Developing just (AI) systems and algorithms

Developing just (AI) systems and algorithms

Developing just (AI) systems and algorithms

What do our tax system, a job application process, and an AI application have in common? They are all systems that aim for a just outcome. The step-by-step process they use is called respectively a tax return, an interview, or an algorithm. Developing fair (AI) systems and algorithms requires filling in the following components:

  • Just treatment
  • Just distribution
  • Just procedure

This article explains each component and describes how it can be implemented within an (AI) system.

Just treatment

In short, a just treatment means:

  • Equal cases are treated equally
  • Unequal cases are treated unequally

The first point is well-known to many people. For the same work within the same organization, everyone should receive the same salary. The second point, that unequal cases should also be treated unequally, is less frequently mentioned. However, it is equally important for fair treatment. For instance, within the mentioned organization, someone with children can take parental leave, while someone without children cannot. When it comes to developing fair (AI) systems and algorithms, recognizing unequal situations is just as important as ensuring equal treatment of equal cases.

Just distribution

If your organization decides to distribute a portion of its profits among its employees, it might seem like a kind gesture, but it can lead to complications. The question arises: what is a fair distribution? Do you consider those who have contractual entitlements? Do you divide it proportionally based on the number of hours someone works per week? Do you factor in the number of years someone has been employed? Do you base it on salary (with a higher salary indicating higher value)? Do you take into account who might need extra support (due to personal circumstances)? Or do you divide it equally among all employees and call it a day?

When you’re not opting for an equal distribution, you have the choice between three distribution principles. You can distribute based on:

  • Acquired rights (contractual agreements)
  • Merit (hours per week, years of service, salary)
  • Need and/or ability to contribute (personal circumstances)

To develop fair (AI) systems and algorithms, first determine with the stakeholders what a just distribution principle is. Incidentally, the same choices apply to negative distribution. For example, when distributing costs, overtime, or taxes, you can distribute them equally among everyone, or apply one or more distribution principles.

A pie chart supplemented with a bar chart to illustrate that developing fair (AI) systems and algorithms requires a lot of attention and coordination.

A pie chart supplemented with a bar chart to illustrate that developing fair (AI) systems and algorithms requires a lot of attention and coordination
Successfully implementing a fair distribution through (AI) systems requires a lot of attention

Just procedure

How do you create support for the chosen distribution principle and/or the chosen treatment of people? For this, you need a fair procedure. Such a procedure revolves around two things:

  1. everyone who needs to have a say has a say
  2. the right information is used

To ensure that everyone can have a say, it’s important that the procedure is known to all parties involved and stakeholders. Make it clear where, when, and in what way(s) people can provide input or share their views, and how the decision-making process is organized. Equally important as communicating the procedure is following it. This applies not only to input but also to the information being used. The procedure outlines the conditions that information must meet to be included in the decision-making process. These conditions could involve requirements related to sources, source citation, verification, checks, and/or timely sharing of information with all involved parties and stakeholders.

Developing robust and just (AI) systems and algorithms

A just procedure is essential for creating consensus and a fair outcome from everyone’s perspective when it comes to decisions about the desired outcomes of (AI) systems and algorithms. For the same reason, the algorithm itself must have a clear and verifiable procedure for making decisions. In the past, the transparency of algorithms has proven problematic, often resulting in undesirable and even illegal outcomes. According to the European Union, building robust and ethical AI systems cannot be achieved without meeting this requirement.

Nick Nijhuis helps organizations become digitally mature, serves as a business innovation lecturer, provides training in moral leadership, and works as a NIMA examiner.

One comment

  1. Pingback: Building robust, ethical AI systems - Moral Leader

Comments are closed.