Page 220 - AI Computer 10

Over-Reliance on Accuracy

Accuracy can be misleading, especially in imbalanced datasets. For example, a fraud detection model predicting
"no fraud" 99% of the time may have high accuracy but fail at catching actual fraud cases.
Ignoring Class Imbalances

Models evaluated on imbalanced datasets may favour majority classes, leading to biased results. Precision,
Recall, and F1 score should be used instead of just accuracy.

Ignoring Calibration of Probabilities
Some models output probabilities, but if these are not well-calibrated, decision-making based on them may be
flawed, leading to poor real-world outcomes.

Lack of Transparency in Metric Computation

Some models report evaluation metrics without disclosing the exact methods, datasets, or assumptions used,
making reproducibility and ethical assessment difficult.
Ignoring Ethical Trade-offs in Metrics

Some applications require ethical trade-offs. For example, in autonomous driving, prioritizing pedestrian safety
might lower overall accuracy, but ethical evaluations should consider human lives over raw performance.

 Make sure that the datasets used to evaluate
an AI model are not biased at source.
Bias  Make sure the metric used to evaluate the
model does not favour any one group or
sector.

 Explain the working behind the chosen

Transparency metric and how it produce results.
 Explain the dataset involved in calculating
the chosen metric.

 Own up to the choice of metric used for
evaluation of a specific model.
Accountability
 Make sure no stakeholder suffers unjustly
due to the choice made.

Addressing the ethical issues during evaluation of an AI model ensures that the model will perform according to
required parameters of performance, reliability, and in general good of humankind.

K Keyey TTermserms

u Model Evaluation
The process of assessing the performance and reliability of an AI model using a test dataset that was not
used during training.
u True Positive (TP)

The AI model predicts an outcome/event to occur that actually occurs.

86
86