A good place to start is to understand your own biases. You can take the

Harvard Implicit Bias Test

⁠

and encourage your team members to do the same.

Conduct a bias review to understand where bias may be present in your current people, practices and products in order to identify ways in which to counter it.

Human bias

For example: A product manager may only seek out feedback and user data that confirms their initial product design decisions, disregarding data that suggests a different approach would be more effective.

For example: A product team may make decisions based on common user feedback. This may only represent the perspective of a small, homogeneous group of users, leading to the exclusion of other user groups, where more effort is required to understand their needs and preferences.

For example: A product team may conduct initial user research and gather feedback that suggests a certain feature is critical to the success of the product. As a result, the final product may fail to be inclusive for a diverse range of users because it is anchored to the initial feedback from a narrow sample of users.

For example: A marketing team may present a new feature in a way that emphasises its potential benefits for a specific user group, leading to the exclusion of other user groups who may not find it relevant or useful.

For example: A product manager may prioritise features and designs that cater to a specific group of users who share the same background or interests as themself, neglecting the needs and preferences of other user groups with different backgrounds or interests. This can lead to a lack of diversity and inclusiveness in the product. Affinity bias can also play out when hiring new teams members too, where people who are more similar to the interviewers are favoured.

For example: A product manager may prioritise features and designs that cater to the needs and preferences of senior executives or stakeholders, neglecting the needs and preferences of entry-level employees, lower-income or marginalised users. This can lead to a lack of diversity and inclusiveness in the product as diversity tends to decrease as you rise up the ranks!

For example: A designer might assume that older users are not technologically savvy and do not require or desire complex features.

For example: A data scientist may only use data from a narrow group of users to train a machine learning model, neglecting to include data from diverse user groups. This can result in a model that has a bias towards the narrow group of users and fails to accurately predict outcomes for other user groups, leading to a lack of representation and exclusion of these groups in the product or service.

When airbags were first introduced, there were higher numbers of fatalities with women and children because all of the crash test dummies were tall males.

Countering bias during user research

Build Rapport: By being aware of affinity bias, you might try to find other ways to help people who are very different to each other find common ground.

Clear Framing: Consider different ways to position your goals and request feedback to reduced the framing effect bias. Ensure you are providing enough guidance and structure so participants know what is expected of them without influencing how they respond.

Hypothesis Testing : Ask your team to forecast what they think they’ll see or hear in a user test or interview, before the event. It can then be helpful to refer back to this afterwards to see what insights surprised them and what assumptions were wrong. This can help to create a concrete discussion about confirmation bias or anchoring bias and help people to stop making assumptions.

Countering bias during recruitment

Actively seek out diverse talent: Don’t just ‘see who applies’.

Use structure and criteria: Don’t just rely on gut instinct.

Diversify the interview panel: Don’t let all decisions be made exclusively by white, middle class, straight, able bodied men.

Anonymise first stage: Consider removing info such as name which can imply gender and ethnicity for the initial screening stage.

Offer adaptations: Create an experience to help people shine, not put them through their paces.

Redefine success: Review what is a critical requirement for the job versus a nice to have, and what could be developed versus required on day one.

Algorithmic bias

For example, even the most advanced artificial intelligence is hindered by the inherently racist data it’s trained on.

⁠

Loading

www.newstatesman.com

⁠

⁠

For example, voice recognition technology can be biased against women, people of colour, and non-native speakers, because it is often trained on a narrow dataset of voices that may not represent the diversity of the population.

For example, On Aug. 9, 2023, a tutoring company agreed to pay $365,000 to settle an artificial intelligence (AI) lawsuit with the Equal Employment Opportunity Commission (EEOC), when the company’s recruitment AI was found to discriminate against people on the basis of age. I

Here are a number of examples of different types of algorithmic bias.

For example, using a bright lens camera to take pictures of different people will not reflect their real skin tones. When the algorithm is trained on these pictures, then the model will fail to appropriately recognise the difference in real skin tones and results will be skewed.

For example: Facial recognition software only trained on white faces fails for non white faces.

⁠

US government tests

⁠

find even top-performing facial recognition systems misidentify people who are black at rates five to 10 times higher than they do people who are white.

For example: If historic pay data is used to forecast future salaries, historic ethnicity and gender pay gaps will prevail.

Amazon used an

AI powered recruiting engine

⁠

to search for top talent. However, it penalised female candidates because the system was rating candidates based on the patterns of resumes submitted to the company over 10 years. As this was a tech company, a sector historically dominated by men, the program taught itself that male candidates were preferable.

A

study

⁠

found, that “broadly targeted ads on Facebook for supermarket cashier positions were shown to an audience of 85% women, while jobs with taxi companies went to an audience that was approximately 75% black. This is a quintessential case of an algorithm reproducing bias from the real world, without human intervention.This underscores the need for policymakers and platforms to carefully consider the role of the ad delivery optimisation run by ad platforms themselves---and not just the targeting choices of advertisers---in preventing discrimination in digital advertising. “

For example: Using postcode as a predictor of an individual's likelihood of criminal offense (or exam results).

For example: An employee well-being survey shows a good average, but when filtered to look at just black employees tells a very different story. If ethnicity is excluded as a data point for analysis this insight could be missed entirely.

For example: The categories in surveys usually enforce historical categorisations that are false, and exclude those who don't fit those categories. For example, gender options that only include male and female.

For example:

Tay

⁠

was an artificial intelligence chatter bot that was originally released by Microsoft Corporation via Twitter on March 23, 2016; it caused subsequent controversy when the bot began to post inflammatory and offensive tweets through its Twitter account, causing Microsoft to shut down the service only 16 hours after its launch.

Mitigate bias in AI: If your solutions leverage AI, prevent bias before launch, by educating everyone about risks of AI model bias and providing a common language to discuss model risk and how to mitigate it.

Review your processes, your data and your technology and ensure that humans who are most impacted by the model are “in the loop” when developing the model.

Resources