Algorithm Bias

The artist sought perfection with every stroke, but each line strayed from its intended path, creating chaos instead of order. This is the paradox of creation: we aim for precision, yet our tools, our data, our very hands introduce deviations we never planned for.

It's been a while, had to grab an ink refill and a lot of content for you all. You ready for me?

So, over these last few weeks I have been inactive on LEMMA, I have had an opportunity to see, observe, and study crucial problems on an individual level, national level and in extent on a global level. What do I have to share? It could pique your interests highly and on the other hand......

But we would never know till we start, would we?

I have been thinking heavily in fact especially when I use ChatGPT to execute some daily prompts I come to find out that the ai system is almost always in support of me, how? I once gave an ai a prompt to judge an event, it judged it as prompted then right after I corrected the verdicts based on my own personal preference, then I found that the ai also corrected its judgement to support my personal preference, you find that this happens a lot when you are working on a mathematical task using ai.

Mhmmmmm....you see it now, don't you?

After detailed research I found out that yes, it's true, my observation was valid, artificial intelligence usually has a tendency of being overly agreeable, what researchers call sycophancy or algorithm bias.

What is algorithm bias?

Picking the words one by one, algorithm - a set of predefined instructions for performing a task, bias - falls in the way of inaccuracy towards a decision or an idea, together algorithm bias occurs when a set of instructions for performing a task works inaccurately or with prejudice toward a specific group or outcome.

To further explain this term I would like to give an example - in Nigeria it is very well known that the Yorubas like pepper so very much, imagine a chef who spent all his life say 10-20yrs preparing spicy meals for the Yorubas on the street, was moved to make food for a general audience (Yoruba, Igbo, French, Japanese) at a 5 star hotel but he doesn't know how to make a meal except a spicy dish, so he makes a spicy dish and in the end, some of the people (Yorubas) are able to eat whereas others find it hard to eat the dish. The chef's training, those 20 years cooking only spicy food, created a bias in what he produces. This is how algorithm bias works: AI learns from its training data, and if that data is limited or skewed “tilted” toward one group, the AI's outputs will be too.

It is important to note that the problem doesn't just come from how you structure your prompt, but also from the data the AI was trained on. When these models are being developed and they go through the machine learning process, it all boils down to how true or unbiased the training data is. If the AI learns from biased historical data. say, resumes from companies that mostly hired men, it will learn those biases. The model architecture, the training data, how we prompt it, and how it's deployed can all introduce bias. In essence algorithm biasedness is an effect that stems from multiple sources, but particularly from the hands of we, the data providers.

What are the effects of algorithm bias?

The effect of algorithm biasedness can best be seen in the HR department, which now uses artificial intelligence to support resume review and hiring, we call it ATS (applicant Training System). Based on the ATS algorithm, which is designed to check for the structure of the resume from titles down to the choice of words, a lot of applications and resumes have been turned down either because they tilt away from the ATS algorithm slightly or majorly, that's one effect of algorithm bias. Another major effect can be found when we prompt our ai agents individually then get results which are in support of our prompt but in reality, may be wrong (this is the sycophancy I mentioned earlier, AI being too agreeable rather than accurate).

I believe that is one of the reasons every ai agent states that ai is subject to mistakes lol.

What are the ways which we can deal with algorithm bias?

One of the major ways in which algorithm bias has been addressed is through fairness metrics, these are measurements that help us detect when AI systems are producing unfair outcomes for different groups. Fairness metrics allow developers to test whether their AI treats everyone equally, catching bias before the system is deployed. Individuals should also take their time to find out what are the best ways to prompt AI to get better results and maybe unbiased responses. Beyond this, we need more diverse teams building AI systems, better testing for fairness across different groups, and transparency about where AI systems might fail or show bias.

Algorithm bias isn't just a technical problem; it's a human one. As AI shapes more of our lives, from hiring to healthcare, we must remember: these systems only learn what we teach them. The responsibility for fairer AI lies with us.