Good practices for Service Developers

Using AI responsibly

Account for human diversity

Organisations are bound by the Non-Discrimination Act

The Non-Discrimination Act obliges authorities and private actors alike to promote equality and equal opportunities.

This means not only listening to and consulting population groups, but also having diversity in the entire design and implementation chain.

Updated: 9/11/2023

AI produces the kind of world that it is taught

At its core, each AI system is based on two non-technological and non-mathematical things:

An identified and defined problem that you want to solve with AI.
Datasets used to train the system.

Both may pose challenges in terms of equality and non-discrimination:

How has the problem been identified and selected?
Who is most affected by the problem? Who benefits most from resolving it?
Can a solution produced with AI be harmful to some?
Are there discriminatory features in the selected training data? Discrimination can arise not only from distorted data but also from the lack of data.

The fields of data processing, mathematics and technology have a history of being heavily gendered and dominated by majority groups, most notably white men.

The situation has changed to some extent, especially during this millennium, but things are still not equal enough. AI systems have particularly been developed by companies and teams where women, older people, minorities and people in difficult situations in their lives are clearly underrepresented.

As AI systems almost inevitably come out looking like their makers and meet their specific needs, the results are often embarrassing, harmful or even useless for different groups.

For example, it has been widely known for a long time that facial recognition systems have problems with ethnic groups, but the issue still persists.

Updated: 9/11/2023

Identify biases in datasets

Human diversity in the real world encompasses a range of variables from ethnicity and gender to occupations, but at the moment, the content of the internet does not reflect this diversity. There are biases in all types of data: text, image and audio. It is precisely these datasets that are used to train AI systems and large language models such as ChatGPT. This is why diversity is an important consideration when cooperating with stakeholders and citizens.

Sometimes a system and the service built on it are specifically intended for a specific group. Of course, the focus must be on that group in that case, but even so, you have to be mindful of the broader social context and the internal diversity of the target group.

Different age groups use technology in different ways. Multiculturalism, individualisation and social exclusion are also increasingly prominent phenomena that can complicate the adoption and application of AI-based technological solutions. So, it is important to study AI-based solutions in interaction with users and to take different needs into account from the start. There is no one-size-fits-all solution.

–Professor Mark van Gils, University of Tampere

Updated: 9/11/2023

The more diverse the development team, the more diverse the end result

The most effective way to ensure that an end result actually serves its purpose and does not unintentionally discriminate against or exclude anyone is to strive for development teams, companies and organisations whose composition reflects the diversity of the population as accurately as possible.

The perspective of maximising inclusion in development is also a success factor because it broadens the potential user or customer base of an AI service or product, making it as wide as possible.

Updated: 9/11/2023

Are you satisfied with the content on this page?

Checklist