When machines learn everything from us, wouldn’t human bias be inevitable in machine learning? For years now, everyone has hoped that machines could help us make unbiased, fairer and more objective decisions. Wasn’t hoping this wrong in the first place? A paper published in Nature ML, has been discussed in the AI community of reddit and they are making a bold claim. Bias isn’t a bug in AI systems, it’s in fact a feature.
Researchers when analyzing ten of the world’s most powerful large language models, have found that biases persist regardless of all the measures taken for safeguarding against them. And what’s even worse is the fact that when attempts where made to “debias” them, it did not prejudice it. What the researchers did to debias is they might have filter out certain words, adjust outputs, or fine-tune responses so that the model appears fair on the surface. Highlight the word surface here, because underneath it bias still prevails. In other words, we haven’t removed it, we’ve just covered it up.
What is the Root of the Problem?
Well… as mentioned above the fact that machines learn everything from us, that’s the problem. What is in the heart of a machine? Millions of datasets- screenshots, videos, words, audios which are generated from humans and from which machines learn constantly. Now what is the likelihood that these datasets contain only facts? That percentage is actually not that much. Because these datasets also contain assumptions, stereotypes, and cultural narratives. So what happens when AI learns from these data is that it not only memorizes it, it also generalizes it. And that is the root of the problem. You might have come across Grok’s white genocide responses, or even the kill the boer. They also exhibited subtle patterns of bias, such as:
- Reinforcing racial and cultural stereotypes in chatbot responses.
- Preferring resumes with Western names over non-Western ones.
- Associating leadership traits more with men than women.
Why Debiasing Fails
So how does one treat bias? Many companies treat bias as something they can filter out. But Bias isn’t random, it is deeply woven inside. Into how language and culture work. Let me just put it out in simple words for you, let’s say a resume says “Captain of the football team”, while this may seem a neutral statement with no room for bias, it can carry certain implications. In many context such as about gender, class, and race.
What are the risks associated with Debiasing?
There are two risks associated with debiasing data:
- Losing useful context: So when one filters out sensitive data, this process can strip the model of nuance and understanding.
- Creating a false sense of fairness: As mentioned above, a debiased model might still show preferences, it is not removed, it’s just covered and hidden.
Can Ethical AI Still Exist?
From the above text, if AI systems cannot be truly unbiased, then is Ethical AI a myth? Not literally… But what this means is we need to rethink what “ethical” means in this context.
What we should be doing instead of chasing the dream of a perfectly neutral machine:
- Make AI systems transparent – Designers should clearly show users how each response is generated and what data the system uses.
- Build in accountability – Teams must include human oversight for all high-stakes decisions to ensure responsibility and ethical judgment.
- Include diverse voices – Developers should actively involve people from different backgrounds when creating and training AI systems.
- Shift focus from “debiasing” to harm reduction – Companies need to understand where AI causes the most damage and prioritize mitigation of the same.
Final Thoughts: Bias Is a Human Problem
The idea that bias is undeniable is not a reason to give up. It is a call to take radical responsibility. If we build machines in our image, we must confront what that image includes. Because in the end AI won’t save us from ourselves. But it might show us who we are. If you’re ready to confront these hard truths, we’ve explored some of the toughest ethical questions about AI in this in-depth article.
