AI倫理

AI ethics asks a practical question: when a machine helps make a decision about people, how should it be built, tested and governed so that it is fair, safe and accountable? The topic covers several recurring concerns. Fairness asks whether a system works equally well for different groups. Privacy asks what personal data is collected and how it is used. Transparency and explainability ask whether people can find out that AI is involved and why it reached a result. Accountability asks who is responsible when it goes wrong. Safety asks whether the system can cause physical, financial or psychological harm. Wider questions include effects on jobs, the environment and the concentration of power. AI ethics is not only philosophy; it shows up as design decisions, test results, documentation, audits and laws.

In practice, ethical review follows a repeatable process. First, define the system's purpose and the people it affects, including people who never use it directly, such as the subjects of a risk score. Second, map what could go wrong in that specific context. Third, measure: test accuracy not just overall but separately for each relevant group, and test for misuse. Fourth, mitigate: change the data, the model, the decision threshold or the human process around it, or decide not to deploy. Fifth, assign accountability and keep monitoring after launch. The NIST AI Risk Management Framework 1.0, released by the US National Institute of Standards and Technology in January 2023 for voluntary use, names these activities govern, map, measure and manage. It lists the traits of trustworthy AI as valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed, and it notes that these must be balanced against one another depending on context.

The case for measuring by group comes from documented failures. In 2018 Joy Buolamwini and Timnit Gebru published Gender Shades, which tested three commercial systems that guessed gender from face photos. The worst group, darker-skinned women, saw error rates up to 34.7%, while the best group, lighter-skinned men, saw a maximum error of 0.8%. Part of the explanation was the data used to build and judge such systems: two widely used face benchmarks they examined were 79.6% and 86.2% lighter-skinned. A system can report high average accuracy while failing badly for a group that was underrepresented in its data.

The second landmark case shows that fairness itself has more than one definition. In May 2016, ProPublica analyzed COMPAS, a commercial tool from Northpointe that scores how likely a defendant is to reoffend, using records for more than 7,000 people arrested in Broward County, Florida in 2013 and 2014. About 61% of people labeled likely to reoffend were rearrested within two years, and only about 20% of those predicted to commit violent crimes did so. Among defendants who did not reoffend, Black defendants had been labeled higher risk 44.9% of the time and white defendants 23.5%. Northpointe disputed the analysis, arguing its scores were equally accurate for both groups. Months later, Jon Kleinberg, Sendhil Mullainathan and Manish Raghavan proved that, except in highly constrained special cases, no risk score can satisfy three common fairness conditions at once. Both sides could be right by their own definition. Choosing which kind of fairness matters is a value judgment that the math cannot make.

Law has begun to turn some ethical lines into rules. The European Union's AI Act takes a risk-based approach: a short list of uses is banned, high-risk uses face requirements for data quality, documentation and human oversight, and many low-risk uses face few or none. Its prohibitions, applying since 2 February 2025, cover manipulative techniques that cause significant harm, exploiting people's age, disability or economic situation, social scoring, predicting crime from profiling alone, building face recognition databases by untargeted scraping, inferring emotions in workplaces and schools, and most real-time remote biometric identification in public spaces by police. The NIST framework, by contrast, is voluntary guidance. The Stanford AI Index 2025 counted 59 AI-related regulations issued by US federal agencies in 2024, more than double the year before.

AI ethics is easy to confuse with neighboring ideas. Law sets minimum requirements that are enforced; ethics also covers what is permitted but still harmful or unfair. AI safety, in its narrow sense, focuses on preventing accidents and dangerous capabilities, often in powerful future systems, while ethics usually deals with harms to people from systems already in use. Responsible AI is the name many companies give to their internal ethics programs, which typically combine principles, review boards and testing. All three overlap, and the same audit can serve each.

The broader costs reach beyond individual decisions. The International Energy Agency's April 2025 report estimates that data centres consumed about 415 terawatt-hours in 2024, about 1.5% of global electricity, and projects consumption more than doubling to around 945 terawatt-hours by 2030, slightly more than Japan uses today, with AI the most important driver. Concentration is another concern: the Stanford AI Index 2025 found that nearly 90% of notable AI models released in 2024 came from industry, so a small number of companies make many of the design choices that affect everyone. Public opinion is divided as well: in the same report, optimism about AI ranged from 83% of people surveyed in China to 39% in the United States.

Doing ethics well has real costs and real limits. Measuring disparities requires knowing which group each person belongs to, which can conflict with privacy. Audits, documentation and human review take staff time. Principles written in general terms, such as fairness or transparency, can be adopted without changing any product, a pattern critics call ethics washing. Trade-offs are unavoidable: making a model more explainable can make it less accurate, and equalizing one error rate can unbalance another. The most useful habit is specific: name the decision, the affected people, the measured error rates for each group, who can appeal, and who is accountable.