harm” took shape, and these are the same meanings now framing the proposed revision. The amendments concentrate on precisely the accountability mechanisms – self-assessment authority, registration requirements and fundamental rights obligations – that the Article 6(3) analysis identifies as capture-consistent. This pattern is consistent with the expectation that regulation shaped through asymmetric consultation may be structurally vulnerable when implementation costs materialise.

Before reaching a democratic deficit conclusion, it is necessary to acknowledge alternative legitimacy channels. The European Parliament introduced several of the Act’s stronger fundamental rights provisions, and electoral accountability operates through MEPs’ responsiveness to public AI concern. The legitimacy deficit identified here is accordingly more specific: the Better Regulation consultation mechanism systematically failed to provide the participatory quality that its own standards claim. The ECA’s finding that only 33% of surveyed participants agreed their input is taken into account (ECA, 2019, Annex III) indicates that even those who participated doubted the process’s responsiveness.

The findings point toward institutional reforms that go beyond procedural adjustments. Three immediate-term reforms would improve throughput quality within the existing framework. Mandatory transparency requirements for how consultation inputs are weighted would respond to the ECA’s finding that only 7 of 23 synopsis reports detailed analytical methods (ECA, 2019). Dedicated civil society capacity grants would address the resource-disparity mechanism that drove citizen participation from 30.8% to under 4%. A formal disagreement register would require the Commission to document and justify instances where coordinated stakeholder demands were overridden. In the medium term, parliamentary authority to require targeted consultations addressing identified participation gaps before legislative proceedings advance would address structural exclusion without fundamental institutional redesign ^[4]. In the longer term, deliberative mechanisms with binding authority – citizens’ assemblies adapted to AI governance – represent the most substantive response, overcoming the resource-and-expertise barriers that self-selection mechanisms cannot address. The evidence establishes that systematically more inclusive participation was structurally attainable and that the AI Act’s consultation architecture failed to achieve it.

This study has documented systematic participation asymmetry across the EU AI Act’s consultation cycle and traced its association with regulatory outcomes. Total responses collapsed 75% between Phases 1 and 2. EU citizen representation fell from 30.8% to under 6% during the decision-critical phases, while industry consolidated at 47%–54% in Phases 2 and 3. A contemporaneous comparison with the DSA consultation establishes that 66% citizen engagement was achievable under the same institutional framework, indicating that the exclusion documented here tracks the AI Act’s consultation architecture rather than the regulatory domain. The proposed Digital Omnibus amendments, targeting precisely the accountability mechanisms identified as capture-consistent, are consistent with the expectation that regulation shaped through asymmetric consultation may be structurally vulnerable when implementation costs materialise.

Resource disparities, institutional power imbalances and definitional flexibility ran across reinforcing levels – demographic, organisational, geographic and epistemic – yielding a consultation record that conditioned regulatory architecture rather than merely accompanying it. Through the co-production framework, these asymmetries shaped not merely policy details but what “trustworthy AI” came to mean. The Article 6(3) self-assessment architecture is the clearest observable expression of this dynamic within the provisions examined: a single but consequential case satisfying all four structural capture indicators, adopted over coordinated civil society opposition backed by more than 150 signatories. Whether comparable dynamics operated across other provisions remains an open empirical question.

The evidence suggests that procedural consultation as currently designed under the Better Regulation framework is insufficient to ground the democratic authorisation that durable AI regulation requires. Systematically more inclusive participation was structurally attainable – the DSA demonstrates this – and the AI Act’s consultation architecture failed to achieve it. Institutional reform moving toward deliberative mechanisms with binding authority is needed before the next major AI governance cycle. Future research should examine whether these patterns replicate across AI Act implementation consultations and comparable regulatory processes in other jurisdictions. Comparative analysis across non-EU AI governance contexts – including jurisdictions with different consultation traditions and varying levels of civil society capacity – would test whether the structural mechanisms identified here operate as general features of technically complex regulation or are specific to the EU Better Regulation architecture.