Page 332 - Artificial Intellegence_v2.0_Class_11
P. 332

Reporting Bias
              When the training dataset's event frequency doesn't precisely reflect reality, this kind of AI bias develops. Consider a
              case where a technology for detecting customer fraud underperformed in a remote location, giving all of the customers
              there an unjustifiably high fraud score.
              It was observed that every historical inquiry in the area had been classified as a fraud case by the training dataset the
              tool was using. Due to the remoteness of the place, fraud case investigators wanted to confirm that each fresh claim was
              false before making the trip to the area. As a result, there were much more fake events than there should have been in
              the training dataset.

              Selection Bias
              When  training  data  is  either  not  representative  or  is  chosen  without  sufficient  randomization,  this  kind  of  AI  bias
              emerges. The study by Joy Buolamwini, Timnit Gebru, and Deborah Raji, in which they examined three commercial
              image recognition software, provides a clear illustration of the selection bias. 1,270 photos of parliament members
              from European and African nations were to be categorized using the AI models. Due to the absence of diversity in the
              training data, the study discovered that all three algorithms performed better on male than female faces and shown
              more pronounced bias against women with darker skin tones, failing on over one in three women of colour.

              Group Attribution Bias
              Group attribution bias occurs when data teams generalize individual truths to whole groups that the individual is or is
              not a member of. This kind of artificial intelligence bias can be found in admissions and recruiting tools that may favour
              applicants who graduated from particular colleges and display prejudice against those who didn't.

              Implicit Bias
              When AI decisions are drawn based on individual experiences that may not be applicable more broadly, this kind of
              bias emerges. For instance, despite their conscious conviction in gender equality, data scientists may find it difficult
              to link women to prominent positions in business if they have picked up on cultural cues about women serving as
              housekeepers. This example is similar to the gender bias in Google Images.
              We should work to create algorithms that are equitable to everyone given the expanding usage of AI in delicate fields
              like banking, criminal justice, and healthcare. Businesses must also make an effort to lessen bias in AI systems. Even a
              rumour that an AI system is biased can drive away customers and harm a company's reputation. On the other hand,
              relying on an AI solution that functions accurately for all genders, races, ages, and cultural backgrounds is much more
              likely to give higher value and appeal to a wider and more diversified group of potential customers.

              How Data-Driven Decisions can be De-biased
              “AI bias does not come from AI algorithms; it comes from people.” Some people may disagree with the statement that
              bias does not come from people but data sets. However, people collect data. A book may reflect the author's bias. Like
              books, datasets also have authors. They are collected
              according to people's instructions or preferences.     Establish processes to test for and
              ML and AI are technologies that are always based on    diminish bias in AI systems.
              algorithms  created  by  humans.  Like  any  man-made      Engage in fact-based conversations
              thing,  these  algorithms  tend  to  take  into  account   about potential biases in human decisions.
              the  biases  of  their  creators.  As  artificial  intelligence
              algorithms  learn  from  data,  any  historical  data  can    Adopt a multidisciplinary approach.
              quickly generate biased AI models that make decisions
              based on unfair data sets. Some steps can be taken to            Invest more in AI bias research.
              manage bias in AI. Here are some of them:


                    330     Touchpad Artificial Intelligence (Ver. 2.0)-XI
   327   328   329   330   331   332   333   334   335   336   337