Stereotypes and Bias

Theory from the field of cognitive linguistics that looks at how people categorize objects or concepts

People learn in categories and each category has a prototype, the membership to this category is then evaluated by the shared characteristics of an item with the prototype

An example is the mum-doctor riddle or yellow bananas

Prototype –> Most typical representative of a concept/category

Stereotype –> Perceived characteristics of an entire group of people

Bias –> Disproportionate weight in favor of or against an idea, thing or group of people, often in a way that is unfair

Male-dominated professions tend to have even more men in their results than would be expected if the proportions reflected real- world distributions

This effect is also seen when people rate the quality of search results or select the best image from a result: They were found to prefer images with genders that match the prototype of a particular occupation

Men are typically perceived with the attributes of indepence, instrumental competence or leadership confidence, which all can be summarized with the term agency

Women are typically perceived with the followeing attributes: concern for others, sociability or emotional sensitivity, which all can be described with the term communality

But according to social role theory, gender stereotypes are dynamic constructs influenced by the actual perceived changes

Unconscious biases are social stereotypes in favor of or against certain groups of people that individuals form outside their own conscious awareness

Biases are of course not limited to gender-specific stereotypes. Biases may exist toward any social group, e.g. ethnicity, age, physical ability, religion, sexual orientation, weight, and many other characteristics

Everyone holds unconscious beliefs about various social and identity groups, which stem from our tendency to organize social worlds and reduce complexity by categorizations

Unconscious biases may influence the decisions and actions we take and can have negative or positive consequences for the person or group concerned

Unconscious bias is far more common than conscious prejudice and often incompatible with one’s conscious values

With the Implicit Association Test

The IAT measures attitudes and beliefs that people may be unwilling to report

The idea is to detect the stregth of the associations of oncepts, evaluation and prejudices

Via translation systems, e.g. GoogleTranslater –> turkish gender-neutral noun is translated according to stereotypical gender in that context

By recruting algorithms –> Amazon’s resume-reading AI system were sexist

With Machine Learning –> since AI only mirrors the societal beliefs of humanity there can be underrepresented groups, confusing causation with correlation or an overgeneralization accroding to the training data

By wordlviews of annotators –> most of the people have specific knowledge, e.g. know how a westernized wedding looks but don’t know how a african wedding looks and therefore label accronglingly

With face recognition software or voice assistents –> underrepresented groups are sometimes not identified as persons or their voice is not recognized correctly

AI bias doesn’t come from the systems themselves, it comes from people and is therefore not by default objectively neutral or fair

No technology is free of its creators

All technology comes from and is designed by people, which means that it’s not automatically more objective than we are

Data and math don’t equal objectivity

Specification of the problem to be solved in ways that affect classes differently

Failure to recognize /address statistical bias

 Reproduction of past societal prejudice and decision-making

 Consideration of an insufficiently rich set of factors

Incomplete or skewed data –> selection bias

Coverage bias –> The population represented in the data set does not match the population that the machine learning model is making predictions about, e.g.: Consumers who instead bought a competing product were not surveyed

Sampling bias –> Data is not collected randomly from the target group, e.g.: surveyer chose the first 200 consumers that responded to an email

Participation bias –> Members of certain groups opt-out of surveys (or labeling tasks) at different rates than members of other groups, e.g.: Consumers who bought the competing product were 80% more likely to refuse to complete the survey; Moral machine experiment

Awareness and understanding of the various causes of algorithmic biases is the first step

More diversity in teams developing AI would be very beneficial

Integration of bias detection strategies in AI development

Creators and operators of AI algorithms need to ask: Will we leave some groups of people worse off as a result of the algorithm’s design or its unintended consequences?

Bias impact statements as a self-regulatory practice of AI creators/companies

Ethical governance standards

Missing Feature Values –> feature value is missing for one or more features with a large number of examples might lead to underrepresentation

Unexpected Feature Values –> look out for uncharacteristic or unuasial feature values because these might lead to inaccuricies or biases

Data Skew –> any data were groups are underrepresented or overrepresented might lead to biases

Countermeasures

• adding the option of male voices

• voices that sound less humalike

• genderless voice

Effects

• Reflecting,reinforcingand spreading stereotypical images of traditional gender roles

• Feminine voice standing for servility and dumb mistakes

• Tolerance of verbal insults towards feminized characters