In this position paper, we advocate for the use of value sensitive design (VSD) as a framework for developing fair AI recruitment tools. As a starting point, we assert that the current paradigm in AI fairness in the hiring context is severely limiting. We then document an ongoing process within the EU-horizon project BIAS, seeking to escape this paradigm by applying VSD to the development of AI applications for candidate selection with diversity and fairness as focal points. In particular, we present case-based reasoning as a case study in the intentional operationalization of stakeholder positions on fairness and detail how such an approach can be further expanded, drawing from the concept of agonistic machine learning. In this endeavor, we hope to contribute to the discourse on the ethical design and use of AI within the labor market and in general.
CEUR ceur-ws.org
The last decades have seen a growing trend toward designing and deploying artificial intelligence (AI) applications for recruitment and selection. However, such tools lack transparency and pose the risk of algorithmic diversity bias,1 reinforcing harmful stereotypes and power structures, and, on the individual level, acting to the detriment of dignity, autonomy, and well-being. Consequently, job candidate profiling is classified as high risk under the EU AI Act.2
As a starting point for this position paper, we assert that the current paradigm in AI fairness in the context of recruitment is severely limiting. Attempts to promote fairness in AI hiring tools typically take diversity metrics as a starting point and modify the training or computational aspects of existing ranking/scoring models to improve metric performance. This approach bypasses the explicit consideration of values and normative assumptions and fails to engage with the question of whether the technical interventions and metrics are truly aligned with ethical fairness goals. It operates within the framework of a rigid, deterministic technical solution whose blueprint implicitly assumes a set of normative stances that can hinder fairness while precluding the innovation of novel technical approaches. The resulting models remain opaque to users and impose a predetermined balance of values, stifling moral agency and accountability and leading to unjustifiable decisions.
This paper documents an ongoing process within the EU-horizon project BIAS, seeking to escape this paradigm by applying Value Sensitive Design to the development of AI applications for candidate selection with attention to diversity and fairness. Section 1 introduces the necessary background aspects of VSD. A description of the application of VSD towards fair AI recruitment tools follows in Section 2, laying the foundation for the pro-active technical approaches described in Section 3.
Value sensitive design (VSD) was developed by Batya Friedman and Peter Kahn in order to
promote alternative design goals and methods built upon respect for human agency and
responsible computing [
Values, Normative Assumptions and Technical Methods. In practice, the development
of new technology under the VSD framework involves the identification of the values
important to each group of primary stakeholders (e.g. fairness) and the corresponding normative
assumptions/stances about what those values mean (e.g. “Everyone should be treated equally,
regardless of their background,” or “People with facing greater structural inequality should receive
more resources to ensure equitable outcomes.”). Finally, an investigation of technical methods
that operationalize these normative stances (e.g. The requirement to meet certain fairness metric
thresholds.) is conducted, with careful consideration of the balancing and/or reconciliation of
competing values. This should all be considered within the sociotechnical context in which the
technology is situated, the relationships it mediates, and the social efects its use could have
over time.
Tri-Partite Methodology. VSD recommends a methodology consisting of conceptual,
empirical and technical investigation. At the conceptual level, researchers are expected to consider
the specific values and normative assumptions in which a technology is situated, clarifying
fundamental issues raised by the technology and design process. It focuses on theoretical
considerations of values, potential conflicts, and trade-ofs. During empirical investigations,
researchers position the design within the relevant social context, seeking to identify, e.g.,
important stakeholders and understand their values. One gathers data from real-world
stakeholders to understand their experiences, needs, and concerns. Technical investigations aim to
identify specific technological approaches suited to supporting certain values, with attention
to potential detriment to others. This includes assessing the capabilities and limitations of
technical solutions. Friedman highlights two subtypes of technical investigation. The first,
retrospective analyses, focuses on “how existing or historical technological processes and
underlying mechanisms support or hinder human values,” [
The primary values of interest are diversity, fairness and the selection of the most capable job
candidate. We concentrate on two sub-types of fairness: 1) Procedural, concerning how the
decision was made; 2) Substantive, concerning the outcome of the decision. VSD defines
stakeholders as those who are or will be significantly implicated by the technology [
Defining Fairness. In the broad social and EU legal context, substantive fairness in relation
to diversity may be viewed in terms of group fairness with respect to hiring outcomes, i.e.,
the fair distribution of jobs amongst demographic groups [
HR practitioner/company: The decision-making process should be aimed toward selecting the
most capable candidate for the job [
One of our main methodologies for empirical investigation consists of co-creation workshops carried out in all nine partner countries. These workshops focus on understanding the views of diferent stakeholders, as well as their desires, needs and opinions regarding what constitutes a fair and useful recruitment tool, making use of mock tools3 to stimulate discussion. Agency and transparency in relation to both users and job applicants, as well as technical robustness, were highlighted as key aspects of a fair and trustworthy decision-making process.4 The outcomes of the decision-making process–algorithmic, human or hybrid–are not fair if they are not justifiable . Participants brought up situations exemplifying unfairness, such as candidate scores being artificially raised by invisibly pasting job ads into their applications or being rejected based on incorrectly parsed data. A disconnect between how outliers are perceived by statistical models and the desirability of candidates with unique profiles also arose as a potential source of unfairness. There was additional concern over how the omission of particular types of candidates by a deterministic model becomes systemic, whereas multiple diferent evaluators may compensate for each other’s particular biases.
Encoded Values and Hindrances to Fairness. Compared to human decision-making, existing tools claim to save time and money for employers while simultaneously making the recruitment process more objective and accurate. VSD rejects the notion that technology can ever be objective; terms like objectivity and accuracy mask normative assumptions behind neutral-sounding terminology, and the repeated discovery of encoded bias in machine learning models has made it clear that such technologies are not objective. This is often explained via a garbage in, garbage out understanding of bias in AI systems, but it is important to note that value-encoding goes beyond training data. For example, ostensibly neutral metrics such as loss and accuracy scores implicitly reinforce the values underpinning the status quo, while predictive modeling treats unclassifiable outliers as data points to be ignored.
The fundamental task–using a pool of candidate data, compute the best candidates—assumes that job suitability is a one-dimensional trait, largely based on a pre-defined notion of what constitutes a high-achieving individual. It is associated with objectivity because suitability is usually evaluated based on educational background, work experience and other “objective” 3These simulated LLM-driven AI tools, allowing users to specify and weight various criteria, providing natural language justifications for candidate rankings, and flagging potential unfair biases. 4User agency over the criteria under which candidates are evaluated, the ability to flag system errors or potential harmful behaviors, an interactive process that allows users to explore candidate data from multiple quantitative and qualitative angles, and the capacity for the tool to draw attention to their own unconscious biases. criteria. But this ignores structural inequalities that can render apparently objective features into grounds for proxy/indirect discrimination. By collapsing each candidate to a small subset of features to be evaluated by a fixed algorithm, job suitability scores can work to the detriment of creating a diverse employee pool with complementary skills.
Fairness Methods. The technical research community has predominantly focused on
operationalizing group fairness [
The Fairness/Accuracy Trade-Of. Research shows that there are unavoidable trade-ofs in
attempting to simultaneously optimize performance and fairness metrics [
Case-Based Reasoning (CBR) forms a foundational AI approach for our proposed Decision Support System (DSS) in recruitment. This system assists HR professionals during the candidate screening process and produces one of three potential outcomes for a candidate applying for a specific job opening: “Shortlist,” “Longlist,” or “Negative.” In a CBR engine, the case-base consists of a database of past candidate data along with the decision that was reached on each candidate’s application. The system compares new candidates to the case-base, retrieves past similar candidates, and a decision is reached by aggregating those made in the past. The features stored in the case-base and the similarity metric are two key points where the designer has direct influence over the operationalization of fairness principles, including the integration of domain knowledge; the model’s computations are grounded in the decision-making processes of HR professionals. The CBR workflow is depicted in Figure 1.
We emphasize that the development of the CBR engine is fully situated within the intended deployment context; the data and design insights were gathered from our industry partner, and the model is bespoke for their own needs and internal processes. At the same time, the model framework, principles and parameters can be readily adapted to new industry contexts. We assert that this is preferable over the development of a generic tool targeting the broadest possible context; one-size-fits-all approaches impose natural limits to accountability, transparency, justified decision-making processes, user agency and the use of domain knowledge. Data. The CBR system depends on careful feature engineering to determine precisely the information on which the model bases its decisions. To efectively capture domain-specific insights, our approach combined workshops and surveys with HR professionals from our industry partner. This mixed methodology collects both quantitative and qualitative observations about the candidate data and the operational context of the recruiting company. The domain insight we gather represents HR experts’ background knowledge about the job types, attributes, and the relevant range of values for those attributes5 for the given sector and company. Another key aspect of our system is capturing the relative importance of each attribute in the decision-making process. The importance of an attribute is context-sensitive and can vary across diferent job types. To ensure the validity of our findings, it was crucial to gather responses from multiple HR personnel employed by our industry partner. This approach helps mitigate potential biases, as diferent evaluators may interpret applicant data in distinct ways. Figure 2 demonstrates the extracted general attribute importance from the survey results.
Model. The importance measured for each feature determines a baseline for their relative weight in computing similarity scores, which can be further adjusted according to domain. This 5E.g. Bachelor’s or Master’s for Educational Level, as opposed to Some higher education and No higher education. is represented by the equation global( , ) = ∑ ⋅ local, ( , ), (1) Where and consist of the feature vectors of two candidates, and and local, respectively denote the weight and job-specific similarity metric attributed to the th feature. For each new candidate, the most similar candidates from the case-base are retrieved by computing similarities using the extracted, engineered features and relative weights. A decision is then rendered by aggregating the decisions made on the retrieved cases, weighted by similarity.
A key element of procedural fairness is that every candidate is subjected to the same process. A weakness of the CBR engine is that its reliability depends on having suficiently many similar candidates within the case-base. This is essentially a version of the outlier problem that plagues all statistical models. We thus define the decision certainty as a metric for procedural/individual fairness; it is inversely proportional to the average distance between a new case and the retrieved cases, as computed using the similarity metric in Eq. 1. As opposed to machine learning, one can use certainty to immediately if the candidate in question is an outlier, so that their application may be considered with further scrutiny rather than systematically rejected, and the case-base can be immediately updated. Whereas a machine =1 learning model may require thousands of such updates to perform reliably on future, similar candidates, a CBR-engine only requires a number proportional to the number of cases retrieved for decision-making ( ≈ 10 ).
Operationalizing Fairness. The setup we describe attempts to operationalize several fairness notions established during our investigations: • Omitting any direct information about protected attributes prevents direct discrimination and limits the possibility of proxy discrimination compared to machine-learning-based methods using raw text data. • The selection of features deemed highly relevant in the candidate selection process and weighted according to user specification promotes job relatedness, user agency, and selection of the most qualified candidate in a well-grounded and explainable decision-making process. • The decision to sort candidates into rough categories instead of rankings acknowledges that job suitability is not one-dimensional. • The certainty metric provides a proxy for robustness and reliability, aspects of procedural fairness, as well as the capability to flag outliers who may otherwise be treated unfairly. • The ability to rapidly update the case-base, and customizability of extracted features and similarity scores further support technical robustness and post-deployment adaptivity, acting as guardrails against the hegemony of static algorithms.
The CBR engine represents a design framework and prototype developed with explicit attention
to stakeholder perspectives on fairness, but we acknowledge its shortcomings; the engine’s
output is still dependent on historical hiring decisions that may be unfair, while features
rated highly by HR professionals such as University Name, District, Foreign Language and
Employment Status could serve as proxies for protected attributes and be used to reinforce
unfair biases in the hiring system rather than mitigate them. We thus conclude this position
paper with a summary of further research directions in proactive design that can complement
CBR and further integrate stakeholder fairness principles.
(Large) Language Models
Deep-learning-based language models extract complex features from raw text data,
achieving unparalleled performance and completing tasks that would be intractable with simpler,
interpretable models such as the CBR engine. The language comprehension capabilities of
large language models (LLMs) such as ChatGPT and DeepSeek further expand the frontier of
possibilities. Such models can ofer qualitative analyses that are not possible using numerical
methods and could further promote fairness. However, what these models gain in capability is
paid for with opaque processes that hinder transparency, agency, and justifiability. Moreover, a
large body of research demonstrates that such models encode harmful stereotypes that can lead
to negative outcomes when integrated into AI recruitment tools (e.g., [
Justifying Decisions. Following [
The fairness perspectives described in this work can be complementary or adversarial, and VSD
asserts that balancing these values must be conducted with care and intention. In the context
of AI tools, a candidate’s data stands as a proxy for the candidate themselves, and ranking
algorithms are typically calibrated to a particular, implicit balance of values, processing
candidate data and returning a deterministic output. Kate Crawford [
Agonistic Recruitment Tools. As an ideal, we imagine a multi-model, multi-human
decisionmaking process in which the aim of an AI recruitment tool is not simply to generate a single
ranked shortlist of candidates, but rather to facilitate a deliberative process by presenting
candidates according to multiple criteria and various value-balances. A CBR engine, bias-mitigated
language models and qualitative LLM components can all be combined in a multi-model
agonistic ecosystem, which can be further expanded to include fairness methods such as fair ranking
and counterfactual data augmentation. Models should incorporate randomness or aim to identify
outliers as means to avoid unfair systematic exclusion and place value on uniqueness. Such an
ecosystem extends the notion of ensemble models, in which the output and capabilities of
multiple models are aggregated to balance the strengths and weaknesses of individual components,
e.g., combining simpler models (CBR) with black-box neural networks to enable a fine-tuned
balance between interpretability and performance. By explicitly establishing numerical proxies
for diferent normative stances, mathematical constructs such as the Pareto front can be used
to select from a set of aggregating schemes that minimize value trade-ofs. Moreover, there
are existing guidelines for exploring the space of possible models according to various value
balances [
This work represents a reversal of a status quo in which fairness metrics act as proxies for
undeifned normative stances and existing models are modified to optimize them. Deep engagement
with stakeholder values should precede technical methods, which can then be built from the
ground up to operationalize and balance those values. To summarize our contributions,
• An explicit set of values and stakeholders is determined and the tri-partite investigation
of stakeholder normative stances and existing technical methods is described.
• The case-based reasoning engine is presented in the context of pro-active design, with
explicit attention towards the operationalization of stakeholder fairness perspectives.
• The CBR engine is situated within a larger technical research and development
environment that engages with fairness aspects that are not fully encompassed within the CBR
framework.
• We draw from the notion of agonistic machine learning as a means to combine and balance
the values underlying individual models and evaluation methods while further promoting
moral agency and fully engaging with the fact that candidates cannot be reduced to a
particular set of numerical features or a single score or deterministic automated decision.
VSD emphasizes progress, not perfection; the ideas described in this article aim to stimulate
further discussion and research.6 Designers bear a moral responsibility for the values embedded
6An EU Horizon sister project, FINDHR, has recently published parallel work detailing their design of fair
recruitment systems based on VSD [
This work is part of the Europe Horizon project BIAS, grant agreement number 101070468, funded by the European Commission, and has received funding from the Swiss State Secretariat for Education, Research and Innovation (SERI). 7Recent political developments demonizing DEI eforts appear to have already influenced ChatGPT. An inquiry with no prior context, dated 12.2.2025, asked, “What is the definition of diversity bias? ” and received the response, “Diversity bias refers to a type of bias that occurs when eforts to promote diversity lead to unintended discrimination, favoritism, or misrepresentation. It can manifest in several ways, such as:…Tokenism – When organizations include individuals from diverse backgrounds for appearance’s sake rather than fostering genuine inclusion. Reverse Discrimination – When attempts to promote diversity result in bias against traditionally dominant groups. Diversity bias can occur in hiring, media representation, education, and decision-making processes…” The ripple efects that polarized politics and the political leanings of AI tech giants could have on AI tools in general–let alone in recruitment–are both harrowing and dificult to quantify.