AI and society

AI is not only a technical subject. Systems that screen job applications, approve loans, recommend medical treatments, decide what news people see, generate realistic fake images or drive cars change people's lives, for better and for worse. As AI becomes more capable and more widely used, questions about fairness, privacy, truth, work, safety and power become engineering questions too, because they depend on how systems are designed, tested and deployed.

This chapter surveys the main benefits and risks of AI, shows how some of them can be measured, describes how governments are responding, and closes with what responsible practice means for the people who build and use AI. The later courses AI-430 (safety and evaluation), AI-440 (law, policy and governance) and AI-460 (fairness and privacy) go much deeper.

The promise

The benefits are real and growing. AI helps doctors read medical images and spot disease earlier, speeds up drug discovery and protein design, forecasts weather more accurately, translates between languages, captions speech for people who are deaf or hard of hearing, describes images for people who are blind, and helps programmers, writers and analysts work faster. In engineering it optimizes energy systems, predicts equipment failures and helps design structures and circuits. Many economists expect AI to raise productivity significantly, though estimates of how much and how soon vary widely.

The same capabilities create risks. A system good at writing persuasive text can write misinformation; a system good at recognizing faces can enable mass surveillance; a system that automates tasks can displace workers. Most of the difficult questions are about how AI is used, by whom, and with what safeguards.

Fairness and bias

Machine learning learns patterns from data, and data reflect the world that produced them, including its inequalities. Bias can enter at many points:

• Historical bias: if past hiring favored one group, a model trained on past hiring decisions learns to favor it too.
• Representation bias: if a face-recognition dataset contains few darker-skinned faces, the model will be less accurate on them. Studies around 2018 found commercial face-analysis systems with error rates far higher for darker-skinned women than for lighter-skinned men.
• Measurement bias: if the label is a poor proxy for what we care about (for example, using past health care spending as a proxy for health need, when some groups have had less access to care), the model optimizes the wrong thing.
• Deployment bias: a model used in a setting different from the one it was built for can fail unevenly.

Bias can be measured by comparing a model's errors across groups.

There are several mathematical definitions of fairness: equal approval rates (demographic parity), equal true positive rates (equal opportunity), equal error rates in both directions (equalized odds), or equal precision. A well-known result shows that when groups have different base rates, a model generally cannot satisfy all of these at once. Choosing which matters most is a value judgment that should involve the people affected, not only the engineers.

Privacy and surveillance

AI is hungry for data, and much of the most useful data is about people. Risks include collecting personal data without meaningful consent, using it for purposes people never agreed to, re-identifying individuals from supposedly anonymous data, and inferring sensitive facts (health, sexuality, politics) from innocent-looking data. Face recognition makes it possible to identify people in public spaces at scale.

The United States has no general federal privacy law. Personal data is protected by sector-specific federal laws (such as HIPAA for health information), by the Federal Trade Commission's power over unfair and deceptive practices, and by a growing number of state laws. Illinois's Biometric Information Privacy Act, which requires consent before collecting biometric data such as face geometry, has produced some of the largest privacy settlements in US history.

Truth, deepfakes and information

Generative AI makes it cheap to produce realistic fake text, images, audio and video. Deepfakes have been used for fraud (for example, cloned voices of executives or relatives asking for money), harassment (non-consensual intimate images), and political disinformation. Detection tools help, but they are imperfect, and their errors interact with how rare fakes are.

Responses include labeling AI-generated content, content provenance standards that cryptographically record how a file was created and edited, platform policies and new laws.

In the United States, the TAKE IT DOWN Act, signed in May 2025, made it a federal crime to publish non-consensual intimate images, including AI-generated ones, and requires platforms to remove them within 48 hours of a valid request. Many states have also passed laws on deepfakes in elections and intimate images.

Work and the economy

AI automates tasks, not whole jobs, but as more tasks become automatable, jobs change. The International Monetary Fund estimated in 2024 that about 40% of employment worldwide, and about 60% in advanced economies, is exposed to AI, meaning a significant share of the tasks involved could be affected. Exposure does not mean job loss: in many roles AI complements workers, making them more productive, while in others it may substitute for them. Unlike earlier waves of automation that mainly affected routine manual work, AI particularly affects knowledge work: writing, coding, analysis, customer service and design.

How the gains and losses are shared depends on choices by companies and governments: retraining, education, social supports and competition policy. For individuals, the most durable skills are those that combine domain expertise with judgment, responsibility and the ability to use AI tools well.

Safety, misuse and advanced AI

As models become more capable, three kinds of risk receive attention:

• Misuse: people using AI to cause harm, for example to write malware, plan cyberattacks, run large-scale scams, or obtain knowledge that helps build biological or chemical weapons. Developers test models for such capabilities and build safeguards; governments increasingly expect them to.
• Malfunction: systems that fail in high-stakes settings, such as a medical, financial or infrastructure system making a serious error, or an agent taking an unintended harmful action.
• Loss of control: the concern, held by some leading researchers, that future AI systems much more capable than people could pursue goals that conflict with human interests in ways that are hard to correct. Others consider this unlikely or distant. The question is actively debated and studied in AI safety and alignment research (AI-430).

In January 2025, the first International AI Safety Report, written by about 100 experts from around the world and chaired by Yoshua Bengio, summarized the scientific evidence on these risks and stressed how uncertain many of them remain. Several governments have created AI safety institutes to evaluate advanced models.

In the United States, the AI Safety Institute at the National Institute of Standards and Technology (NIST) was renamed the Center for AI Standards and Innovation in 2025, with a focus on evaluating capabilities and national security risks and on developing standards.

Environmental impact

Training and running AI uses large amounts of electricity, and some data centers use substantial water for cooling. The International Energy Agency projects that global data center electricity use will more than double between 2024 and 2030, driven largely by AI. The impact depends heavily on where the electricity comes from and how efficiently it is used. At the same time, AI can help reduce emissions elsewhere, by optimizing grids, buildings and industrial processes. Engineers can measure and reduce AI's footprint with efficient models, efficient hardware and data centers (DCE-200, DCE-460), and by choosing the smallest model that does the job.

How governments are responding

Governments are trying to capture AI's benefits while managing its risks, and they have taken different approaches. Internationally, the OECD AI Principles (2019, updated 2024) set out shared values such as transparency, robustness and accountability. The European Union's AI Act, in force since August 2024, is the most comprehensive law so far: it bans some practices outright (such as social scoring by governments) from February 2025, sets obligations for general-purpose AI models from August 2025, and imposes strict requirements on high-risk systems in areas such as employment, credit and critical infrastructure. An amendment that entered into force in July 2026 postponed the high-risk requirements to December 2027 and August 2028. The EU AI Act matters outside Europe because it applies to anyone offering AI systems in the EU market.

In the United States

The United States has no comprehensive federal AI law. Federal policy has been set mainly by executive action and by applying existing laws, and it has changed with each administration:

• In January 2025, Executive Order 14179 revoked the previous administration's 2023 order on AI safety and directed agencies to remove barriers to American AI leadership. America's AI Action Plan followed in July 2025.
• Executive Order 14365 (December 2025), "Ensuring a National Policy Framework for Artificial Intelligence", set out to limit what the administration considers burdensome state AI laws; an AI Litigation Task Force was created in January 2026, and in March 2026 the White House sent Congress non-binding recommendations for a national framework that would pre-empt some state laws.
• Existing laws still apply to AI: the Federal Trade Commission acts against deceptive AI claims and unfair practices, and anti-discrimination laws in employment, housing and credit apply whether or not a decision is automated. NIST's AI Risk Management Framework (2023) is a widely used voluntary standard.
• States have been active. California's SB 53 (2025) requires developers of frontier models (trained with more than 10²⁶ FLOP) to publish safety frameworks and report critical safety incidents. Colorado replaced its 2024 AI Act in May 2026 with a narrower law on automated decision-making in areas such as employment, housing, lending and health care, expected to take effect in 2027. Many states have laws on deepfakes, chatbots and government use of AI.

AI law is changing quickly. Anyone deploying AI in a regulated area must check the current rules, and AI-440 tracks them in detail.

Responsible AI in practice

Laws set minimums. Engineers and organizations that build and deploy AI can do much more, and professional engineers have an existing duty to put public safety and welfare first. Common responsible-AI practices include:

• Define the purpose, the people affected and what could go wrong before building, and decide whether AI is the right tool at all.
• Document data sources, consent and known limitations; publish model documentation ("model cards").
• Test for performance and fairness across groups, for robustness and for misuse, before and after deployment.
• Keep meaningful human oversight for high-stakes decisions, with a way for people to contest them.
• Protect privacy by collecting only what is needed and securing it.
• Monitor in operation, report incidents and fix problems.
• Be honest about what a system can and cannot do.