Artificial intelligence at work: fixing data is not enough to prevent discrimination

Algorithms can reproduce inequalities in hiring and promotion. Tackling them requires transparency, human oversight and ways to challenge decisions. Outcomes also depend on how companies introduce and use these systems

Spread the love

Link to the research article: Lazazzara A., Za S., Georgiadou A., A taxonomy framework and process model to explore AI-enabled workplace inclusion, Journal of Business Research, Volume 201 (2025). https://doi.org/10.1016/j.jbusres.2025.115697.

Artificial intelligence (AI) is transforming the world of work, including how work is organised. Recruitment, performance evaluation, talent management, training—even career development. Decisions once made exclusively by managers are increasingly supported, and sometimes automated, by algorithmic systems.

Technology providers make an appealing promise: AI can make these decisions fairer, more objective and less influenced by prejudice and stereotypes. Yet as the adoption of data-driven systems grows, cases of discrimination are emerging from the very algorithms designed to eliminate it.

AI systems do not emerge in a vacuum. They are trained on historical data that reflect the decisions, hierarchies and biases of real organisations. If a company has historically promoted more men than women into leadership roles, an algorithm may treat that pattern as the norm and reproduce it.

A recent study involving the University of Milan, “Gabriele d’Annunzio” University of Chieti-Pescara and Nottingham University Business School proposes a different perspective. Asking whether data contain biases is not enough. We also need to examine how interactions between people and algorithms change patterns of inclusion and exclusion.

The study addresses a further crucial question: not simply whether AI can support or hinder inclusion, but how it redefines inclusion itself within contemporary organisations.

Its starting point is straightforward: AI is more than a technology; it is an organisational actor. It interacts with people, guides decisions, establishes standards and, in some cases, introduces new assessment criteria.

Discussing “algorithmic inclusion” therefore means more than ensuring that a system does not discriminate. Algorithmic inclusion is the form of inclusion that emerges from using AI in human resource management. It concerns how algorithmic decisions affect fairness, participation and a sense of belonging at work, and depends on how AI is trained, designed and used.

A process, not a single event: when AI assists, augments or automates

To understand whether and how AI can support or hinder workplace inclusion, the researchers propose viewing its adoption as a process: a sequence of events that tends to recur whenever organisations introduce AI for inclusion purposes.

Drawing on an analysis of 25 empirical studies, the authors developed what they call a meta-causal network: a network of causes, effects and conditions describing how AI adoption unfolds and the dynamics it creates around inclusion.

The model begins with “non-agentic” technologies—systems that do not make autonomous decisions—and moves towards increasingly complex systems with greater agency, or capacity to act independently. Each stage brings a different level of risk and therefore requires different mitigation strategies.

The process starts with a company’s reasons for introducing AI into human resource management. There are three main motivations: improving recruitment, supporting inclusion within the organisation or enhancing overall performance.

When the aim concerns existing employees, workers have more contextual information to help them interpret why AI is being introduced. External job applicants, by contrast, lack that knowledge and rely on signals conveyed by the technology itself to assess how inclusive the organisation really is.

AI’s effects on inclusion depend crucially on the type of technology adopted and its degree of autonomy.

When companies use Assisting AI—tools that support activities without making independent decisions, such as teleoperated robots, CV-ranking systems or simple pre-screening algorithms—the risk of introducing additional discrimination is low.

Indeed, many studies report positive effects: more women shortlisted at the initial recruitment stages, a stronger sense of fairness among unemployed applicants and greater attention to groups that are often overlooked, such as part-time workers.

Only rarely have assistive technologies produced negative consequences. In these settings, particularly sophisticated interventions are not generally required: human oversight remains dominant, and the risk of discrimination is limited.

The picture changes when organisations adopt Augmenting AI. Here, technology does more than organise information: it generates recommendations, identifies patterns and substantially influences the criteria used to make decisions.

This intermediate stage produces mixed outcomes. Some machine-learning systems improve women’s representation in recruitment or reduce stereotypes in job advertisements. Others disadvantage women, older workers or people with darker skin tones, inadvertently reproducing biases embedded in historical data.

Because Augmenting AI learns from past data, its reliability depends on the quality of those data. The study therefore emphasises the importance of data mitigation: rebalancing categories, removing sensitive variables and correcting imbalances in datasets. Without these measures, AI risks amplifying inequalities rather than correcting them.

There is, however, another less visible but crucial effect. When an algorithm begins to help define—even indirectly—who is considered “deserving” or “suitable” for an organisation, the meaning of inclusion itself also changes.

In settings that use Augmenting AI, people begin to question what fair treatment really means, which criteria guide technological recommendations and who has the authority—human or artificial—to establish assessment standards. Inclusion becomes something negotiated rather than taken for granted.

The final stage is the adoption of Automating AI, where systems do more than help define procedures: they make decisions autonomously about promotions, performance evaluations, candidate screening and behaviours to reward or correct.

At this level, inclusion becomes particularly sensitive. Some studies show that, when properly designed, automated algorithms can significantly reduce certain forms of hiring bias by making decisions more consistently and less subjectively.

Yet workers’ perceptions are complex. People tend to consider feedback from a machine less fair than the same assessment delivered by a human. Paradoxically, they may also tolerate discrimination more readily when it is attributed to an algorithm rather than a manager.

At this stage, inclusion depends on more than data or algorithms. It also depends on how the company governs and communicates its use of AI.

The authors show that highly agentic technology requires specific mitigation strategies: transparency about how systems operate, opportunities to challenge automated decisions, training for human resource managers and a shared understanding of why and how AI is being used.

Figure 1 — The process model of AI-enabled workplace inclusion

Inclusion can therefore be strengthened or weakened by the technical, organisational and cultural choices made by companies using AI. The central challenge, the study suggests, is governing its adoption while recognising that each type of technology brings new opportunities and risks for workplace inclusion.

“Fixing the data” is not enough: a multilevel perspective

One of the study’s most significant conclusions is that algorithmic inclusion cannot be left solely to technical specialists or developers.

Even with “clean” data, an algorithm can produce exclusion when introduced into an organisation characterised by mistrust, limited transparency or authoritarian leadership. Conversely, imperfect systems can produce inclusive outcomes when implemented in environments where people feel heard, can experiment without fear and have opportunities to question automated decisions.

This also changes how inclusion itself is understood. The authors propose a multilevel perspective in which AI-enabled inclusion comprises three interconnected dimensions.

Figure 2 — The three levels of algorithmic inclusion

At the organisational level (inclusion at work), the focus is on how AI shapes organisational culture and shared experiences of belonging.

An algorithm used to allocate rewards, for example, may be perceived as fair only if the company clearly explains how it works, which data it uses and how errors can be corrected. Likewise, an automated monitoring system can become an “algorithmic cage” if it tracks every activity without leaving room for autonomy.

At the individual level (inclusion in work), the focus is on workers’ personal experiences and their relationship with AI-supported tasks. Inclusion depends on being able to access and interact with digital tools without feeling excluded, whether because of limited digital literacy or aspects of identity that are less well recognised in digital environments.

At the human–machine interaction level (inclusion of work), the focus is on how people and intelligent systems collaborate in everyday activities.

An AI assistant that suggests how to respond to customers, for example, can support autonomy and learning, while a system that automatically imposes decisions can leave workers feeling “replaced”.

When AI provides support, useful feedback and opportunities to develop new skills, workers tend to view it as an ally rather than a threat.

Designing for inclusion across these three levels means creating a sociotechnical environment that reduces the risks of marginalisation and algorithmic bias while increasing workers’ willingness to integrate AI into their daily tasks. This strengthens human–machine collaboration and supports sustainable AI adoption.

AI alone does not determine whether inclusion increases or decreases. It can support inclusion when the organisation adopting it pursues that goal deliberately, transparently and responsibly.

Spread the love
Avatar photo

Alessandra Lazazzara

Alessandra Lazazzara is an associate professor of organization studies in the Department of Social and Political Sciences at the University of Milan. Her research focuses on people and work within organizations, including human resource management, diversity and inclusion, and the ways digital technologies reshape professional practices, skills and relationships. She chairs the interdepartmental teaching board for the degree programme in Management of Organizations and Labour.