Computing ethics and societal impact

Intuition Beginner

Computing ethics examines the moral responsibilities that arise from the design, development, and deployment of computing technologies. Every piece of software reflects choices made by its creators, choices about what to optimize for, whose needs to prioritize, what data to collect, and how to handle errors. These choices have consequences for real people, and computing professionals have a responsibility to consider those consequences.

Consider a social media algorithm designed to maximize engagement. The algorithm learns that angry, polarizing content generates more clicks and shares than calm, nuanced content. It begins promoting increasingly extreme views, contributing to social division and political radicalization. The engineers who built the algorithm did not intend to polarize society. They optimized for a metric (engagement) without considering the broader consequences. This is an ethical failure, not a technical one.

This pattern, where a technically correct system produces harmful outcomes because it optimizes for the wrong objective, is called "reward hacking" in AI research or "Goodhart's law" in economics: when a measure becomes a target, it ceases to be a good measure. Engagement was intended as a proxy for user satisfaction, but optimizing for engagement produced outcomes that reduced user well-being. The lesson is that technical systems require not just correct implementation but correct objectives, and choosing objectives is an ethical decision.

Privacy is a central concern in computing ethics. Every website you visit, every purchase you make, every location your phone records generates data about you. Companies collect, aggregate, and analyze this data to build detailed profiles of your behavior, preferences, relationships, and vulnerabilities. This data can be used to serve you targeted advertisements, but it can also be used to manipulate your behavior, discriminate against you, or surveil you.

Data collection often happens without meaningful consent. When you click "Accept All Cookies" on a website, have you truly consented to having your browsing behavior tracked across dozens of sites? The legal fiction of consent through lengthy privacy policies that nobody reads does not satisfy the ethical requirement of informed, voluntary agreement.

Algorithmic bias occurs when automated systems produce outcomes that systematically disadvantage certain groups. A hiring algorithm trained on historical hiring data may learn to favor candidates from demographic groups that were historically preferred, perpetuating past discrimination. A facial recognition system trained primarily on images of light-skinned faces may fail to correctly identify dark-skinned faces, leading to false arrests. A credit scoring algorithm may deny loans to qualified applicants from certain neighborhoods (redlining by algorithm).

Bias enters algorithms through several channels. The training data may reflect historical prejudices. The choice of features may encode assumptions that disadvantage certain groups. The optimization objective may prioritize accuracy for the majority at the expense of minority groups. The evaluation metrics may fail to measure performance across different demographic subgroups.

Privacy-enhancing technologies provide technical tools for protecting privacy. Differential privacy adds calibrated noise to data releases, providing mathematical guarantees that individual records cannot be inferred. Federated learning trains models on decentralized data without centralizing it. Homomorphic encryption allows computation on encrypted data. Secure multi-party computation enables collaborative analysis without sharing raw data. These technologies demonstrate that privacy protection and data utility are not always in conflict: with careful engineering, it is possible to extract insights from data while protecting individual privacy.

Informed consent in the digital context is deeply problematic. A typical internet user encounters hundreds of privacy policies and cookie consent dialogs per week. Nobody reads these documents: the average privacy policy takes 10-20 minutes to read and requires a college-level education to understand. The result is "consent fatigue": users click "accept all" without understanding what they are agreeing to. Some ethicists argue that meaningful consent in the digital context requires radical simplification: short, clear notices; genuine choice (not just accept-or-leave); and the ability to change one's mind after the fact. Others argue that consent is fundamentally inadequate as a privacy safeguard because the power imbalance between individuals and technology companies makes truly voluntary agreement impossible.

Accessibility is an ethical obligation. Software should be usable by people with disabilities. Screen readers, keyboard navigation, high-contrast modes, and captioned videos are not optional features. They are essential for millions of people who rely on technology for education, employment, and social participation. Designing software that excludes people with disabilities is a form of discrimination.

Intellectual property in the digital age raises complex ethical questions. Should software be patentable? Is it ethical to download copyrighted material without permission? How should open source contributions be credited? What are the ethical obligations of companies that build products on open source software without contributing back?

The tension between proprietary and open approaches reflects deeper values. Proprietary software companies argue that intellectual property protection incentivizes innovation by allowing creators to profit from their work. Open source advocates argue that restricting access to knowledge hinders progress and creates artificial scarcity. In practice, most software ecosystems include both: companies build proprietary products on open source foundations, and many open source projects are funded by companies that benefit from them. The ethical challenge is ensuring that this symbiosis is fair and sustainable.

The digital divide refers to the gap between those who have access to computing technology and those who do not. In an increasingly digital world, lack of internet access, lack of computing skills, and lack of affordable devices exclude people from education, employment, healthcare, and civic participation. Bridging the digital divide is not merely a technical challenge but a matter of social justice.

The digital divide has multiple dimensions. The access divide separates those with internet connectivity from those without, disproportionately affecting rural, low-income, and elderly populations. The skills divide separates those who can effectively use digital tools from those who cannot. The participation divide separates those who create digital content and shape digital culture from those who merely consume it. During the COVID-19 pandemic, the digital divide became acutely visible: students without home internet access could not participate in remote learning, widening educational inequality.

Automation and employment raise questions about the responsibility of technologists to workers displaced by their creations. Automation has historically created more jobs than it destroyed, but the transition can be devastating for individual workers and communities. The collapse of manufacturing employment in the American Rust Belt, driven in part by automation, devastated communities and contributed to social and political instability. Ethical technology development should consider transition support for displaced workers: retraining programs, social safety nets, and gradual deployment that gives communities time to adapt.

Open source ethics involve questions about the obligations of developers who create freely available software. Open source software powers the internet (Linux, Apache, OpenSSL), but the maintainers of critical infrastructure are often unpaid volunteers. When a vulnerability is found in widely used open source software (like the Log4j vulnerability in 2021, which affected millions of systems), the burden of fixing it falls on volunteers. The ethical question is whether companies that build profitable businesses on open source software have an obligation to contribute back, either through code contributions, financial support, or employee time dedicated to maintenance.

Professional ethics in computing are codified by organizations like the ACM (Association for Computing Machinery) and the IEEE (Institute of Electrical and Electronics Engineers). The ACM Code of Ethics states that computing professionals should contribute to society and human well-being, avoid harm, be honest and trustworthy, be fair and take action not to discriminate, respect the work required to produce new ideas, respect privacy, and honor confidentiality.

These principles are not abstract ideals but practical guidelines. "Avoid harm" means testing software for safety-critical applications rigorously. "Respect privacy" means collecting only necessary data and protecting it with strong security measures. "Be fair" means evaluating algorithms for discriminatory outcomes before deployment. The code provides a framework for navigating the ethical decisions that computing professionals face daily.

Whistleblowing presents difficult ethical dilemmas. When a computing professional discovers that their employer's product is causing harm (a social media algorithm promoting self-harm content, a safety-critical system with known defects, a surveillance tool being used against dissidents), they face a conflict between loyalty to their employer and responsibility to the public. The ethical course of action depends on the severity of the harm, the availability of internal remedies, and the protections available to whistleblowers.

Sustainability and green computing are increasingly important ethical concerns. Data centers consume enormous amounts of electricity for computation and cooling. The information and communication technology sector is responsible for approximately 2-4% of global carbon emissions, comparable to the airline industry. Training a single large language model can emit as much carbon as several transatlantic flights. Ethical computing requires considering the environmental impact of technology and striving for efficiency: optimizing algorithms to reduce computation, using renewable energy for data centers, and designing hardware for longevity and recyclability.

Dual-use technology raises questions about the responsibility of developers for how their creations are used. Encryption protects dissidents from authoritarian governments but also shields criminals from law enforcement. Social media connects communities but also enables coordinated harassment. Facial recognition helps find missing persons but also enables mass surveillance. The dual-use nature of computing technology means that ethical evaluation must consider both beneficial and harmful applications, and developers must grapple with the possibility that their work may be used in ways they did not intend.

Visual Beginner

Ethical issue	Description	Example
Privacy	Protecting personal data from unauthorized collection and use	Location tracking without consent
Algorithmic bias	Systematic disadvantage of certain groups by automated systems	Hiring algorithms favoring one demographic
Accessibility	Ensuring software is usable by people with disabilities	Websites without screen reader support
Digital divide	Unequal access to computing technology and skills	Rural areas without broadband internet
Intellectual property	Respecting ownership of creative and innovative work	Using open source without attribution
Surveillance	Mass monitoring of communications and behavior	Government tracking of online activity

Worked example Beginner

A company develops an AI-powered resume screening tool. It trains the model on 10 years of hiring data from the company. The model learns to score resumes based on patterns in the historical data.

After deployment, an internal audit reveals that the model systematically scores resumes from women lower than equivalent resumes from men. Investigation shows that the historical hiring data reflects a pattern of gender discrimination: over 10 years, the company hired fewer women than men with identical qualifications.

The model learned this pattern and perpetuated it. It did not use gender as an explicit feature, but it picked up on correlated features: women's colleges, gaps in employment history (correlated with maternity leave), and certain extracurricular activities.

The ethical analysis involves multiple stakeholders. Applicants who are unfairly rejected suffer harm (lost opportunities, diminished self-esteem). The company suffers reputational damage and misses out on qualified candidates. Society suffers when discrimination is automated and scaled.

The remediation requires several steps. First, audit the training data for historical bias. Second, remove or transform features that serve as proxies for protected characteristics. Third, evaluate the model's performance separately for each demographic group. Fourth, implement ongoing monitoring to detect bias that may emerge as the model is used. Fifth, involve diverse stakeholders in the design and evaluation process.

This case illustrates several general principles. Historical data reflects historical inequities: if a company discriminated in the past, training a model on that data automates and perpetuates the discrimination. Proxy features can smuggle in protected attributes even when they are explicitly excluded. Auditing must be ongoing, not one-time, because models can develop bias as the data distribution shifts over time. Diverse teams are more likely to identify bias because they bring different perspectives on what constitutes harm.

Worked example: content moderation. A social media platform uses an automated content moderation system to flag hate speech. The system is trained on labeled examples of hateful and non-hateful content. After deployment, the system disproportionately flags posts by Black users discussing racism as "hate speech."

The analysis reveals that the training data labeled discussions of racial discrimination as hateful because the labelers (predominantly from outside the affected communities) found the language uncomfortable. The system learned to conflate discussions of racism with racism itself. This is an example of how bias in the labeling process, not the algorithm, produces discriminatory outcomes. The remediation requires more diverse and contextually aware labeling, consultation with affected communities, and regular audits of moderation decisions across demographic groups.

Check your understanding Beginner

Formal definition Intermediate+

Ethical frameworks for computing. Several philosophical traditions provide frameworks for analyzing computing ethics.

Utilitarianism evaluates actions by their consequences: the right action is the one that produces the greatest good for the greatest number. Applied to computing, a utilitarian might argue that a data breach affecting millions is worse than one affecting thousands, regardless of the intentions behind the vulnerability.

Deontological ethics (Kant) evaluates actions by whether they follow universal moral rules, regardless of consequences. A deontologist might argue that lying to users about data collection is wrong even if it leads to better products, because it violates the duty of honesty.

Virtue ethics (Aristotle) evaluates the character of the agent rather than the action or its consequences. A virtue ethicist might ask: what kind of person, what kind of company, would design a system that exploits user data?

Care ethics emphasizes relationships and responsibilities to specific others. A care ethicist might prioritize the needs of vulnerable users (children, the elderly, people with disabilities) over abstract principles of efficiency or profit.

Fairness definitions in machine learning

Formal fairness criteria attempt to make the concept of fairness mathematically precise. Several definitions exist, and they are mutually incompatible in general.

Demographic parity: the outcome should be independent of the protected attribute. $P(\hat{Y}=1|A=0)=P(\hat{Y}=1|A=1)$.

Equalized odds: the prediction should be equally accurate across groups. $P(\hat{Y}=1|Y=y,A=0)=P(\hat{Y}=1|Y=y,A=1)$ for $y\in \{0,1\}$.

Individual fairness: similar individuals should receive similar predictions. $d(\hat{Y}(x_1),\hat{Y}(x_2))\le L\cdot d(x_1,x_2)$ for a task-specific metric $d$.

Chouldechova (2017) and Kleinberg, Mullainathan, and Raghavan (2016) independently proved that these fairness criteria cannot all be satisfied simultaneously when base rates differ across groups, except in degenerate cases. This impossibility result means that fairness requires trade-offs and value judgments, not just better algorithms.

Privacy frameworks

Informational self-determination: individuals should control the collection, use, and dissemination of their personal information.

Contextual integrity (Nissenbaum, 2004): information flows are appropriate when they conform to the informational norms of the context. Sharing health information with a doctor is appropriate; sharing it with an employer is not, even if the same information is involved.

Differential privacy (Dwork, 2006): a mechanism $\mathcal{M}$ satisfies $ϵ$-differential privacy if for any two datasets $D$ and $D^{\prime }$ differing in one record, and any set of outputs $S$: $P(\mathcal{M}(D)\in S)\le e^{ϵ}\cdot P(\mathcal{M}(D^{\prime })\in S)$. This provides a mathematical guarantee that the presence or absence of any individual's data has a bounded effect on the output.

The parameter $ϵ$ quantifies the privacy loss. When $ϵ=0$, the output is completely independent of the input (perfect privacy but no utility). As $ϵ$ grows, more information about individuals leaks. Practical deployments typically use $ϵ$ between 0.1 and 10. The US Census Bureau adopted differential privacy for the 2020 Census, adding noise to population counts to protect respondent confidentiality while preserving statistical accuracy for policy decisions.

Responsible AI principles

Major organizations have proposed principles for responsible AI development. These typically include: fairness (avoiding discrimination), transparency (enabling understanding of AI decisions), accountability (establishing responsibility for AI outcomes), privacy (protecting personal data), safety (preventing harm), and inclusiveness (ensuring AI benefits everyone). The challenge is translating abstract principles into concrete engineering practices.

Algorithmic auditing is an emerging practice for evaluating deployed AI systems for fairness, accuracy, and compliance. An audit examines the training data for bias, evaluates model performance across demographic subgroups, tests for adversarial vulnerabilities, and assesses the system's societal impact. Algorithmic audits can be conducted internally (by the developing organization), by independent third parties, or by regulators. The proposed EU AI Act would mandate audits for high-risk AI systems, creating a regulatory framework similar to financial auditing.

Impact assessments, modeled on environmental impact assessments, evaluate the potential societal effects of a technology before deployment. They consider who benefits, who is harmed, what data is collected, how decisions are made, what errors are possible, and what recourse exists for affected individuals. Canada's Algorithmic Impact Assessment Tool, developed by the Canadian Digital Service, provides a structured questionnaire that helps government agencies evaluate the risks of automated decision systems.

Key result: fairness impossibility Intermediate+

Theorem (Chouldechova, 2017). When base rates (the true prevalence of the outcome) differ across groups, it is impossible to simultaneously satisfy calibration (predicted probabilities are accurate across groups) and equal false positive/negative rates.

Proof sketch. Let $p_g$ be the base rate for group $g$ (the proportion of positive outcomes). Calibration requires that among individuals predicted positive with probability $s$, the actual positive rate is $s$ for all groups. Equal FPR requires $P(\hat{Y}=1|Y=0,A=a)=P(\hat{Y}=1|Y=0,A=b)$ for all groups $a,b$.

From Bayes' theorem, the false positive rate can be expressed in terms of the base rate, the positive predictive value (which calibration sets equal to the score), and the score distribution. When base rates differ, the algebra forces a trade-off: equalizing FPR requires different score thresholds across groups, which violates calibration (using the same threshold for all groups).

This result shows that algorithmic fairness is not a purely technical problem. Choosing which fairness criterion to prioritize is a moral and political decision that cannot be resolved by mathematics alone.

Exercises Intermediate+

Domain evidence Master

The COMPAS recidivism algorithm. ProPublica's 2016 investigation of the COMPAS algorithm, used by US courts for sentencing and bail decisions, found that Black defendants were nearly twice as likely as white defendants to be incorrectly classified as high-risk for reoffending. The algorithm achieved similar overall accuracy across racial groups but distributed errors differently: higher false positive rates for Black defendants and higher false negative rates for white defendants. Northpointe (the algorithm's developer) argued that the system satisfied equal predictive accuracy. This case demonstrated that different fairness criteria conflict and that the choice of metric is a moral decision.

The Volkswagen emissions scandal. Volkswagen programmed software in 11 million diesel vehicles to detect when the car was being emissions-tested and activate pollution controls only during testing. On the road, the cars emitted up to 40 times the legal limit of nitrogen oxides. This was not a bug but a deliberate design choice, raising the question: who is responsible when a product is designed to deceive? The scandal led to criminal charges, billions in fines, and a loss of trust that extended beyond Volkswagen to the automotive industry and software engineering profession.

The Theranos case. Elizabeth Holmes founded Theranos on the promise of revolutionizing blood testing with technology that could run hundreds of tests from a single drop of blood. The technology never worked as claimed, but Holmes and her team continued to raise funds, partner with pharmacies, and test patients on unreliable devices. Thousands of patients received inaccurate medical results. The case illustrates the ethical importance of honest reporting in technology development: the gap between what technology promises and what it delivers can have life-or-death consequences.

Deepfake technology. Deepfakes use generative AI to create realistic but fake images, audio, and video. While the technology has legitimate applications (film production, education), it has been used to create non-consensual pornography, political disinformation, and financial fraud. A 2023 deepfake video of a Ukrainian president announcing surrender went viral before being debunked. The technology creates an "information liar's dividend": as deepfakes become more convincing, genuine evidence becomes easier to dismiss as fabricated, undermining trust in all visual and audio media.

Advanced results Master

Surveillance capitalism

Shoshana Zuboff coined "surveillance capitalism" to describe an economic system where human experience is treated as raw material for commercial extraction. Companies like Google and Facebook provide free services in exchange for behavioral data, which they use to predict and influence behavior for profit.

The business model depends on opacity: users do not know what data is collected, how it is analyzed, or how it is used to target them. This asymmetry of knowledge creates a power imbalance that traditional market mechanisms (informed consumer choice) cannot correct.

Zuboff identifies four key features of surveillance capitalism. First, the extraction of behavioral data extends beyond what users voluntarily provide to include tracking of online behavior, location data, and even emotional responses. Second, this data is used to build "prediction products" that forecast what users will do, enabling behavioral modification. Third, these prediction products are sold to advertisers in "behavioral futures markets." Fourth, the system depends on what Zuboff calls "experience extraction," where every aspect of human experience is treated as potential data. This represents a fundamental shift from earlier forms of capitalism: the commodity being extracted and traded is human behavior itself.

Environmental impact of computing

Training a single large language model can emit as much carbon as five automobiles over their lifetimes. Data centers consume approximately 1% of global electricity, and this share is growing. Bitcoin mining consumes more energy than many countries.

The environmental cost of computing raises questions about the ethics of resource consumption. Is the benefit of a marginally better AI model worth its carbon footprint? Should companies be required to report the energy consumption of their AI systems? How should the environmental costs of computing be weighed against its benefits?

Strubell, Ganesh, and McCallum (2019) quantified the carbon footprint of NLP model training: training BERT produced approximately 1,438 pounds of carbon dioxide, roughly equivalent to a round-trip transatlantic flight. Training a larger model with neural architecture search produced 78,467 pounds. These figures do not include the energy consumed during inference (serving millions of queries), which can exceed training costs for popular models. The concentration of AI research in well-funded organizations (Google, Meta, Microsoft, OpenAI) means that access to the computational resources needed for state-of-the-art AI is increasingly restricted, raising equity concerns alongside environmental ones.

Autonomous weapons and AI in warfare

Lethal autonomous weapons systems (LAWS) select and engage targets without human intervention. Proponents argue they may be more precise and less prone to atrocities than human soldiers. Opponents argue that delegating life-and-death decisions to machines is fundamentally immoral, that accountability gaps make violations of international humanitarian law impossible to enforce, and that an arms race in autonomous weapons would destabilize international security.

The Campaign to Stop Killer Robots, a coalition of NGOs, has advocated for an international treaty banning LAWS. The debate centers on whether meaningful human control over the use of force is a requirement of international law and human dignity.

The accountability gap is a central concern. Under international humanitarian law, individuals are responsible for war crimes. If an autonomous weapon commits a violation, who bears responsibility? The commander who deployed it? The engineer who designed the targeting algorithm? The manufacturer? The state? The machine itself cannot be held morally or legally responsible because it lacks intentionality and moral agency. This gap between capability and accountability is a fundamental ethical challenge.

Right to explanation and algorithmic transparency

The European Union's General Data Protection Regulation (GDPR) includes a right to explanation: individuals have the right to understand how automated decisions that significantly affect them are made. This right creates tension with the complexity of modern ML models, which are often opaque even to their creators.

Explainable AI (XAI) aims to make model decisions interpretable. Techniques like SHAP values, LIME, and attention visualization provide post-hoc explanations. However, these explanations may be incomplete or misleading, and the fundamental tension between model complexity and interpretability remains.

The interpretability-accuracy tradeoff is a central challenge. Simple models (linear regression, decision trees) are inherently interpretable but may be less accurate on complex tasks. Deep neural networks achieve high accuracy but are effectively black boxes. This tradeoff raises the question: in high-stakes domains (criminal justice, healthcare, loan approvals), should we accept lower accuracy for the sake of interpretability and accountability? The GDPR implicitly answers yes by requiring that automated decisions can be explained, but enforcing this requirement for complex models remains an open challenge.

Digital sovereignty

Digital sovereignty refers to the ability of nations and communities to govern their own digital infrastructure, data, and technology policies. The concentration of cloud computing, social media, and AI capabilities in a small number of American and Chinese companies raises concerns about technological dependency and cultural imperialism.

The European Gaia-X project, India's data localization requirements, and China's Great Firewall represent different approaches to digital sovereignty. Each reflects different values and priorities regarding openness, security, and autonomy.

The debate over TikTok in the United States illustrates the tension. Critics argued that TikTok's parent company, ByteDance, could be compelled by the Chinese government to share user data, creating a national security risk. Supporters argued that banning TikTok violated free expression and set a dangerous precedent for government control of communication platforms. The underlying question is whether any country should depend on foreign-controlled platforms for its citizens' communication, entertainment, and information.

Content moderation and free speech

Social media platforms face the impossible task of moderating billions of posts per day across hundreds of languages and cultural contexts. Content moderation raises questions about free speech, cultural sensitivity, and the concentration of power. A small number of companies (Meta, Google, X/Twitter, TikTok) control what billions of people can see and say online.

The " moderation paradox" is that platforms are criticized both for moderating too little (allowing hate speech, misinformation, and harassment) and for moderating too much (censorship, political bias, restricting free expression). Section 230 of the US Communications Decency Act shields platforms from liability for user-generated content, giving them broad discretion over moderation decisions. Critics argue this creates unaccountable power; defenders argue it enables the open internet.

Algorithmic amplification complicates the picture. Platforms do not merely host content; their recommendation algorithms actively decide which content to promote. When YouTube's recommendation algorithm promoted conspiracy theories, the harm was not just that the content existed but that the algorithm chose to amplify it. This shifts the ethical question from "should this content be allowed?" to "should this content be promoted?" and raises the possibility of regulating recommendation algorithms rather than content itself.

AI alignment and existential risk

Some researchers argue that advanced AI systems could pose existential risks to humanity. The argument is that if AI systems become significantly more capable than humans (artificial superintelligence), they may pursue goals that are misaligned with human survival, not out of malice but out of indifference. The "orthogonality thesis" (Bostrom, 2014) holds that intelligence and goals are independent: a superintelligent system could pursue any goal, including goals incompatible with human flourishing.

The alignment research community works on techniques for ensuring that advanced AI systems pursue goals consistent with human values. Approaches include inverse reinforcement learning (inferring human preferences from behavior), debate (having AI systems critique each other's outputs), and constitutional AI (specifying principles that guide AI behavior). Whether these approaches will scale to superintelligent systems is unknown.

Connections Master

Connections to law

Computing ethics intersects with law in areas including intellectual property (software patents, copyright, fair use), privacy (GDPR, CCPA, HIPAA), accessibility (ADA, Section 508), consumer protection (right to repair, warranty), and competition (antitrust in big tech). Law provides enforcement mechanisms for ethical principles, but law lags behind technology, creating gaps where ethical obligations exceed legal requirements.

The GDPR, which took effect in 2018, is the most comprehensive privacy regulation in the world. It requires organizations to obtain explicit consent for data processing, provide data portability, enable the right to erasure, report data breaches within 72 hours, and conduct privacy impact assessments for high-risk processing. Violations carry fines of up to 4% of global annual revenue. The GDPR has influenced privacy legislation worldwide, including Brazil's LGPD, India's Personal Data Protection Bill, and China's Personal Information Protection Law. However, enforcement remains inconsistent, and the "consent fatigue" from ubiquitous cookie banners has not meaningfully improved user understanding or control.

Connections to philosophy

Computing ethics engages with core philosophical questions. What is privacy, and why does it matter? What is fairness, and how should it be measured? What is autonomy in a world of algorithmic recommendation? What is the nature of consent in digital environments? These questions require both technical understanding (how the systems work) and philosophical depth (why the principles matter).

Floridi's concept of the "infosphere" reframes the digital world not as a tool humans use but as an environment humans inhabit. This ontological shift has implications for how we think about privacy (not just data protection but protection of the self), identity (not just authentication but the construction of the person), and agency (not just user interfaces but the conditions for autonomous action). The information philosopher argues that digital technologies are not merely instruments but constitutive of human reality, making computing ethics not a specialized domain but a central concern of philosophy itself.

Connections to social science

Understanding computing's societal impact requires social science methods: empirical studies of how people use technology, qualitative research on lived experiences of affected communities, and statistical analysis of disparate impacts. Computer science alone cannot address these questions; interdisciplinary collaboration is essential.

Safiya Noble's "Algorithms of Oppression" (2018) demonstrated through empirical research that Google's search results reinforced racist and sexist stereotypes. Searches for "Black girls" returned sexualized content, while searches for "White girls" returned wholesome results. This was not the result of deliberate programming but emerged from the interaction of search algorithms with societal patterns of racism and sexism in web content. Noble's work showed that algorithmic bias cannot be understood without understanding the social context in which algorithms operate.

Ruha Benjamin's "Race After Technology" (2019) introduced the concept of the "New Jim Code," describing how technology can encode and reinforce racial inequality while appearing to be neutral or even progressive. Examples include predictive policing algorithms that target Black neighborhoods (because historical policing data reflects biased enforcement), beauty AI that rates lighter skin as more attractive (because training data reflects societal beauty standards), and automated hiring tools that disadvantage candidates with non-white names.

Connections to economics

The economics of technology platforms shape their ethical implications. Network effects (the value of a platform increases with the number of users) create natural monopolies, concentrating power in a small number of companies. The freemium model (free services funded by advertising) incentivizes data collection and attention manipulation. The gig economy (Uber, DoorDash) relies on algorithmic management of workers, raising questions about labor rights, fair wages, and worker autonomy.

The attention economy, where human attention is the scarce resource being competed for, creates perverse incentives. Platforms compete for engagement, and engagement is maximized by content that provokes strong emotional responses: outrage, fear, tribalism. This economic logic systematically favors polarizing content over accurate or constructive content. The ethical challenge is not just that individual actors make bad choices but that the economic structure of the attention economy incentivizes harmful outcomes.

Historical and philosophical context Master

The origins of computing ethics

Norbert Wiener's "Cybernetics" (1948) and "The Human Use of Human Beings" (1950) were among the first works to consider the ethical implications of computing and automation. Wiener warned that machines capable of making decisions previously reserved for humans would create profound social challenges.

Joseph Weizenbaum's "Computer Power and Human Reason" (1976) argued that certain decisions should never be delegated to computers, regardless of their technical capability. His critique of AI was not that it was impossible but that it was inappropriate for tasks requiring human judgment, empathy, and moral reasoning.

The development of professional codes

The ACM Code of Ethics, first adopted in 1992 and updated in 2018, provides principles for ethical conduct in computing. It emphasizes contributing to society, avoiding harm, honesty, fairness, respect for intellectual property, privacy, and confidentiality. The code is not enforceable by law but sets expectations for professional behavior.

The IEEE Code of Ethics and the Software Engineering Code of Ethics (jointly developed by ACM and IEEE) provide complementary frameworks. These codes reflect the growing recognition that computing professionals have responsibilities analogous to those of engineers, doctors, and lawyers.

The Cambridge Analytica scandal

In 2018, it was revealed that Cambridge Analytica had harvested personal data from millions of Facebook users without consent and used it to target political advertising in the 2016 US presidential election and the UK Brexit referendum. The scandal demonstrated how personal data could be weaponized for political manipulation and catalyzed public awareness of data privacy issues.

The mechanism was straightforward. A researcher created a personality quiz app on Facebook. When a user took the quiz, the app not only collected the user's data but also data from their friends (a feature of Facebook's API at the time). Through this viral mechanism, data from approximately 87 million users was collected. Cambridge Analytica used this data to build psychological profiles and target political ads to users identified as persuadable, a technique they called "behavioral microtargeting."

The aftermath included regulatory action (GDPR enforcement, FTC fines of $5 billion), platform changes (Facebook restricted third-party data access), and a broader public conversation about the ethics of data-driven advertising. It also highlighted the inadequacy of relying on user consent alone as a safeguard against data misuse: the 87 million affected users had not consented to having their data collected; only the quiz-takers had clicked "accept."

The philosophical significance of computing

Luciano Floridi argues that information technologies are reshaping human identity and social reality in ways comparable to the Copernican, Darwinian, and Freudian revolutions. The "infosphere," the totality of information entities and processes, increasingly defines human experience. Understanding this transformation requires philosophical engagement with computing, not just technical expertise.

The question of whether machines can be moral agents, whether they can bear responsibility for their actions, connects computing ethics to the philosophy of mind and moral philosophy. If an autonomous vehicle kills a pedestrian, who is responsible: the manufacturer, the programmer, the owner, or the machine itself? These questions will become more pressing as AI systems become more autonomous.

Floridi's concept of the "fourth revolution" follows Copernicus (humans are not the center of the universe), Darwin (humans are not separate from animals), and Freud (humans are not fully conscious of their own minds). The fourth revolution, driven by information technology, reveals that humans are not disconnected from artificial agents but share an information environment with them. This ontological shift challenges anthropocentrism and requires new frameworks for understanding human identity, agency, and value in a world where artificial agents increasingly participate in social, economic, and political processes.

The open source movement and ethics of collaboration

The open source movement raises distinctive ethical questions. Open source software (Linux, Python, Apache) forms the infrastructure of the internet and most computing systems. The ethical principle of sharing knowledge freely conflicts with the economic principle of intellectual property. Richard Stallman's Free Software Foundation argues that software should be free (as in freedom, not price) because restricting users' ability to study, modify, and share software is unethical.

The practical ethics of open source include maintaining sustainable funding models (most open source maintainers are unpaid or underpaid), ensuring inclusive governance (open source communities have historically been dominated by Western, male contributors), and addressing the responsibility gap when critical infrastructure depends on volunteer-maintained software. The 2014 Heartbleed vulnerability in OpenSSL, a critical security library maintained by a small team of volunteers, demonstrated the fragility of this model: a bug in software protecting roughly two-thirds of the internet's secure web servers went undetected for two years.

Bibliography Master

Primary sources

• Wiener, N. (1950). The Human Use of Human Beings: Cybernetics and Society. Houghton Mifflin.
• Weizenbaum, J. (1976). Computer Power and Human Reason: From Judgment to Calculation. W.H. Freeman.
• Nissenbaum, H. (2004). "Privacy as contextual integrity." Washington Law Review, 79(1), 119-157.
• Dwork, C. (2006). "Differential privacy." ICALP, 1-12.
• Chouldechova, A. (2017). "Fair prediction with disparate impact: A study of bias in recidivism prediction instruments." Big Data, 5(2), 153-163.
• Kleinberg, J., Mullainathan, S., and Raghavan, M. (2016). "Inherent trade-offs in the fair determination of risk scores." ITCS.

Secondary sources

• Quinn, M.J. (2020). Ethics for the Information Age (7th ed.). Pearson.
• Floridi, L. (2014). The Fourth Revolution. Oxford University Press.
• Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs.
• Noble, S.U. (2018). Algorithms of Oppression. NYU Press.
• Vallor, S. (2016). Technology and the Virtues. Oxford University Press.
• Benjamin, R. (2019). Race After Technology. Polity Press.
• Bostrom, N. (2014). Superintelligence. Oxford University Press.
• O'Neil, C. (2016). Weapons of Math Destruction. Crown.
• Eubanks, V. (2018). Automating Inequality. St. Martin's Press.
• Crawford, K. (2021). Atlas of AI. Yale University Press.
• Strubell, E., Ganesh, A., and McCallum, A. (2019). "Energy and policy considerations for deep learning in NLP." ACL, 364-369.