Page 2 - AI Vol 2: Risks of AI
P. 2
01 GENERAL RISKS OF AI
01A. BIAS
A primary concern raised about the use of AI inadvertently learn and replicate these biases. For
systems is its potential to exhibit bias in decision instance, if an AI system is trained on employment
making. AI bias, a reflection of discriminatory data that historically favors a particular
patterns or unequal representations in the data demographic, it may continue to replicate this
or its application, poses significant risks. It favoritism, despite changes in societal norms or
can inadvertently perpetuate existing societal legal standards. The repercussions of training AI
prejudices, leading to unequal treatment of systems on biased data are significant, especially
individuals or groups. Recent experiences with for public agencies. Decisions based on such data
Google Gemini have also demonstrated that can lead to discriminatory outcomes, such as
programming protocols intended to prevent bias, unfair resource allocation, biased hiring practices,
can also lead to the generation of inaccurate data. or unequal service provision.
AI is fundamentally shaped by the data it is trained Not all AI bias is caused by the underlying biases
on. This data, typically vast in scope, is generally that are learned by the model’s training data. Bias
all human generated and it does not always can also occur due to the way the AI algorithm
provide a neutral or unbiased representation of processes and prioritizes different inputs,
reality. AI systems learn by identifying patterns sometimes at the direction of the user. In many
and correlations in their training data. If this data AI systems, decisions are made based on a set
includes historical prejudices or societal biases – of features or attributes considered relevant. For
whether in terms of race, gender, socioeconomic instance, in employment decisions, factors such
status, or other characteristics – the AI is likely to as years of experience, education level, or past job
2 | VOLUME 2 RISKS OF AI | LOZANOSMITH.COM
