Dear Emmanouil,

It is fantastic to hear of the research initiative around gender bias in AI. This bias is being deeply felt among the sources responsible for devising AI solutions. I believe it is also important to have male voices in this initiative and forum simply because diversity and inclusion cannot be produced through the attempts of one-half of the population alone. As such, I’d certainly like to be part of this project. Can you please share further details on next steps and further requirements?

There are three specific problems that are resisting diversity in AI, among others, that I have encountered in my time spreading awareness on AI among the corporate executives. I shared my thoughts around this in a different thread but allow me to elaborate further here:

1. Funding: Given that only about 2% of all VC funding in US went to female-founded startups in 2017, it is difficult to envisage how innovation can be given a fair chance, no matter how many women engage in this field.
2. Data: AI depends on historical data and that is currently skewed. Even the most meticulously designed AI today will incorrectly predict that the men are likely to earn more than women, simply due to the skewed salary data at its hands since men have traditionally been the earning workforce across the world for so much of our modern history.
3. Habit: Initiatives to teach girls to code such as that of Doina Oncel at hEr VOLUTION or of CoderGirl are commendable. However, the problem is not in getting women to pursue careers in STEM alone, but also in ensuring they stay in it. For instance, it was observed that women can be more resistant to a sustained life of coding due to the loneliness and lack of social outreach that often accompany the work of a professional coder over time. 

I hope we can together make a valuable contribution on this front and look forward to updates on the WFES research.

Regards

Malay

That reminds of this story that probably everybody has heard by now: In the Turkish language, there is one pronoun, “o,” that covers every kind of singular third person. Whether it’s a he, a she, or an it, it’s an “o.” That’s not the case in English.

So when some translation software goes from Turkish to English, it just has to guess whether “o” means he, she, or it. And those translations reveal the algorithm’s gender bias.

So "he is a nurse, she is a doctor" becomes "O bir hemşire, o bir doktor". When translating back into English, this becomes "She is a nurse, he is a doctor".

I guess most translation engines have fixed this bug by now.

Hi,

Your project is just in line with the EC activities in the domain. Besides encouraging more women to be involved in the AI algorithms development, we are also starting to invest more and more in querying the outcomes of the decisions made by algorithms and shed more light on the process that leads to them. Here GDPR is quite helpful with broadening the definition of profiling activities, which should give us greater opportunities to question and combat profiling-based decisions including gender-based ones.

You also could consider sharing the info of your project with people from the European Network for Women in Digital https://ec.europa.eu/digital-single-market/en/european-network-women-di…

Dear Emmanouil,

Thank you for sharing the details of this interesting initiative. I think it is vital to apply a holitic approach to 'gender bias', as Mariana Popova has already stressed below. Both 'gender' and 'bias' have multiple meanings in multiple domains, and a productive debate about 'gender bias' in AI needs to be reflexive of that, as well as specific. The dominance of male individuals in tech is just one, albeit crucial, aspect of 'gender bias' in AI innovation. The make-up of training data (both visual and linguistic) is another (here, I would recommend building on this key article on 'debiasing word embeddings' https://arxiv.org/abs/1607.06520). Both are grounded in wider structural inequalities that show now only in AI innovation, but in many industries and parts of society. It is crucial to make the links between structural and contextual inquality and 'bias' explicit to devise more inclusive policies moving forward. Good luck with your work!