AI Ethics

AI ethics is the field that applies moral principles to the design, development, and use of artificial intelligence (AI), asking which systems should be built, how they should behave, and who is accountable for their effects. Its central concerns are fairness, privacy, transparency, safety, accountability, and human oversight, and it draws on philosophy, law, social science, and engineering. In practice, AI ethics shapes how teams collect data, evaluate models, and decide which products to ship, and it underpins regulation such as the EU AI Act and California's AI laws.

Core Principles#

International frameworks converge on a similar set of principles. The OECD AI Principles, adopted in 2019 and updated in 2024, were the first intergovernmental standard on AI; UNESCO's Recommendation on the Ethics of Artificial Intelligence was adopted by all 193 member states in 2021; and the NIST AI Risk Management Framework (2023) turns similar values into characteristics of trustworthy AI that organizations can measure:

• Fairness and non-discrimination: Systems should not perform worse for, or disadvantage, people because of skin tone, gender, age, or other protected traits. Bias in AI usually enters through unrepresentative data or labels, and fairness in AI supplies the metrics that test for it.
• Privacy and data governance: Personal data should be collected lawfully, with consent where required, and processed no more than the purpose needs. Techniques such as differential privacy and federated learning reduce how much a model reveals or centralizes. In video analytics, privacy by design can mean anonymizing people before footage is stored: the Ultralytics object blurring solution runs YOLO26 detection on every frame and blurs each detected object, such as every person, in place.
• Transparency and explainability: People should know when they are dealing with an AI system and be able to learn how it was built, the subject of transparency in AI, and decisions that affect them should be explainable with explainable AI (XAI) methods.
• Safety and robustness: Systems should work reliably under the conditions they meet in deployment, resist manipulation such as adversarial attacks, and fail safely.
• Accountability and human oversight: A named person or organization answers for each system's outcomes, and humans can monitor, override, or shut it down.

From Principles to Law#

The EU AI Act turns these principles into obligations scaled to risk. It entered into force in August 2024; its bans on prohibited practices have applied since February 2, 2025, and its transparency duties since August 2, 2026. After the 2026 Digital Omnibus amendment, the requirements for high-risk uses listed in Annex III, such as hiring, credit scoring, and education, apply from December 2, 2027. The Act sorts each use of AI by the obligations it triggers:

Several prohibitions concern computer vision directly: untargeted scraping of facial images from the internet or CCTV to build recognition databases, emotion recognition in workplaces and schools, biometric categorization that infers traits such as race or political opinions, and, with narrow exceptions, real-time remote biometric identification in public spaces for law enforcement. Organizations can also certify their AI management practices against ISO/IEC 42001, an international standard published in 2023.

AI Ethics vs. Related Terms#

• AI safety: Focuses on preventing accidents, misuse, and loss of control as systems grow more capable. AI ethics also asks whether a system should exist and whom it serves.
• AI governance: The policies, processes, and laws through which organizations and governments put ethical principles into practice.
• Responsible AI: The term companies usually use for their own practice of AI ethics, and largely a synonym.
• Algorithmic bias: One specific ethical problem, systematic error that produces unfair outcomes, within the broader field.

Ethical Issues in Computer Vision#

• Healthcare: In medical image analysis, a model validated on one hospital's scanners and patients can miss findings for populations it never saw, which is why ethics frameworks call for disaggregated evaluation and clinician oversight.
• Synthetic media: Deepfakes enable fraud, harassment, and disinformation, and the EU AI Act requires deployers to disclose them.
• Emotion recognition: In 2022, Microsoft retired Azure Face features that inferred emotional states and attributes such as gender and age, citing the lack of scientific consensus on how to define emotions and doubts that inferences from facial expressions generalize across regions and demographics.
• Consent and dataset provenance: Many face and person datasets were scraped without the knowledge of the people pictured; Microsoft took down its MS-Celeb-1M dataset of about 10 million photos of 100,000 people in 2019.
• Facial recognition and surveillance: In June 2020, IBM stopped offering general-purpose facial recognition and Amazon paused police use of its Rekognition service, a moratorium it later extended indefinitely. In September 2024, the Dutch Data Protection Authority fined Clearview AI €30.5 million for building a database of billions of face photos scraped from the internet.