AIDA

Machine Learning, Human Unlearning: The transition into the digital age raises concerns about the discriminatory biases of artificial intelligence against racial and ethnic minority groups. How can the EU combat bias in algorithms to make sure that the racial discrimination of our current society is not replicated into the digital world?

Special Committee on Artificial Intelligence in a Digital Age

By Kasper Feremans (BE)

Introduction

Since the recent developments in Artificial Intelligence (AI)and the widespread use of chatbots, deep fakes, and image processing software, we are more reliant on AI than ever before. In 2021, global private investment in AI reached a peak of over 120 billion dollars. The growth in popularity of AI has also brought a lot of  controversy. 

Google released an automatic image labelling feature in their Google Photos app, but had to shut it down soon after its release as it had a tendency to  label black people as gorillas. The cause of this was the underrepresentation of black people in the data that the Machine Learning Algorithm (MLA) was trained on. 

The need for extreme amounts of training data is directly linked to the popularity and efficiency of Deep Learning algorithms. This is a subset of machine learning algorithms with the main benefit that the efficiency of the algorithm keeps  growing with the amount of data it can use to train with. 

The training data of an algorithm is very important as in 2016 the Chatbot Tay, which was trained on popular Tweets started generating  very controversial and even racist tweets. Numerous studies have found a more subtle racial and ethnic bias in all types of AI increasing the need for trustworthy and explainable AI.

Key Terms & Concepts

  • Artificial Intelligence: Artificial intelligence (AI) refers to systems designed by humans that, given a complex goal, act in the physical or digital world by perceiving their environment, interpreting the collected structured or unstructured data, reasoning on the knowledge derived from this data and deciding the best action(s) to take (according to predefined parameters) to achieve the given goal.
  • Machine Learning: Machine learning (ML) is the scientific study of algorithms that learn through experience. ML algorithms build a model based on training data in order to make predictions or decisions without being explicitly programmed to do so. 
  • Trustworthy AI: Trustworthy AI has two components: (1) it should respect fundamental rights, applicable regulation and core principles and values, ensuring an “ethical purpose” and (2) it should be technically robust and reliable since, even with good intentions, a lack of technological mastery can cause unintentional harm. 
  • Bias: Bias is a prejudice for or against something or somebody that may result in unfair decisions. Many current AI systems are based on machine learning data-driven techniques. Therefore, a predominant way to inject bias can be in the collection and selection of training data.

Key Actors & Stakeholders 

The European Commission

With AI falling under the area of shared competences, the European Commission, the EU’s executive arm, is allowed to legislate on the matter, even though financing and implementing the legislation is left to Member States. Member States may adopt legally binding acts on a national level if the European Commission has not yet exercised its competence, or has explicitly decided not to.  With the Artificial Intelligence Act (AIA) still awaiting a lengthy deliberation process until implementation, it is worth considering whether Member States should take legislative matters into their own hands until the AIA is ratified. 

AI watch

AI Watch is the artificial intelligence website of the European Commission’s Joint Research Centre (JRC), which presents the outputs of the activities of the European Commission in Trustworthy AI. They also work in research and policy support on the development and uptake of trustworthy AI and algorithmic systems which are relevant for digital services in various sectors. They also keep track of the global  AI landscape and the position of the EU within the field.

High-Level Expert Group on Artificial Intelligence (HLEG AI)

The HLEG AI is a group of 50 experts in the field of AI from different sectors of society (academia, corporate and governmental) which provide advice on the European Commission’s AI strategy. One of the projects the HLEG AI has worked on is a document that lists seven key requirements that AI systems should meet in order to be trustworthy, putting forward a human-centric approach on AI.

The European AI Alliance

The European AI Alliance is an initiative of the European Commission to establish an open policy dialogue on Artificial Intelligence. Since its launch in 2018, the AI Alliance has engaged around 6000 stakeholders through regular events, public consultations and online forum exchanges. They were initially created to steer the work of the HLEG AI but continue to represent stakeholders in AI development.

Key Conflicts

Most deep learning algorithms are black box algorithms, which means that they are highly unexplainable and are viewed in terms of their inputs and output rather than their internal workings. This makes subtle biases a lot harder to detect as the user has no visibility into how the model is making its decisions. A trustworthy AI is necessarily also an explainable AI, as we can only trust an AI if we know its decision making process. Once we have an explainable AI, we can try to balance the algorithm to improve fairness and resolve any unwanted biases through many different tests and techniques. Since explainable AI is harder to develop, the performance of explainable AI has not yet reached the level of deep learning algorithms. 

As the EU Agency for Fundamental Rights (FRA) highlights, the question of bias in algorithms is still heavily underresearched and lacks evidence-based assessments. Every Machine Learning algorithm is a direct reflection of its training data. More data generally equals a better performance. Having said that, is it fair to consider gender or ethnic background when developing an AI which predicts recidivism risk to help courts decide whether a prisoner should be granted parole? 

We should also consider how the biases of society hide in data. Imagine we create a dataset for a face-recognition AI by asking 1000 random people for a picture. Because the selection of data is completely random, minorities will be very likely  underrepresented in the data. The algorithm will thus receive fewer pictures of minorities and will consequently have a harder time recognising their faces. This shows that we have to ensure a fair representation in every step of the process to develop ethical AI.

This is in a certain sense a shortcoming of General Data Protection Regulation (GDPR) as even though data about someone’s ethnic background might be protected, this does not mean that the algorithm is not biassed. In a segregated neighbourhood, access to someone’s geographic location can be a proxy for ethnic background. This only hides the bias of an algorithm instead of removing it, proving the need for further regulation.

Lastly, note that an AI always has some bias as we need this bias to extrapolate information from given data. Imagine an algorithm without any bias or preconceived notions: this algorithm will not be able to generalise beyond the training data. Trying to ask this algorithm a question it has not been explicitly told the answer to is useless, as it would give you an answer only if the algorithm considers it the most probable answer. Our algorithm would not prefer any answer over another, as this would implicate a bias of one answer over another.

Measures in Place

The General Data Protection Regulation (GDPR)

The General Data Protection Regulation is an EU regulation on information privacy in the EU and the European Economic Area (EEA). The GDPR’s goals are to enhance individuals’ control and rights over their personal information and to simplify the regulations for international business. This regulation ensures the right to not have your data used to train algorithms without your consent. 

The Digital Services Act (DSA)

The Digital Services Act is a regulation in EU law regarding illegal content, transparent advertising, and disinformation. It aims to regulate the rights and obligations of intermediaries, which are websites that allow anyone to upload content to their platforms. Examples of such intermediaries are Instagram, Youtube and Facebook.

The Digital Markets Act (DMA)

The Digital Markets Act is an EU regulation that aims to make the digital economy fairer and more contestable. The main objective of this regulation is to regulate the behaviour of the so-called “Big Tech” firms within the European Single Market and beyond. With regards to AI development, the DMA limits the sources of training data these “Big Tech” firms can use.

The Artificial Intelligence Act (AIA)

The Artificial Intelligence Act is a proposed EU regulation on artificial intelligence in the EU. It was proposed by the European Commission on 21 April 2021 and has not yet been enacted. The proposed Act aims to classify and regulate artificial intelligence applications based on their risk to cause harm. This classification includes four categories of risk (“unacceptable”, “high”, “limited” and “minimal”), plus one additional category for general-purpose AI. Importantly for the detection of bias in algorithms, the AIA provides legal clarification on the processing of sensitive data that could lead to discrimination (Art. 10 (5)). 

Food for Thought

The issue of unfair representation within AI algorithms is not reserved only for racial biases. In fact, the bias hidden in many algorithms reflects the bias in our society. Women are  also disadvantaged by algorithms as for example amazon had to take down its hiring algorithm as it discriminated against women. It is important that we are able to balance an AI in a way that ensures everyone gets treated fair. This then becomes an ethical question. How do we define fairness and what goal do we strive towards? As contradictions always arise whenever trying to balance different value systems.

Further Research 

  • ‘The historical evolution of AI’ by the Joint Research Center of the European Commission: A short report on the historical evolution of AI and an amazing introduction to the topic.
  • ‘’Ethics Guidelines on trustworthy AI’ by The High Level Expert Group on AI of the European Commission: This guideline was published together with a recommendation of regulation on AI which inspired parts of the new AI act. It also gives a great overview of the ethical considerations regarding AI.
  • ‘Are we automating racism?’ by Vox: A very interesting Youtube video explaining the key conflict of the topic with a lot of extra examples.