The Amplifier Effect: Human-Factor Risks of AI-Suggested Correlation and Auto-Propagation in Multi-Framework GRC Self-Assessment

I. INTRODUCTION

• The CRA, NIS2, DORA, and EU AI Act depend on truthful, consistent organisational answers as cyber-resilience duties expand.The paper describes a shift from periodic external audit toward continuous self-assessment as the default evidence of compliance.Source: §I · p. 1
• Multi-framework control mapping and embedded AI assistance reduce repeated work but create a measurement problem for human-factor effects.The paper calls the resulting property the amplifier effect: one human action drives many compliance assertions without proportional additional review.Source: §I · p. 1
• The amplifier effect is a platform-design risk: coarse attestation, ungated propagation, and AI suggestions that are hard to distinguish from verified fact.In GRC, the error cascade ends in regulatory assertions, turning a general design failure into a compliance failure.Source: §I · p. 1

A. Concrete amplification mechanism

• CYBERFORT and CYBER-BRIDGE use human-authored and AI-suggested question correlations to propagate an answer and its evidence to correlated questions.The result is many compliance claims across one or more frameworks from a single user action.Source: A. Concrete amplification mechanism · pp. 1–2

B. Why the human-factor lens matters

• Amplification converts answer-level biases into organisation-level compliance posture without an intermediate review gate, while the UI presents propagation as a productivity feature.When AI partly authors the correlation graph, automation bias can affect which propagations receive human scrutiny.Source: B. Why the human-factor lens matters · p. 2

C. Contributions

• The paper defines the amplifier effect, identifies four code-level mechanisms, proposes six exposure dimensions, and assesses thirteen representative tools.The mechanisms are transitive correlation, answer synchronisation, evidence propagation, and AI-suggested new correlations.Source: C. Contributions · p. 2
• The authors do not run a controlled user study; they outline a five-question protocol for consortia with production self-assessment data.The protocol is described as a next deliverable for the CYBERFORT and CYBER-BRIDGE consortia.Source: C. Contributions · p. 2

A. Self-Assessment Reliability

• Self-assessment can reflect social-desirability bias, Dunning–Kruger, expert–non-expert gaps, and optimism under organisational incentives,.Automation bias predicts over-reliance on system-supplied answers, particularly when an AI assistant presents them.Source: A. Self-Assessment Reliability · p. 2
• Independent question review bounds reliability one answer at a time; when one answer drives many, the same biases affect many answers at once.Source: A. Self-Assessment Reliability · p. 2

B. The Multi-Framework GRC Data Model

• Across enterprise IRM, compliance automation, and EU SME projects, a single control or evidence object can satisfy multiple obligations.The frameworks named include ISO/IEC 27001, NIS2, NIST CSF, SOC 2, GDPR, DORA, and the CRA.Source: B. The Multi-Framework GRC Data Model · p. 2
• Continuous Control Monitoring replaces periodic self-attestation with telemetry for some controls, but works less well for organisational controls.The paper locates the amplifier effect where CCM does not reach.Source: B. The Multi-Framework GRC Data Model · pp. 2–3

• Security self-report instruments HAIS-Q and SeBIS measure awareness and intended behaviour, while documented social-desirability ceilings and awareness-campaign limits motivate examining answer amplification [17].Source: §III · p. 3
• Cognitive forcing functions can reduce AI overreliance, while the accountability–automation-bias trade-off informs the paper’s recommendation R2.The paper uses favourability bias for an AI assistant’s tendency to produce overly favourable answers when prompted to advocate for the user.Source: §III · p. 3
• The amplifier effect extends compliance critique by linking favourable answers and amplified consequences to a cheaper local outcome than per-question scrutiny,.Source: §III · p. 3
• The paper uses Gartner’s segmentation to select thirteen tools and adds a human-factor dimension to comparisons of GRC functionality and total cost of ownership.Source: §III · p. 3

A. Walk-Through

• One saved answer is the sole human action in a chain that operates across questions linked in a stored correlation graph.The answer is a value plus an evidence document tied to one question in one assessment.Source: A. Walk-Through · p. 3
• Transitive correlation groups questions together, so linking qa to qb and qb to qc places all three in one group.Source: A. Walk-Through · p. 3
• Answer synchronisation writes the same value to every question in the group across the organisation’s open assessments.One save can therefore record up to |G(q)| answers.Source: A. Walk-Through · p. 3
• The original evidence document is copied onto every synchronised answer, substantiating each claim without further review.Source: A. Walk-Through · p. 3
• Accepted AI-suggested correlations enlarge the group, increasing the reach of the next answer without review of the propagation itself.Fig. 1 shows the correlated assessments and the AI-suggestion action in CYBERFORT.Source: A. Walk-Through · p. 3, Figure 1 · p. 3

B. Numerical Sensitivity

• For |F| = 4 frameworks, when mean cross-framework correlation degree is 3, A ≈10: one human attestation supplies the evidentiary basis for ten compliance claims.For CRA-anchored question banks also covering ISO/IEC 27001 and NIS2, mean degree is empirically 2–5.Source: B. Numerical Sensitivity · p. 3

C. A Worked Example

• In the CYBERFORT example, transitive links group four CRA, ISO/IEC 27001, and NIS2 questions, including vulnerability-management and disclosure questions.Source: C. A Worked Example · p. 3
• CYBER-BRIDGE’s assessment view also shows tenant assessments receiving the answer, assigned policy, and evidence description through propagation.Source: Figure 2 · p. 4
• A Yes answer and OSV-Scanner report are copied to the other three questions, so a report covering only one of three product components silently carries the scope error into all four claims.The platform records separate audit-trail entries per database row, although the human-factor surface is one decision.Source: C. A Worked Example · pp. 3–4

D. The CYBER-BRIDGE Variant

• CYBER-BRIDGE retrieves the K questions most similar to the current question, drafts answers for them, and updates K answers together when accepted.This retrieval assistant adds a second amplification path alongside the explicit correlation graph.Source: D. The CYBER-BRIDGE Variant · p. 4

V. SIX-DIMENSION SCORING FRAMEWORK

• The six-dimension framework scores exposure from NONE to STRONG and aggregates the six scores as a rounded mean, with higher scores better.The aggregate maps NONE, LIGHT, MODERATE, and STRONG to 0, 1, 2, and 3.Source: §V · p. 4
• The dimensions cover attestation surface, AI-assistance depth, sign-off granularity, favourability-bias mitigation, training-transfer measurement, and self-report–telemetry corroboration.For AI-assistance depth, STRONG requires human rewriting of AI drafts, while NONE permits suggestion, propagation, and persistence without a review gate.Source: §V · p. 4

VI. THIRTEEN-TOOL COMPARISON

• The comparison spans thirteen tools across six dimensions; scores draw on public product pages and market analysis, with direct code inspection for the two EU SME projects.The scores are indicative design-risk ratings based on inspected development versions, which may change as the projects mature.Source: §VI · p. 4, Table I · p. 5
• Enterprise IRM suites cluster toward the top through deep CCM integrations and mature sign-off workflows, but remain weak on training-transfer measurement and favourability-bias mitigation.Source: §VI · p. 4
• Compliance-automation tools maximise CCM but offer weak sign-off granularity and no favourability-bias mitigation, allowing one user to drive a complete framework attestation.Source: §VI · p. 4
• CYBERFORT and CYBER-BRIDGE score lowest because their answer propagation is designed to reduce SME workload, the feature penalised by the amplifier dimensions.The authors frame this as a design trade-off, not a verdict on the platforms.Source: §VI · p. 4
• Their scanners corroborate technical controls but not organisational controls such as training, leaving those self-reports unchecked.Source: §VI · p. 4
• The narrower concern is that weak per-answer sign-off combined with AI-written answers can leave many compliance claims resting on little human attestation.The authors note that regulation assumed a human meaningfully stood behind each claim.Source: §VI · pp. 4–5

VII. MEASUREMENT PROTOCOL

• Q1 tests for favourable-answer skew, especially where questions lack objective telemetry, as a headline optimism-bias signal.It reports the empirical distribution of Answer.value across users.Source: §VII · p. 5
• Q2 treats short answer dwell time on favourable correlated answers as a rubber-stamp signal, while Q3 uses heavy-tailed evidence reuse as a single-point-of-failure signal.Q2 computes updated_ati − created_ati; Q3 counts how many Answers link to each Evidence object.Source: §VII · p. 5
• Q4 compares answer-value and dwell-time distributions between logged AI-assisted and unassisted answers to distinguish acceleration-consistent patterns from bias-consistent patterns.A leftward dwell-time shift without an answer-distribution shift is consistent with acceleration without bias change; shifts in both are consistent with automation bias.Source: §VII · p. 5
• Q5 probes favourability bias by prompting each platform’s LLM with neutral and SME-favourable instructions on ten representative CRA questions and measuring rubric divergence.The probe uses Mistral 7B for CYBERFORT and phi-4 for CYBER-BRIDGE.Source: §VII · p. 5
• The five questions are mechanically reproducible once instrumentation is in place; Q1–Q3 use existing tables, Q4 may need per-answer AI logs, and Q5 is a one-day probe.Source: §VII · p. 5

A. A Usability-Heuristic Reading

• Silent propagation obscures when one answer becomes many claims, while missing confirmation and undo controls weaken user control and error prevention.AI-suggested correlations presented like human-authored ones also erode users’ understanding of what the system knows.Source: A. A Usability-Heuristic Reading · p. 5

B. Design Recommendations

• Require per-target acknowledgement for changed existing answers or mandatory targets sourced from non-mandatory questions, and log AI actions with user, prompt, and output hash.The log is intended to support Q4 and external audit.Source: B. Design Recommendations · p. 5
• Both projects log cyber-range completion but not behavioural transfer; a pre/post-training phishing simulation rate would raise the transfer-measurement score from LIGHT to MODERATE.Source: B. Design Recommendations · p. 6

C. Threats to Validity

• Only CYBERFORT and CYBER-BRIDGE received direct code inspection; the eleven commercial-tool scores rely on public documentation and may change with trials or vendor input.The structural amplification factor is an upper bound, with realized amplification dependent on deployment correlation-graph density; the protocol is not yet executed.Source: C. Threats to Validity · p. 6

D. Ethical and Deployment Considerations

• The production-data protocol assumes aggregate-only publication, no individual- or tenant-identifiable rows leaving the consortium, and prior tenant notification through data-sharing agreements.Q5 uses synthetic vignettes rather than tenant answers and poses no additional disclosure risk.Source: D. Ethical and Deployment Considerations · p. 6

IX. CONCLUSION

• The paper’s next step is to run the protocol on CYBERFORT’s first cohort and pair log analysis with user research on propagation visibility and attestation behaviour.The authors present the amplifier effect as a structural GRC property that becomes tractable to measure once named.Source: §IX · p. 6