LIBE 2

Resulting from skewed input data, flawed target variables, or biased trainings implemented into an algorithm’s design, AI bias embeds real-life discrimination and human rights violations into Artificial Intelligence, thus damaging consumers’ trust as well as the future of innovation. To strengthen the EU’s ethical technological development, it is critical to develop human-centric AI frameworks aligned with the EU’s humanitarian values. While AI legislation has been developing rapidly, it is encountering significant points of contention in terms of technological innovation and transparency, as well as an unwillingness for binding regulations by the private sector. Ultimately, with biased AI often reinforcing existing social problems, it is important for European legislators and companies alike to critically examine and rectify potential social repercussions of automated decision-making.

big dataThis term is used for datasets that are too large or complex for traditional data processing software to analyse. The process of discovering patterns and extracting information from such large datasets is called data mining, often facilitated by machine learning., data mining, Machine Learning (ML)The practice of autonomous improvement of an Artificial Intelligence system through the processing of large datasets and without the need for additional programming input., deep learningA particular Machine Learning technique that uses structures called ‘neural networks,’ inspired by the human brain., Artificial Intelligence (AI)Based on human intelligence, the term Artificial Intelligence (AI) refers to a set of technologies that generate advanced automation of tasks, enabling machines to update their experience and adjust to new inputs. Through the use of machine learning and other data processing strategies, AI technologies can perform tasks such as visual perception, speech recognition, and semi-autonomous decision-making., Bias (implicitImplicit bias: an unconscious attitude or belief held towards a group of people.; data;When allowed to interfere with Artificial Intelligence, implicit biases on the human end of pattern recognition algorithms can result in a data bias, which in this setting refers primarily to ML data skewed to over- or under-represent certain social groups. AIAI bias or algorithm bias is a phenomenon that occurs when an algorithm produces systemically prejudiced results due to erroneous assumptions or data biases in the machine learning process.), black-box metaphorThis metaphor demonstrates the fact that the internal workings of an AI system are practically not observable, similar to an impenetrable system, due to the extremely complex processes involved. With the process of getting from input to output unclear, severe problems arise in the process of detecting, understanding the origins of and mitigating data bias., explainabilityA key term in AI research that stands out as the opposite to black-box systems, referring to the ability to explain in human terms the internal mechanics of a machine or deep learning system. This is often used interchangeably with interpretability, i.e., the ability to interpret the predictions of ML models..

‘To become a global leader in innovation in the data economy and its applications’ stands out as the primary target of the European data strategy. As a key point of the EU’s digital future, AI is attracting significant attention due to not only its promising technological contributions, but also its uncontrolled development. Despite the multidisciplinary benefits of AI, several experts and human rights advocates are concerned about instances of discriminatory and biased behaviour by AI systems against persons of a particular gender, race, socio-economic background or sexual orientation, with allegations targeting local enterprises and leading brands alike over the years. For instance, Google Vision Cloud reportedly practised discrimination against people of colour by labelling images differently depending on skin colour, as either an ‘electronic device’ for white users, or a ‘gun’ for black ones. 

©Google Vision Cloud (2020)

With similar allegations against Amazon’s recruiting system for discrimination against women, and even against the US healthcare system for favouring white over black patients, AI bias is seen to magnify existing discrimination in society, further disadvantaging underrepresented groups. Thus, the EU needs to follow a human-centric approach focused on protecting human rights, in order to defend social equality and restore citizens’ trust in AI systems.

The European Commission (EC) is the executive arm of the EU, responsible for proposing legislation and managing EU policy. For matters related to AI and digital technologies in general, the EC acts through its Directorate-General for Communications Networks, Content and Technology (DG CONNECT), the department responsible for the EU’s Digital Agenda and the Digital Single MarketThe Digital Single Market is a policy framework proposed by the European Commission as a part of its Digital Agenda for the European Single Market, also called the ‘Internal Market’. The Digital Single Market covers digital marketing, e-commerce, and the information and communications technology (ICT) sector..

As a relatively new and rapidly growing policy area, AI, along with its socio-economic implications, lies at the intersection of various shared EU competence areas, i.e., research and development (R&D), consumer protection, justice and security, and the single market. This means that the EC and Member States can both initiate legislations, with EU laws taking precedence. It is especially important for Member States to create their own legislation on areas adjacent to AI bias but not covered by the shared EU competences, such as national health systems.

The European Parliament (EP) is the legislative branch of the EU, responsible for debating the legislative proposals of the EC, particularly through its IMCO, ITRE, JURI and LIBE committees, which regularly discuss AI and its implications, alongside the recently established a Special Committee on Artificial Intelligence in a Digital Age (AIDA). With its own initiative reports, the EP can also formally request that the EC bring forward legislative proposals on topics of shared EU competence.

The European Economic and Social Committee (EESC) is an EU advisory body that unites various economic interest groups of the Internal Market. In its advisory capacity, the EECS offers opinions on the EC’s policy frameworks and puts forward proposals regarding the Digital Single Market and its stakeholders.  

The High-Level Expert Group on Artificial Intelligence (AI HLEG) is a team of 52 experts appointed by the EC and tasked with implementing AI-related strategic plans and policy recommendations.

The Council of Europe (CoE) stands out as one of the leading international organisations for human rights. Its Ad Hoc Committee on Artificial Intelligence (CAHAI) is responsible for addressing prominent AI-related concerns and drafting a humane framework regarding the development of AI. 

Non-Governmental Organisations (NGOs) and researchers on human rights play a crucial role in monitoring the implementation of AI regulations policies regarding ethical and social issues, as well as influencing the legislative process through networks such as the EU Agency for Fundamental Rights (FRA) and the Human Rights and Democracy Network (HRDN).

As captured in the words of EC President Ursula von der Leyen, ‘the algorithm is as smart as the data you feed it,’ meaning that the rapid development of AI does not free it from the implicit biases of its human creators. AI bias mirrors the opinions and prejudices in society, inheriting existing biased practices through poorly selected, unconsciously or intentionally biased training data. With human input being a necessary part of AI research, unbiased algorithms need unbiased data sets that address real-life discrimination. 

Algorithm experts suggest that this issue runs even deeper than simple human bias, with groups facing AI discrimination often being significantly underrepresented in certain key environments, such as women in technological research groups, and largely overrepresented in others, such as people of colour in the American prison system. With the quantity and diversity of data playing a large role in the eventual workings of an AI algorithm, these systems end up being adversely influenced by extraneous factors, such as the subject’s race and socio-economic status, thus reinforcing existing social inequalities. This self-reinforcing effect is built into the way AI and ML work, thus posing a continuous challenge to ethical AI usage.

Alongside the human factor being critical in the development of AI, the algorithms themselves are also major deciding factors, with some being more bias-prone than others. Thus, the transparency of AI systems and their training datasets is vital in detecting human rights violations. While advocates of AI transparency insist that publishing the source code and training data of critical AI algorithms need not harm companies and may be essential for identifying high-risk issues, multiple AI developers contend that increased transparency requirements may facilitate replication, thus reducing innovation and investment incentives. 

Evidently, corporate secrecy laws, along with the pre-existing ‘black-box’This metaphor conceptualises the fact that the internal workings of an AI system are practically not observable, similar to an impenetrable system, due to the extremely complex processes involved. With the process of getting from input to output unclear, severe problems arise in the process of detecting, understanding the origins of and mitigating data bias. intricacy of AI algorithms, render systems unaccountable and obstruct the assessment of bias and the correction of errors, thus being seen as stymying progress. On the other end, opponents of explainabilityA key term in AI research that stands out as the opposite to black-box systems, referring to the ability to explain in human terms the internal mechanics of a machine or deep learning system. This is often used interchangeably with interpretability, i.e., the ability to interpret the predictions of ML models. argue that it requires a decrease in complexity, thus sacrificing efficacy and innovation, despite the notable progress made by ethical AI actors. Since AI algorithms rely on highly complex models and data points that evolve with time, even waiving trade secrecy and fully auditing these systems might not yield a complete solution. 

At the same time, the internal opaqueness of AI is amplified by a lack of credible external evaluation, whereby consumers and users of AI systems are often unable to understand beforehand when and to what extent an algorithm’s decision is biased, despite efforts at voluntary labellingThis term refers to the manufacturing practice of providing additional information on product labels, i.e., beyond what is required by national and EU regulations. This form of self-regulation applies particularly to information of specific interest to customers in newly developed technologies.. Ultimately, the shifting balance between transparency and competitive advantages for businesses remains a prominent challenge in today’s AI technology that creates highly branched and complex issues.

With competitive market forces in the AI primarily driving the search for new applications and markets, rather than incentivising non-discrimination, experts are increasingly concerned that self-regulation may not be enough to tackle the ethical challenges posed by the development of AI and ensure accountability. Thus, various NGOs have called for increasing regulatory and supervisory power to government agencies, with the World Economic Forum issuing guidelines for such audits, especially relating to facial recognition AI. Even then, the challenge would be shifted to creating national or EU-wide AI safety guidelines and certification models that meet nuanced sectoral expertise requirements to enable innovation, a task that has historically troubled lawmakers.