BY: Ekaansh Dixit
Artificial intelligence is gradually impacting both simple and complex jobs in our constantly changing technological world. Our growing dependency on AI algorithms and judgments is reflected in the increasing allocation of tasks to AI systems that once used to be within the capabilities of human thought and intelligence alone. It is plausible to argue that AI algorithms frequently influence decisions in a variety of industries.
But as AI becomes more prevalent, concerns about the algorithm’s impartiality and fairness emerge. This affects not just the reliability and precision of AI, but also the confidence that people have in it before relying on it mindlessly for all other issues.
Common Misconception: AI’s neutrality
Even before proceeding with the discussion of AI’s fairness, there lies a common misconception that algorithms are naturally unbiased. However, the reality is a bit different and lies in the very origin of algorithms. AI bias is not spontaneous or on behalf of the algorithm, because AI algorithms themselves are not self-prompted. They are the very human creation shaped by the choices, assumptions, and limitations of their designers.
Hence, it is correct to assume that AI algorithms are just an amplified version of human thoughts. If the data fed to the algorithms contains deep-rooted societal inequalities, the model will learn it as the only truth. We will now discuss such other causes and sources through which AI becomes biased in its opinions.
Root Cause of the Problem: What are the sources of AI bias?
- Bias in the Data
Data is the fuel of AI. AI algorithms are trained upon a dataset given by its developer known as “Training Data”. This is the data on which an AI algorithm develops its basic knowledge and intelligence and if the very data on which it is trained is filled with biases, prejudices and assumptions then the algorithm doesn’t just reflect them in the outcome, it amplifies them.
- Such a source of bias was seen in Amazon’s Hiring AI algorithm which took 10 years to be trained. However, a slight mistake made the company scrap the whole system just 3 years after its deployment, because it was trained wholly on résumés of men and hence the algorithm learned to downgrade any résumés containing the word “women”.
- The COMPAS criminal‑risk assessment tool used in U.S. courts labelled Black defendants as “high-risk” at nearly twice the rate of white defendants with similar profiles.
- Bias from Humans
People, through their chats, can significantly influence the way AI responds. Many AI models, even after deployment, are trained through the conversations they have with the users, making them vulnerable to have ranging opinions. If the user presents biased, prejudiced or racist comments and thoughts to the model, the model will reproduce those ideas.
- A classic example was seen in Twitter’s online chatbot released in 2016 called Tay. It was made to mimic the speech of a 19-year-old American girl and learn from the conversations of the user. However, this move took a big turn as within hours the users persuaded her into posting racist, sexist, and anti-Semitic content, including denying the Holocaust and praising Hitler.
This leaves us with a question, is it rational to think that AI can achieve absolute fairness in its judgement?
The answer here would be that the question has a deeper meaning to it, which we will discuss about further
Fairness: A difficult concept to define
Fairness is not a single, universal concept. Even if we remove all technical sources of bias, we cannot ensure that the AI model or algorithm remains unbiased and fair because there are several fairness metrics- demographic parity, equal opportunity, predictive equality, and such more which may contradict each other’s definition making it almost impossible for an algorithm to be fair in all criteria at once.
And this brings us to a question: “Is it right to prioritize one criterion over the other?”.
To understand this, it should be evident to realise that “fairness” is not a technical value but a philosophical one requiring value judgement. If the model is aimed to achieve equality in every criterion, it may contradict one or the other criterion. Such as giving equal opportunities to every person in a community will contradict achieving equality by favouring the lower class a bit more than the upper classes.
Thus, the concern of fairness in AI is not only a scientific challenge but also a political and moral one.
Can Technology Ever be Fair?
The short answer here is: Not perfectly.
The more meaningful answer would be that it can be made much more fair in its decision than it is today
This is because AI is a concept, a technology and an intelligence that is still a protostar in its developing. Every day or the other there is new facts and information which are discovered about Artificial Intelligence hence highlighting its constantly evolving nature.
Thus, technology may never be perfectly fair – but it can be made more just, more transparent, and less harmful.
Conclusion
The question “Can technology ever be truly fair?” forces us to deal with the limits of AI and the flaws of the societies that shape it. It isn’t that algorithms create biases out of thin air, but they are just a mirror of the world that we present to them. If human societies contain inequality, the data we produce will carry traces of that inequality, and the algorithms trained on that data will reproduce it.
It is only until the societies change and become more neutral, could we think about AI being rational and unbiased in its opinions.
