AI learns another bias from us, thinks improvement equals addition

"Adding sugar makes coffee better"- GPT-3. Don't @ me.
Amal Jos Chacko
Is AI biased.jpg
Is AI biased?


Humans pride themselves to be the most intelligent of animals, but are we as intelligent as it gets; are our decisions and acts as rational and logical as they can be? Well, not quite. 

It is established that human beings are prone to biases- prejudices towards something or someone. Confirmation bias- the unconscious tendency to seek information to confirm what we already believe; and Self-serving bias- an inclination to hold ourselves responsible for our successes yet faulting external factors for our failures- are two well-known examples.

Making things worse in an attempt to improve

Linguistic bias is another of these- a propensity to use phrases and words that reinforce a set of assumptions. Research shows we tend to associate the concept of ‘improvement’ more closely with addition, rather than subtraction.

This causes us to lean towards overcomplicating things while making decisions, even with good intentions. 

A study by an international research team from the Universities of Birmingham, Glasgow, Potsdam, and Northumbria University in Cognitive Science builds on this research. “We found that the same bias is deeply embedded in the English language,” said Dr. Bodo Winter, Associate Professor in Cognitive Linguistics at the University of Birmingham. 

“The word ’improve’ is closer in meaning to words like ‘add’ and ‘increase’, rather than to ‘subtract’ and ‘decrease’. So when somebody at a meeting says, ‘Does anybody have ideas for how we could improve this?’, it will already, implicitly, contain a call for improving by adding rather than subtracting,” he added. 

Linguistic addition bias can worsen things rather than improve them if ignored. Adding to excessive bureaucracy instead of subtracting is an example. The research further finds that verbs of change, such as ‘to modify’, ‘to revise’, or ‘to enhance’ are similarly perceived.

What makes it even more interesting is that the reverse of the bias holds too. Words related to addition are found more often in ‘improvement’ contexts rather than subtraction-related ones, indicating addition bias at multiple levels of the English language structure and use.


AI- amplifier of human flaws

In this day and age where our experiences and traits echo in Artificial Intelligence systems that are increasingly becoming prevalent, it is no surprise that even chatbots have inherited this bias. 

When asked what it thought of the word ‘add’, GPT-3, the predecessor of ChatGPT replied: “The word ‘add’ is a positive word. Adding something to something to something else usually makes it better”. 

GPT-3 cited examples of adding sugar to coffee to improve its taste and making new friends making us happier. Perhaps we could all do with the latter. 

Dr. Winter observes the need to take an extra second to consider our choices when asked to suggest improvements. “The positive addition bias in the English language can influence our decisions and mean we are predisposed to add more layers, more levels, when in fact we might benefit from simplifying.”, he concluded.