Fairness & responsibility

AI systems sort job applications in advance, score loan requests and support doctors with diagnoses. That can make decisions faster and sometimes fairer, because humans have biases too. But a system that learns from data also learns the imbalances in that data. And unlike with a human, this often goes unnoticed until thousands of decisions have already been made.

Bias comes from the data

A model knows the world only through its training data. If the data is skewed, so is the model. The best known example comes from Amazon: from 2014 the company trained a hiring system on ten years of resumes, from a period in which mostly men had been hired. The system learned to downgrade resumes that contained the word 'women', for example in the name of a club. Amazon stopped the project before it ever sorted real applications, but the principle stands. Think of training data as a bag of marbles: if one kind heavily outnumbers the rest, the system will draw that kind almost every time, no matter how neutrally the algorithm itself is built.

Who carries the responsibility?

The answer is clear: always people and companies, never 'the AI'. An algorithm cannot be liable, just as a car is not liable, but its manufacturer or driver is. In the EU the AI Act has been in force since 2024, and its duties take effect in stages over several years. It sorts systems by risk: high risk applications such as hiring or credit scoring face strict requirements on data quality, documentation and human oversight. Whoever builds a system is liable for the product, and whoever deploys it for the deployment. 'The computer decided' is not a valid excuse, legally or morally.

How fairer systems are built

Fairness does not happen by itself, it is engineered and measured. Good teams audit their training data for imbalances and deliberately rebalance it. They test the finished system separately for different groups, because a high overall accuracy can hide miserable accuracy for a small group. The gap between two such rates is given in percentage points: 70 percent approvals in one group and 52 percent in another are 18 percentage points apart. Those teams also document what the system can and cannot do, and have humans confirm important decisions. And because data and usage change, all of this is checked continuously, not once.

Your part in it

Even when you are on the receiving end you are not powerless. In the EU, the General Data Protection Regulation gives you the right, for purely automated decisions with significant effect, to demand human review, state your point of view and contest the decision. Ask questions when a rejection seems odd: was an automated system involved, and what criteria did it use? Report obvious misjudgements, because operators of high risk systems have to monitor how those systems are used and report serious incidents to the authorities. And in your own daily life the same rule applies as for companies: if you pass on AI results, you take responsibility for them.