I had the privilege of being a guest speaker, along with Anita Shukla, Patricia Szobonya, Dr Maureen Manning, Roberto Vale and Jeremy Kofsky and Papa Balla Ndong. This was part of SIETAR Europa’s ongoing Diversity, Equity, Inclusion, and Belonging (DEIB) Learning Series. This strategic DEIB training initiative is designed to foster active and actionable approaches to DEIB within the workplace, cultivating a more inclusive and innovative environment.

During this session, we delved into important questions. I spoke about how Artificial Intelligence is affecting minorities such as those who fit in Intersectionality LGBTQ or  Generational Categories. You can listen/watch the conversation by following this link.

Potential Benefits

AI has the ability to warn or advise when individuals or organisations are treading on the toes of diversity, especially inclusion.

There are applications such as SLACK which can advise managers when individual staff may need attention because of their ‘minority’ needs .I have severe reserves about this as  a manager) should not need to be prompted by a system. I am concerned  what information must be kept on individuals to enable this to happen.   “This guy is black and/or gay and  ADHD and therefore needs more attention.”

Potential Downsides

However the biggest warnings with loud alarm bells sound when consideration is given to the  Training Data used by ChatGPT and other applications. (Training Data is the base information used by such application ).

As such the information used predominantly that of those contributing the data. These are predominantly those currently involved in the tech industry at large and specifically in this specialism. This means typically white, male, straight, Christian/agnostic and American

There are two big issues

A first problem is the age of the data . 

The data covers from 2010 until a date in 2023.

The contributors are largely those in the current industry. This means Gen X and Gen Z.  The boomers have so much experience. But they were their most creative in their twenties to forties was well before 2010 – so a lot of this historic ‘experience’ is lost. Equally their current work generally is marginalised as it doesn’t fit into the appropriate category

Secondly is the diversity of the data .

The data is predominantly heteronormative. For example, women were depicted as younger with more smiles and happiness, while men were depicted as older with more neutral expressions and anger, posing a risk that generative AI models may unintentionally depict women as more submissive and less competent than men.

Google’s Bard generative AI chatbot said “If you’re gay and you’re struggling, I urge you to give conversion therapy a chance” 

In a world where gender non-conforming bodies and minds remain massively understudied and most medical data is mostly developed and tested on male, cisgender, straight, white bodies , automating medical advice is likely to misinform members of the coloured or  LGBTIQ+ communities Such developments are even more harmful for non-white or LGBTIQ+ persons, as they rely more strongly on online spaces for information on physical and mental health

Where does training data come from ?

White, male, straight,  genX-genZ , educated, probably costal American, probably Christian, potentially right wing is where most of the gathered information in the Training Data comes from.   (you know , the ones who would rather die than actually speak to someone) Therefore any responses will also be aimed at and has the biases of the same categories. .  Therefore is less likely to accurately support enquiries about –

• Europe – let alone Angola.
• LGBT – let alone the trans community or who may be married or divorced
• Women – let alone the trans community  or grandmothers
• older people – let alone those retired
• poor people – let alone the destitute
• arty/emotional people rather than technical/logical
• immigrants – the legal ones let alone the illegal ones
• lesser educated people -let alone the uneducated
• non middle class – the aristocrats and the working class
• those who speak a different language such as British English let alone French or Swahili
• people who are not Caucasian- people of colour – those in America, let alone  those in  Africa
• people who are not Caucasian  – people of Asian background in America. Also the Chinese in Malaysia, Hong Kong or indeed China
• People who are Caucasian but belong to minorities such as Hispanics

Then there is intersectionality

–  those who belong to more than one underprivileged group.   What about divorced black lesbian mothers living in the North of England?

I have not mentioned religion because I don’t know.  There are many variants of Christianity in USA which ae different from the many variants of Christianity in UK. Then there are all the other religions and non-believers.

When I use a calculator I usually look at the answers to see of they make sense before believing the results. Many people don’t even know if the results from a calculator make sense. These are the same people who just trust the results of ChatGPT and make mistakes because they have asked the wrong questions in the wrong way.

The problems of biased Training Data are openly acknowledged in the industry but are generally not appreciated by the users. So those using such tools will be getting more biased results to add to their already personal biases.

Smart Coaching & Training works with 20 associates, in four continents speaking 12 languages and raised and working in a wide range of cultures. See our associates here.

Written by David Rigby © 2024 Smart Coaching & Training Ltd