• AI systems deeply influence decisions affecting human lives, rights, and welfare.
• Ethical considerations ensure AI enhances rather than harms society.

Key risks:

• Bias and discrimination
• Lack of transparency
• Privacy violations
• Accountability gaps

📚 Recent literature: (Floridi et al. 2018)

• Short answer: Because we are functional, stable adults.
• Longer, more cynical answer: Because Responsible AI is sustainable AI — and sustainability is good for your organization.

Adapted by Montreal AI Ethics Institute
from Ethics Guidelines for trustworthy AI by EU High-level expert group on AI.

🤔 What does fairness mean to you?

5. Justice as Fairness

   Definition: Philosophical approach by John Rawls emphasizing equal basic rights and opportunities.

   Example: Designing AI systems that do not infringe upon individuals’ rights and provide equal access to benefits.

   Reference: (Rawls 1971)

Fairness involves impartial and equitable treatment by AI systems:

• Procedural fairness: Ensuring decision-making processes are transparent and consistent.
• Distributive fairness: Ensuring outcomes do not unfairly disadvantage particular groups.

📚 Recent literature: (Mehrabi et al. 2022)

Transparency ensures stakeholders clearly understand how AI decisions are made.

• Explainability: Technical clarity about how algorithms or models reach outcomes.
  • Q: How does this model work?
  • A (e.g.): It encodes semantics in the \(5^{th}\) layer
• Interpretability: Stakeholder ability to make sense of decisions.
  • Q: Why was this decision made?
  • A (e.g.): This particular patient has asthma and is above 65

Relevant literature: (Barredo Arrieta et al. 2020)

Why it’s ethically essential:

• 🧠 Promotes trust by reducing mystery and perceived manipulation.
• ⚖️ Enables accountability—who is responsible when things go wrong?
• 🧪 Supports scrutiny and correction of errors and bias.

(Cheong 2024)

From Theory to Practice:

1. Explainability: Can people understand the model’s outputs?
   • Example: Explaining why an AI flagged a claim as fraud.
2. Uncertainty Awareness: Are limits and confidence levels disclosed?
   • Example: “Prediction accuracy: 72%, based on limited demographic data.”
3. Audience-Appropriate Communication:
   • Managers need different explanations than data scientists or customers.

🔍 Liao and Vaughan (2023) proposed a framework for contextual transparency based on user roles and goals.

1. Ethical Black Box (EBB)

• Records key AI decisions and data flows for audit and review.
• Think: “flight recorder” for algorithms.

🔗 “The Case for an Ethical Black Box” (2017) and the Case for an Ethical Black Box

2. Microsoft Transparency Toolkit (2024)

• Provides templates for explaining model purpose, input/output, performance, and risks.
• Encourages transparency as part of everyday AI deployment.

🔗 Download the report (PDF)

3. Community Audits – CRASH Project

• Algorithmic Justice League invites public reporting of AI harms.
• Builds a culture of external, participatory transparency.

🔗 The AJL CRASH Project

We will see ways to implement Explainable AI – and examples of how it goes wrong – in Module 4

Ethical decision-making means balancing competing duties.

• The prima facie duty theory is widely used in biomedical ethics to provide first principles for ethical decisions in health care. These principles are:
  • Respect for patient autonomy: Continuously present in classical medical ethics, often through informed consent. How much user autonomy is there when the user is using autonomous systems?
  • Nonmaleficence: Often, this is juxtaposed with the Hippocratic oath “to do no harm.” Is it wrong to take away an employee’s work?
  • Beneficence: This can be a subjective measure – the risks should be commensurate with the benefits.
  • Justice: This is often described in terms of “fairness”, which is poorly (if ever) defined. Is a utilitarian view of fairness preferable to one focused on each individual?

📚 Recent literature: (Anderson and Anderson 2011)

Choice pervades ethics. Consider this scenario:

Possible actions:

Action A: Deploy a new, unproven LLM with potential benefits but uncertain risks.

Action B: Continue standard deployment with known, limited effectiveness.

Ethically relevant feature:

Risk of new unknown harm to the user (potentially present in A, but not in B).

Inferred principle:

There is at least a prima facie duty not to expose the user to undue harm.

What if there is a risk in inaction?

E.g., the Ethics Guidelines for Trustworty AI from the EU High-Level Expert Group for AI are:

Key insights from Luxton (2014)

• The therapeutic alliance is central to effective care.
• AICPs (AI Care Providers) may simulate empathy, but they lack authenticity.
• Users often form attachments to AICPs — even when they know it’s a machine.

Ethical Implications

• Is it ethical to simulate empathy if it creates dependency or deception?
• Should AICPs disclose their machine identity clearly and frequently?
• Risks: Emotional overattachment, misinterpreted capabilities, lack of appropriate termination protocols.

• AICP System Competence
  • AICPs must not exceed their tested scope of use.
  • Design must include clinical safeguards for risk scenarios (e.g., self-harm disclosures).
  • Ethical concern: Many AICPs today are marketed as capable of far more than they’re proven to do.
• Liability Questions
  • Who is responsible if an AICP fails? The developer? The deploying organization?
  • Increasing autonomy challenges traditional legal frameworks.
  • Ethical agents? Or just tools? (cf. Sullins, “When is a robot a moral agent?”)

📌 7 key recommendations (from Table 3 of the paper):

1. Design to follow the appropriate ethical codes and guidelines consistent with domain of use
2. Identify and provide specifications of use and limits of autonomy of [AI] systems to end users
3. Test safety and ethical decision making of [AI] systems in diverse ethical situations encountered in all applicable clinical contexts
4. Include capability for data logs/audit trails to track and explain [AI] decision making
5. Provide built-in safeguards to assure that systems are only able to provide services within established boundaries of competence and domain of use
6. Consider level of human realism of [AIs] including appearance and other behavioral (sic) characteristics as appropriate for intended application
7. Consider cultural sensitivity and diversity in design of [AIs]

See IEEE Ethically Aligned Design (EAD; 2019)

• Purpose & Scope:
  • Vision for embedding values into autonomous and intelligent systems.
  • Aims to influence standards (P7000 series), certification (ECPAIS), policy, and legislation.
• Three Pillars:
  1. Universal Human Values: Human rights, societal flourishing, and sustainability.
  2. Political Self‑Determination & Data Agency: Empower individuals with data control and support democratic values.
  3. Technical Dependability: Ensure AI is safe, reliable, explainable, auditable, and certifiable
• Eight General Principles:

• Impact & Adoption:
  • Led to 14 active standardization efforts, a new Certification Program (ECPAIS), educational consortia (EADUC), and global policy influence (EU, UN, OECD)
  • Encourages embedding ethics from conception to decommissioning of AI

• Recognizes the transformative potential of GenAI in healthcare, education, and science.
• Highlights critical policy challenges: misinformation, bias, intellectual property, and labour.

Key risks include:

• Amplification of mis/disinformation and erosion of trust in public discourse.
• Perpetuation of bias and discrimination through training data.
• Legal uncertainty around use of copyrighted material and ownership of AI-generated outputs.
• Labour displacement and job-task exposure, especially in high-skill sectors.

Ongoing efforts:

• The G7 Hiroshima AI Process, in collaboration with the OECD, promotes international AI governance.
• OECD.AI Observatory supports governments with policy tools and expert analysis.

📚 Recent literature: (Lorenz, Perset, and Berryhill 2023)

• Scope & Aim:
  • First major regulatory framework for AI across sectors in the EU.
  • Focuses on aligning AI use with “EU values”, i.e., trust, safety, and fundamental rights.
• Risk-Based Approach:
  • Prohibited: Social scoring, manipulative systems, some uses of real-time biometric surveillance.
  • High-risk: AI in critical areas (e.g. hiring, education, law enforcement) must comply with requirements for:
    • Data governance, Documentation & traceability, Transparency, Human oversight, Accuracy, robustness, cybersecurity
• Enforcement Model:
  • Inspired by product safety law: places compliance responsibility on providers.
  • Relies heavily on technical standards and self-assessments (not public regulators).

See The EU AI Act but also EU ethics guidelines for trustworthy AI

• Scope & Mandate:
  • Governs all federally developed/procured automated decision systems since 1 April 2020 that impact individuals’ rights, benefits, or obligations.
  • Applies to both fully and semi‑automated systems in service (e.g., permits, hiring, benefits).
• Core Requirements:
  • Mandatory Algorithmic Impact Assessments (AIA) before procurement or deployment.
  • Ensure transparency (explainability), human oversight, and procedural fairness.
  • Public commitments: quality assurance, recourse mechanisms, reporting.
• Oversight & Updates:
  • Bi‑annual policy reviews; current (\(4^{th}\)) cycle targets strengthening client protections, peer review, and documentation of banned applications.

See here

• Artificial Intelligence & Data Act (AIDA) – part of Bill C‑27 (2022).
  • Focused on high-impact AI systems affecting safety, rights, and health.
  • Obligations for businesses: risk management, transparency, record-keeping, ongoing monitoring.
  • Introduced a new AI & Data Commissioner for oversight.
• Legislative Status:
  • Passed second reading April 2023; stalled by Jan 2025 prorogation.
• Pan‑Canadian AI Strategy:
  • Initial CA$125 M launch (2017); Phase II (2022) adds >CA$400 M for research, computing, and commercialization.
  • Supports Amii, Mila, Vector, CIFAR Chairs, superclusters, computing infrastructure.

• Generative AI Guide (2025) issued by Treasury Board introduces FASTER principles:

  • Fair, Accountable, Transparent, Educated, Relevant use for federal institutions .

• Voluntary Code of Conduct (2023) – advanced generative AI guidance for private and public sectors during Bill C‑27 legislative delay .

• Scope of Guide vs Directive:

  • Generative AI guide applies to tools used in government everyday work (emails, drafting, service chatbots, etc.) and complements the Directive.

• Best Practices Covered:

  • Risk assessment; data privacy (PII); watermarking; disclaimers; human review; cybersecurity; environmental impacts.

• Mission & Process
  • Developed collaboratively (2017–18) by ~500 citizens, experts, policymakers via public deliberation in Montréal.
  • An open-source charter guiding ethical AI to serve human well-being & social justice.
• 10 Ethical Principles

• Outcome & Reach
  • Endorsed by 300+ organizations (e.g., Montréal, MILA, CHUM, McGill).
  • Supported initiatives: summer schools, grants (e.g. HAND‑AI), toolkits, civic-engagement projects.

📚 Reference: The Montréal Declaration for a Responsible Development of Artificial Intelligence

• Developed at 2017 Future of Life Asilomar Conference by ≥100 experts in AI ethics, safety, law, and philosophy.
• Scope:
  • Short-term: Trustworthy systems—emphasize transparency, accountability, privacy, and avoidance of arms races.
  • Long-term: Manage advanced AI risk—value alignment, human oversight, and caution around superintelligence.
• Selected Principles:
  1. Value Alignment: AI must align with human values.
  2. Transparency & Explainability.
  3. Responsibility: Designers accountable for impacts.
  4. Human Control & Non-Subversion.
  5. Avoid AI arms races in lethal systems.
• Impact: Signed by thousands—including leading voices like Bengio, Russell, Hawking.

• Human-centered commitment – designing, building, and releasing AI for people.

• Six Guiding Principles: 1. Fairness, 2. Reliability & Safety, 3. Privacy & Security, 4. Inclusiveness, 5. Transparency, 6. Accountability

• Ecosystem for Trust:

  • Responsible AI Standard to embed principles into development lifecycles.
  • Tools like the Responsible AI dashboard, scorecards, A/B testing, and open‑source toolkits (SmartNoise, Counterfit) to operationalize principles :contentReferenceoaicite:3.

📚 Reference: Microsoft Responsible AI (2025)

From IBM

From IBM

From IBM

• AI systems reflect and amplify societal structures—including existing biases.
• Fairness and accountability require deliberate choices about data, labels, proxies, and oversight.
• Technical principles (e.g. transparency, robustness) must be supported by institutional mechanisms (e.g. regulation, audits, redress).
• Global efforts (e.g. EU AI Act, Canada’s AIA, Asilomar, Montreal Declaration) offer complementary models of ethical governance.
• Responsible AI is not just a design problem—it’s a political, legal, and social challenge.

Ethics is not a “final step”—it must be embedded across the AI lifecycle.