9. Evaluating the Social Impact of Computing

Understanding computing’s wider effects

Computing innovations can improve communication, education, health care, scientific research, business, emergency coordination, and political participation. They can also create surveillance, privacy loss, misinformation, harassment, labor displacement, cyberattacks, environmental costs, and concentrated power.

The central question is not simply whether a technology works. Ask:

• Who benefits, and under what conditions?

• Who may be harmed or excluded?

• Are the effects intended or unintended, immediate or long-term, individual or collective?

• Are harms reversible, or could they be difficult to undo?

• Is there a less harmful way to achieve the same public purpose?

A technology may increase an organization’s efficiency while imposing hidden costs on workers, users, or communities. Benefits may also be distributed unevenly when people differ in income, access, skills, language, disability, or social power.

A concise evaluation should consider seven connected dimensions:

• Technical: reliability, safety, security, and performance.

• Economic: income, employment, productivity, market power, and access to resources.

• Social: relationships, health, education, participation, safety, and inequality.

• Ethical: privacy, autonomy, fairness, dignity, accountability, and human rights.

• Cultural: language, identity, traditions, representation, and community control.

• Environmental: energy, materials, operating costs, and electronic waste.

• Legal and civic: compliance, democratic participation, and meaningful remedies.

Takeaway: Responsible evaluation asks how a technology changes power, opportunity, and risk—not only how efficiently it performs.

Mapping stakeholders, benefits, and risks

Begin by identifying everyone who may be affected, including people who never use the system. Stakeholder analysis should include:

• direct users who operate or rely on the system;

• nonusers who experience its effects indirectly;

• developers and owners who design, sell, or control it;

• workers and communities whose jobs, services, or environments may change; and

• regulators and the public, who have responsibilities related to safety, rights, and accountability.

Ask whose interests shaped the design and whose interests may have been overlooked. A system designed only around the organization that purchases it may fail the people who must live with its decisions.

A benefit–risk comparison can organize the evidence:

• What benefits are expected, and who receives them?

• What harms are possible, and who bears them?

• How likely and severe are the harms?

• How long might they last, and can they be reversed?

• What safeguards are available, such as audits, encryption, human review, accessibility features, or appeal processes?

• Is a smaller deployment, opt-in design, nonautomated process, or pilot program safer?

Do not rely only on average outcomes. A small average improvement may be unacceptable if it creates severe harm for a vulnerable group.

Takeaway: A responsible decision compares alternatives and distributional effects, not merely overall efficiency.

Access, inequality, and digital inclusion

The digital divide concerns more than whether someone has an Internet connection. Meaningful access depends on quality, availability, affordability, suitable devices, skills, and security. Inequality can occur between countries, urban and rural areas, income groups, generations, genders, languages, and people with and without disabilities.

It can also occur within a household. Someone may have a smartphone but lack a computer, enough data, private access, reliable service, or the skills needed for online education and employment.

Suppose a government moves all public-benefit applications online. The policy might reduce administrative costs and increase convenience for connected users, but it could exclude people who lack broadband, devices, digital skills, accessible interfaces, language support, or reliable identification documents. More inclusive implementation could provide:

• public access locations;

• telephone and in-person alternatives;

• accessible design;

• multilingual support; and

• assistance from trained staff.

Digital inclusion means expanding people’s real opportunities rather than making essential services conditional on resources they do not possess.

Takeaway: Measure access by people’s ability to use technology effectively and safely, not by the mere existence of a network.

Bias, fairness, and automated decisions

Computing bias can enter at every stage of a system:

1. Problem definition: The system may encode a narrow or unjust goal.

2. Data collection: Some populations may be missing, underrepresented, or misclassified.

3. Labels and measurements: Categories may reflect historical prejudice or unreliable proxies.

4. Model design: Optimization may prioritize overall accuracy while ignoring unequal error rates.

5. Deployment: Users may apply the system in settings different from those in which it was tested.

6. Feedback loops: Automated decisions may create new data that reinforces the original pattern.

For example, a hiring model trained on an organization’s past decisions may learn to favor characteristics associated with previously selected employees. Removing race or gender from the data does not automatically remove bias if other variables act as proxies.

Algorithmic fairness is not one universal mathematical rule. Equal error rates, equal selection rates, and equal predictive accuracy can conflict, especially when groups have different underlying rates. People must decide which outcomes are ethically appropriate.

A system should be treated as a socio-technical system because outcomes depend on interactions among data, software, institutions, developers, users, and social conditions. Responsible practices include:

• involving affected communities in problem definition and testing;

• measuring performance across relevant demographic and contextual groups;

• documenting data, assumptions, limitations, and intended uses;

• testing for disparate impacts before and after deployment;

• providing human review for high-impact decisions;

• giving people notice, explanations, correction procedures, and appeals;

• monitoring continuously; and

• stopping or redesigning the system when harms cannot be adequately controlled.

Takeaway: Fairness requires lifecycle accountability; it cannot be guaranteed by removing one sensitive variable or performing one initial audit.

Cybersecurity, encryption, and layered protection

Cybersecurity protects three core properties of systems and data:

• Confidentiality: information is available only to authorized parties.

• Integrity: information and systems are not improperly altered.

• Availability: authorized people can access systems and data when needed.

Common threats include phishing, malware, ransomware, credential attacks, injection attacks, social engineering, insider threats, and supply-chain attacks. Encryption protects data by transforming plaintext into ciphertext with an algorithm and key. It can protect information in transit, at rest, and, with specialized methods, in use.

Encryption is essential but not sufficient. Weak passwords, stolen keys, phishing, insecure software, poor access controls, and untrained staff can still compromise a system.

A layered cybersecurity program should combine:

• strong authentication and access controls;

• secure software development and vulnerability management;

• encryption and careful key management;

• staff training against phishing and social engineering;

• backups and recovery procedures;

• incident detection and response planning; and

• clear roles and accountability.

Takeaway: Security is an ongoing risk-management process, not a single tool or one-time technical installation.

Crowdsourcing and meaningful participation

Crowdsourcing can gather information across large areas, identify creative solutions, accelerate data classification and research, give communities a voice, and reduce the time or cost of some tasks. However, participation is often self-selected. Results may overrepresent people with Internet access, available time, particular skills, or strong opinions.

A responsible project should address:

• representation: Are the participants similar to the people affected?

• quality: How will contributions be checked?

• privacy: What personal information is collected and protected?

• consent: Do participants understand how their contributions will be used?

• compensation and recognition: Is substantial labor being rewarded appropriately?

• intellectual property: Who controls the resulting ideas or content?

• moderation and safety: How will abuse, misinformation, or harmful material be handled?

The number of participants is not enough to establish meaningful participation. The process should explain how contributions influence decisions and should report important limitations.

Takeaway: Participation is responsible when it is informed, protected, fairly recognized, and relevant to the people affected.

A repeatable method for responsible decisions

Responsible computing is an ongoing governance process. A city considering an AI system to prioritize housing-inspection requests could use the following sequence:

1. Define the public purpose. Aim to identify dangerous buildings more quickly, rather than merely reducing inspection costs.

2. Identify stakeholders and rights. Consider tenants, landlords, inspectors, neighborhoods, people with disabilities, city officials, and residents with limited Internet access. Address safety, privacy, due process, and equal treatment.

3. Examine data and assumptions. Check whether historical records underreport problems in neighborhoods where residents had less ability to file complaints. Underreporting could make a high-risk area appear safer.

4. Test outcomes and alternatives. Compare error rates across neighborhoods and demographic groups. Consider transparent rules, community reporting, or additional inspectors.

5. Add safeguards. Possible measures include independent audits, public documentation, human review, emergency overrides, appeals, privacy protections, and continued service for people who cannot use an online portal.

6. Monitor and revise. Measure real-world outcomes, investigate complaints, publish performance information, and suspend the system if harms exceed acceptable limits.

This method applies beyond housing inspections. The same reasoning can guide decisions about education, employment, health care, public benefits, policing, finance, and other high-impact uses of computing.

Final takeaway: Define the purpose, identify affected people, inspect assumptions, compare alternatives, add safeguards, and keep monitoring results after deployment.