Artificial Intelligence: Decoding AI For Internal Audit Pankaj Kumar Ca, Cfa, Isa
Artificial Intelligence: Decoding AI For Internal Audit Pankaj Kumar Ca, Cfa, Isa
Artificial Intelligence: Decoding AI For Internal Audit Pankaj Kumar Ca, Cfa, Isa
Intelligence
Decoding AI for Internal Audit
Pankaj Kumar
CA, CFA, ISA
Refers to computer systems or machines that can
perform tasks and make decisions that usually
require human intelligence. It's like teaching a
computer to think and learn like a person.
Pankaj Kumar 2
Complexity of decision-making - Unlike traditional software, AI systems may have a level of
autonomy and decision-making that is not explicitly programmed.
Data-centric nature – Unlike traditional software, Auditing an AI system requires assessing the
quality, representativeness, and potential biases in the data used.
Why AI is Interpretability and explainability - Unlike some traditional software where the code and
difficult to decision logic are typically transparent, AI systems, such as deep learning models, can be highly
complex and difficult to interpret.
Audit Bias and fairness considerations: AI systems can inherit biases present in the training data or
develop biased behavior due to complex interactions between variables. Unlike traditional
software auditing AI involves evaluating and addressing potential biases, ensuring fairness across
different demographic groups, and mitigating any discriminatory impacts.
Continuous learning and adaptation: Unlike traditional software AI systems can learn and adapt
over time, making it challenging to conduct a static audit.
Pankaj Kumar 3
Apple's Credit Card Gender Bias: In 2019, Apple launched its
Apple Card in partnership with Goldman Sachs. However, it was
reported that the credit limit algorithm used to determine
credit limits showed bias against women. Several cases emerged
where women with similar or higher creditworthiness than their
male counterparts received significantly lower credit limits. This
Some error incident highlighted the potential for bias in AI algorithms used
for credit assessments.
led to YouTube's Recommendation Algorithm: YouTube's
failure of AI recommendation algorithm has been criticized for promoting
and amplifying content that is extreme, misleading, or
offensive. The algorithm is designed to optimize user
engagement, leading to the inadvertent spread of biased or
harmful content. The amplification of misinformation and
radicalization through recommendation algorithms has been a
subject of concern and scrutiny.
Pankaj Kumar 4
Facial Recognition Technologies: Facial recognition technologies
have faced numerous instances of bias. In 2020, it was revealed
that facial recognition systems developed by major tech
companies, including IBM, Microsoft, and Amazon, exhibited
higher error rates when identifying individuals with darker skin
Some error tones and women compared to lighter-skinned males. Such
biases can have serious implications in areas like law
led to enforcement, surveillance, and hiring processes.
Pankaj Kumar 5
Define the scope and objectives: Clearly define the scope of the audit, including the
specific AI system or systems to be audited, as well as the objectives and goals of the
audit. Identify the key areas of focus, such as performance evaluation, bias
assessment, transparency, or compliance with regulations.
AI Audit stakeholders.
approach Assess data quality and preprocessing: Evaluate the quality, relevance, and
representativeness of the data used to train and test the AI system. Assess the data
preprocessing techniques employed, including data cleaning, normalization, and
feature selection. Identify any potential biases or data-related issues that could
affect the system's performance.
Evaluate performance metrics: Define appropriate performance metrics based on
the nature of the AI system. Assess the system's accuracy, precision, recall, F1 score,
or other relevant metrics. Compare the system's performance against benchmarks or
industry standards. Identify any performance gaps or areas for improvement.
Pankaj Kumar 6
Analyze fairness and bias: Assess the AI system for potential biases or unfairness.
Examine the system's behavior across different demographic groups or sensitive
attributes. Analyze the impact of biases on decision-making and outcomes. Identify
and address any unfair biases to ensure fairness and non-discrimination.
approach
transparency and explainability.
Assess ethical and regulatory compliance: Evaluate the AI system's compliance with
continued
relevant ethical guidelines, regulations, and industry standards. Ensure the system
respects privacy, security, and data protection. Assess if it adheres to requirements
such as fairness, accountability, and transparency.
Pankaj Kumar 7
UNDERSTAND AI DESIGN IDENTIFY AND INVOLVE ADOPT AND ADAPT
AND ARCHITECTURE TO ALL STAKEHOLDERS EXISTING FRAMEWORKS
Key to
SET PROPER SCOPE AND REGULATIONS
success
FOCUS ON TRANSPARENCY
THROUGH AN ITERATIVE
PROCESS
Pankaj Kumar 8
Conclusion
Auditors must be ready to tackle the
challenges of auditing AI. Despite the
lack of a clear definition, AI is rapidly
growing across various domains. This
presents an opportunity for the business
and IT community to step up, prepare,
and establish strong governance for
auditing AI systems. By addressing tough
questions and proactively adapting to the
evolving technological landscape,
auditors can ensure effective and
responsible AI auditing.