Connecting the dots in trustworthy Artificial Intelligence: From AI principles, ethics, and key requirements to responsible AI systems and regulation

This is not an empirical study. It is a conceptual review and synthesis paper. The authors did not test an intervention, comparator, or outcome measure in the traditional sense. Instead, they:

Reviewed and consolidated existing AI ethics principles from governments, industry, and academia (e.g., EU, OECD, UNESCO, IEEE).

Analysed seven technical requirements for trustworthy AI: human agency and oversight; robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental wellbeing; and accountability.

Proposed a framework connecting these requirements to three pillars (lawful, ethical, robust) and four axes (global principles, philosophical ethics, risk-based regulation, technical requirements).

Introduced the concept of a "responsible AI system" that can be audited, and discussed regulatory sandboxes as a practical implementation tool.

There are no numerical outcomes, effect sizes, or statistical tests. The paper's "findings" are conceptual: a structured taxonomy and a set of recommendations for building and auditing trustworthy AI.

No human participants were studied. The paper analyses documents, principles, and regulatory frameworks. The "sample" consists of:

Published AI ethics guidelines from 84 documents (as cited in prior work by Jobin et al., 2019, which the authors reference).

Regulatory frameworks from the European Union (AI Act), OECD, UNESCO, and others.

Philosophical and technical literature on AI ethics, fairness, transparency, and accountability.

The "setting" is academic and policy-oriented: the authors are based in Spain and Belgium, with expertise in computer science, philosophy, and finance.

No measurements were taken. The authors used qualitative methods:

Literature review and synthesis of existing AI ethics principles.

Conceptual analysis of each requirement (What, Why, How).

Mapping of requirements to regulatory frameworks.

Discussion of auditing processes and regulatory sandboxes.

There are no instruments, scales, or quantitative metrics. The paper does not report any data collection or statistical analysis.

**Study design:** This is a conceptual review and framework synthesis. It is not a systematic review, meta-analysis, or empirical study. The authors do not follow PRISMA guidelines or report a search strategy. They do not state inclusion/exclusion criteria for the documents they analysed.

**What they did:**

1. Identified seven technical requirements for trustworthy AI, drawn from prior work (e.g., EU High-Level Expert Group on AI, 2019).

2. For each requirement, they answered three questions: What is it? Why is it needed? How can it be implemented?

3. They connected these requirements to three pillars (lawful, ethical, robust) and four axes (principles, philosophy, regulation, requirements).

4. They discussed auditing processes and regulatory sandboxes as practical tools.

5. They debated future directions and the role of regulation.

**What this design can and cannot prove:**

**Can prove:** The paper can provide a structured, logically coherent framework for thinking about trustworthy AI. It can synthesise existing ideas and highlight gaps or tensions between different approaches.

**Cannot prove:** The paper cannot prove that any particular AI system is trustworthy, that the framework works in practice, or that following the requirements leads to better outcomes. It offers no empirical evidence, no test of the framework against real-world systems, and no comparison with alternative frameworks. It is a normative argument, not a scientific finding.

**Major methodological weaknesses:**

No systematic search or selection criteria for included documents.

No quantitative analysis or empirical validation.

No discussion of how to weigh competing requirements (e.g., transparency vs. privacy).

The framework is presented as comprehensive, but it is not tested against real-world AI failures or case studies.

The authors do not address potential conflicts between requirements (e.g., fairness vs. accuracy).

Since this is a conceptual paper, "findings" are the proposed framework and its components. Key points include:

**Seven technical requirements for trustworthy AI:**

- Human agency and oversight: Humans should be able to understand, supervise, and override AI decisions.

- Robustness and safety: AI systems should be reliable, secure, and resilient to errors or attacks.

- Privacy and data governance: Data collection, storage, and use should respect privacy rights and be governed by clear policies.

- Transparency: AI systems should be explainable, and their decision-making processes should be open to scrutiny.

- Diversity, non-discrimination, and fairness: AI should not perpetuate bias or discrimination, and should be accessible to diverse populations.

- Societal and environmental wellbeing: AI should benefit society and minimise environmental harm.

- Accountability: There should be clear responsibility for AI outcomes, with mechanisms for redress.

**Three pillars of trustworthy AI:**

- Lawful: Compliance with applicable laws and regulations.

- Ethical: Adherence to ethical principles and values.

- Robust: Technically and socially reliable.

**Four axes of a holistic vision:**

- Global principles for ethical AI (e.g., from OECD, UNESCO).

- Philosophical perspectives on AI ethics (e.g., deontology, utilitarianism, virtue ethics).

- Risk-based regulation (e.g., EU AI Act's categories: unacceptable, high, limited, minimal risk).

- Technical requirements (the seven listed above).

**Responsible AI system:** The authors define this as an AI system that can be audited against the seven requirements, subject to regulatory oversight, and held accountable for its outcomes.

**Regulatory sandboxes:** The authors propose these as controlled environments where AI systems can be tested under regulatory supervision before deployment, allowing for iterative improvement and risk mitigation.

**Diverging views on AI future:** The paper discusses tensions between "AI optimism" (AI will solve major problems), "AI pessimism" (AI poses existential risks), and "AI pragmatism" (focus on incremental, regulated progress). The authors argue that regulation is key to reconciling these views.

Not applicable. There are no numerical effects, effect sizes, or confidence intervals. The paper does not report any quantitative results.

The authors acknowledge several limitations, and additional critical points are noted:

**What the authors acknowledge:**

The framework is a "multidisciplinary vision" and not a definitive solution.

Implementing the requirements in practice is challenging and context-dependent.

Regulatory sandboxes face challenges, including resource requirements and potential for regulatory capture.

The paper does not provide a detailed auditing methodology or specific metrics.

**What a critical reader would note:**

**No empirical validation:** The framework has not been tested against real-world AI systems. There is no evidence that following these requirements leads to trustworthy outcomes.

**Lack of specificity:** The requirements are broad and abstract. For example, "transparency" can mean many things (e.g., explainability, interpretability, openness of code). The paper does not resolve these ambiguities.

**No trade-off analysis:** The paper does not address how to balance competing requirements. For example, transparency may conflict with privacy (e.g., revealing personal data to explain a decision). Fairness may conflict with accuracy (e.g., equalising outcomes may reduce predictive performance).

**No discussion of power dynamics:** The framework assumes that AI developers and regulators are acting in good faith. It does not address how power imbalances, corporate interests, or political pressures might undermine trustworthiness.

**No consideration of cultural differences:** AI ethics principles vary across cultures (e.g., individual privacy vs. collective good). The paper does not address how to reconcile these differences.

**No practical guidance for individuals:** The paper is aimed at policymakers, developers, and regulators. It offers no actionable advice for someone running a personal experiment with AI tools.

For someone running their own n=1 experiment, this paper is not directly applicable because it does not test an intervention. However, you can use its framework to audit AI tools you use in your experiments (e.g., AI-based sleep trackers, fitness apps, or decision-support tools). Here is how:

### What to test

**Specific intervention:** Use the seven requirements as a checklist to evaluate an AI tool you rely on. For example, if you use an AI-powered sleep tracker, test whether it meets each requirement:

- Human agency: Can you override or correct its recommendations?

- Robustness: Does it work reliably across different conditions (e.g., different bedtimes, sleep environments)?

- Privacy: Does it clearly explain how your data is stored and shared?

- Transparency: Can you understand how it calculates your sleep score?

- Fairness: Does it work equally well for your demographic (e.g., age, gender, skin tone for optical sensors)?

- Societal wellbeing: Does the company have environmental or social responsibility policies?

- Accountability: Is there a clear process for reporting errors or seeking redress?

### Minimum meaningful duration

**Duration:** One week to one month. This is long enough to test the tool under varied conditions and identify recurring issues (e.g., frequent errors, opaque recommendations). For a thorough audit, you might need longer (e.g., 3 months) to assess robustness across seasons or lifestyle changes.

### What to measure (specific metrics)

**For each requirement, define a measurable outcome:**

- Human agency: Number of times you could override the tool vs. times it ignored your input. Score 0–10 for perceived control.

- Robustness: Frequency of errors (e.g., misdetected sleep stages, app crashes). Percentage of days with missing data.

- Privacy: Does the privacy policy clearly state data retention period? Yes/No. Can you delete your data? Yes/No.

- Transparency: Can you explain how the tool calculates its main metric (e.g., sleep score)? Score 0–10 for understandability.

- Fairness: Does the tool's accuracy vary by your age, gender, or skin tone? (You may need to compare with a gold standard, e.g., polysomnography.)

- Societal wellbeing: Does the company publish a sustainability report? Yes/No.

- Accountability: Is there a customer support channel that responds within 48 hours? Yes/No.

### Key confounds to control for

**Confound 1: Your own expectations.** If you believe the tool is trustworthy, you may overlook flaws. Keep a daily log of issues and rate them objectively.

**Confound 2: Changes in the tool.** The app may update during your test period, altering its behaviour. Note the version number and date of updates.

**Confound 3: External factors.** Your sleep quality, stress, or health may change independently of the tool. Measure these as covariates (e.g., daily stress rating 1–10, caffeine intake, exercise).

**Confound 4: Comparison with a gold standard.** Without a reference (e.g., a validated sleep diary or polysomnography), you cannot assess accuracy. If possible, use a second, independent method for comparison.

### What a positive result would look like

**Positive result:** The tool scores ≥7/10 on each of the seven requirements, with no major failures (e.g., no data breaches, no unexplained errors, clear accountability process). You can explain how it works to a friend, and you feel confident relying on its recommendations.

**Negative result:** The tool fails on one or more requirements (e.g., no transparency, frequent errors, unclear privacy policy). You would then consider switching to a more trustworthy alternative or supplementing the tool with manual tracking.

**Bottom line:** This paper is a map, not a destination. Use it to navigate the trustworthiness of AI tools in your life, but do not expect it to tell you which tool is best. For that, you need empirical testing—and that is where your n=1 experiment comes in.