The technological development of AI raises major challenges and questions of ethics. By ignoring them, we take the risk of eroding public confidence and exacerbating social inequalities.
Consequently, how to design ethical AI systems that minimize biases and ensure transparency, two of the most critical challenges?
Take into account the existence of bias in AI
Machine Learning models on which AI are based learning from data provided by humans. They can reproduce and amplify existing biases, thus perpetuating prejudices, inequalities and stereotypes. These biases can have several origins: they can arise from disparities within the population (for example, the majority of nurses are women), be systemic (such as the correlation between certain geographic areas and ethnic origins), come from our history and the trivialization of racism, sexism or homophobia, or even result from measurement errors (people discriminated against sexual orientation, being less inclined to declare it). For example, facial recognition algorithms used by police in the United States in 2010 were very efficient on white men (1% error), but much less to women of color (35 to 38% error), sometimes leading to false crime accusations (1).
Synthetic data for fairer and inclusive models
Thanks to the use of synthetic data, it is possible, as in the European Achilles (2) project, to generate more diverse and representative data sets to supply AI models. These data provide a solution to the problems linked to the lack of data diversity and the discriminatory biases that result from it. They make it possible to artificially fill the gaps in real data. In the field of dermatology, for example, synthetic samples have been created to compensate for the lack of images of patients with dark skin, thus reducing biases of diagnostic models. Solutions such as those developed in the European Mammoth (3) project, gathered within an intuitive and easy -to -use toolbox, make it possible to identify and correct biases, both in data and models.
Transparency and confidence, keys to an ethical AI
Another challenge is transparency. This is essential to establish a relationship of trust between AI systems and their users. Transparency is not limited to explaining how a decision was made. It includes the entire process, from the design of the model to its deployment. AI Act, which constitutes the first legal framework governing the use of AI on a European scale, adopted last July, recommends that each AI system is accompanied by clear and accessible documentation, such as “model cards”. These documents make it possible to describe the tests carried out, the limits of the system, and the populations for which the model is optimized. They facilitate audits and strengthen the confidence of users and regulators. Increased transparency, however, brings other size challenges linked to data protection and intellectual property. These challenges will have to be met to guarantee an ethical AI.
Build a legal framework for an equitable AI
Regulations such as the General Data Protection Regulations (GDPR) and AI Act of the European Union play a crucial role in supervising the development of ethical systems. It is essential that regulations are more focused on equity and transparency standards, for greater confidence and to promote adoption by users. For example, current regulations do not always specify how to measure the equity of a model or which transparency standards to adopt. Initiatives such as the European Achilles project seek to align technical developments on the ethical and legal standards of these regulations. By involving multidisciplinary teams – technical experts, lawyers and sociologists – the project aims to establish practical and robust standards. This collaboration is essential to guarantee the consideration of user needs, technical constraints and legal implications from the design phase.
Reconcile innovation, energy sobriety and ethics
The other challenge is to design more sustainable systems with the priority of reducing the energy consumption of models while allowing effective use of AI models, without compromising ethics. In terms of developments, the objective is to ensure that principles such as equity, transparency and confidentiality are integrated into each stage of the AI life cycle, from the collection of learning data to deployment. In this perspective, AI systems must be designed taking into account the real needs of end users. An AI intended for health, for example, should integrate feedback from practitioners and patients from the design phase. Tools such as Mammoth Toolkit (4) or OECD recommendations catalogs (5) offer concrete executives to guide the developers in this direction. Finally, legislative executives such as the GDPR and AI Act of the EU evolving rapidly, it is essential for companies to anticipate legislation and invest in legal teams capable of interpreting these regulations and translating them into technical requirements.
Ethical AI is not an option, but a necessity. By minimizing biases and ensuring transparency, it is possible to build trust systems, beneficial for all.
By Ahmad Montaser Awal, research director at IDNOW, and André Carreiro, Achilles project coordinator and researcher at the Fraunhofer Aicos Institute
1. Https://www.radiofrance.fr/franceculture/etats United States-la-reconnaissance-faciale-accusee-de-favoir-les-biactes-570065
2. Https://www.achilles-project.eu/
3. https://mammoth-ai.eu/
4. https://github.com/mammoth-eu
5. OECD, Ethical IA tools Catalog: https://www.oecd.org/fr/publications/2019/06/artificial-intelligence-in-society_c0054fa1.html
