Digitalisation and diversity

14 May is once again Diversity Day across Germany. Many companies are organising events to promote diversity, give an insight into their strategies and present what diversity means to them. Most people think directly of HR and equal opportunities projects. But diversity means much more than that. What is often forgotten: Digitalisation cannot be driven forward without diversity awareness. At the same time, the development of AI in particular faces difficult tasks in this area.

On the term: social diversity in the context of digitalisation

Social diversity generally refers to the diversity of people in all its forms, including age, gender, cultural background, religion, sexual orientation, physical or mental characteristics, etc. It influences digitalisation in an ambivalent but central way because it determines who uses digital technologies and how they help shape and benefit from them. It influences digitalisation in an ambivalent but central way because it determines who can use, help shape and benefit from digital technologies and how.

Diversity also means that people in companies have very unequal prerequisites when it comes to using technology. Some employees have a high level of digital expertise, stable access and the ability to quickly adapt new tools, while others have difficulties using basic applications or even gaining access to the relevant infrastructure. This creates inequality, which is often described as a digital divide and often exacerbates existing disparities.

At the same time, diversity also characterises the way in which digital solutions are developed. If development teams, institutions or platform operators are not themselves diverse, there is a risk that technologies will reproduce implicit assumptions about "typical" users and systematically disadvantage certain groups. This can be seen, for example, in user interfaces that are difficult to access, algorithmic distortions or offerings that do not reflect the realities of many people's lives. A lack of consideration for diversity can therefore lead to exclusion by design.

On the other hand, social diversity also holds considerable potential for digitalisation. Different perspectives, experiences and problem situations can promote innovation because they lead to more diverse approaches to solutions and enable the development of more robust, inclusive technologies. However, the prerequisite for this is that this diversity not only exists, but is actively integrated into digital processes, for example through participatory approaches, inclusive design or targeted skills development.

Overall, social diversity in digitalisation is therefore not inhibiting or promoting per se, but rather represents an area of tension: Without suitable equalisation and integration mechanisms, it reinforces inequalities and makes participation more difficult; however, if it is consciously addressed and integrated, it can become a decisive driver for innovation, resilience and broad social support for digital transformation.

Learning experiences: AI is not yet colourful enough

There are several well-documented cases from the business world that show how problematic it can be when AI systems are developed without sufficient awareness of diversity. These examples are so revealing because the problems did not arise from "bad intentions", but from biased data, homogeneous development teams or a lack of reflection on different user groups.

• Recruiting tools: An international online giant made headlines by developing an AI system for the automated assessment of job applications that was trained on historical recruitment data. As this data was heavily biased towards men (typical for the tech industry), the system systematically began to rate applications from women less favourably. Terms such as "women's" (e.g. in "women's chess club") were weighted negatively. The project was eventually cancelled because the bias could not be reliably corrected.
• Facial recognition: In studies, software systems used by companies worldwide showed significantly higher error rates for people with dark skin colour, especially women. The main reason for this was that the training data sets mainly consisted of images of light-skinned people. In practice, this can lead to serious problems, for example in security applications or identity checks.
• Algorithms in the financial sector: There were also unforeseen consequences in the credit card system of technology companies in cooperation with banks. Customers reported that women were given significantly lower credit limits than men despite having a comparable financial situation. The exact functioning of the algorithm was not transparent, but the case triggered a broad debate about algorithmic discrimination and a lack of traceability.

These cases show a common pattern: AI systems reproduce existing social inequalities if diversity is not taken into account in data, development and application. The problems rarely arise on the surface, but rather deep in the assumptions about what is considered "normal". This is precisely why diversity awareness is now increasingly seen as a key prerequisite for responsible AI development.

Preventive measures for better AI

In Germany, attempts are being made to ensure diversity in AI development through a combination of regulation, guidelines, funding and institutional structures. The reason is simple: diversity cannot be "built in", but must be taken into account throughout the entire development process - from data to application.

A central framework is European regulation, in particular the AI Act. This law obliges companies to systematically assess the risks of AI systems, especially in so-called high-risk applications such as recruiting, lending or healthcare. This also means that training data must be as bias-free as possible and that discrimination risks must be assessed. Diversity is therefore indirectly safeguarded here by reducing bias and demanding fairness.

At a national level, the German government's AI strategy provides an important framework. It explicitly emphasises that AI should be "people-centred" and "oriented towards the common good". This results in funding programmes that strengthen interdisciplinary teams, social perspectives and ethical reflection in development - precisely the factors that anchor diversity in practice.

Another important building block is ethical guidelines, for example from the Data Ethics Commission. It has formulated specific recommendations on how algorithmic systems should be designed, including transparency, traceability and non-discrimination. Although such guidelines often do not have the character of direct legislation, they have a significant influence on standards, certifications and corporate practice.

Institutions such as the German Research Centre for Artificial Intelligence and the Fraunhofer-Gesellschaft also play a role by driving forward research into "fair AI", explainable AI and inclusive data sets. Diversity is operationalised technically here, for example through methods for bias detection or through deliberately more diverse training data.

This is complemented by concrete measures in practice: companies are increasingly being encouraged to set up diverse development teams, carry out external audits and pursue so-called "ethics by design" approaches. This also includes involving different user groups at an early stage, for example through participatory development or test phases with heterogeneous target groups.

Despite these approaches, we must remain realistic: The measures create framework conditions and incentives, but they do not guarantee perfect implementation. Precisely because diversity is a social issue, it remains an ongoing challenge to translate it into technical systems. Germany is therefore focusing less on a single "solution" and more on a governance model that interlinks regulation, research and practice in order to recognise and limit distortions as early as possible.

Diversity-conscious digitalisation and digital literacy in SMEs

For SMEs, the uncomfortable truth is that discrimination through AI rarely arises in the finished system, but almost always in data, assumptions and a lack of expertise in dealing with the results. This is precisely why the most effective lever lies not only in technology, but also in organisation and digital literacy.

A sensible approach is to think about the topic on three levels:

1. Digital expertise as a foundation (digital literacy)
   Employees need to understand what AI can - and cannot - do. Without this understanding, results are quickly adopted "blindly". In concrete terms, this means building up a basic understanding of training data, critically scrutinising AI results and developing sensitivity to discrimination risks - for all employees, not just in IT teams.
   
2. Consciously select and check data
   Many problems arise from distorted data. SMEs can do more here than is often thought by checking data sets for bias (e.g. only certain customer groups), scrutinising proxy variables (e.g. postcode as an indirect indicator of income or origin) and manually checking samples.
   
3. Think processes instead of just technology
   Discrimination is often caused by uncontrolled use, not just by the model itself. SMEs should therefore define clear rules of use (where may AI decide, where not?), incorporate "human-in-the-loop" (humans review critical decisions), document decision-making processes.
   
4. Include a variety of perspectives
   Even without large teams, SMEs can take diversity into account by involving different employees in development/introduction, obtaining feedback from different user groups and incorporating external perspectives. This massively reduces the risk of blind spots.
   

5. Orientation towards guidelines
   SMEs should also be guided by existing guidelines such as the AI Act and also develop their own guidelines, compliance with which must be monitored.

Overall, it is clear that diversity is not an add-on when it comes to digitalisation and AI development. Human diversity is a basic prerequisite for ensuring that AI is not developed based on social distortions and consequently reinforces them. Fair AI is not created by better algorithms alone, but by competent people, conscious data and clear rules.