Page 64 - CT_AI_Class-8

UNDERSTANDING BIAS Let us consider a scenario.

Computer scientist Joy Buolamwini tested AI technology used by major tech companies to A school is developing an AI system to recommend after-school clubs to students. The system uses
recognise faces. She found that while the systems worked well for light-skinned men, they made past data about which students have enjoyed various clubs. The following training data is used:
more errors when identifying women and people with darker skin tones. The reason was clear: the Club Boys Girls
systems were trained on datasets that contained far more images of certain groups than others.
In one demonstration, the computer failed to detect her face until she wore a white mask. This Football 60 20
highlighted how AI systems can be biased and unfair if the training data does not include a wide Art 15 55
representation of all groups.
The AI program studies the data and learns that most boys enjoy playing football and most girls
Bias in Datasets enjoy art. Now, imagine a new girl student joins the school.

Bias in datasets occurs when certain groups are overrepresented or underrepresented, leading Think: fact bits
to unfair outcomes. Bias in datasets occurs when certain groups are overrepresented or  Which club will the AI probably recommend to new students A study by the MIT Media
underrepresented, leading to unfair outcomes. The consequences of bias include unequal model after studying the given data? Lab found that facial
performance, discrimination and loss of trust. recognition systems had a
 Is the AI making the recommendation based on ability or 34.7% error rate for dark-
data pattern? skinned women, highlighting

 Do you think every student gets an equal chance through the need for balanced
this AI recommendation system? training data.

 How can the data be improved to help the AI make better
and fairer suggestions?

Ask
AGENT
OrangeAI
If an AI makes a biased decision, who should be held responsible—the AI or the humans who trained it?

Study
Bee is a wrist-worn AI device that records conversations and creates
summaries, reminders and personal insights. It automatically
transcribes tasks and sends the information to a companion app for ENSURING FAIRNESS
ai easy review. A company once created an AI system to help recruit new employees. The system learned from

past hiring data, which showed that most of the engineers hired in the past were men. Because of
lens
this pattern, the AI started to prefer resumes that looked similar to those of male applicants. This
AI Systems and Their Use is a good example of how AI can pick up human biases from historical data if the developers are

AI systems are already being used in many areas of life, such as recommending videos, approving not careful.
loans, assisting doctors and selecting job applicants. However, if the data used to train these AI This situation shows that AI can unintentionally favour certain groups, even though it is supposed
systems is biased, the AI could make unfair or incorrect decisions.
to treat everyone fairly. To prevent this, developers need to ensure that AI systems are trained
For example, imagine an AI system that helps companies choose job candidates. If the past hiring with balanced and diverse data.
data mostly includes men, the AI might learn to associate men with being better candidates, even AI should be designed to benefit everyone, not just a particular group, so that all applicants have
though this is not true. This can result in the AI unfairly favouring men over women, even if both an equal chance, regardless of their gender, race or background. This helps create a fairer and
are equally qualified for the job. This shows how important it is to ensure that the data used to more inclusive system.
train AI is balanced and fair, so that the system makes decisions based on facts, not biases.

62 Artificial Intelligence (CT & AI)-VIII