Send me a message

Daniel Hulme
Daniel Hulme


Does AI represent me?

Posted on 14 July 2017

I often browse the discovery mode of my “TED talks” app to venture into the land of topics I wouldn’t usually explore. After a few clicks to generate more suggestions, I could not skip the title saying, “Why do ambitious women have flat heads?”. The speaker was Steve Shirley, on paper, and Dame Stephanie Shirley, in reality. She founded a software company called Freelance Programmers in 1962 which created jobs for women in science with dependents who needed more flexible work conditions. The company was a great success. Starting with only £6 of capital, it reached a value of £150 million by 1980 and today is part of Sopra Steria Group. To combat prejudice and gender imbalances, Dame Shirley used her male pseudonym in her business communications to propel her company to success. This is, undoubtedly, an inspiring story for women to follow careers in science. However, at the height of its success, and as a result of its female first preferences, only 3 of the companies 300 employees were male. Something about this doesn’t sound quite right. The pro-female policy ended with the passing of the Sex Discrimination Act in 1975.

If we aspire for equality, or rather, equity and inclusivity, then avoiding reverse bias is key. Women have famously hidden behind pseudonyms to have their work recognised throughout history. Inclusivity is not restricted to gender discrimination; religion, ethnicity, race and sexual orientation are all issues that people have long fought to achieve acceptance for in wider society.

If we aspire for equality, or rather, equity and inclusivity, then avoiding reverse bias is key.

Seventeen years into the new millennium, we like to think these issues have been relegated to the history books, but in reality, the same issues are still abundant – just in shiny new outfits. We believe Artificial Intelligence will save the day by eliminating the prejudice that human intervention usually leads to. Technological advancement and the increasing availability of data have empowered humans to create remarkable algorithms that can look deeper into our lives than we ever could before. Hiring for a new position will be based solely on the expertise of a candidate rather than their race, gender, age or religious beliefs. Promotions and salary increases will be given to the people who have the highest merit rather than the right connections. AI will dive into the data with its magic wand and present the right decision to us – sounds like a fairy tale come true, doesn’t it? So what’s out there to make all this magic happen?

AI will dive into the data with its magic wand and present the right decision to us – sounds like a fairy tale…

Natural Language Processing (NLP) algorithms can be used to capture the knowledge that disseminates throughout human culture – continuously learning from the data we produce. Matrix factorization methods have been the rockstar on the scene of NLP algorithms. These use low-rank approximations to decompose matrices of words and capture statistical information about them. This method is used in topic modelling where abstract topics can be discovered in a collection of word documents. Twitter, for example, provides an ocean of data where topic modelling can be applied to dig deep into what people are discussing. The downside of matrix factorization is that the number of topics needs to be defined prior to the analysis. Additionally, this method does not account for the sequence in which words are being said.

The Stanford Natural Language Processing Group recently developed the more robust GloVe algorithm, which analyses different words and suggests how they are associated based on their similarity. Tomato and Leek are close pals since both are vegetables. Steam and Ice also get along since they’re associated with water. This is a powerful technique – having 5 different programming languages on your CV along with industry experience can launch you to the front of the start line ahead of tonnes of other candidates for a software engineering position without your future employer ever having to look at any CVs. So far, this sounds like something from a sci-fi movie!

BUT! As the saying goes “You cannot escape from your past”. Our supposedly unbiased, accurate and reliable AI models have been developed by learning from data that was produced by humans, and as a result has adopted the human biases we set out to rid of. Some 67 years ago, Alan Turing posed the famous question “Can machines think?”. If we were to pose the same question of today’s computers, the answer to this question would most likely be yes. Our fancy algorithms learn from what humanity has generated, and our prejudices still exist. Words like surgeon and promotion are associated with men, while words representing ethnic groups are related to immigration or cheap labour. The fact that AI learns from existing (and very human) data puts us in danger of not fully avoiding biases. Our good intentions may have created additional problems rather than simply providing an “easy” solution to the existing ones.

A snapshot of initial statistics based on internally generated Slack data. The aim was to begin to understand how the behaviour of women differs from that of men on our main online communication platform.

As a technology company devoted to self-organisation, non-hierarchical structure, employee happiness and data science, Satalia strives for AI that makes difficult decisions without succumbing to the bias of humans.  Our team is made up of people from 17 countries (a number only just smaller than a number of desks we have in our UK office) and 22% of us are women – both aspects of our culture we’re proud of, and continuously investing in.

Satalia’s largest internal project is ‘Semantic’, which focuses on the development of a suite of organisational analysis tools. A fair salary scheme and our pay review process are some of the most challenging problems we are currently tackling. Naturally, we wanted to develop a set of robust data science based methods to untangle us from the damaging human biases that typically lace these types of decisions. Collecting external data about the market and exploring our internal data to show the development of each individual and their strategic position within the company are only the very first steps we have taken. Deep discussions have been at the heart of our “inclusivity” interest group where everyone has offered a perspective on the problem.

How do we build AI without bias?

Being aware of the pitfalls of AI has enabled us to consider that women may not feel as confident to request higher salaries, or appear as trustworthy, according to our NLP algorithm. Another approach we explored is the variation in behaviour between men and women (Figure 1). Identifying inspiring individuals, introverts and how they conduct themselves using NLP techniques has given a reflection of the behaviour that companies rarely explore. One conversation is all it took to start all this amazing pioneering work to develop an organisational culture where AI would be representative of each and every person.

Recognising nested associations in this type of technology is the first step towards avoiding blind trust in machines. Attending big AI and machine learning conferences proves this point – the most influential and numerous attendees are white middle aged males. If inclusivity and fair representation in AI are what we strive for in the world, then is it the right strategy to put all the responsibility into the hands of a homogenous group of people? Let’s educate young people, let’s question our actions at every step, let’s collaborate and engineer a future that fits all!


Related topics: AIEqualityInclusivity

Something to add?

Let's take the discussion offline

Dessie Petrova

Dessie Petrova

Start a conversation

Send me a message

Dessie Petrova
Dessie Petrova


Message sent

Thank you for contacting us, we will come back to you as soon as possible