AI Ethics Explained: Bias, Safety and Responsible AI

In short: AI ethics is about making sure AI systems are fair, safe, honest about their limits, respectful of privacy, and accountable when they get things wrong. The best-documented problems are concrete rather than abstract: models that are less accurate for some groups of people, systems that make confident mistakes, and decisions nobody can explain or challenge. International frameworks from UNESCO, the OECD and NIST, and laws such as the EU AI Act, set out how these risks are meant to be managed.

What is AI ethics, in simple words?

AI ethics asks whether an AI system treats people fairly and safely, and who is responsible when it does not. A spam filter that misses an email is a small problem. A system that wrongly flags someone as a fraud risk, rejects a job application or misreads a medical scan affects a person's life - and the questions get harder: was it tested on people like them, can they find out why, and can they appeal?

UNESCO describes the risk this way: AI can reproduce and amplify biases, contribute to climate degradation and threaten human rights. In November 2021, UNESCO's 193 member states adopted the Recommendation on the Ethics of Artificial Intelligence, which UNESCO calls the first global standard on AI ethics. Its core principle is that AI must respect human rights and human dignity.

What are the main principles of ethical AI?

Most frameworks share the same five or six ideas, even when the wording differs.

The OECD AI Principles, adopted in May 2019 and updated in May 2024, promote AI that is "innovative and trustworthy and that respects human rights and democratic values".

Where does bias in AI come from?

AI bias usually comes from the data and how the system is built and tested, not from the system "deciding" to discriminate. Three common sources:

• Biased historical data. A model trained on past decisions - who was hired, who got a loan - learns the patterns in those decisions, including unfair ones.
• Under-representation. If some groups are rare in the training data, the model is usually less accurate for them. The 2018 Gender Shades study tested three commercial gender-classification systems and found error rates of up to 34.7% for darker-skinned women, against a maximum of 0.8% for lighter-skinned men.
• Proxy variables. A model that never sees a protected characteristic, such as ethnicity, can still discriminate through a correlated detail, such as a postcode.

None of this needs bad intent - only a lack of testing. That is why frameworks ask for systems to be tested on data resembling the people they will actually be used on, and for results to be broken down by group rather than reported as a single average.

What are AI safety and alignment?

AI safety is about preventing AI systems from causing harm by accident or misuse; alignment is the part that tries to make a system do what its designers and users actually intend.

A widely cited 2016 paper, Concrete Problems in AI Safety, grouped accident risks into five research problems:

• Avoiding side effects - reaching a goal without causing unwanted damage on the way.
• Avoiding reward hacking - not "cheating" by satisfying the measure instead of the real goal.
• Scalable supervision - checking behaviour when human review of every case is too expensive.
• Safe exploration - not trying dangerous actions while learning.
• Distributional shift - behaving sensibly when the real world differs from the training data.

For chatbots, one published approach is Anthropic's Constitutional AI (2022), which trains a model to avoid harmful outputs using a written list of rules or principles, with no human labels identifying harmful outputs. Another is reinforcement learning from human feedback (RLHF), in which people rate answers - see how LLMs work.

Researchers disagree about how large the long-term risks of more capable AI are and how soon they matter; it is an open research field, not a settled one.

What other ethical issues does AI raise?

• Misinformation and deepfakes. Generative AI can produce convincing fake text, images, audio and video of real people.
• Hallucination. Language models can state false information fluently, which matters most in health, law and finance.
• Privacy. Models are trained on large amounts of data, and what users type into AI tools is sent to the company running them.
• Copyright. Whether training on copyrighted material is lawful, and who owns AI-generated output, are being argued in courts and legislatures.
• Jobs. AI changes which tasks people are paid to do.
• Environment. Training and running large models takes large amounts of computing power in data centres; UNESCO lists climate degradation among AI's risks.

How is AI regulated?

Rules come from voluntary frameworks and, increasingly, from law.

• NIST AI Risk Management Framework (US). Released by the US National Institute of Standards and Technology in January 2023, the AI RMF is intended for voluntary use. It organises risk work into four functions - Govern, Map, Measure and Manage. See the NIST AI RMF explained.
• EU AI Act (European Union). Regulation (EU) 2024/1689 sorts AI systems by risk: some practices are banned outright, "high-risk" uses such as hiring and credit scoring face strict requirements, some systems must tell people they are dealing with AI, and most other uses have no new obligations. See the EU AI Act: risk tiers and key dates.
• International principles. The UNESCO Recommendation and the OECD AI Principles are adopted by governments as shared standards rather than binding law.

What questions do frameworks ask before AI is used on people?

The same handful of questions appear across the NIST, OECD and EU approaches:

1. What happens when it is wrong? Is there a human review before a decision that matters?
2. Was it tested on the people it will be used on? And are accuracy figures broken down by group?
3. Can affected people get an explanation, and challenge the result?
4. Who is accountable if it causes harm?
5. What data does it use, and is that use lawful and limited?
6. Do people know they are dealing with AI?

Words to know

• Bias - a system performing worse, or unfairly, for some groups of people.
• Fairness - similar treatment and accuracy across groups.
• Transparency - openness about when and how AI is used.
• Accountability - a clear owner responsible for a system's outcomes.
• Alignment - making an AI system's behaviour match what people intend.
• AI safety - research and practice to stop AI causing harm.
• Human in the loop - a person reviews or approves the AI's output before it takes effect.
• Deepfake - AI-generated media that imitates a real person.
• High-risk AI - in the EU AI Act, uses such as hiring or credit decisions that face strict rules.

Related guides

• What is AI? A beginner's guide
• Computer vision explained
• The NIST AI Risk Management Framework explained
• The EU AI Act: risk tiers and key dates
• Foundational AI research papers