Power, Ethics, and Governance

The most advanced AI is built by a small number of companies, depends on chips made by an even smaller number of suppliers, and runs in data centers owned mostly by a handful of tech giants, so control is concentrated at every layer, from the factory to your screen.

The AI supply chain

The layers, in plain terms

• Chips: one U.S. company, Nvidia, designs most of the specialized chips used to train AI. Most of the world's most advanced chips are manufactured by one company in Taiwan, TSMC, using machines that only one Dutch company, ASML, can make. This makes chips a matter of national security.
• Data centers and cloud: enormous warehouses of computers, mostly owned by Amazon, Microsoft, and Google, plus a few others. Big tech companies are spending hundreds of billions of dollars a year building them.
• Frontier models: the most capable AI models come from a handful of labs, mainly OpenAI, Google DeepMind, Anthropic, Meta, and xAI in the U.S., and DeepSeek, Alibaba, and a few others in China (Lesson 3, Section 7).
• Apps: thousands of companies build products on top of those models, often paying the model makers for access.

Who controls the companies?

AI labs have unusual structures, and control matters:

• Big public companies (Google, Meta, Microsoft) answer to shareholders; some founders hold special voting shares that give them outsized control.
• OpenAI began as a nonprofit and has restructured so that a nonprofit foundation oversees a for-profit company, a structure debated in court and by state attorneys general.
• Anthropic is a "public benefit corporation," legally required to weigh a stated public mission alongside profit.
• Investors, including the biggest tech companies, often own large stakes in the AI labs, linking them together.

These structures shape what companies prioritize: speed, safety, profit, openness, or public benefit. Their stated missions are a starting point; watching their actions is the real test.

The open-source question

Some companies release their models openly for anyone to download (Lesson 2, Section 7), spreading control more widely. Supporters say this prevents a few companies from controlling AI and allows independent research. Critics note that open models can't be recalled and their safety features can be removed. It's one of the central power debates in AI.

Because the most advanced AI costs enormous sums and depends on scarce chips, power over it is concentrated in a few companies and, geopolitically, mainly in the United States and China, which raises concerns about competition, democracy, and a world split between AI "haves" and "have-nots."

Why AI tends to concentrate

• Cost: training the largest models costs hundreds of millions to billions of dollars.
• Chips and electricity: both are scarce, and the biggest players buy most of the supply.
• Data and users: companies with billions of users (search, social media, phones, email) have both data and a built-in audience.
• Talent: top AI researchers command very high salaries, concentrated in a few firms.
• Network effects: more users produce more feedback, which improves the product, which attracts more users.

Why it matters

• Competition: fewer rivals can mean higher prices, less choice, and less pressure to protect users. Antitrust regulators in the U.S. and Europe are examining AI partnerships and investments.
• Influence over information: AI increasingly shapes what people read, believe, and decide. A few companies' choices about what AI says can ripple across billions of conversations.
• Political influence: AI companies now spend heavily on lobbying and political campaigns.
• Dependence: businesses, schools, and governments that build on one company's AI can become locked in.

The nations: a U.S.–China race

The United States and China lead AI development by a wide margin. Both governments treat AI as vital to economic strength and national security.

• Chip export controls: since 2022, the U.S. has restricted sales of the most advanced AI chips and chipmaking equipment to China. China has responded by building its own chip industry and producing competitive models, like DeepSeek's, with fewer resources.
• Controls on AI models themselves: in June 2026, U.S. Commerce Department export controls led Anthropic to temporarily suspend access to two of its most advanced models; the controls were lifted at the end of June and access was restored in July. The episode showed how directly governments can now affect access to specific AI systems.
• Data centers as strategy: countries compete to host data centers, and wealthy nations in the Middle East and elsewhere are investing heavily.

The "AI divide"

Most countries don't build frontier AI; they use what others build. Many languages, cultures, and regions are underserved (Lesson 8, Section 3). Some countries are investing in "sovereign AI" (their own models and data centers) to avoid depending on foreign companies. International bodies, including the United Nations, have begun efforts to include more countries in shaping AI's future.

Different parts of the world are taking very different approaches: the European Union has a broad, risk-based AI law; the United States relies on existing laws, state rules, and a federal push toward lighter, national standards; China regulates tightly around content and state priorities; and many other countries are finding their own paths.

The European Union: the AI Act

The EU AI Act, passed in 2024, is the world's most comprehensive AI law. It sorts AI by risk:

In 2026, after pressure from businesses and some governments worried about competitiveness, the EU agreed to postpone the high-risk rules. Supporters of the law see this as a setback; others see it as sensible adjustment.

The United States: a patchwork and a federal push

• No comprehensive federal AI law. Existing laws on discrimination, consumer protection, and fraud still apply to AI (Lessons 7 and 8).
• Federal direction since 2025: the administration revoked the previous administration's AI executive order, issued an "AI Action Plan" focused on innovation, infrastructure, and U.S. leadership, and in December 2025 signed an executive order aiming to challenge state AI laws it considers burdensome, including a Justice Department task force to litigate against them. Congress has debated, but not passed, a national law that would override state rules.
• States acting on their own: despite federal pressure, states have introduced well over a thousand AI-related bills and passed many of them. California requires the largest AI developers to publish safety frameworks and report serious incidents, with whistleblower protections. Other states have laws on deepfakes, companion chatbots, hiring tools, and more.
• A few federal laws target specific harms, such as the TAKE IT DOWN Act on non-consensual intimate images (Lesson 4).

China: tight control, fast rules

China was among the first countries to regulate generative AI, in 2023. Its rules require AI services to register with authorities, pass security reviews, label AI-generated content (strengthened in 2025), and ensure content reflects state-approved values and doesn't undermine state authority. China also invests heavily in AI as a national priority. Its approach combines rapid development with strong government control over information.

Others and international efforts

• United Kingdom: a lighter, sector-by-sector approach, with a government AI Security Institute that tests advanced models.
• South Korea passed a broad AI law that took effect in 2026; Japan passed a law promoting AI development with light oversight.
• International summits (UK 2023, South Korea 2024, France 2025, India 2026) have produced voluntary commitments, though countries increasingly disagree on priorities.
• The Council of Europe adopted the first binding international treaty on AI and human rights in 2024.
• The United Nations created an international scientific panel on AI and a global dialogue on AI governance in 2025.

The core debate

Polls in the U.S. have generally found more Americans worried that government won't go far enough in regulating AI than worried it will go too far. How to balance that with innovation is one of the defining policy questions of this decade.

AI models were trained on vast amounts of books, articles, art, music, and code, much of it copyrighted and used without permission, and courts, lawmakers, and licensing deals are now deciding whether that was legal and whether creators should be paid.

The central legal question: fair use

U.S. copyright law allows some uses of copyrighted work without permission, called fair use, judged on factors like the purpose of the use, how much is used, and whether it harms the market for the original. AI companies argue training is "transformative," like the student learning from books. Creators argue it's copying on an industrial scale to build products that compete with them.

What's happened so far

• Mixed early rulings (2025): in two separate authors' cases, federal judges suggested that training AI on lawfully obtained books could be fair use, but one judge sent a case to trial over the use of pirated books.
• A record settlement: in September 2025, Anthropic agreed to pay authors $1.5 billion to settle claims over pirated books used in training, among the largest copyright settlements in U.S. history.
• Pending giants: The New York Times' lawsuit against OpenAI and Microsoft, filed in 2023, is among many cases still unresolved.
• Licensing deals: many AI companies now pay news publishers, social media sites, music labels, and image libraries for access to their content.
• Output ownership: as Lesson 10 noted, U.S. law generally doesn't grant copyright to purely AI-generated works.

Creators' concerns beyond the courtroom

• Consent: most creators were never asked.
• Credit: AI outputs rarely name the works they drew on.
• Compensation: AI products earn money partly from creators' work.
• Competition: AI can produce work "in the style of" living artists, competing with them directly.
• Voice and likeness: new state laws protect performers from unauthorized digital replicas.

AI runs in data centers that use large and fast-growing amounts of electricity and, in many places, water for cooling, raising concerns about climate, local utility bills, water supplies, and who bears the costs, while AI may also help solve some environmental problems.

How much energy?

Sources: IEA, "Energy and AI"; Our World in Data. Estimates vary between researchers.

What about my chatbot question?

A single typical text question to a chatbot uses a small amount of electricity: recent estimates put it around a third of a watt-hour, similar to running an LED bulb for a couple of minutes. Very long requests, image and video generation, and "thinking" agents use much more. The big totals come from training huge models and from billions of daily requests worldwide.

Water

Many data centers use water to keep computers cool, much of it evaporated. A 2024 U.S. federal laboratory report estimated U.S. data centers directly consumed about 17 billion gallons of water in 2023, with use expected to rise sharply. Power plants that supply the electricity use more water still. Impact depends heavily on location: a data center in a drought-stricken region matters far more than one in a water-rich area. Some newer designs use far less water.

Local impacts

• Electricity bills: utilities are building new power plants and power lines for data centers, and debates are growing over whether households end up paying part of that cost.
• Fossil fuels: in some places, rising demand has delayed closing coal or gas plants, or prompted new gas plants.
• Noise, land, and water: communities in several states have pushed back against new data centers; others welcome the tax revenue and jobs.

The other side of the ledger

• AI companies are among the largest buyers of wind, solar, and new nuclear power, and some are funding new reactor projects.
• AI is helping improve weather and climate forecasting, manage power grids, discover new materials for batteries and solar panels, and cut energy waste in buildings and industry.
• Models and chips are becoming more efficient, doing more with less energy per task, though total use is still rising.

Militaries worldwide are adopting AI for intelligence analysis, logistics, cyber operations, and increasingly for weapons that can select and attack targets with less human involvement, raising profound questions about human control over life-and-death decisions.

How militaries use AI

• Intelligence: sorting satellite images, intercepted communications, and video far faster than people can.
• Logistics and maintenance: predicting equipment failures and planning supplies.
• Cyber operations: both defending networks and attacking them.
• Decision support: suggesting targets and courses of action to commanders.
• Autonomous systems: drones, vessels, and defenses that can navigate, identify, and in some cases attack with little human input.

What's happening now

• In the war in Ukraine, both sides use huge numbers of drones, increasingly with AI that lets them complete attacks even when radio links are jammed.
• Investigative reports in 2024 described AI systems used by Israel's military to generate large numbers of potential targets during the war in Gaza; the Israeli military disputed aspects of those reports and said humans made the final decisions. The reports intensified global debate about AI and civilian harm.
• The U.S. military has policies requiring "appropriate levels of human judgment" over the use of force, and in 2025 awarded contracts to several leading AI companies to develop AI for national security.
• In 2024, the U.S. and Chinese presidents agreed that humans, not AI, should control decisions to use nuclear weapons.

The push for international rules

For more than a decade, countries have discussed "lethal autonomous weapons systems" at the United Nations. In September 2026, three years of expert talks ended without a treaty, though the final report affirmed that human control and judgment are required to comply with international law. A growing number of countries, about 76, now support formal treaty negotiations; major military powers, including the United States and Russia, have resisted binding limits. The next key decision point is a UN review conference in November 2026.

When AI causes harm, responsibility can fall on the company that built the model, the business that deployed it, or the person who used it, and the law is still sorting out who owes what, which is why transparency about how AI systems work and fail is so important.

Who might be responsible?

A case that set a tone

In 2024, a Canadian tribunal ruled against Air Canada after its website chatbot gave a grieving customer wrong information about bereavement fares. The airline argued the chatbot was responsible for its own words. The tribunal disagreed: the company is responsible for information its chatbot provides. It's a small case with a big principle: businesses can't blame their AI.

Why accountability is hard

• The black box: even builders can't fully explain a model's decisions (Lesson 2, Section 4).
• Many hands: a harm may involve a model maker, an app builder, and a user, each pointing at the others.
• Invisible decisions: people often don't know AI was involved in denying them a job, loan, or benefit.
• Old laws, new facts: courts are still deciding how existing product liability, discrimination, and free-speech laws apply.
• Speed: AI changes faster than lawsuits and regulations move.

Tools for transparency

• Model and system cards: documents describing how a model was built, tested, and where it fails. Major labs publish these for new models.
• Independent testing: government AI safety and security institutes and outside researchers evaluating models before and after release.
• Incident reporting: public databases tracking AI failures, and new state laws requiring companies to report serious incidents.
• Whistleblower protections: letting employees report safety concerns without retaliation.
• Disclosure to people affected: notice when AI is used in decisions about them, and a right to explanation and appeal.

AI's future isn't only being decided by tech companies and governments; ordinary people shape it through voting, public comments, local hearings, workplaces, purchases, and simply understanding enough to ask good questions, and there are real examples where public pressure changed outcomes.

It has worked before

• After public outcry, the UK scrapped its exam-grading algorithm within days in 2020 (Lesson 8).
• Research and advocacy led face-recognition companies to fix accuracy gaps, and several cities to limit police use (Lessons 7 and 8).
• Hollywood writers and actors won AI protections in their contracts in 2023 (Lesson 10).
• Parents and advocates helped drive new state laws protecting children from harmful companion chatbots (Lesson 10).
• Community opposition has reshaped or stopped data center proposals in several places (Section 5).

Ways to have a say

Why AI literacy is itself a form of power

Those who understand a technology can question it, use it wisely, and demand better. When only experts and companies understand AI, only they shape it. Every person who learns the basics, as you've done across these lessons, widens the circle of people who can take part in deciding what kind of future AI helps build.