IEB · Session 10 · Exam Preparation

Ethics &
Responsibility

Navigating the moral landscape of AI scale, bias, and the environmental impact of digital transformation.

The Ethics of Scale, Scope, and Learning

Origin: Iansiti & Lakhani (2020), Chapter 8

Iansiti & Lakhani argue that the same mechanisms that make digital firms powerful also make their ethical failures unusually consequential. Scale lets a single system affect millions of people. Scope lets data and functionality travel across products, partners, and contexts. Learning lets algorithms adapt and optimize continuously, sometimes toward objectives managers did not fully anticipate.

The theoretical point is not simply that digital firms can do harm. It is that digital operating models change the magnitude, speed, and distribution of harm. A biased or insecure process in a traditional firm may affect a local customer base; a biased platform rule, recommender system, or ad-targeting model can restructure opportunities across an entire ecosystem.

Market Concentration

Digital returns favor giants. This leads to platform monopolies that can stifle competition and choice (Iansiti & Lakhani, 2020).

Digital Amplification

A single algorithmic error (bias or leak) affects millions instantly due to the speed and scale of the operating model.

Privacy Erosion

The "AI Factory" requires massive data. The pressure to feed the machine often conflicts with individual privacy rights.

Five Ethical Challenges of Digital Business

Origin: Iansiti & Lakhani (2020); Session 10 slides

The session slides summarize Iansiti & Lakhani's core ethical categories as platform control, cybersecurity, competition fairness and equity, digital amplification, and algorithmic bias. These are not isolated risks. They reinforce one another because platforms collect data, use it to learn, use learning to control ecosystems, and then scale those decisions across networked markets.

Challenge Theoretical Meaning Managerial Question
Digital amplification Algorithms optimize engagement, reach, or conversion and can intensify misinformation, bias, or harmful content. What objective is the algorithm optimizing, and what harms scale with that objective?
Algorithmic bias Data, labels, model objectives, and deployment contexts reproduce or intensify unequal treatment. Whose data defines "normal," and who bears the errors?
Cybersecurity Digital operating models concentrate sensitive data, making breaches more damaging. What data is collected, who can access it, and what happens if the system is compromised?
Platform control Open interfaces, APIs, and ecosystems create innovation but also loss of control over misuse. How open should the platform be, and who is accountable for third-party behavior?
Fairness and equity Network effects and learning effects concentrate power in hub firms and reshape value distribution. Does the platform sustain ecosystem health or extract value from dependent complementors?

Algorithmic Bias & Fairness

Origin: Iansiti & Lakhani (2020), Chapter 8; Session 10 slides

AI models inherit and transform bias from the social world around them. Iansiti & Lakhani emphasize that bias can appear even when no actor intentionally discriminates. Once embedded in a scalable operating model, small local biases become system-level consequences.

Key Ethical Risks

  • Selection bias: Training data does not represent the relevant population or context, so the model performs unevenly across groups.
  • Labeling bias: Human labels reflect stereotypes, diagnostic habits, or cultural assumptions, which the model then learns and amplifies.
  • Feedback loops: Biased predictions shape future data, making the system appear more accurate while reinforcing the original pattern.
  • Lack of recourse: When a black-box decision is wrong, affected people may not know why it happened or how to challenge it.

Why Bias Cannot Be Fully Removed

No training dataset covers every possible situation, and every model is built for a purpose. Choosing an objective, such as accuracy, engagement, cost reduction, or fairness, is already an ethical choice. The managerial task is therefore not to promise neutrality, but to make the objective, data, trade-offs, and error distribution visible and contestable.

Generative AI Risks: Hallucination, Opacity, and Accountability

Origin: Session 10 slides; Hannigan, McCarthy & Spicer (2024); Koenecke et al. (2024)

The Session 10 slides add risks that are especially important for generative AI. LLMs produce plausible sequences of words, not guaranteed truth. The practical risk is epistemic: managers, lawyers, doctors, or students may treat confident language as reliable knowledge.

!

Hallucinations

LLMs can produce false claims, fake citations, or fabricated details in polished language. The risk increases when users lack domain expertise.

?

Algorithmic Opacity

Users often cannot inspect training data, model logic, or confidence boundaries. This weakens accountability and informed consent.

§

Liability

If an autonomous system causes harm, responsibility may be distributed across model provider, deploying firm, user, and regulator.

Model Autophagy Disorder and Data Limits

The slides also discuss the risk that future models train increasingly on AI-generated content rather than high-quality human-generated data. If synthetic outputs feed future training loops, model quality can degrade. This connects ethics to knowledge production: organizations need provenance, curation, and quality controls for training data.

Labor Displacement, Data Labor, and Intellectual Property

Origin: Session 10 slides; Dell'Acqua et al. (2023)

AI ethics is also about who gains and who pays. The slides highlight both visible labor displacement and less visible labor exploitation in the AI value chain. Generative AI appears automated, but it often depends on manual data labeling, content moderation, prompt evaluation, and outsourced platform work.

Concept Meaning Exam Use
Jagged frontier AI performs some professional tasks extremely well and others surprisingly poorly. Managers should not automate by job title; they should map task-by-task capability and risk.
Centaurs Professionals divide tasks between human and AI depending on where each is stronger. Useful where human judgment must decide when to rely on AI.
Cyborgs Professionals integrate AI throughout the workflow, constantly blending human and machine input. Useful for creative and iterative work, but risky when users cannot detect AI errors.
Data labor exploitation Low-paid workers label data, moderate harmful content, or train systems without sharing much value. Shows that AI's ethical footprint extends beyond the adopting firm.
Intellectual property Training data may include copyrighted or creator-owned material without clear consent or compensation. Connects AI strategy to legal risk, legitimacy, and stakeholder responsibility.

AI & Environmental Impact

Origin: Marabelli & Davison (2025); Session 10 slides

Marabelli & Davison argue that AI's environmental impact is a strategic and ethical issue, not merely a technical efficiency problem. Generative AI is resource hungry across its lifecycle: model training, inference, cooling, data-center construction, chip production, and continuous retraining. The harm is also unevenly distributed: intensive users are often wealthy organizations and Global North consumers, while climate consequences often fall hardest on poorer and more climate-vulnerable communities.

CO2

Carbon Footprint

Data centers require large amounts of electricity. If the grid is fossil-fuel based, AI use increases greenhouse gas emissions.

H2O

Water Usage

Cooling data centers can require large water withdrawals, creating ethical concerns when facilities are located in water-stressed regions.

AI

Greening via AI

Strategic opportunity: Using AI to optimize supply chains and reduce waste, potentially offsetting its own footprint.

Environmental Concepts Missing from the Earlier Page

  • Net-zero and greenwashing: Firms may claim climate responsibility while AI use increases energy and water demand elsewhere.
  • Global North / Global South inequality: The benefits of GAI are concentrated among rich users, while climate burdens are distributed unevenly.
  • Secondhand environmental effects: AI-driven nudging can increase consumption, delivery trips, and waste beyond the data-center footprint.
  • Jevons paradox / rebound effect: Efficiency gains may lower the cost of AI use, increasing total demand and total environmental impact.
  • "Bazooka to kill a fly": Managers should ask whether a resource-intensive AI solution is proportionate to the problem.

Strategic Opportunities

Marabelli & Davison also emphasize that AI can be used strategically for environmental benefit: smart-city traffic optimization, smart agriculture, weather and flood forecasting, energy-grid optimization, and waste management. The ethical question is whether these benefits outweigh the direct and indirect environmental costs.

Firm Responsibility: Keystone Strategy and Information Fiduciaries

Origin: Iansiti & Lakhani (2020), Chapter 8; Session 10 slides

Iansiti & Lakhani argue that hub firms in digital ecosystems have responsibilities beyond narrow shareholder value because their decisions shape the health of entire networks. A platform can become a de facto governor: it sets rules, controls access, ranks visibility, allocates attention, and defines what behavior is rewarded.

Keystone Strategy

A hub firm sustains its own long-term performance by sustaining the health of the ecosystem it depends on.

i

Information Fiduciary

Firms that hold sensitive information should treat users' data in a trustworthy way and avoid using it against their interests.

Regulatory Governance

The slides connect responsibility to DSA-style platform transparency and AI Act-style risk classification and obligations.

Novo Nordisk — Moral Complexity

Ethical synthesis for the pharmaceutical case:

  • The Ethics of Efficiency: The 9,000 layoffs mentioned in your synopsis are an ethical "Internality" of AI adoption. Does the gain in R&D speed justify the human cost?
  • Bio-data Privacy: Partnering with OpenAI means Novo must ensure that sensitive patient data used for training models is rigorously protected and never "leaks" into the public model.
  • Healthcare Inequity: If the AI is trained on data from Western clinical trials, the drugs discovered might be less effective for global populations, creating a global health bias.
  • Hallucination and Scientific Reliability: Generative AI can produce plausible but false scientific suggestions. Novo needs validation against biological ground truth, not just persuasive outputs.
  • Environmental Proportionality: If AI is used at large scale in R&D, Novo should ask whether compute, energy, and water use are proportionate to the expected health benefit.
  • Partner Responsibility: Novo cannot outsource ethics completely to OpenAI. It remains accountable for data governance, bias, auditability, and the effects of AI-generated decisions in its own value chain.

Likely Oral Exam Questions