Responsible AI

Responsible AI means building and using AI in a way that is fair, safe and honest with people, with someone who answers for it when it goes wrong. For a business it comes down to questions: does the tool treat all customers equally, do people know they are dealing with AI, is personal data protected, and does a person check important decisions? In the EU, part of this is law.

Which principles does it cover?

There is no single list, but the main ones overlap. The OECD AI Principles (May 2019, updated May 2024), the first intergovernmental standard on AI, name five values: inclusive growth and well-being; human rights and democratic values, including fairness and privacy; transparency and explainability; robustness, security and safety; and accountability. In April 2019 an independent expert group appointed by the European Commission published seven requirements for trustworthy AI, from human oversight to environmental well-being. Microsoft works with six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability.

How do principles become practice?

Microsoft's Responsible AI Standard (June 2022) breaks each principle into goals and requirements for its teams, such as impact assessments, data governance and human oversight under accountability. Applying it changed products:

• Microsoft limited access to its facial recognition service to managed customers and partners, and retired features that claimed to read emotions, gender or age from a face. For emotions, it cited the lack of scientific consensus on what emotions are.
• After a March 2020 study found that speech-to-text across the tech sector made nearly twice as many errors for some Black speakers as for white speakers, Microsoft brought in a sociolinguist and collected more varied speech data.

Other common tools have their own entries: model cards, red teaming, tests for bias and human-in-the-loop checks.

What does the law require?

In the EU, part of this became law with the AI Act. The Act recalls the seven requirements of 2019 in its recitals. Since 2 February 2025 it bans, among other practices, emotion recognition at work and in education, except for medical or safety reasons. Since 2 August 2026 a chatbot must make clear that people are dealing with AI, unless that is obvious. Strict duties for high-risk uses, such as screening job applicants, apply from 2 December 2027, a date set by the AI Omnibus of July 2026.

Where does it fall short?

Principles are voluntary and can change. In February 2025 Google rewrote the AI principles it had published in 2018, and the new version no longer contains its pledge not to pursue AI for weapons or for surveillance that violates internationally accepted norms, CNBC reported. Principles can also pull against each other. Kleinberg et al. proved in 2016 that three common fairness conditions for risk scores cannot all hold at once, except in special cases.

Example

Suppose a recruitment agency uses an AI tool to pre-sort CVs. Working responsibly, it compares per group, such as men and women or age brackets, who gets through; it tells applicants that AI is involved; it keeps CV data no longer than needed; and a recruiter reviews every rejection. CV screening is high-risk under the AI Act, so from 2 December 2027 the agency must also follow the tool's instructions for use and have a competent person oversee it.