AI has made its way into most people’s lives in one way or another, but how much do you know about the tool and more specifically, AI bias?

Real-world examples

Most AI tools learn by analysing data, using a mix of patterns and algorithms to determine answers. However, if they are taking this information from humans, with their own biases and prejudices, what’s to stop AI from developing its own? 

The answer is nothing. 

Amazon had developed a recruitment tool which quickly had to be stopped due to the tool’s sexist views’. The AI was trained on data submitted over ten years, but most of the data came from male applicants. Developed in 2014, by 2015 it became clear that the system was not rating candidates from a gender-neutral perspective. The data it was built on was gathered mostly from men. 

Another famous example of AI bias was an American algorithm called ‘Compas’. The Correctional Offender Management Profiling for Alternative Sanctions. COMPAS was proven to incorrectly label black defendants as more likely to re-offend. COMPAS wrongly flagged black defendants at almost 2x the rate as white people (45% compared to 24%). This system was used across hundreds of US courts. Systemic racism in a very demonstrable form.

As Microsoft researcher Kate Crawford said, “Histories of discrimination can live on in digital platforms, and if they go unquestioned, they become part of the logic of everyday algorithmic systems”. Situations like this demonstrate the need for continuous fact-checking. As AI becomes more ingrained in our society, we must work to prevent bias-based errors like this from occurring.

So what does this mean? 

You may think that AI as a machine has no bias, but sadly this is not true. There are many types of AI bias and it’s hard to pinpoint the main forms. This is because each programme learns through different methods. However, there are some types of bias more recognised than others. 

The five most recognised types of AI bias are:

Selection Bias: This occurs when the training data does not accurately represent the real-world population. Leading AI to make an incorrect assumption. If a facial recognition model is trained primarily on one skin tone, it may struggle to recognise those with other skin tones, leading to discrimination.

Confirmation Bias: This happens when the AI relies too heavily on previous information it has learnt. This means it reaffirms previous historical bias found in the data or bias that it has learnt before. Such as the stereotype that women are less qualified to work in tech. This alongside Stereotyping bias was a huge factor in causing the Amazon scandal.

Stereotyping Bias: Also called Prejudice bias. This is a bias that is trained into the machine and reflects societal prejudices, stereotypes and assumptions.  Many virtual assistants, such as Siri or Alexa, are designed with female voices and programmed to be polite and accommodating. Reinforcing the stereotype that women should be helpful and submissive which can unintentionally perpetuate gender biases in society.

Measurement Bias: Similar to sample selection bias, measurement bias occurs when the data collection is not completed, or there is an error during collection. For example, if a university sends out a course satisfaction survey, but only contacts those who stayed enrolled for the full three years and not members who dropped out or changed courses, they’re not getting the full picture.

Out-Group Homogeneity Bias: This is where AI generalises data groups on which it has less data. Grouping smaller groups even though they may have very few similarities. For example, Facial recognition systems have difficulty distinguishing individuals from racial or ethnic minority groups. This is due to a lack of diversity in the training data.

These are also addressed in our why AI isn’t replacing marketers article. 

Why is AI bias such a pressing issue?

AI bias is such an important issue because it can often go unnoticed, especially with how integrated AI bias is in modern society. If bias is left unchecked it can play into social stereotypes and enhance misinformation. =

In America, an AI that was being used across many U.S. health systems prioritised ‘healthier white patients’ over ‘sicker black patients’ when it came to providing care management. This is because it was incorrectly trained on the wrong data, being trained on cost data not care needs. Additionally, algorithms may incorrectly predict health risks for groups of people that it has less information on such as ethnic minorities. 

Researches in America also combed through 50,000 records and discovered “software guiding care for tens of millions of people systematically privileges white patients over black patients”. This happened consistently for chronic health conditions such as diabetes and kidney issues. The paper that raised this issue does not identify the company behind the algorithm, claiming that “the company has confirmed the problem and is working to address it”.  

This happened because the algorithm was not trained to take into account a person’s race when addressing their health issues. Its skewed results showed that even relatively “race-neutral formulas can still have discriminatory effects when they lean on data that reflects inequalities in society.” The skewed data highlights the importance of encouraging people from all ethnicities to take part in building AI, Algorithms, and research. 

 

What should we be doing?

Not all hope is lost! It is sometimes possible for AI to reduce bias. Bias reduction techniques (such as cross-checking, fact-checking and verification) can be trained into the tools. Making AI a more neutral perspective than some humans. However, as long as it makes decisions based on human data, there will be some form of bias. 

To reduce AI bias, we need to diversify the data we are using to train AI models. Focusing on being careful to not exclude or limit any data on categories such as ethnic groups, location, age and gender. There is also a need for continuous monitoring with AI tools. Checking for anomalies, data misinterpretation and misrepresentation. 

Another way to help reduce AI bias is by diversifying AI creators. Because AI is based on mostly historical data, it’s easy to amplify already existing biases. By encouraging people of all backgrounds to deepen their interest in AI building, we can ensure as many perspectives as possible to help reduce AI bias. 

 

Ways to reduce AI bias in your work.

1. Be aware of the tool you are using, who created it, what biases they may have
2. Not everything AI produces is fact, if it looks off, fact-check it. If it is contributing to anything of value, fact-checking should always be part of your AI use process
3. Consider your prompts, you can mitigate some forms of bias by taking the time to consider your prompt engineering. 
4. Add context, supply the tool with as wide a range of data as possible and encourage it to think and consider different perspectives.
5. Ask your AI to ‘explain your thought process’ and cite its sources so you can see how it gathered its data. 

Specificity is key, AI can only do so much. Sometimes AI will have a different idea of ‘diversity’ to you as will other humans, so you may have to spell it out for the machine.