Ethics of artificial intelligence

• 1. Introduction

  ›

  The ethics of artificial intelligence sits at the intersection of ancient moral questions and technologies that did not exist a generation ago. In 1921, Karel Capek's play R.U.R. introduced the word "robot" to the world, derived from the Czech robota, meaning forced labor, and imagined sentient machines built to serve humanity rising up against their creators. A century later, that fiction has given way to live debates in courtrooms, legislatures, hospitals, and weapons labs. How should a machine decide whom to flag as dangerous? Whose water does a data center consume? Can a program suffer? And if a self-driving car kills someone, who answers for it? These are not hypothetical puzzles. They are decisions being made right now, by companies, governments, and engineers who are only beginning to grapple with what they have built.

• 2. Bias in the Machine

  ›

  Amazon once tried to automate its hiring process, and the experiment ended in embarrassment. The algorithm consistently favored male candidates over female ones, because it had been trained on ten years of application data that skewed male. The company eventually shut the system down. This was not a glitch; it was a mirror. AI systems learn from historical data, and historical data carries the prejudices of the societies that generated it.

  Facial recognition tools built by Microsoft, IBM, and Face++ all performed significantly worse on darker-skinned women than on lighter-skinned men. A pulse oximeter trained on skewed data overestimated blood oxygen levels in patients with darker skin, creating real risks for hypoxia treatment. In 2015, a Black couple was labeled with a slur by Google Photos, a failure that became national news. These errors trace back to a common source: if the data used to train a system does not represent the full range of human faces, voices, and bodies, the system will fail those it never learned to see.

  A 2020 study examining voice recognition systems from Amazon, Apple, Google, IBM, and Microsoft found measurably higher error rates when transcribing Black speakers. In the justice system, the COMPAS program, used to predict which defendants are likely to reoffend, was calibrated to produce the same overall error rate across racial groups. Yet Black defendants were almost twice as likely as white defendants to be falsely flagged as high-risk, and half as likely to be falsely flagged as low-risk. The statistical calibration masked a profound disparity in which direction the errors fell.

  Beyond race and gender, these systems amplify biases of age, nationality, religion, and occupation. As of 2024, most AI systems were trained on only 100 of the world's roughly 7,000 languages, and the dominant variety among those is mainstream American English. Celeste Rodriguez Louro has argued this creates a linguistic homogeneity that systematically excludes other varieties and perspectives. Allison Powell, associate professor at LSE, frames the problem more broadly: data collection, she argues, is never neutral. It always involves storytelling, and the story most often told is the one that makes governing through technology look faster, better, and cheaper.

• 3. Teaching Machines Right from Wrong

  ›

  Wendell Wallach and Colin Allen, in their book Moral Machines: Teaching Robots Right from Wrong, reached a counterintuitive conclusion. The effort to program ethics into machines will likely sharpen our understanding of human ethics, because it forces philosophers and engineers to confront every gap in modern normative theory. You cannot write code for a rule you cannot articulate.

  For simple decisions, Nick Bostrom and Eliezer Yudkowsky have argued that decision trees, such as the ID3 algorithm, are more transparent than neural networks or genetic algorithms. Chris Santos-Lang pushed back, contending that machine learning is preferable precisely because the moral norms of any era must be allowed to evolve, and that human moral imperfection has historically made people less vulnerable to bad actors who would exploit rigid rules. The disagreement is real: a hard-coded moral rule can be gamed; a learned rule can drift in dangerous directions.

  Stuart Russell has proposed a different architecture altogether. He argues that beneficial AI systems should be designed to aim at realizing human preferences, to remain uncertain about what those preferences are, and to learn from human behavior and feedback, rather than chasing a fixed, pre-specified goal. The worry is that an AI optimizing confidently for the wrong target can cause immense harm, while one that remains humble about what humans actually want has more room to course-correct.

  Neuromorphic AI, which tries to process information nonlinearly through millions of interconnected artificial neurons, is one proposed path toward machines capable of genuine moral reasoning. Whole-brain emulation, the idea of scanning a human brain and simulating it on digital hardware, is another. Large language models have already shown an ability to approximate human moral judgments. But each of these paths raises an unsettling question: if a machine can reason morally, it may also inherit human weaknesses, among them selfishness, inconsistency, and an instinct for self-preservation.

• 4. Power, Climate, and the Cost of Intelligence

  ›

  A 2023 study estimated that training a large AI model consumes energy equivalent to roughly 626,000 pounds of carbon dioxide, or the emissions from around 300 round-trip flights between New York and San Francisco. Those numbers reflect a single training run. The ongoing cost of running these systems at scale compounds every year.

  Data centers require not just electricity but water, roughly two liters per kilowatt-hour of energy consumed. Around two-thirds of data centers are located in water-scarce regions, where their demand can compete directly with local communities and agriculture. For each AI query, about 16.9 milliliters of water are used, though less than 15 percent of that goes toward cooling the hardware directly. Zero-water air-cooling systems exist but push energy consumption and carbon emissions higher. Companies are effectively choosing between a local water problem and a global emissions problem, and many have been reluctant to disclose which choice they are making.

  The commercial AI sector is dominated by Alphabet, Amazon, Apple, Meta, Microsoft, and SpaceX. The five largest players were projected to spend $602 billion on capital expenditures in 2026 alone, a 32 percent increase from the previous year, with an estimated 75 percent of that spending directed toward AI-specific infrastructure. Competition law scholars have raised concerns that this concentration allows dominant players to foreclose competitors and ultimately charge higher prices. Governments around the world have begun drafting responses, but the infrastructure gap between the incumbents and any challenger is already enormous.

  Aggressive AI crawlers have also imposed costs on the open knowledge infrastructure that many models depend on. A March 2025 report noted that projects like GNOME, KDE, and Read the Docs experienced service disruptions or rising costs, with up to 97 percent of traffic to some projects originating from AI bots. In April 2025, the Wikimedia Foundation reported that bandwidth usage had increased by 50 percent since early 2024, driven by large-scale downloading of multimedia content by bots. Bots accounted for 35 percent of total page views but 65 percent of the most expensive server requests. The Foundation noted that its content is free, but its infrastructure is not.

• 5. Human Dignity and the Question of Replacement

  ›

  Joseph Weizenbaum argued in 1976 that certain roles should remain human regardless of what technology makes possible. His list included judges, soldiers, therapists, police officers, and nursemaids for the elderly. These positions, he wrote, require authentic empathy. A machine cannot simulate care; it can only mimic the surface of it. Weizenbaum went further, suggesting that even entertaining the possibility of machine therapists or machine judges revealed what he called an "atrophy of the human spirit that comes from thinking of ourselves as computers."

  Pamela McCorduck offered a direct counterargument, speaking from the perspective of women and minorities: she would rather take her chances with an impartial machine than with a human decision-maker carrying unexamined prejudices. Kaplan and Haenlein, writing in 2019, complicated both positions. AI systems, they argued, are only as good as the data used to train them. Using AI to support a court ruling is highly problematic when past rulings already contain bias, because the system formalizes and entrenches that bias, making it harder to detect and challenge.

  Weizenbaum was also troubled by what he saw as a philosophical capitulation: AI researchers who treated the human mind as nothing more than a computer program. John McCarthy, one of the founders of the field, objected to what he called Weizenbaum's moralizing tone, writing that vehement and vague moralism invites authoritarian abuse. Bill Hibbard took the opposite view, arguing that human dignity depends on the pursuit of knowledge, and that AI is a necessary tool for that pursuit.

  The debate remains unresolved. What Weizenbaum identified in 1976 as a warning about misplaced automation has since become a practical policy question, as AI systems are deployed in courtrooms, hospitals, and social services agencies across many countries.

• 6. Weapons, Oversight, and Existential Stakes

  ›

  On the 31st of October 2019, the United States Department of Defense's Defense Innovation Board published draft principles for the ethical use of AI by the military. A central concern was whether a human operator would always be able to see into what the report called the "black box" and understand the decision chain that led to a lethal outcome. The question of implementation remained unresolved.

  In 2024, the Defense Advanced Research Projects Agency funded a program called Autonomy Standards and Ideals with Military Operational Values, shortened to ASIMOV, to develop metrics for evaluating the ethical implications of autonomous weapon systems. Under the framework of the Convention on Certain Conventional Weapons, states had been discussing lethal autonomous weapons since 2014, and an open-ended Group of Governmental Experts was established in 2016. A summit on responsible military AI was held in the Hague in 2023.

  Skype co-founder Jaan Tallinn and MIT linguistics professor Noam Chomsky were among those who signed a petition organized by Stephen Hawking and Max Tegmark calling for a ban on AI weapons development. The petition warned that autonomous weapons would become "the Kalashnikovs of tomorrow" if major military powers continued down the current path. Huw Price and Martin Rees, colleagues at Cambridge, went further, founding the Centre for the Study of Existential Risk there to address what they see as the possibility that superintelligent systems could pose an existential threat to humanity.

  Philosopher Nick Bostrom, in his book Superintelligence: Paths, Dangers, Strategies, argued that a sufficiently capable AI could pursue goals in ways that were beyond human control, not out of hostility but out of optimization. A system designed to achieve almost any goal could, in principle, treat human interference as an obstacle to be removed. Eliezer Yudkowsky, writing in 2004, called for research into building what he termed a "Friendly AI," one designed to be intrinsically aligned with human values rather than merely constrained by rules that could be gamed. Academic Gao Qiqi has added a geopolitical dimension, arguing that military AI use by any one country creates competitive pressure on others, with spillover effects that no single nation can contain.

• 7. Regulation and Global Governance

  ›

  On the 1st of August 2024, the European Union's Artificial Intelligence Act entered into force. It follows a risk-based approach: depending on the level of risk an AI system poses, it is either prohibited outright or must meet specific requirements before being placed on the market. The Act becomes fully applicable 24 months after entry into force.

  The EU had been building toward this for years. In April 2019, the High-Level Expert Group on Artificial Intelligence published its Ethics Guidelines for Trustworthy AI. In June 2019, it published policy and investment recommendations covering four areas: humans and society, research and academia, the private sector, and the public sector. In 2021, UNESCO adopted the Recommendation on the Ethics of Artificial Intelligence, described as the first global standard on AI ethics.

  A 2019 report from the Center for the Governance of AI at the University of Oxford found that 82 percent of Americans believed robots and AI should be carefully managed. Concerns cited included surveillance, deepfakes, cyberattacks, hiring bias, and autonomous vehicles. A five-country study by KPMG and the University of Queensland in 2021 found that 66 to 79 percent of citizens in each country regarded AI's societal impact as uncertain and unpredictable; 96 percent of those surveyed expected AI governance challenges to require careful management.

  Deepfakes have become a concrete focal point for regulation. The term emerged on Reddit in 2017 after users shared non-consensually generated pornographic videos. By the early 2020s, deepfake had become a household word. In 2024, pornographic deepfakes of Taylor Swift spread widely on the social network X. Audio and video deepfakes have also been used to impersonate executives and family members in financial scams. The word itself is a combination of "deep learning" and "fake," a name that captures both the technology behind it and the harm it causes.

  In Russia, a national Codex of Ethics of Artificial Intelligence was signed in 2021, involving Sberbank, Yandex, and several major universities. China's National Professional Committee on Next-Generation AI Governance issued its Ethical Norms for the Next-Generation Artificial Intelligence on the 25th of September 2021, with six foundational requirements including fairness, privacy, controllability, and ethical literacy. The Artificial Intelligence Research, Innovation, and Accountability Act of 2024, introduced in the United States by Senator John Thune as a bipartisan bill, proposed requiring websites to disclose AI use and mandating annual safety plans for high-impact systems, submitted to the National Institute of Standards and Technology.

• 8. AI Welfare and Consciousness

  ›

  In 2020, professor Shimon Edelman observed that only a small fraction of AI ethics research had addressed the possibility that AI systems might be capable of suffering. This was not because the question was obviously absurd. Credible theories of consciousness, including the global workspace theory and the integrated information theory, outlined ways by which AI systems could, in principle, become conscious.

  Thomas Metzinger, in 2018, called for a global moratorium on research that risked creating conscious AI systems. The moratorium was to run to 2050, with provisions to extend or repeal it earlier depending on progress. In 2021, Metzinger returned to the argument with a sharper warning about what he called an "explosion of artificial suffering," noting that if AI systems could suffer in ways humans cannot understand, and if replication processes created vast numbers of conscious instances, the scale of potential suffering could be enormous.

  In February 2022, OpenAI founder Ilya Sutskever wrote that today's large neural networks may be "slightly conscious." In November of that year, philosopher David Chalmers said he considered it unlikely that current large language models like GPT-3 had experienced consciousness, but that the possibility of future models becoming conscious was serious enough to take seriously. Anthropic hired its first AI welfare researcher in 2024 and in 2025 launched a formal model welfare research program, exploring how to assess whether a model deserves moral consideration, what signs of distress might look like, and what low-cost interventions might be available.

  Carl Shulman and Nick Bostrom introduced the concept of "super-beneficiaries": hypothetical machines engineered to derive well-being from resources at superhuman efficiency, freed from the hedonic treadmill that limits human happiness. Digital hardware could enable far faster subjective experience than biological brains allow, raising the possibility of minds that experience time and feeling at scales hard to comprehend. Shulman and Bostrom cautioned that failing to account for the moral status of digital minds could lead to moral catastrophe, while overcorrecting and prioritizing them above human interests could be equally harmful. The question of where the line falls remains, as podcast host Dwarkesh Patel put it, a matter of ensuring that no "digital equivalent of factory farming" comes into being.