Ethics and Artificial Intelligence: Fairness (5/7)

In 2019, the European Commission's High-Level Expert Group on Artificial Intelligence (AI HLEG) published its "Ethics Guidelines for Trustworthy AI". To this day, it remains one of the foundational documents on the ethical requirements for artificial intelligence.

The Guidelines set out seven key requirements for trustworthy AI, which must be met at every stage of a system's lifecycle - from design and deployment to practical use in a real-world environment.

These requirements are:

1. Human agency and oversight - read our article here

2. Technical robustness and safety - read our article here

3. Data governance and privacy - read our article here

4. Transparency - read our article here

5. Diversity, non-discrimination and fairness

6. Societal and environmental well-being - read our article here

7. Accountability - read our article here

Each of these plays a key role in making AI not only technologically advanced, but also ethically sound, sustainable and beneficial to society. In this series, we examine each principle in turn, exploring its significance for legal practice and for the future of the legal profession.

Having already covered the first four principles, we now turn to the fifth: diversity, non-discrimination and fairness. This is especially important for the legal sector, where every decision can have serious consequences for the lives and rights of citizens.

What is algorithmic bias?

Algorithmic bias arises when artificial intelligence systems make decisions that systematically favor one group at the expense of another. It can emerge at several levels:

• The design of the algorithm - the way the model's parameters, criteria and objectives are defined can introduce implicit prejudices. For example, if a recruitment algorithm gives excessive weight to prior professional experience, this can reinforce historical discrimination against women or minorities.

• The training data - if it contains incomplete, historically loaded or inaccurate information, the model will reproduce it and may even amplify it. This is the "garbage in, garbage out" principle: biased data leads to biased results.

Algorithmic bias is especially dangerous because it often remains hidden behind a facade of objectivity. A decision made by a computer system appears neutral, but in reality it may be the product of deeply embedded human prejudices, transferred through the data or the design.

Because AI has the power to influence critical decisions - who gets a job, how a credit score is determined, how the risk of reoffending is assessed in court. Any bias in the system can lead to unfair, discriminatory and even unlawful outcomes.

A simple, everyday example of algorithmic bias can be seen in the use of various AI systems for generating images. Below is a collage of AI-generated pictures that are supposed to depict a "successful person":

The principle of fairness

The principle of fairness in the European Commission's Guidelines covers not only equality, non-discrimination and solidarity, but also the protection of groups at risk of exclusion. It requires AI systems to guarantee equal moral status and dignity for all people, without privilege or discrimination.

This goes beyond simply avoiding discrimination on the basis of objective criteria. Fairness means that AI systems must not produce unjust outputs, especially where this could affect vulnerable groups such as women, people with disabilities, ethnic minorities, children, consumers and others at risk of social exclusion.

The substantive dimension

According to the Guidelines, the fair development and use of AI must ensure:

• a fair distribution of benefits and costs,

• the absence of unfair bias, discrimination or stigmatization,

• inclusive data that represents different population groups,

• guaranteed equal access to education, goods, services and technology,

• no deception, manipulation or impairment of freedom of choice for end users,

• a balance between means and ends, through the principle of proportionality.

If unfair bias is avoided, AI systems can even help to increase fairness in society.

The procedural dimension

Fairness in AI also includes the ability of individuals to contest decisions made by automated systems. This requires:

• a clearly identified entity responsible for the decision,

• transparency and explainability of the decision-making processes, and

• effective legal redress for the parties affected.

Risks where fairness is lacking

A lack of fairness in the use of AI can lead to serious problems across various fields:

• Recruitment and hiring - automatic rejection of women's applications because the position has historically been held mostly by men (a real example from Amazon).

• Healthcare services - an erroneous assessment of an elderly patient's need for treatment due to statistical anomalies.

• The justice system - risk-of-reoffending assessment algorithms that assign higher scores to members of ethnic minorities without objective grounds (a real example from the US).

• Border control - according to research, facial recognition systems that perform better for white men (around 1% errors) than for dark-skinned women (up to 35% errors).

The principle of fairness and diversity is key precisely for:

1. Protecting vulnerable groups Without this principle, AI - under the guise of objectivity - can reproduce systemic prejudices directed at the most vulnerable, an outcome we are obliged to prevent.

2. Increasing public trust When a system respects everyone, including those who are excluded, and provides an opportunity for feedback, it is perceived as reliable, ethical and legitimate.

3. Upholding fundamental human rights The principle protects personal autonomy, access to services and the fundamental freedoms that lie at the foundation of a society governed by the rule of law.

4. Regulatory compliance and sustainability One of the central pillars of the regulations on high-risk AI systems (the AI Act) is the prevention of discrimination and the provision of equal access. This makes fairness mandatory for certain types of AI systems.

Significance for Bulgarian legal practice

Within Bulgarian law, the principle of fairness carries particular weight, since every decision can directly affect fundamental rights and legitimate interests. Even when attorneys and lawyers do not develop their own artificial intelligence systems, the use of applications such as ChatGPT calls for critical thinking and human judgment.

The most important thing is not to assume that AI-generated texts or analyses are entirely objective. Lawyers must always check whether they contain stereotypes, discriminatory wording or the exclusion of particular groups of citizens. This ensures that AI serves as an assistant to legal professionals, rather than becoming a source of biased or unfair decisions.

🟢 Try Praven Intelekt now! 🟢