Have you ever thought on gender approach when concerning artificial intelligence? Well, recently gender bias in artificial intelligence (AI) has been under (some!) fire.
According to recent reports, AI systems have been adopting gender bias. The fact is that, even if AI appears to be neutral, these machine learning systems are developed by humans and based on natural language processing (NLP), which is based on human language and therefore reflect some of the human bias. Simply put, these models tend to internalize and perpetuate human bias, namely gender bias.
How artificial intelligence can perpetuate gender bias
Artificial intelligence has been a hot topic for some time now and many even state that it’s on the verge of revolutionizing and (possibly) reshaping businesses and governments. So far so good. But there are some problems to solve yet.
The truth is that digitalization is taking a big part in everyone’s life, most of the time one has no awareness of it. As it is, currently, algorithms and artificial intelligence have a huge impact on society and people’s choices.
As it goes, AI promises to overcome some of the human limitations like processing speed and thought and even to open up a wide range of possibilities for how people live and work.
This being said, some reports claim that artificial intelligence might have implications for human bias. While AI systems are, by themselves neutral, the fact is that they might embed some of the human cognitive constraints, such as gender equality.
These machine learning systems tend to absorb the information they’re fed up with and, subsequently, they learn from it. So, AI and automation are not developed in a gender-responsive way, this means that AI is likely to reproduce (and even) to reinforce existing stereotypes and discriminating social norms.
Where gender bias in AI occurs?
The origin of the gender bias in artificial intelligence is clear: it’s not other than human biases. The question now is where does it happen? Basically, in NLP programs. The fact is that NLP is the main ‘ingredient’ for most AI systems, mostly known as Alexa (Amazon), Siri (Apple) or Google Home, for instance (to name just a few).
These systems have displayed some gender biases, whether it was on computer vision systems for gender recognition or on word embeddings. Among the listed bias, there are reports on higher error rates when recognizing women, namely women with darker skin tones.
When speaking of word embeddings, it’s important to understand how computer programs process text and the used algorithms for that. In short, words are represented by lists of numbers, defined as word embeddings, which encode information about a specific word (namely, its meaning, usage, and properties), and are then used as inputs in natural language processing models. Word embeddings usually represent words both as a sequence or a vector of numbers. This means that embeddings tend to process information considering the context in which a specific word appears. So, for two similar words, both embeddings will be similar between them (mathematically speaking). This means that AI systems can relate gender to a specific word, let’s say: man for king and woman for queen.
The big issue is for those cases in which AI analogies create some linguistic context that perpetuates gender biases. For instance: a man appears as a computer programmer and a woman as a housewife, or a man appears as a doctor and a woman as a nurse.
What causes this gender bias in AI?
According to Harvard Business Review, there are several factors that might lead to gender bias in AI, such as:
1. An incomplete or skewed training dataset
If some demographic categories are missing from the training data, models won’t be able to scale accurately when applied to new data including some of those categories.
2. Labels used for training
Most AI systems use supervised machine learning, which ultimately label the training data to teach the model how to behave. So a possible misclassification and bias towards a particular gender category will be encoded into the model, thus perpetuating the bias.
3. Features and modeling techniques
For way too long, text-to-speech technology and automatic speech recognition (speech-to-text technology), for instance, tend to use typically male voices. This happened because of the way that speech was analyzed and modeled, which indicated that it was more accurate for speakers with certain characteristics (taller speakers with longer vocal cords and lower-pitched voices).
This means that the measurements used as inputs for machine-learning models or the actual model training itself can lead to bias and therefore must be taken into consideration.
How to approach gender bias in artificial intelligence
The only way to tackle and overcome gender bias in AI is by making an extra effort to minimize risks. And this starts within the machine learning teams. Here are some ideas:
- Ensure that developers within the AI developing team have different backgrounds;
- Make sure that training samples are as diverse as possible, in terms of gender, ethnicity, age, sexuality, etc.;
Spur the machine-learning teams to measure the accuracy levels:
- Individually for diverse demographic categories;
- Raise the machine-learning team awareness to identify those categories that might be treated unfairly;
- Collect more training data associated with sensitive groups to diminish an unfavorable treatment. And after that apply up-to-date machine learning de-biasing techniques to reduce recognition errors for the primary variable and to include additional penalization for producing unfairness.
On the way to gender equality even in AI
Gender equality or gender bias (as you prefer) has a long way to go yet. And the AI industry it’s not an exception. It also has to work towards equality in its approach and perspective. The first step for it is to diversify the workforce creating these new technologies and reduce some stigmas within our society that transfer to AI.
Also, future research work on AI must consider a broader representation of gender variants (not only female, but also transgender, non-binary, etc.) that might help increase the understanding of how to handle expanding diversity and to reduce bias.
