Position Paper: If Innovation in AI Systematically Violates Fundamental Rights, Is It Innovation at All?

AI summary

Overview

Research area: AI safety and ethics, with a focus on technology policy, EU regulation and fundamental rights governance.

Technical level: Beginner-Friendly to Intermediate. The paper is a legal and policy position paper rather than a technical or experimental study; it requires no machine learning background and is written for readers interested in governance, law and responsible AI.

Scope: A single-sentence scope: the paper argues that well-designed regulation, exemplified by the EU AI Act, is a precondition for rather than an obstacle to responsible AI innovation.

What This Paper Is About

The paper challenges the widely repeated claim that regulation and innovation are opposites, a framing the authors call a false dichotomy. Using historical examples from pharmaceuticals, aviation and welfare systems, plus recent AI harms such as synthetic misinformation, bias and unaccountable decision-making, the authors argue that the absence of well-designed regulation has already produced substantial damage. Their goal is to show, through the EU AI Act as a case study, that risk-based and adaptive governance can accelerate responsible innovation rather than delay it.

Key Contributions

1. Reframes the regulation-versus-innovation debate as a false dichotomy. The paper argues that no field with major public implications has flourished without a regulatory framework, and that failures attributed to regulation are usually design flaws or the absence of regulation rather than an inherent tension between rules and progress.

2. Redefines "innovation" as inherently responsible innovation. Drawing on the OECD's concept of responsible innovation and the Oslo Manual's expanded guidelines, the paper rejects the notion of innovation that is devoid of accountability and ethical consideration, and asks whether systems that systematically violate rights can be called innovative at all.

3. Presents the EU AI Act as a model that addresses the Collingridge Dilemma. The paper details the Act's risk-based classification and its adaptive mechanisms — regulatory sandboxes (Art. 57), real-world testing (Art. 60), SME support (Art. 62) and the Fundamental Rights Impact Assessment (Art. 27) — as instruments that act early enough to prevent harm while remaining flexible enough to sustain innovation.

4. Maps perceived regulatory burdens onto actual benefits. Table 1 in the paper sets out a structured comparison of common criticisms (regulation blocks innovation, high compliance costs, excessive bureaucracy, slows AI adoption) against corresponding mechanisms and advantages (sandboxes, SME support, dedicated support channels, consumer trust, legal certainty, ethical leadership and long-term advantage).

Main Findings

• The deregulatory argument overlooks the Collingridge Dilemma. Collingridge observed that "when change is easy, the need for it cannot be foreseen; when the need for change is apparent, change has become expensive, difficult and time consuming." The paper notes Collingridge did not claim the dilemma is insurmountable and advocated proactive, evidence-adaptive governance.

• Historical precedents show regulation fostering rather than hindering progress. The thalidomide scandal, in which over 10,000 babies were born with severe physical abnormalities, led to the Kefauver-Harris Amendments in the USA mandating rigorous clinical trials. In aviation, the Federal Aviation Act of 1958 established the Federal Aviation Administration; where a passenger in the USA or Canada in 1959 faced a 1-in-25,000 chance of a fatal crash per departure, regulatory developments including Advisory Circular AC 25.1309-1 in the 1980s contributed to an over 1,000-fold improvement in aviation safety, with recent odds of a fatal accident in the USA or EU dropping to 1 in 29 million.

• Opaque automated decision-making has already caused major harm. The Dutch SyRI scandal saw a 2020 court ruling that an automated fraud detection system violated human rights due to lack of transparency and disproportionate targeting of low-income communities, ultimately leading to the government's resignation.

• AI harms are described as real, ongoing and escalating. Examples cited include deepfake audio mimicking a candidate during the 2023 Slovak elections, synthetic non-consensual sexual content including deepfake pornography and child sexual abuse material (CSAM), the "Almendralejo case" involving minors in Spain, and AI-generated explicit images of Taylor Swift.

• Bias is socio-technical, not purely algorithmic. The paper states that human and machine decisions often reinforce each other, creating compounded discrimination, and cites a comprehensive literature review identifying 152 concrete measures for addressing and reducing biases, including bias detection algorithms, diverse and inclusive datasets and rigorous auditing.

• Unaccountable decision-making involves a deeper risk than automation. The authors argue symbolic human oversight can be worse than no oversight, creating a false sense of legitimacy, and identify "induced irreflexivity" — where automated systems press humans to decide without questioning authorship or consequence — as the deeper risk. Moral outsourcing shifts ethical responsibility from humans to AI systems.

• Regulatory sandboxes will be mandatory across the EU. Article 57 of the AI Act makes sandboxes mandatory in each EU Member State by August 2026. Spain has already launched its first national sandbox for high-risk AI systems. The paper notes 57 countries currently operate 73 fintech sandboxes, and that sandboxes offer legal certainty, risk mitigation, no fines for good-faith participants, cross-sectoral collaboration and mutual learning.

• SME support is prescribed in four forms under Article 62. Priority access to sandboxes, awareness and training, dedicated communication channels, and facilitation of standardization. Microenterprises (fewer than 10 employees, less than €2 million turnover) may implement a simplified quality management system under Article 63.

• Compliance has driven rights-protective innovation before. The paper cites GDPR catalyzing privacy-enhancing technologies including differential privacy and federated learning, and AI regulation accelerating watermarking technologies for identifying AI-generated content and state-of-the-art measures to safeguard copyright during training of general-purpose AI models.

• The paper acknowledges but rejects alternative views. It engages with the Draghi report's warning about excessive regulation hindering EU digital competitiveness, arguments from Ridley and Braben on freedom from bureaucratic and ethical constraints, Castro et al.'s call for optimism-driven targeted governance, Gikay's critique of the Act's rigid risk classification, and Wachter's identification of legal loopholes.

Methodology in Plain English

This is a position paper, not an empirical study — the authors state no experiments, datasets or benchmarks. Their approach is argumentative and comparative. They begin by identifying the dominant claim that regulation hinders innovation, then test it against historical precedents in high-risk sectors (pharmaceuticals, aviation, welfare systems) where regulation followed or accompanied major harms and was later associated with safety improvements. They apply the Collingridge Dilemma as an analytical framework for the timing problem in AI governance. They then use the EU AI Act as an illustrative case study, walking through specific articles and mechanisms for sandboxes, real-world testing, SME support, impact assessments, transparency, accountability and AI literacy, and organising perceived burdens against actual benefits in a comparison table. Finally, they present and respond to counterarguments from named critics of AI regulation.

Why This Matters

Impact on research: The paper contributes to the policy and AI ethics literature by arguing for a conceptual shift — treating regulation as constitutive of innovation rather than opposed to it — and by connecting legal instruments such as the FRIA to anticipatory design and regulatory learning. It frames "unaccountable decision-making" and "induced irreflexivity" as governance problems requiring structured distrust and reflexive human oversight.

Real-world applications:

• Regulatory sandbox design: Informing how EU Member States implement the mandatory sandboxes required by August 2026, including Spain's existing national sandbox for high-risk AI systems.
• Small business compliance: Guiding SMEs and start-ups on priority sandbox access, training, dedicated communication channels and simplified quality management systems under Articles 62 and 63.
• Rights impact assessment practice: Supporting organisations in conducting Fundamental Rights Impact Assessments under Article 27, including evaluation of the severity, duration and reversibility of rights interferences.
• AI literacy programmes: Informing training in high-risk domains, drawing on resources such as the EU Living Repository on AI Literacy and its FAQ guidance.

Industry relevance: The paper argues that early compliance can be a competitive advantage, citing the Porter hypothesis, a first-mover advantage for firms that anticipate regulatory expectations, and Apple Inc. as an example where CEO Tim Cook positioned user data privacy as both business strategy and ethical imperative. It quotes Lucilla Sioli, head of the European Commission's AI Office: "You need the regulation to create trust, and that trust will stimulate innovation." The paper also notes that complying with intellectual property law has become one of the most pressing legal and operational challenges for the AI industry, as illustrated by a growing wave of litigation over training data and copyright, and that the Code of Practice on General-Purpose AI was endorsed by most major technology companies except Meta.

Future Directions

• Resolving the timing problem in practice: The paper leaves open how proactive governance can be operationalised so that intervention occurs early enough to prevent harm yet flexibly enough to accommodate evolving AI capabilities, given that the Collingridge Dilemma itself does not prescribe a solution.

• Making human oversight substantive rather than symbolic: The authors raise the question of how to design oversight systems that preserve human reflexivity, clear override guidelines and audits for cumulative bias in combined AI-human decision-making, rather than reducing oversight to a token gesture.

• Building adaptive and hybrid rights-assessment frameworks: The paper points to emerging scholarship on hybrid frameworks that assess the severity, duration and reversibility of rights interferences, and to practical implementations using iterative, context-aware evaluation and continuous feedback loops — suggesting further work on turning abstract rights into measurable criteria.

• Testing whether regulation delivers the claimed benefits: The paper presents its perceived-burdens-versus-benefits comparison as not exhaustive, and asserts that proactive regulatory strategies can reinforce institutional trust and sustain the long-term viability of the AI sector; whether these advantages materialise, and how they compare across the EU's approach and the deregulatory direction described in the USA, remains an open question the work raises.

Target Audience

Policymakers and regulators working on AI governance; legal scholars and practitioners in technology law and fundamental rights; AI ethics and responsible-AI researchers; compliance, risk and public affairs teams at AI developers and deployers; SME and start-up founders navigating the EU AI Act; and civil society organisations concerned with bias, discrimination and accountability in automated decision-making. The paper is also accessible to general readers with an interest in technology policy, since it avoids technical machine learning detail and builds its argument from historical precedent and regulatory design.

Authors’ abstract

Artificial intelligence (AI) now permeates critical infrastructures and decision-making systems where failures produce social, economic, and democratic harm. This position paper challenges the entrenched belief that regulation and innovation are opposites. As evidenced by analogies from aviation, pharmaceuticals, and welfare systems and recent cases of synthetic misinformation, bias and unaccountable decision-making, the absence of well-designed regulation has already created immeasurable damage. Regulation, when thoughtful and adaptive, is not a brake on innovation -- it is its foundation. The present position paper examines the EU AI Act as a model of risk-based, responsibility-driven regulation that addresses the Collingridge Dilemma: acting early enough to prevent harm, yet flexibly enough to sustain innovation. Its adaptive mechanisms -- regulatory sandboxes, small and medium enterprises (SMEs) support, real-world testing, fundamental rights impact assessment (FRIA) -- demonstrate how regulation can accelerate responsibly, rather than delay, technological progress. The position paper summarises how governance tools transform perceived burdens into tangible advantages: legal certainty, consumer trust, and ethical competitiveness. Ultimately, the paper reframes progress: innovation and regulation advance together. By embedding transparency, impact assessments, accountability, and AI literacy into design and deployment, the EU framework defines what responsible innovation truly means -- technological ambition disciplined by democratic values and fundamental rights.