Computing, Ethics and Society

Computing ethics asks how people ought to design, build, and use computing systems. Technical questions ask whether a system can do something. Ethical questions ask whether it should, whose interests it serves, and what responsibilities follow. A system's effects reach beyond its users to communities and the environment.

Technology and its setting

A computing system is socio-technical: software works within organizations, incentives, rules, and human habits. A recommendation system can help people discover music. An objective that rewards watch time alone can encourage increasingly extreme videos. The objective and the surrounding business model matter alongside the code.

Designers help shape behavior through defaults and available choices. A public-by-default photo app exposes posts more widely than a friends-only default. Calling the technology neutral misses the decision already built into the interface. Autonomy, the ability to make one's own choices, depends partly on whether those choices are understandable and practical.

Map stakeholders and effects

A stakeholder affects or is affected by a system. Direct stakeholders use it or appear in its data, including a bystander recorded by smart glasses. Indirect stakeholders are affected without using it or appearing in its data: workers whose jobs change or neighbors of a data center.

A stakeholder analysis follows five steps:

1. Name the technology and the decision being made.
2. List direct and indirect stakeholders, including people with little influence.
3. For each group, record benefits, harms, and values at stake.
4. Identify who decides and who carries the risks.
5. Look for unintended effects, long-term effects, and effects at scale.

Values include privacy, safety, security, fairness, autonomy, accountability, transparency, access, trust, and sustainability. A location-sharing app can improve family safety while reducing privacy. A ride-hailing app can make booking convenient and work flexible while contributing to traffic and unstable earnings.

Work through an exam-proctoring choice

A college considers AI proctoring for remote exams. Students gain travel-free testing, but recording their homes threatens privacy. False cheating alerts threaten fair assessment. Teachers gain support detecting cheating and take on the work of checking alerts. Family members caught on camera are direct stakeholders because their images enter the data, even though they never chose the service. The vendor earns revenue; the college chooses the system. Students carry much of the risk, so consultation must give their concerns real weight.

A dual-use technology has beneficial uses and can also be misused to cause harm. Face recognition can unlock a phone or identify people at a peaceful protest for intimidation. Crowdsourcing draws contributions from many people online, as Wikipedia, open-source projects, Galaxy Zoo, and crowdfunding do. Its scale brings knowledge and resources, but also quality problems, vandalism, unpaid labor, and low-paid tasks.

Scale changes the seriousness of small risks. A flaw affecting 0.1% of 1,000,000 users harms 1,000 people. Software can spread that flaw to everyone at once, so a low individual risk still deserves attention.

An ethical framework gives a structured reason for judging a choice. Several frameworks can examine the same facts and reveal different concerns. A strong argument is accurate, consistent, and attentive to everyone affected; disagreement does not make every argument equally good.

Four ways to reason

Utilitarianism, associated with Jeremy Bentham and John Stuart Mill, is a form of consequentialism. It seeks the greatest overall well-being for everyone affected. List feasible options, estimate each group's gains and losses, and compare their combined effects. A small gain to many people does not make a major harm to a vulnerable group disappear from the analysis.

Deontology, associated with Immanuel Kant, judges actions by duties. Kant asks whether you could want everyone to follow the same rule and whether you treat people as ends in themselves, never merely as tools. Rights-based reasoning similarly gives privacy and free expression weight even when ignoring them would benefit others.

Virtue ethics, traced to Aristotle, asks what habits a choice shows and builds: honesty, courage, care, fairness, and responsibility. A professional can ask what a competent engineer acting in good faith would do.

For justice, John Rawls's veil of ignorance asks you to choose rules without knowing your own place under them. Would the rule still seem fair if you were the bystander or the wrongly flagged user? Care ethics draws attention to relationships and the needs of people who are vulnerable or dependent.

Apply the approaches

An employer proposes recording keystrokes and taking screenshots every 10 minutes. Consequence-based reasoning weighs useful security information against stress, lost trust, and capture of private information. Duty-based reasoning asks whether workers received meaningful notice and choice. Virtue ethics asks whether the employer acts respectfully and honestly. Fairness asks why managers decide while workers carry the monitoring burden.

A concrete alternative is to assess work by results or limit monitoring to work accounts with clear notice. Several frameworks supporting that alternative strengthens the case. If they disagree, identify the conflict and explain which concern you give greater weight.

Keep law and ethics distinct

Law specifies obligations in a jurisdiction. Ethics asks what is right. A data sale can be legal yet violate reasonable expectations. Some whistleblowing can be defended ethically while still breaking a law. Legal compliance is one part of a decision, alongside evidence about harm, rights, and responsibility.

Personal data relates to an identified or identifiable person. It includes names, photos, email addresses, location, IP addresses, and device identifiers. US discussions often use personally identifiable information (PII). Health, biometrics, financial details, precise location, religion, political views, and sexual orientation can be especially sensitive because misuse can cause serious harm.

Collection goes beyond what you type

Active collection happens when you submit a form or post a photo. Passive collection records activity such as clicks, searches, device type, and time on a page. Metadata describes other data: who called whom, when, and for how long. Repeated calls to a cancer clinic can reveal something sensitive even without the conversation.

A cookie is a small piece of information stored by a website in a browser. First-party cookies belong to the visited site; third-party cookies come from another domain and can enable tracking across sites. Email pixels can record that a message was opened. Device fingerprinting combines browser settings, fonts, and screen characteristics to recognize a browser. Advertising IDs, permissions, and sensors in cars, speakers, and wearables create other trails.

Organizations use data to deliver and improve services, personalize content, detect fraud, and target advertising. Data brokers buy, combine, and sell profiles, often without a direct relationship with the people described. Both free and paid services can collect and share data.

Combining and identifying

Aggregation brings pieces together; inference draws a conclusion from patterns, such as estimating a life event from shopping changes. Re-identification links supposedly anonymous records to people. Removing names leaves other clues intact.

Latanya Sweeney's study using 1990 US census data estimated that 87% of the population, about 216 million of 248 million people, was likely unique by five-digit ZIP code, gender, and full birth date. A later study using 2000 census data estimated about 63%. Uniqueness singles out a record; a link to a named list supplies an identity. Sweeney demonstrated that second step in 1997 by linking released state-employee health data to a voter list and finding a governor's record.

Researchers also matched some Netflix users' released movie ratings to public IMDb reviews after Netflix's 2006 release. A 2013 study of about 1.5 million mobile-phone users found that four place-and-time points singled out about 95% of people in that dataset. Other information, such as a known home and workplace, can then reveal names. The percentage describes that study, not every dataset.

Privacy protects context

Surveillance systematically monitors people. It includes workplace tracking, face-recognition camera networks, and hidden stalkerware. A chilling effect occurs when people avoid speech, searches, or activities because they believe they are watched.

Privacy matters even when a person has done nothing wrong. Data can be breached, misunderstood, or used in a new setting. Helen Nissenbaum's contextual integrity asks whether information flows fit the context: sharing symptoms with a doctor fits medical care; passing them to advertisers changes that relationship.

Privacy protection starts with the purpose of collection. A delivery app needs an address, contact details, and payment arrangements. Its service does not require a contact list or continuous background location. Data minimization asks for the least information that meets the purpose, reducing what a later breach can expose.

Turn principles into design choices

The OECD Privacy Guidelines, introduced in 1980 and revised in 2013, organize responsible data practice into eight principles:

Informed consent is specific, informed, freely given, and easy to withdraw. A long policy or a take-it-or-leave-it demand can prevent real understanding or choice. Dark patterns steer decisions through interface design, such as a bright acceptance button beside a hard-to-find settings link.

Privacy by design and by default puts protection in place early: protective defaults, minimal collection, limited retention, restricted access, and encryption. A retention limit answers when data will be deleted, not just why it is stored today.

Reduce identifiability carefully

Pseudonymization replaces direct identifiers with codes while preserving a way to link them back. Anonymization aims to prevent identification. A coded record with a separate lookup key is still linkable. Age bands, combined totals, and carefully designed releases can lower exposure. Differential privacy adds calibrated random noise to statistics to limit what they reveal about individuals; it was used for the 2020 US Census results.

For a community survey, collecting age bands instead of full birth dates reduces detail. Publishing only totals reduces it further. A rare combination in a small group can still expose someone, so the release design matters as well as the removal of names.

Law, trade-offs, and personal choices

The EU's GDPR, Brazil's LGPD, and India's Digital Personal Data Protection Act of 2023 are examples of broad data-protection laws. The US also has sector laws, including HIPAA for covered health information and COPPA for children under 13 online, and state laws such as California's CCPA. Rights such as access, correction, deletion, and objections depend on the applicable law.

Privacy can conflict with convenience, safety, or investigation. Some COVID-19 contact-tracing apps used decentralized Bluetooth designs that kept contact logs on phones, showing how a design can reduce a trade-off.

Review app permissions, social-account defaults, and third-party tracking. Restrict location to use of an app when that meets its need. Public addresses, birth dates, and travel plans can aid identity theft or burglary; copies of posts can outlive deletion. Private browsing limits local history and session traces, but does not hide activity from websites or all network observers. HTTPS still protects the contents of encrypted traffic in transit.

Social engineering exploits human trust and pressure. An attacker can appeal to authority, urgency, fear, curiosity, or helpfulness to get a person to reveal information or take an unsafe action. A message can be polished, personal, and still deceptive.

Recognize the route into an attack

Phishing impersonates a trusted sender to obtain clicks, credentials, files, or payments. Spear phishing targets a person with personal details; smishing uses text messages; vishing uses voice calls. Business email compromise uses a seemingly legitimate workplace payment request. QR codes can hide phishing links, and a familiar-sounding AI-cloned voice can impersonate a colleague.

Consider: "Dear customer, your account has been suspended. Verify your details within 24 hours: paypa1-secure-login.com." The threat and deadline create pressure, the greeting is generic, and the domain substitutes the digit 1 for a letter. Open the service through its known app or address, report the message, and delete it.

Read the destination

For a simple web address, inspect the host after https:// and before the path. The registrable domain is the public suffix, such as .com or .co.uk, plus the label immediately before it. In login.example.com.account-check.net/example.com, the domain is account-check.net. The recognizable name in the path cannot change who controls the host.

Hover over or long-press a link to inspect its actual destination. A padlock or HTTPS protects the connection to that destination; a phishing site can have an encrypted connection too. A trusted name elsewhere in an address is no substitute for checking the host.

Malware and related attacks

Malware is software made to harm or exploit devices or people.

Malicious attachments, fake downloads, pirated software, extensions, USB drives, and unpatched flaws can deliver malware. Slowness, pop-ups, changed settings, unknown apps, and messages sent without your knowledge can be warning signs, though one symptom alone can have other causes.

Credential stuffing tries leaked username-password pairs at other services. SIM swapping moves a victim's phone number to an attacker's SIM, putting text-message codes at risk. An evil-twin Wi-Fi hotspot imitates a real network. Fake shops and tech-support calls use similar trust tricks.

Act quickly after exposure

If credentials were entered, change that password from a trusted device and change it anywhere it was reused. Enable MFA. Contact the bank promptly if payment information was involved, run a security scan where appropriate, and report the incident to the service or IT team. Well-made phishing can fool experienced people; quick reporting helps contain the harm.

Authentication proves identity. Its factors are something you know, such as a password; something you have, such as a security key; and something you are, such as a fingerprint. Multi-factor authentication (MFA) combines different factor categories. A password plus a PIN is still one category.

Make passwords hard to guess and reuse attacks fail

Length and unpredictability create strength. With independent random choices, possibilities equal the number of available symbols raised to the password length. Eight random lowercase letters give ${26}^8\approx 2.1\times {10}^{11}$ possibilities. At an example rate of 10 billion offline guesses per second against a stolen database, covering that space takes about 21 seconds. Six random words from a 7,776-word list give ${7776}^6\approx 2.2\times {10}^{23}$ possibilities, about 700,000 years at that rate. These are full-space estimates under the stated rate, not predictions for every attack.

Quotes and patterns such as "P@ssw0rd!" are predictable. A password manager generates and stores long random passwords and helps avoid filling them into a lookalike domain. Protect the manager with a strong master password and MFA. Use a unique password for each account so one leak cannot unlock others.

US NIST SP 800-63B-4 guidance from 2025 requires at least 15 characters for a password used alone, or at least 8 when it is part of MFA. Systems should support at least 64 characters, reject common or breached passwords, and avoid mandatory character mixtures and scheduled changes. Change a compromised password; monthly predictable tweaks offer little protection.

Choose stronger sign-in methods

Text codes improve on a password alone but face phishing and SIM-swap risks. Authenticator-app codes avoid the SIM dependency, yet can still be phished. Repeated push prompts can cause MFA fatigue. Never approve a login you did not start.

Passkeys and FIDO2 security keys resist phishing by binding authentication to the real website's domain. A passkey uses a cryptographic key pair and can be unlocked with a device PIN, fingerprint, or face.

Protect devices and stored work

Updates close known security flaws; automatic updates reduce the period of exposure. A zero-day is exploited before a developer has released a fix. Unsupported devices eventually lose this protection.

Encryption in transit protects a connection, as HTTPS does. End-to-end encryption keeps message contents readable by the communicating endpoints, rather than the messaging provider. Full-disk encryption and a strong screen lock protect a lost device's stored data. A VPN protects the connection to the VPN provider and shifts trust to it; it does not stop phishing or malware, and HTTPS still protects encrypted content beyond that connection.

Use 3-2-1 backups: three copies on two storage types, with one offsite. A laptop original, an unplugged external drive, and cloud backup fit the pattern. Keep a copy offline or unchangeable and test restoration, since ransomware can damage reachable copies. Lock screens, use official app stores, remove unused extensions, limit permissions, and erase devices before passing them on.

Intellectual property (IP) combines creations such as software, inventions, and brands with rights controlling their use. These rights reward creative effort while leaving room for access, learning, and further invention. Different rights protect different aspects of the same product.

Compare four forms of protection

For a photo-editing app, copyright protects its code, a patent might protect a new technical image-processing invention, a trademark can protect its name and icon, and trade secrecy can protect unpublished model settings. A trade secret requires steps such as access restrictions or non-disclosure agreements. It cannot stop someone independently discovering the same information.

Copyright protects expression

Copyright covers writing, music, images, films, and software. Under the Berne framework, protection does not depend on registration or a copyright symbol. The protected expression differs from an underlying idea or algorithm. Two programmers can independently implement the same sorting method. Copying protected code requires permission or an applicable exception.

Owners control activities such as copying, distribution, adaptation, and public display or performance. Berne sets a general minimum of the author's life plus 50 years; the US and EU member states commonly use life plus 70. Special categories have different terms. Expired protection places a work in the public domain.

US fair use considers purpose and character, the work's nature, the amount taken, and the effect on the original's market. Commonwealth fair dealing systems specify permitted purposes. Giving credit, taking fewer than 10 seconds, or avoiding profit is not a universal permission rule. Quotation, criticism, reporting, teaching, and parody depend on the relevant exception and circumstances.

Patents and digital restrictions

A patent lets an owner exclude others from making, using, or selling the invention within the territory of protection. Under TRIPS, the available term is at least 20 years from filing. The inventor discloses how the invention works. A patent does not itself authorize sale: another patent or a separate approval may still matter.

Software-related patent rules vary. The US grants some software-related patents; the European Patent Convention excludes programs "as such" while allowing qualifying technical inventions.

Software piracy copies or distributes software without permission. Digital rights management (DRM) restricts copying through technology. Some laws also restrict bypassing DRM, including the US Digital Millennium Copyright Act. A file being visible online tells you where it can be found, not what you may legally do with it.

A software license grants permission to use copyrighted software under stated conditions. Paying for software usually buys rights to use it, not its copyright. A proprietary end-user license agreement (EULA) often limits modification and redistribution. Freeware costs nothing but may still withhold source code and modification rights.

Freedom and price are separate

Free and open-source software provides rights to run, study, modify, and share code. The Free Software Foundation describes four freedoms covering running, studying and changing, redistributing, and sharing modified versions. The Open Source Initiative's ten criteria include source access, redistribution, derived works, and freedom from discrimination against people or fields of use. Open-source software can be sold.

Follow the conditions, including the trigger

MIT requires retaining copyright and permission notices. Apache 2.0 includes an explicit contributor patent license. On redistribution, include its license, retain applicable notices, mark changed files, and carry forward relevant NOTICE-file attributions if present.

GPL obligations principally arise when software is distributed. An organization can use and modify GPL software internally without publishing it. AGPL addresses modified software offered over a network. Merely compiling with a GPL compiler or writing in a GPL editor does not license the resulting program under GPL.

Suppose a startup includes an MIT date library in an app it sells. It can keep its own source private while supplying the MIT notices. Copying a GPL-3.0 image library into that distributed app instead requires GPL-compatible distribution of the combined program and corresponding source. The team can comply, select a suitably licensed alternative, or obtain a commercial license if the author offers dual licensing.

Check compatibility before combining

License compatibility means all applicable conditions can be met together. MIT code can enter a GPL project while keeping MIT notices. GPL code cannot be redistributed in a combined program released only under permissive terms. Apache 2.0 is compatible with GPLv3, but not GPLv2-only.

A public repository without a license grants no general permission to reuse its code. Ask the holder for permission. Teams use license scanners and a software bill of materials (SBOM) to track dependencies and their terms.

For your own work, permissive licensing supports wide reuse; copyleft preserves source-sharing obligations for covered downstream versions. Open source allows independent inspection, but maintainers still need support. Log4Shell, found in December 2021 in the widely used Log4j library, showed how much infrastructure can depend on volunteer-maintained code.

A creator can grant reuse permission in advance with a Creative Commons (CC) license. These licenses suit text, images, music, video, and similar works. For software, Creative Commons recommends software-specific licenses; CC0 is an exception it considers suitable for code.

Read the license components

The six licenses are CC BY, CC BY-SA, CC BY-NC, CC BY-NC-SA, CC BY-ND, and CC BY-NC-ND. All require attribution. CC0 waives rights as far as possible, placing the work in effect in the public domain.

A CC BY-NC-SA 4.0 photograph cannot be used on a company's product sales page under that license. A volunteer group's free newsletter can use it with attribution; a cropped and recolored adaptation must retain BY-NC-SA terms. Credit should give TASL: title, author, source, and license, plus changes. For example: "'Harbor at Dawn' by M. Okafor, Wikimedia Commons, CC BY-NC-SA 4.0, cropped." Link the source and license where possible.

Separate access, permission, and honesty

Public-domain works can be reused without copyright permission. They include expired works, CC0 dedications, and material never protected by copyright, such as US federal government works within the US. Check the actual edition: a recent translation of an old work may have its own protection.

Open-access publications are free to read and often carry a reuse license. Open educational resources grant permissions that support teaching and adaptation. Free access alone does not tell you the exact reuse terms.

Plagiarism presents another person's work or ideas as your own without credit. Copyright infringement uses protected expression without permission or a valid exception. Copying an old public-domain poem into an essay as your own is plagiarism even though copyright has expired. Publishing a protected photograph with full credit can still infringe copyright if no permission or exception permits it.

There is no general rule that changing 10% of a work makes it yours. An adaptation still depends on the original's rights. Likewise, absence of a copyright symbol does not remove protection.

Check AI outputs and the people represented

The US Copyright Office's 2025 approach protects human authorship; prompts alone generally do not establish authorship of generated expression. Human creative contributions can still matter. Rules in other countries differ, and the legality of training on protected material continues to be disputed in courts and legislatures.

Generated text, images, or code can reproduce protected work, so examine the output and its reuse conditions. Follow applicable disclosure requirements at school or work. When sharing media, consider privacy and consent as well as copyright. A deepfake imitating someone's face or voice without consent can damage their safety and reputation.

An automated decision system uses an algorithm to make or support choices about people, such as loan, hiring, or medical-care decisions. Algorithmic bias is a systematic pattern of errors or outcomes that unfairly disadvantages groups. It can arise without deliberate discrimination.

Follow the data through the system

Machine learning finds patterns in training data. If past decisions reflect unequal treatment, a model trained to copy them can repeat that treatment at large scale. Removing an explicit gender or race field leaves other features that may act as proxies, indirectly carrying the same information.

Learn from distinct failures

In 2018, Reuters reported that Amazon abandoned an experimental résumé-screening system trained on ten years of submissions, mostly from men. It penalized wording such as "women's chess club captain." Excluding a gender field could not remove all of the signals learned from historical data.

Joy Buolamwini and Timnit Gebru's 2018 Gender Shades study tested three commercial gender-classification systems. Errors reached 34.7% for darker-skinned women, while the highest rate for lighter-skinned men was 0.8%. Two common benchmarks were 79.6% and 86.2% lighter-skinned. This concerns gender classification, a different task from identifying a person. Face-recognition errors have separately contributed to wrongful arrests, including Robert Williams's 2020 arrest in Detroit.

Obermeyer and colleagues studied a US health-care algorithm in 2019. It predicted future costs as a proxy for need. Less had been spent on Black patients at the same level of illness, so cost made them appear healthier. Correcting the bias would raise the share of Black patients identified for extra care from 17.7% to 46.5%. A more balanced sample alone cannot fix a label that measures the wrong goal.

Look beyond a single score

A predictive-policing loop can send patrols where recorded crime is high. Extra patrols find more offenses, which increases recorded crime there and attracts more patrols. The records partly measure where police looked, so the loop can grow even when actual offense rates elsewhere are similar.

Handwritten rules can exclude people too: a form rejecting apostrophes or requiring two names blocks valid users. Speech recognition can work unevenly across accents, and generated images can repeat occupational stereotypes.

Overall accuracy can hide group failures. If 90% of users receive 97.8% accuracy and 10% receive 70%, then $0.9\times 97.8+0.1\times 70\approx 95.0$ percent overall. Break results down by affected groups and by the errors that matter to them.

A fairness assessment asks who gets selected and who receives each kind of error. First define a positive decision and the real outcome. In hiring, "positive" can mean an interview invitation, while "actually positive" means qualified. A confusion matrix cross-tabulates predictions and reality.

Choose the right denominator

TPR starts with people who actually qualify. Precision starts with people selected. Their denominators answer different questions.

Work through two groups

Group A has 100 applicants: 54 TP, 6 FN, 8 FP, and 32 TN. Group B has 100: 28 TP, 12 FN, 6 FP, and 54 TN.

• Selection: A selects 62 out of 100 (62%); B selects 34 (34%).
• TPR: A has $54\mathrm{/}60=0.90$; B has $28\mathrm{/}40=0.70$.
• FPR: A has $8\mathrm{/}40=0.20$; B has $6\mathrm{/}60=0.10$.
• Precision: A has $54\mathrm{/}62\approx 0.87$; B has $28\mathrm{/}34\approx 0.82$.

Demographic parity requires equal selection rates, so it fails. The selection-rate ratio $34\mathrm{/}62\approx 0.55$ is below the US employment four-fifths screening guideline of 0.8, a flag for possible adverse impact. Equal opportunity requires equal TPR, also absent: B misses 30% of qualified people; A misses 10%. Equalized odds requires equal TPR and FPR. Predictive parity requires equal precision.

Calibration within groups means predicted probabilities match observed rates within each group: roughly 70% of people assigned a 70% risk have the outcome. It differs from matching precision at one selection threshold.

Measures express different priorities

Different base rates and imperfect predictions can make fairness aims conflict. In the 2016 COMPAS debate, ProPublica reported false-positive rates of 44.9% for Black defendants and 23.5% for white defendants. The developer and other researchers emphasized calibrated scores. Those claims can coexist: calibration does not ensure equal error rates. When groups have different base rates, matching both error rates generally conflicts with matching precision in imperfect systems.

Reduce harm throughout use

Ask whether automation is appropriate. Choose representative data and valid labels, audit group results, adjust thresholds or models, and document limits in model cards and dataset documentation. Monitor after launch.

Affected people need notice, meaningful reasons, and a way to contest decisions. Human reviewers need time, evidence, and authority to disagree; automation bias can turn review into a rubber stamp. Consult affected communities.

The EU AI Act regulates AI by risk, including certain prohibited social-scoring practices and high-risk uses in hiring and credit. New York City requires independent audits for covered automated hiring tools.

The digital divide separates people who can benefit from digital technology from those who cannot. It exists between countries and within them. Counting connected people is useful, but a connection alone does not show whether someone can complete a job application, learn effectively, or obtain health care.

Distinguish three levels

Meaningful connectivity involves adequate speed, affordability, regular availability, and a suitable device. A shared phone on an expensive data plan can count as internet access while making a long application impractical.

ITU estimated that about 6 billion people used the internet in 2025, roughly three-quarters of the world, while about 2.2 billion remained offline. Its estimates were about 94% online in high-income countries versus 23% in low-income countries, and 77% of men versus 71% of women. These describe broad populations; they do not identify every individual's circumstances.

Trace causes and effects

Device and data costs can block low-income households. Remote locations can lack infrastructure. Age, education, language, disability, and confidence affect use. Gender norms, censorship, and internet shutdowns can restrict participation. As tools change, new gaps emerge in access to AI, the skills to evaluate it, and the languages it supports.

During COVID-19 school closures, lack of home connectivity and suitable devices contributed to a homework gap. Online-only banking, appointment booking, job applications, and public services can deepen existing disadvantages. A service may be technically available while requiring equipment, language skills, or confidence its intended users lack.

Match the response to the barrier

Imagine a city moving bus-pass renewals entirely online. List who could be excluded: residents without devices, people facing data costs, people who cannot read the site's language, older users needing help, and disabled users facing inaccessible controls. The causes require different responses.

Infrastructure such as rural broadband, community networks, mobile service, and satellite links addresses coverage. Affordable plans, subsidies, and refurbished devices address price. Libraries and community centers offer shared access and help. Training builds skills; local-language content makes services understandable. Accessible interfaces remove disability-related design barriers. Phone and in-person routes preserve access to essential services while other changes take effect.

Giving everyone a smartphone would address only part of this problem. Evaluate success by what people can actually complete and gain, alongside device and connection counts. Test the service with the people expected to depend on it.

Accessibility means people with disabilities can perceive, understand, navigate, and operate technology. WHO estimates that about 1.3 billion people, around 16% of the population, have a significant disability. A temporary broken arm or a noisy environment can also make an interface hard to use.

Start with how people use the system

Screen readers turn text and structure into speech or braille. Magnifiers enlarge content. Captions and transcripts convey audio. Voice control, switches, and keyboards provide alternatives to a mouse. Reading aids and text-to-speech help some people with dyslexia. Real headings, labels, and buttons let these technologies interpret the interface.

W3C's WCAG 2.2, first published in October 2023, organizes web accessibility into POUR:

Success criteria have levels A, AA, and AAA; AA is a common organizational and legal target.

Convey meaning in more than one way

Meaningful images need purpose-based alt text. "chart.png" identifies a file; "Online sales doubled between 2022 and 2024" conveys the chart's message, with the detailed data available in a table. Decorative images use empty alt text so a screen reader skips them. An icon-only button needs an accessible name such as "Search."

Avoid color-only instructions. Add words or symbols to errors and required fields. Videos need captions; audio needs transcripts. Captions also help in noisy spaces and when learning a language, an example of the curb-cut effect: disability-focused improvements benefit others too.

Check contrast and operation

WCAG AA requires at least 4.5:1 contrast for normal text and 3:1 for large text, at least 18 point or 14 point bold. For lighter luminance $L_1$ and darker luminance $L_2$, calculate $(L_1+0.05)\mathrm{/}(L_2+0.05)$. White on #767676 gives approximately $1.05\mathrm{/}0.231=4.54$, enough for normal text. White on #999999 gives about 2.85:1, below even the large-text threshold.

Every function needs a keyboard route, a visible focus indicator, and freedom from focus traps. Give people enough time. Avoid flashing more than three times per second. Label fields clearly, explain errors in words, and keep navigation consistent.

Test with people, not only tools

An automated checker finds some defects, but it cannot establish that a whole service is accessible. Test keyboard operation and screen-reader behavior and involve disabled users. Build accessibility in early, alongside visual design. Universal design aims for broad usability without separate adaptations. UN disability-rights commitments, including CRPD Article 9, recognize equal access to information and communication technology; legal requirements vary by country.

A device's life cycle runs from mining through manufacture, transport, use, and disposal. Embodied emissions come from production and transport; operational emissions come from electricity during use. CO2e, carbon dioxide equivalent, puts different greenhouse gases on a common warming-impact scale.

For many smartphones and laptops, manufacturing accounts for most lifetime emissions. Servers running continuously can have a much larger operational share. Electricity is only one part of a device's footprint.

Work from power to energy to emissions

Power is the rate of energy use. Convert watts to kilowatts before using hours:

A laptop drawing 50 W uses 0.050 kW. For eight hours, $0.050\times 8=0.4$ kWh. Over 250 working days, $0.4\times 250=100$ kWh. At an example grid carbon intensity of 0.4 kg CO2e/kWh, operational emissions are $100\times 0.4=40$ kg CO2e. Manufacturing is additional. Grids differ greatly, so identical energy use can cause different emissions.

Measure a data center's overhead

A data center holds servers, storage, and network equipment, plus cooling and power systems. Power usage effectiveness (PUE) is

If total energy is 1,500 MWh and IT uses 1,000 MWh, PUE is 1.5. Overhead is 500 MWh, one third of the total. The ideal ratio is 1.0; efficient large facilities report about 1.1. PUE shows overhead, not whether computation is useful, algorithms are efficient, or electricity is clean.

IEA's 2025 report estimated data-center electricity at 415 TWh in 2024, about 1.5% of global electricity, and projected about 945 TWh by 2030 in its base case, with AI the main driver. Ireland's data centers used 21% of metered electricity in 2023. A modest global share can still create major local pressure.

Cooling can consume water directly, while electricity generation uses water indirectly. Water scarcity changes the importance of the same volume. Cooling choices can trade water savings against electricity use; air cooling often uses more electricity while avoiding evaporative water consumption.

Count demand as well as efficiency

AI training is energy intensive. Repeated model use, called inference, can exceed training energy in total once a service handles many requests. Proof-of-work cryptocurrencies such as Bitcoin deliberately require extensive computation. Ethereum's 2022 move to proof of stake reduced its energy use by more than 99%.

The rebound effect occurs when cheaper, more efficient tasks encourage extra use that offsets savings. If extra demand more than cancels savings, total energy rises: backfire, or the Jevons paradox.

Computing can also reduce harm through climate models, smart grids, precise irrigation, route planning, and video calls replacing travel. Compare the whole system with a realistic alternative, including long-term changes in demand.

E-waste is discarded electrical and electronic equipment, from phones and servers to cables and appliances. The Global E-waste Monitor 2024 reported 62 million tonnes in 2022, about 7.8 kg per person and 82% above 2010. Only 22.3% was documented as properly collected and recycled; the report projected 82 million tonnes by 2030.

Why disposal causes harm

Electronics contain useful materials such as copper, gold, cobalt, and rare earth elements, alongside hazardous substances such as lead, mercury, cadmium, and some flame retardants. Informal cable burning and acid extraction expose workers and nearby communities to toxic pollution. Some used electronics travel to lower-income countries through illegal waste exports disguised as secondhand goods.

Early replacement increases that burden. Glued batteries, costly repairs, unavailable parts, special screws, and marketing can shorten use. Ended security support can make otherwise functioning hardware unsuitable for connected work. The right-to-repair movement seeks parts, manuals, and repairable designs; some governments have enacted related requirements.

Compare longer use with replacement

Prefer reducing purchases, reusing, repairing, and finally certified recycling. Recycling recovers only part of the material and itself requires energy.

Take example manufacturing emissions of 300 kg CO2e and annual operation of 40 kg CO2e. Over three years:

Average emissions are $420\mathrm{/}3=140$ kg per year. Over six years:

The annual average becomes $540\mathrm{/}6=90$ kg. The longer-lived device has a larger total because it provides more years of service, but a lower annual footprint, about 36% less. Compare equivalent service periods and include the manufacturing cost of any replacement.

Before transfer or recycling, back up and securely erase data. A factory reset of an encrypted device or certified business data destruction can prevent exposure. Send batteries to appropriate collection points; damaged lithium-ion batteries can start fires in household waste handling.

Improve the service as well as the device

Efficient algorithms and code reduce processor work. Avoid storing and transferring data without a purpose. Choose an AI model suited to the task instead of automatically using the largest. Carbon-aware computing schedules flexible work for times and locations with cleaner electricity. Low-carbon hosting and software that still works on older devices can reduce pressure to replace hardware. Measure results so a claimed improvement can be checked.

Greenwashing presents environmental performance more favorably than the evidence supports. Ask which emissions, stages, and activities a claim includes. A "carbon neutral" claim can rely on offsets while actual emissions remain high. Renewable-energy claims may cover only part of an organization's operations. ACM's public-good commitment includes environmental sustainability, linking procurement and software decisions to professional responsibility.

A code of ethics records a profession's values and responsibilities. It guides developers, designers, data scientists, and IT staff when a law or job instruction leaves a choice unresolved. A code is not legislation. Professional bodies enforce membership rules, and clients and employers can expect professional standards beyond legal minimums.

Connect the ACM principles to action

The ACM Code of Ethics and Professional Conduct was updated in 2018. It has four sections: general ethical principles, professional responsibilities, leadership principles, and compliance. The public good is its central concern.

Professional responsibilities include quality in processes and products (2.1), maintaining competence and conduct (2.2), respecting applicable rules (2.3), thorough impact and risk evaluation (2.5), working within competence (2.6), authorized access with a narrowly framed public-good exception (2.8), and robust, usable security (2.9). Leadership places the public good first (3.1) and gives infrastructure systems special care (3.7).

Respond to a risky release

A developer is told to ship a health questionnaire stored without encryption and accessible to the whole company. Sensitive answers could expose users to serious harm. Privacy, harm prevention, risk evaluation, and usable security are directly relevant.

The developer should describe the risk in writing, propose encryption and restricted access with a concrete time estimate, and ask for a decision. If overruled, escalate the serious concern through appropriate channels and keep records. A manager's instruction does not remove the developer's responsibility.

Resolve conflicts proportionately

Confidentiality and loyalty can conflict with public safety. Whistleblowing reports wrongdoing to someone able to act. Establish facts and preserve appropriate records; raise the issue through a manager or ethics channel; seek trusted professional or legal advice; and consider reporting to a regulator when serious harm persists and internal channels have failed or are unsafe. Legal protection differs by country.

Other codes share these duties. IEEE emphasizes public safety, health, welfare, privacy, ethical design, and sustainability. The ACM/IEEE-CS Software Engineering Code has eight principles: Public, Client and Employer, Product, Judgment, Management, Profession, Colleagues, and Self. National societies have related codes. Principles guide judgment; they do not provide a mechanical answer to every conflict.

An ethical decision needs facts, reasons, and an action someone can carry out. Begin by separating what is known from what needs checking. A technical failure can involve designers, operators, managers, and regulators; shared responsibility still leaves each person responsible for their own contribution.

Use a repeatable analysis

1. Establish the facts and uncertainties.
2. Identify stakeholders and their interests.
3. Name the values and conflicts.
4. List feasible options, including designs that reduce the conflict.
5. Evaluate options with ethical frameworks, code principles, and applicable law.
6. Choose and justify a response, stating trade-offs and what evidence would change it.
7. Act, assign responsibility, and review the results.

The publicity test asks whether you could defend the decision publicly. Reversibility asks whether you could accept being the affected person. The harm test asks whether serious harm could be avoided. These checks support the analysis; feeling comfortable is not itself evidence that a decision is sound.

Learn from failures

Six known Therac-25 accidents between June 1985 and January 1987 caused massive radiation overdoses, deaths, and serious injuries. Fast operator edits could trigger race-condition faults, where timing changes how software operations interact. The error "MALFUNCTION 54" did not explain the danger.

Volkswagen's affected cars could emit nitrogen oxides at up to 40 times US limits on the road. About 11 million vehicles worldwide were affected. Engineers and managers faced prosecution, including a US prison sentence for an engineer.

Lion Air Flight 610 crashed in October 2018 and Ethiopian Airlines Flight 302 in March 2019; 346 people died. MCAS had initially been omitted from pilots' manuals. Boeing issued additional information after the first crash, so the information available changed between the accidents.

The problem of many hands occurs when responsibility is spread across a system and each participant assumes someone else has handled the danger. Commercial pressure and over-trust in automation can strengthen this failure.

Apply the method to a data sale

A fitness app proposes selling detailed location histories. A long accepted policy does not establish meaningful consent. Homes, workplaces, and clinic visits expose users and families; deleting names leaves linkable routes. Options include raw sales, name removal, privacy-protected aggregate statistics with opt-in consent, or another revenue source. Compare revenue with foreseeable privacy harm, apply ACM harm and privacy principles, and check the law. Choose a concrete option, record its reasons and safeguards, and review the resulting disclosures and complaints.