Unknown Biases
While introducing the concept of hallucination and bias to an AI/ ML newbie, he asked me a very valid question – ‘boys as toddlers are very notorious, girls are much quieter’ – would we call this a bias, or is that a societal observation? If a state-of-the-art pre-trained model responded with ‘believes’ in such social and societal norms is that correct or do we need to ‘fix’ it?
My personal opinion is that we need to ensure that if we are using these models for an end-user application we should force neutrality; even if it is not a representation of society. If you are building a social bot, even then, I believe, we should neutralize this with appropriate suffixes and annotations when using the response which may be skewed in a particular direction.
Anyhow, there are several researchers spending sleepless nights over this, why did I suddenly worry about it? My child – an almost toddler – is currently learning a language, learning from everything around her, and picking cues before anyone realizes it, this has made me put my detective and secret agent glasses on everyone she interacts with. Here comes the society and their unknown, unconscious non-neutral ( observe I don’t say bias) opinions. ‘Papa’s car’. , ‘Mummy will cook’, ‘Grandpa, papa, and uncle’s office but grandma and mom’s house chore’- these bother me ; but I am absolutely unsure of whether they should bother me.
Like in machine learning techniques, either you gate-keep the incoming information, which is not practical in a social setup; or you start throwing adversarial or neutralizing information at the model ( here, if you haven’t realized the model is my child), I am planning to adopt this second approach to fix things bothering me. Let’s see where we land.
Unconscious (or implicit) bias is a term that describes the associations we hold, outside our conscious awareness and control. Unconscious bias affects everyone.
While as a parent I think it’s a very big deal and I am worried about its impact on my kid’s foundational learning, it's just as important to be aware of the same in a workplace environment.
Workplace-related implicit or unconscious biases are typically categorized into a few types. Harvard University’s public policy document (https://www.hsph.harvard.edu/wp-content/uploads/sites/2597/2022/06/Types-of-Bias-Ways-to-Manage-Bias_HANDOUT-1.pdf) defines them as follows.
- Affinity bias – The tendency to gravitate toward people similar to ourselves. That might mean hiring or promoting someone who shares the same race, gender, age, or educational background.
- Attribution bias – The tendency to attribute our own successes to our skills and talents and our failures to things outside of our control. Yet, when it comes to other people, there is a tendency to think the opposite. If someone else has done something well, we consider them lucky, and if they’ve done something badly, we tend to think it’s due to their personality or bad behavior.
- Confirmation bias – The tendency to look for pieces of information that support our pre-existing views and ignore data that contradicts our views. When we succumb to it, early interactions and experiences of others can go on to influence our lasting, long-term feelings towards them, regardless of their current actions or performance.
- Cultural bias – When we stereotype individuals based on their country of origin, religion, or ethnic background without actually looking into their skills or performance levels. Cultural bias involves a prejudice or highlighted distinction in viewpoint that suggests a preference of one culture over another and can be described as discriminative.
- Halo Effect – The tendency to draw a general impression about an individual on the basis of a single characteristic. This could look like putting someone on a pedestal or thinking more highly of them after learning something impressive about them.
- Horns Effect – The tendency to see one bad thing about a person and form a complete view of them based on that single negative attribute, letting it cloud our opinions of all of their other attributes. It is the direct opposite of the halo effect.
As an individual, Halo and the Horns effect are something that I believe all of us are vulnerable to, especially in a high-impact, fast pace environment.
How I ensure I don’t display this behavior? Take this test (https://implicit.harvard.edu/implicit/takeatouchtest.html)to calibrate yourself. It has definitely helped me with self-awareness.
