IEEE Technology and Society Magazine - September 2023 - 104

Opinion
to identify and address biases that may be present
in the system. This can help ensure that the system
is making decisions that are fair and unbiased. It's
also important to be able to explain how a decision
was made by an AI system. This can help users
understand the reasoning behind the decision and
can provide important context for how the system
works [43].
Therefore, transparency is critical for ensuring
that AI systems are making decisions that are fair,
unbiased, and trustworthy. It can also help build
user trust and provide important context for how
the system works. To increase transparency in DCAI
systems, it is important to open-source the code and
data, provide clear explanations for model outputs,
and disclose training data sources. This helps identify
potential biases and errors, build user trust, and
provide context for the system's decision-making
process.
Data diversity and representative
The use of diverse and representative data sets is
important to reduce biases in AI systems. When AI
systems are trained on biased or unrepresentative
data, they may make biased decisions, which can
have negative impacts on marginalized groups [44].
By using diverse and representative data sets, AI systems
can be trained to make more fair and unbiased
decisions that better reflect the real world. There are
several techniques for improving data diversity and
representativeness, including:
Data augmentation
This involves creating additional training data
by applying transformations to the existing data set.
For example, flipping images horizontally or adding
noise to audio data. This can help increase the diversity
of the data set and make it more representative
of the real world [45].
Active learning
This involves selecting the most informative samples
for human labeling, which can help improve
the representativeness of the data set. By selecting
samples that are most likely to reduce model uncertainty
or increase diversity, active learning can help
ensure that the labeled data set is more representative
of the real world [46].
104
Crowd-sourcing
This involves outsourcing the labeling of data to
a large group of people, which can help improve
the diversity and representativeness of the data set.
Crowdsourcing can help ensure that the labeled
data set reflects a wider range of perspectives and
experiences, which can help reduce biases.
Ongoing monitoring and evaluation of data sets
are important to ensure continued diversity and
representativeness. Over time, the real world may
change, and new types of data may become more
relevant [45], [47]. If data sets are not regularly monitored
and evaluated, they may become less diverse
and less representative of the real world, which can
lead to biased AI systems. By regularly monitoring
and evaluating data sets, AI systems can be trained
on the most current and representative data, which
can help reduce biases and ensure fair and equitable
decision-making [48].
Fairness and bias testing
AI models need to be tested for fairness and biasedness
to ensure that they are not making decisions
that unfairly advantage or disadvantage certain
groups of people. Without testing, it is possible for
AI models to perpetuate social biases and inequalities
[43]. By testing for fairness and biasedness,
AI models can be improved and made more equitable,
ensuring that they serve all users fairly and
impartially.
Selecting appropriate fairness metrics and performing
comprehensive bias tests are important to
ensure that AI models are evaluated fairly and accurately.
Fairness metrics should align with the specific
context of the AI model and the potential impact
on different groups of people. Bias tests should be
performed to identify potential biases in the data
and ensure that the model does not unfairly discriminate
against certain groups. By selecting appropriate
fairness metrics and performing comprehensive
bias tests, AI models can be evaluated in a thorough
and fair manner, helping to reduce biases and promote
equity. There are several techniques for testing
data-centric approaches, including:
Sensitivity analysis
This involves testing how sensitive the model is to
changes in the input data, which can help identify
potential biases and sources of error.
IEEE Technology and Society Magazine
IEEE Technology and Society Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - September 2023
