AI and fairness: bias, copyright, privacy and the environment
Get a balanced, plain-English view of the biggest debates about AI, so you can form your own opinion.
AI is useful, but it raises real questions. Here are the big four, in brief:
1. Bias. AI learns from human writing, which includes unfair stereotypes. It can repeat them, for example in assumptions about jobs and gender. That matters most when AI is used to make decisions about people, like hiring.
2. Copyright. AI was trained on huge amounts of writing, art and music, often without asking the creators. Many artists and authors think that's unfair. AI companies argue it's similar to how people learn. Courts and governments, including in the UK, are still deciding.
3. Privacy. AI companies collect a lot of data, including what people type into chatbots. Laws like UK GDPR give you rights, and you can change settings (lesson 3).
4. Environment. Training and running AI uses a lot of electricity and water in data centres. Companies are investing in cleaner energy, but demand is growing fast.
There are no simple answers. Knowing the debates helps you use AI thoughtfully and make sense of the news.
Bias
What happens: models learn patterns from human-created data, including stereotypes and imbalances. Without care, they can reflect them, for example assuming a nurse is a woman or an engineer is a man, or performing less well on some accents and dialects.
Why it matters: in casual use, it's mostly an annoyance. In decisions about people (recruitment, lending, insurance, policing), it can cause real unfairness.
What's being done: labs test and reduce bias; laws like the UK Equality Act still apply to AI-assisted decisions; regulators expect organisations to check for discrimination.
Copyright
The debate: AI models were trained on large amounts of copyrighted text, images and music, usually without permission or payment.
- Creators argue this uses their work to build products that compete with them, without consent or pay.
- AI companies argue training is a transformative use, similar to how people learn from what they read, and that strict rules would hold back innovation.
Where it stands: court cases are ongoing in several countries. Some AI companies now pay for licensed content. The UK government has consulted on changes to copyright law for AI, with strong objections from the creative industries. Expect this to keep making headlines.
Privacy
AI services process huge amounts of personal data: training data scraped from the web, and the conversations people have with chatbots. Concerns include how data is collected, how long it's kept and whether it's used for training. UK GDPR gives you rights to access and delete your data, and the ICO regulates this.
Environment
Training large models and answering billions of questions needs large data centres that use significant electricity and, for cooling, water. Estimates vary widely, and individual chat messages use relatively little energy, but total demand is growing fast. Companies are investing in renewable and nuclear power and more efficient chips; critics say growth is outpacing those efforts.
Other debates you'll hear
- Jobs: covered in the next lesson.
- Misinformation and deepfakes: lesson 6.
- Safety of powerful AI: how to make sure increasingly capable systems stay reliable and under human control.
- Concentration of power: a few large companies control the most capable models.
Bias in more depth
Bias can enter at several points: data (who and what is over- or under-represented), labelling (the judgements of the people rating answers), design choices (what the system is optimised for) and deployment (how outputs are used). There are also competing definitions of fairness: equal accuracy across groups, equal outcomes, or equal treatment. These can conflict mathematically, so choices involve values, not just engineering.
Well-documented examples from earlier AI systems include facial recognition performing worse on darker-skinned faces, and a recruitment tool that learned to downgrade CVs mentioning women's activities. Language models show subtler patterns. Mitigations include more representative data, targeted testing, post-training to reduce stereotyping, and human oversight of consequential decisions.
In the UK, the Equality Act 2010 applies regardless of whether a decision involves AI, and data protection law gives people rights relating to solely automated decisions with significant effects. Regulators including the ICO and the Equality and Human Rights Commission have published guidance.
Copyright in more depth
The core legal questions are: (1) is copying works to train a model an infringement, or covered by an exception such as fair use (US) or text and data mining exceptions (which in the UK are currently limited to non-commercial research)? (2) Can model outputs infringe when they closely reproduce training material? (3) Who, if anyone, owns AI-generated output?
Major lawsuits by authors, news publishers, artists, music labels and photo libraries are working through courts in the US and UK, with mixed early rulings. Meanwhile, licensing deals between AI companies and publishers have multiplied. The EU AI Act requires general-purpose model providers to publish summaries of training content and respect copyright opt-outs. The UK government's consultation on whether to allow commercial text and data mining with an opt-out prompted a strong campaign from creative industries; the policy outcome is still developing. Check our news for the latest.
Privacy in more depth
Issues include: web scraping of personal information for training; memorisation (models occasionally reproducing personal data from training); retention and human review of conversations; and inferences (AI deducing sensitive traits from innocuous data). European regulators have taken action against AI providers over GDPR compliance in several cases. Strong practices include data minimisation, opt-outs from training, clear retention limits and enterprise agreements for workplace use.
Environment in more depth
Energy use falls into training (a large one-off cost per model) and inference (answering requests, which adds up because of enormous volume). The International Energy Agency has projected that data centre electricity demand could roughly double in the second half of the 2020s, with AI a major driver. Water is used for cooling, which matters in drought-prone areas. Per-query estimates vary by an order of magnitude depending on the model and hardware, and efficiency improves each year, but total consumption is still rising. Responsible choices include using smaller models for simple tasks and avoiding unnecessary generation of images and video.
Regulation
The EU AI Act takes a risk-based approach, with obligations phased in from 2025 for banned practices, general-purpose models and high-risk uses. The UK has favoured a principles-based approach led by existing regulators, plus the AI Security Institute, which tests advanced models. The US approach has shifted between administrations, with significant activity at state level. International efforts focus on safety testing of the most capable models.
Forming your own view
Useful questions: Who benefits and who bears the costs? What's the evidence, and how recent is it? What would happen under the alternatives? Being informed lets you make choices (which tools you use, what you share, what you support) that fit your values.
Try this prompt
Give me the strongest arguments on both sides of [an AI debate, for example AI training on copyrighted work], then tell me what the main open questions are. Stay neutral.
Replace the [brackets] with your own details
On haystak
Quick check · 1 of 3
Why can AI show bias?
Next step
Where AI is heading, and what it means for jobs
A calm, evidence-based look at where AI is going and how to prepare, whatever your job.