Artificial intelligence forces humans to face their own cognitive biases. The way of conceiving and programming machines is subjective, based on the experiences and social environment of the programmer.
As AI is now ubiquitous in both our personal and professional lives, it can have a major impact on society, especially concerning discrimination. Since human bias is at the source of the discrimination, a way to reduce it may be to teach the engineers who are doing the AI programming.
A Data-based Revolutionary Technology
What is artificial intelligence and how is it biased? Imagine a situation in which there is a need to develop a programme that is capable of determining whether an animal is a dog or a cat, based on a photo. In the past, developers would code “if”, “when” and “else” statements using criteria such as height, fur, and color to determine the nature of the species in the photo. Today, as Sonia Abecassis Le Lan of IBM explains, artificial intelligence uses sets of data provided to the machine to teach it what a cat or a dog should look like. With thousands of photos of cats and dogs, the machine learns to recognise the two different species thanks to similarities between the photos. The machine is able to solve the problem by answering statistically which solution has the biggest probability of being the right one. It could, for example, conclude that the photo has a 95% probability of representing a dog. This is called machine learning and it is the foundation of artificial intelligence.
An Impressive Yet Flawed Technology
Since machine learning and algorithms are fed with important amounts of data, their predictions are subject to bias because of a lack of diversity or irrelevance in the data. When using for example a facial recognition algorithm, if the database does not include enough diversified entries, it may not be able to correctly identify the different faces. Jeremy Patrick Schneider from IBM Interactive compares AI to a child who only knows what it has been taught: if the child has lived his whole life in a room, it will only know how to behave in that one room. When going outside for the first time, the child will not be able to react accordingly to various situations, because of his lack of experience. The scope of information that is fed into the machine is thus important in order to deal with multiple situations that might occur. Another problem linked with artificial intelligence, according to Schneider, is that the technology has spread very quickly, not giving enough time for scholars, politicians and others to thoroughly and rigorously test it and implement regulations to restrict biases.
From Humans to Machines: How Our Biases Affect AI
The technology itself is not the one to be blamed, the humans who created it are. Every human being has unconscious biases that are developed through their life: their culture, education, and experiences combine together to create cognitive biases. Machines merely reproduce them. Mickael Dell’ova at Ubisoft gives the example of a well-intentioned colleague who wanted to make an inclusive video game by adding a lesbian character as the main protagonist of their triple A grand strategy game. His colleague thought it would be inclusive if the lesbian character had stereotypical short hair and a Perfecto leather jacket. This cartoonish representation, despite the colleague’s best intentions, shows a clear unconscious bias. Such biases are mostly accidental, but being aware of them could help reduce their frequency.
If machine bias is ignored, products will be biased as well.
The Changes That Can Be Made
The first improvement that could be made is to give artificial intelligence access to diversity when fed with data. Irène Balmès underlines that algorithm designers need to carefully select the data given to AI and to clean it of bias through multiple tests and checks. More diversified teams would be a real asset in bias detection. Team diversification is only one solution. Educating the teams by providing training on ethical matters can help curb biases. As Balmès says, it is unfortunate that there is not more communication between the fields of engineering and social science. Technology has always been questioned from a philosophical perspective. As concerns artificial intelligence, it is fundamental to think about the ethical issues as well and include the political perspective. Finally, to aid in the creation of diverse team members, the recruiting process should be reviewed, from the job description down to the hiring phase.
What To Remember
Machines have biases. They merely reproduce the cognitive biases that algorithm designers have, which are deeply rooted in our society because of systemic discrimination. But this can change with more diversity and inclusion initiatives in the technology industry. We would like to take this as an opportunity to reflect on our own judgments and ways of thinking, and how they can affect others around us.
The above article has been derived from a webinar organised by LGBT Talents, with the following speakers: Sonia Abecassis Le Lan, Irène Balmes, Jeremy Patrick Schneider and Mickaël Dell’ova.
License and Republishing
The Choice - Republishing rules
We publish under a Creative Commons license with the following characteristics Attribution/Sharealike.
- You may not make any changes to the articles published on our site, except for dates, locations (according to the news, if necessary), and your editorial policy. The content must be reproduced and represented by the licensee as published by The Choice, without any cuts, additions, insertions, reductions, alterations or any other modifications.If changes are planned in the text, they must be made in agreement with the author before publication.
- Please make sure to cite the authors of the articles, ideally at the beginning of your republication.
- It is mandatory to cite The Choice and include a link to its homepage or the URL of thearticle. Insertion of The Choice’s logo is highly recommended.
- The sale of our articles in a separate way, in their entirety or in extracts, is not allowed , but you can publish them on pages including advertisements.
- Please request permission before republishing any of the images or pictures contained in our articles. Some of them are not available for republishing without authorization and payment. Please check the terms available in the image caption. However, it is possible to remove images or pictures used by The Choice or replace them with your own.
- Systematic and/or complete republication of the articles and content available on The Choice is prohibited.
- Republishing The Choice articles on a site whose access is entirely available by payment or by subscription is prohibited.
- For websites where access to digital content is restricted by a paywall, republication of The Choice articles, in their entirety, must be on the open access portion of those sites.
- The Choice reserves the right to enter into separate written agreements for the republication of its articles, under the non-exclusive Creative Commons licenses and with the permission of the authors. Please contact The Choice if you are interested at contact@the-choice.org.
Individual cases
Extracts: It is recommended that after republishing the first few lines or a paragraph of an article, you indicate "The entire article is available on ESCP’s media, The Choice" with a link to the article.
Citations: Citations of articles written by authors from The Choice should include a link to the URL of the authors’ article.
Translations: Translations may be considered modifications under The Choice's Creative Commons license, therefore these are not permitted without the approval of the article's author.
Modifications: Modifications are not permitted under the Creative Commons license of The Choice. However, authors may be contacted for authorization, prior to any publication, where a modification is planned. Without express consent, The Choice is not bound by any changes made to its content when republished.
Authorized connections / copyright assignment forms: Their use is not necessary as long as the republishing rules of this article are respected.
Print: The Choice articles can be republished according to the rules mentioned above, without the need to include the view counter and links in a printed version.
If you choose this option, please send an image of the republished article to The Choice team so that the author can review it.
Podcasts and videos: Videos and podcasts whose copyrights belong to The Choice are also under a Creative Commons license. Therefore, the same republishing rules apply to them.
