What You Can Do to Overcome Discrimination by AI
Bias in AI is one of those topics everyone has an opinion about, but which very few people understand deeply. If you work in AI, you’ve probably heard statements like “we can remove all bias” or “just have people review the AI’s decisions”. These ideas sound reassuring but can lead to risky assumptions when you’re building or using AI systems.
Let’s break down the three biggest misunderstandings and see what you can do instead.
- Andrea JaegerAI Officer
- Shayda ShwanAI Consultant
Misunderstanding 1: Bias is the Same as Discrimination
Although these two concepts are related, they’re not identical. Bias is a mental shortcut we all use to navigate the world quickly. It’s everywhere around us, and our brains rely on it to process information efficiently. Because AI learns from human behaviour and historical data, those biases inevitably show up in AI systems, too.
Discrimination is different. It’s the harmful outcome of bias: treating people differently, disadvantaging them. This can take the form of disadvantaging or excluding individuals, but it can also involve granting unjustified advantages to certain people or groups. Both forms result in unfair treatment and are considered discrimination.
People face discrimination for many reasons, including:
- Their religion or belief
- Their political affiliation
- Their ethnicity or nationality
- Their gender or sexual orientation
- Their marital status
- If they have a disability or chronic illness
- Their age
- Their working hours
- The type of employment contract they’re on
- If they take care‑related leave
Understanding the difference helps you design AI systems that recognize where bias comes from and prevent it from turning into discriminatory outcomes.
Misunderstanding 2: We Should Eradicate Bias in AI
It’s a noble goal and yet an impossible one. Every dataset contains assumptions, gaps and historical patterns. Every AI system makes trade‑offs. People themselves are generally biased, so expecting AI to be perfectly neutral sets you up for disappointment.
A more realistic approach is to accept that the goal isn’t to eliminate bias entirely, but to manage it responsibly. You can begin by making your fairness criteria explicit rather than leaving them unspoken or assumed. You should also evaluate how different groups actually experience the system, so that you can see where disparities show up. Along the way, be sure to document any trade-offs you make – because every design choice affects fairness in some way. In the end, you may not be able to remove all bias, but you can make it visible, measurable and controllable.
Misunderstanding 3. People – the Humans in the Loop – Will Naturally Remove Bias from AI
While humans do help, they’re not fail-safe. People bring their own assumptions and blind spots and often work under time pressure. In addition, humans are susceptible to automation bias: the tendency to over‑trust or defer to AI outputs, even when they are incorrect or incomplete. A human reviewer who processes thousands of AI‑supported decisions a day will not catch every subtle pattern and may unknowingly reinforce bias rather than correct it.
So, what can you do?
Start by understanding who may be affected when your AI is in production. Make sure you understand the origin and limitations of your data. Test your models across different user groups and combine automated checks with well-designed human oversight.
But most importantly, treat fairness as a design principle, not an afterthought. That is why it is essential to make a conscious design decision about the role of the human and the level of AI autonomy having a human in the loop, human on the loop, or human out of the loop.