Why fighting against AI algorithm biases must be a proactive and collective effort
Artificial intelligence (AI) powers humans and societies. It accelerates our actions and decision-making process, it gives us access to precious information and analyses trends and behaviors at a speed and with a detail level no human has ever matched before.
There is no doubt about the need to legally frame the use of AI algorithms by companies and organizations. Designed by human beings, AI algorithms are often made of imperfections that can have high and sometimes dangerous consequences, such as discriminations and prejudices.
Defined by the CNIL as a series of steps making it possible to obtain a result from the elements supplied as input, algorithms can be composed of biases, that are systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others, according to Wikipedia. Unfortunately, a human cannot always figure out on his/her own whether or not he/she is victim of algorithmic biases, because they can be due to multiple reasons such as oversights, misconception as well as a too narrow understanding of datasets. That is why, without any kind of regulation or sensitization, AI algorithms tend to repeat much of the prejudice human beings can have about culture, gender, religion and more.
For a considerably long time, AI was mainly a research field with limited applications and results. However, for a few years, it has been become much more accessible and even mainstream. Its democratization reveals the importance of limiting and mastering algorithm biases to a maximum, since this technology is increasingly present in our day-to-day life.
Across the world, no country already developed a strong, efficient and technical way to audit companies and organizations regarding their algorithm biases. Indeed, identifying biases would be an insanely hard exercise, because their roots can be unlimited and approximative, and their investigation would be very humanly and timely intensive.
Hence, instead of putting efforts into developing ex-post regulations to sanction companies and organizations which developed biased algorithms, why don’t we invest in developing ex-ante prevention methods and policies to sensitize humans?
Algorithms can be as subjective as humans
Before even thinking about preventing algorithm biases, it is primordial to understand where they come from. Aurélie Jean, a French specialist in computer science explains that algorithms are not guilty, humans are: “We are not fully and explicitly aware of the biases which when revealed, become morally unacceptable or legally punishable. It became critical for developers and users, to be more responsible towards the technologies that we are using and/or developing. We tend blindly trust algorithms to make decisions for us because we think that mathematics is impartial; but the reality is more complex, and the algorithms are not guilty.”
Hence, it seems important to analyze the steps in the design of an algorithm where there is the biggest amount of human thinking and decision-making, because this is where biases may arise.
First of all, human decision takes place in the selection of the data on which the algorithm is trained, more specifically in machine learning. An algorithm trained on a set of data that is non-representative or non-diverse leads to the appearance of biases. Biases do not decrease with the volume of data used to train the algorithm, but with the diversity of the given data sample. Indeed, the data collected to train an algorithm may have been preferentially sampled, and if not, the data set can reflect existing societal inequalities that the data scientist or the developer who is building the algorithm is not even aware of.
Second, human decision appears in the way the algorithm is tested once trained, meaning how the algorithm which has been trained on a training dataset, is now evaluated against a test dataset. In order to make a thorough testing process, the algorithm must be compared to realistic scenarios made of representative and diverse data as it is for the training data set.
To finish, human decision takes place in the design of the algorithm’s decision models. The chosen decision model has to take into consideration the way data has been collected and processed until now, to balance the output. For instance, while developing a chatbot, designers must account for the interaction with the social-learning abilities of the algorithm.
These potential roots of biases are not the only ones, but they can be highly useful to understand where human decision takes place in the design of an AI algorithm.
Fighting against AI algorithm biases must be a collective effort
Given that AI algorithm biases seem to come from multiple and mainly human-based roots, fighting against them must arise from a collective and multidisciplinary ex-ante effort at the international level, not only from public institutions but from private organizations and individuals as well.
At the policy level, a collaboration between government advisors and experts in anthropology, law, psychology, sociology and business is needed to address these new tech challenges. It would be more than useful to see prevention policies from public institutions, to provide public and private organizations with tools to understand and prevent algorithms biases, like the CNIL does in France for the GDPR.
Regarding educational policies, Aurélie Jean suggests that democratizing how to code into educational policies would bring more diversity into public and private tech teams. Indeed, she explains that a more diverse tech ecosystem (in terms of gender, ethnicity, social and economic origins etc.) would allow to gather different and complementary opinions in the design of algorithms.
I’d personally add that it seems primordial to develop multidisciplinary tech formations, in which future data scientists and developers would be able to get sociology and economy classes in addition to their tech ones, to develop a useful open-mindedness in the tech industry about how data can be interpreted and algorithms designed.
At the corporate level, companies could make their employees aware of the existence of algorithm biases and how to prevent them. Learning sessions about that topic would be great to raise awareness. In addition, companies could be required by public regulation institutions to appoint an Algorithm Ethic Officer among employees, like the Data Protection Officer for the GDPR. This person would be in charge of ensuring the ethical, diverse and objective aspects of algorithms design.
***
With the popularization of artificial intelligence, the risks related to algorithm biases must be addressed in an urgent way. It seems to be the right time to sensitize as many citizens and stakeholders as possible to raise awareness at the international level. We must all together democratize what algorithm biases are and how to prevent them, in order to make anyone able to recognize them and deal with them. We found some biases roots, but many others must be investigated, at all scales of our society.
Sources:
Https://medium.com/@aureliejeanphd/algorithms-are-the-wrong-usual-suspect-6f6e73b6408d
https://en.wikipedia.org/wiki/ethics_of_artificial_intelligence
https://ai.google/responsibilities/responsible-ai-practices/
https://www.technologyreview.com/2017/10/03/241956/forget-killer-robotsbias-is-the-real-ai-danger/
https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-iduskcn1mk08g
