King's College London

Since I last wrote for the Global Institute for Women’s Leadership on the inherent gender bias in Google Translate, some progress has been made. Google has added a new feature that allows for gender-specific translations for terms in gender-neutral languages. So now when you type “o bir doktor” in Turkish, you’ll get two translations: “she is a doctor” and “he is a doctor”. This moves us on from replicating gender biases that are common in real life, which Google has recognised as a problem.

But is this enough for dealing with gender bias in AI generally? The short answer is No. This is only a surface-level fix for a much larger and deeper social issue. The problem is not with software applications, it’s with the data. Data which comes from people.

Artificial Intelligence (AI) refers to the collection of tools and techniques designed to extend human capabilities of information acquisition, analysis, processing, and decision-making. AI based software trains on data sets in order to understand the world and perform its intended tasks. With almost 2.77 billion users, social media platforms provide big data which is cheap, easily available and gives information on users’ behavioural patterns including popular opinions and trends. The amount of unstructured data and the speed with which it is created on these platforms requires sophisticated AI algorithms for processing. The future of social media will depend entirely on AI, which in turn will determine the direction of future innovation in business and technology.

The problem lies in the fact that social media generates data based on user-generated content, including people’s often irrational and biased opinions. And it’s not just subtle or unconscious biases in opinions; the biggest contributor to data that AI algorithms draw on is hate speech. If biases aren’t tackled, this can corrupt the way software makes decisions, effectively immortalising hate speech in cyberspace.

At Swinburne University of Technology, we carried out a research project to understand the nature of hate speech on Twitter and how it contributes to gender bias. Our experiment analysed different kinds of hate speech, identified by an extensive list of keywords, in approximately 400,000 tweets in English on one day (10 October 2018). We performed sentiment analysis on the tweets and categorised the data on a scale from one (non-threatening) to five (threatening) based on the strength of hate exhibited. Within each category, we further analysed the tweets based on three types of hateful content which are not mutually exclusive – whether they were sexist, racist and/or religiously targeted – to understand the various forms of biases present in hate speech.

In our results, the bulk of the Twitter data (~95%) showed sexism as the most common form of hate speech. In the highest negative sentiment (threatening), racism was at the top, with sexism closely behind. Most notably, 83% of the hate speech we detected was mostly non-threatening sexist and racist jokes. This is consistent with research that shows the extent to which women experience sexist behaviour cloaked as humour in their daily lives.

Research has also revealed that tolerance for sexist jokes only increases the acceptance of gender discriminatory norms in society. It’s difficult to confront sexism in real life when it’s disguised as a joke, but it’s also difficult for AI to identify and discount it. When it comes to AI algorithms learning from social media data, a software programme cannot understand something intended as humour – however insidious – and would not categorise these tweets as offensive “jokes” that should be actively guarded against. The impact of this can be devastating. Given our reliance on AI-based systems in this supposed era of “big data”, the danger is that when AI algorithms are provided with this type of pattern in data, it creates a sexist social reality in cyberspace.

The words most often associated with the hate speech we analysed were either “she” or the use of feminine derogatory words b***h or c**t, even when targeted at men. The numerous sexist hashtags such as #FeminismIsCancer, #LiesToldByFemales, #IHateFemalesWho, #ToxicFeminism, #FakeMetoo, #boyswillbeboys, #FeministMafia, #Feminazis, #MenAreAwesome and #WarOnMen made things worse by encouraging people to retweet or comment, even if intended as humour.

Looking at the individual tweets within each category for sexism, the highest percentage of “body shaming” tweets were targeted at visible women, primarily from politics and the entertainment industry. In our analysis, political events with trends and hashtags on Twitter were associated with higher rates of increasingly threatening hate speech.

With more than 500 million tweets per day, the nature of hate speech is largely dependent on socio-political events or trends going on around the world. Even without any significant major trending on the day of our Twitter data analysis, sexist and misogynistic tweets constituted the main part of the observed data. Similar experiments would likely produce even more threatening examples of sexist hate speech if replicated around the time of the Gillette advertisement on “toxic masculinity” and the associated viral hashtags #GilletteAd and #GilletteBoycott.

Our experiment using Twitter data shows the extent to which biased and prejudiced user-generated content from ordinary people resonates and replicates continually on social media. It is time that we acknowledge our role in promoting hate online and contributing our biases to digital history. AI programmes are only learning from us. It is not AI that is biased, it is us.

Dr Muneera Bano is a lecturer in software engineering at Swinburne University of Technology in Australia and one of the Superstars of STEM at Science and Technology Australia.