Avoiding bias and understanding the real-world consequences of AI-supported decision-making are critical to address fairness and assign accountability. Existing approaches often focus either on technical aspects, such as datasets and models, or on high-level socio-ethical considerations-rarely capturing how these elements interact in practice. In this paper, we apply an information flow-based modeling framework to a real-world recruitment process that integrates automated candidate matching with human decision-making. Through semi-structured stakeholder interviews and iterative modeling, we construct a multi-level representation of the recruitment pipeline, capturing how information is transformed, filtered, and interpreted across both algorithmic and human components. We identify where biases may emerge, how they can propagate through the system, and what downstream impacts they may have on candidates. This case study illustrates how information flow modeling can support structured analysis of fairness risks, providing transparency across complex socio-technical systems.
Artificial Intelligence (AI) is rapidly finding its way into all areas of society, shaping decision-making
processes across personal, business, public, and private sectors[
Current eforts to address these challenges can be broadly categorized into two approaches with
their own perspectives strengths and weaknesses when it comes to characterizing and addressing these
challenges. The first, technical perspective, focuses on aspects like bias detection in datasets, fairness
metrics, and model explainability [
The second, socio-technical perspective is found in high-level guidelines, Fundamental Rights Impact
Assessments (FRIA), legal and socio-technical frameworks [
The abstraction gap forming in between the socio-technical and technical spheres prevent a holistic
understanding of how technical details and stakeholder efects interact. The problem is not merely one
of methodological focus, but one of lacking bridge structures across the gap. While eforts exist on both
sides of the gap to capture something of the other side, significant challenges remain [
To address this gap, we have developed the Information Flow Model (IFM) methodology, which models socio-technical decision systems as flows of information. IFM represents information states as sites and the transformations between them as channels, mapping how inputs are reshaped into outputs and decisions. This provides a form of structural transparency: it makes visible how technical operations and human judgments interact within the broader organizational process, and how biases and responsibilities can propagate through these interactions to produce downstream impacts. In this way, IFM models bridge the gap between technical and socio-technical approaches, as both model-level metrics and higher-level organizational and stakeholder considerations can be connected and traced across the same structural substrate.
Contribution. This paper presents a real-world case study applying the IFM methodology to a recruitment process that combines AI-supported and human decision-making. We make three contributions: 1. We demonstrate how IFM enables joint modeling of technical components (such as automated matching systems) and surrounding social processes (such as recruiter interpretation and client selection). 2. We show how structuring the decision process as flows of information through sites and transformations supports systematic identification of where biases may be introduced, how they propagate, and what impacts they may have on stakeholders. 3. We illustrate how the IFM bias impact analysis clarifies socio-technical structures of decision systems, makes information dependencies explicit, and traces how roots of discrimination arise and propagate across both algorithmic and human steps when the method is performed in a participatory setting.
Taken together, the case highlights IFM’s potential as a pragmatic tool for system-level analysis and design in real-world contexts.
Structure of the Paper The rest of the paper is structured as follows: Section 3 gives a brief technical description of IFM. Section 4 gives a short methodological overview. Section 5 presents the results of the use-case study, presenting information flow models and analysis of bias and impact. We discuss the outcomes of the case-study and framework as well as directions for further study and development in Section 6.
The challenges of AI fairness and accountability are not defined by a single gap but by multiple
separations forming this abstraction gap. These include divergences of method, abstraction, and
perspective, which fragment the field into disconnected approaches that struggle to speak to one
another [
Bridging the gap thus requires a combination of holistic, structural and situated perspectives. A
structural perspective connects decision-making, dependencies and consequences. A holistic perspective
gives an interconnected whole, not be just a collection of details, and ensures that the we are not creating
small framing traps [
Impact assessments such as AIAs and FRIAs foreground governance, rights, and accountability, but stop
at the surface of system documentation and do not address how technical operations actually unfold
[
Our proposed IFM methodology aims to address these separations not by replacing existing approaches, but by providing a structural substrate in which they can connect. At its core, IFM models decision-making as networks of information sites and channels, where each transformation detail it’s information dependencies and outputs. The IFM graph form a mesh which might be arbitrarily detailed down to technical components but still span over the same socio-technical frame. Through paths can be traced downstream to stakeholder impacts or upstream to structural causes. This structural mesh ensures a combination of holistic closures and capture of technical properties.
While we will here present IFM in particular, the real intended strength of IFM is less as a standalone tool and more as a connective framework, providing the scafolding through which technical, organizational, and stakeholder perspectives can meet in the middle, bridging the abstraction gap.
Situatedness is essential in order to understand real consequences of unfairness such as particular
stakeholder impacts and the challenges involved to tackle these risks [
The IFM method employed in this paper and in general uses participatory methods, interviews and workshops as the primary method to obtain a situated model capturing the dynamics and processes as described and understood by those who interact with it, more than using system specification or other objective descriptions.
Participatory approaches can be understood in several ways: as acknowledgment of the democratic
right to participate [
The landscape of participation is characterized by diferent levels and types [ 21], each serving distinct
purposes and ofering varying degrees of stakeholder involvement. However, participation also presents
notable challenges that must be carefully navigated. Critically, participation washing - where the
participation of stakeholders are used merely to justify an otherwise obtained design or model - does
not constitute efective participation [
Understanding how to achieve efective participation is essential for meaningful engagement. It
is argued that for participatory design, risk mitigation, or audit processes, participants who do not
necessarily possess specialized technical skills must first develop suficient understanding of the
sociotechnical system [
With IFM we partly flip this perspective. The goal of the IFM modeling process is to understand how potential AI components are embedded within the broader socio-technical system and it is not necessarily so that anyone possess this understanding at the start of the modeling process. Instead each stakeholder knows their own role, be it a technical role, a user or other stakeholder. The main aim of the IFM participatory process is to let each stakeholder communicate their situated perspective and thereby shape their part of the decision environment. An initial lack of understanding of the whole is not just expected but informative, showing the role of potential transparency or explainability mismatches. As modeling progresses the developed model shift towards being the object of the study, a role it can take just to the degree is properly capture the situated perspectives of the involved stakeholders.
Studying how decision-making occurs in socio-technical systems requires a clear and systematic representation of how information flows through the decision process. The Information Flow Methodology (IFM) ofers a structured way to describe these processes, enabling the identification of potential biases and the evaluations of how outcomes afect stakeholders. The IFM model was developed loosely inspired by the works of Barwise [23, 24] and uses similar components.
This section provides a conceptual overview of the IFM and its core components (Section 3.1), followed by descriptions on the two structured properties - bias and impact, in IFM (Section 3.3). A more in depth treatment of the IFM method can be found in the IFM Methodological guide [25].
The IFM Model provides a detailed description of the decision-making process by specifying the information available for each decision and the resulting outcome. This specification enables a thorough analysis of the factors that might influence a decision outcome and how the resulting outcomes might, in turn, impact other decisions.
The IFM consists of three primary components: 1. Sites: Represent the states or repositories of information within a decision process, such as inputs, intermediate results, or outputs. 2. Channels: Represent the transformations, operations, or decisions that move information from one site to another. 3. Networks: Represent sites and channels connected to form directed graphs, capturing entire decision processes from initial inputs to outcomes.
Any decision process that begins with some initial information and ends with a decision can be described using these elements. The decision process itself constitutes a network, each sub-decision a channel, and the inputs, output decision, and any intermediate result would be sites.
Following this conceptual overview, we extend the model with a formal definition to enable more rigorous analysis.
A Network =< , , , , , ℛ > where is a set of Sites representing sources of information, is a set of Channels representing directed transformations between sites. : → and : → are functions assigning types to sites and channels respectively. = (, ≤ ) is a type system where is a set of type classes and ≤ is a partial ordering defining subtype relationships. ℛ is a set of inference rules determining properties and relationships based on types.
A Site ∈ represents a source of information on which a decision can be made. While sites are unique in , the information they contain can be further described by assigning them one or more types. Types will be further described below and are not unique to a site. A Channel ∈ is a directed transformation between one or more input sites ( ) ⊆ with one or more output sites ( ) ⊆ . Like sites, channels are unique in but can be assigned channel types via or inferred types via .
A valid network forms a directed acyclic graph (DAG) where: • A site in can be the output of at most one channel in . • The output of a channel cannot serve as an input to an upstream channel (no cycles). • ( ) = { ∈ | is not an output of any ∈ }. • ( ) = { ∈ | is not an input to any ∈ }.
• ( ) = ∖ (( ) ∪ ( )) (intermediate outputs of ).
A network can be seen as an abstract channel : ( ) → where ⊆ ( ). Similarly, any channel can be further defined as a network. This allows a network to be nested within another larger network ′ where it participates as a channel. This nesting provides the possibility of abstraction, as parts of large or complicated networks can be abstracted as channels representing sub-networks. It also allows causal reasoning about the properties that networks have as channels.
Alternatives are used to represent uncertainty or options within the formal description. Alternatives are defined as a set of possible values, possibly including the absence of a value (denoted ). Alternatives will be denoted either by specifying the set of alternatives prefixed by a question mark "?", for instance ?{1, 2}, which denotes that 1 or 2 might apply, or by applying the ? operator to a symbol ? =?{, }, denoting that might not be present. While alternatives could be applied to any property such as sites, channels or types, we will primarily handle them on a network level. A network or subnetwork containing alternatives represents multiple possible configurations, where each configuration corresponds to one specific network without alternatives.
In the IFM model, both bias and impact are treated as structured properties of the modeled decision process. They reflect diferent but connected forms of misalignment that may afect stakeholders or violate regulatory norms and obligations.
Defintion of Bias , in this context, refers to a systematic deviation from the described function of the decision process. Such biases might result in decreased accuracy through false positives and false negatives and the efects might be limited or disproportionate to specific subsets of input, such as favoring particular groups. Bias may propagate downstream through dependent channels and sites. If its influence is halted or unused, the propagation path ends. Once bias is introduced in the information lfow it might spread downstream, towards outputs. Any downstream channel that depends on a biased site can be treated as a potential proxy for that bias. Conversely, if biased information is not used further or the bias is controlled, the propagation path terminates. This makes it possible to trace the scope and efect of bias and mitigating eforts across the network, in a transparent and formally bounded way.
Definition of Impact here refers to the real-world normative consequences or risk thereof of decisions, especially unintended harms afecting stakeholders or sub-groups. Impacts strongly relate to bias as bias are sources of unintended decisions and information in the information flow, but harmful impacts can be caused also by the system working as intended if the design have not taken riskmitigation into account. As impacts occur at the downstream ends of the DAG, understanding their causes requires tracing their paths upstream towards contributing decisions and information. Where the lfow of bias downstream and the tracing of impact upstream intersect, the bias may a contributing cause to the impact. Since some impacts only occur for particular subgroups, when analyzing contributing paths we might need to follow identifiers or proxy variables connected to this sub-group through the information flow. This determines if the proxy identifier contributes to the impact.
To analyze fairness-relevant risks in an AI-supported recruitment process, we applied the Information Flow Model (IFM) methodology[25]. The approach enables structured reasoning about how information is represented, restructured, and filtered as it moves through human and algorithmic components of decision-making. Our goal in this case study was to construct such a model grounded in semi-structured interviews, and use it to identify potential sources of bias and trace their downstream impacts.
The recruitment process under analysis includes both manual and automated steps: recruiters interpret client needs, generate or refine job descriptions, use AI-based matching tools to identify candidates, and conduct interviews before forwarding recommendations to the client. These activities involve multiple actors and complex collection and use of information — making it a particularly good case for IFM modeling.
The methodology was applied in five main steps, adapted to the use-case: 1. Sketching the initial model. Three semi-structured interviews were conducted with stakeholders (two recruiters, one developer). During the interviews, the two interviewers independently produced informal sketches of the information flow based on the participants’ descriptions of how decisions were made and what information was used. This interview process is intentionally dynamic and focus on the areas which seems most relevant to both the participants and modelers. Interviews also capture more general views of the stakeholders perspective which help greatly to ground further analysis. 2. Formalization and refinement. The sketches were consolidated and refined channel per channel. Channels were reviewed to ensure valid inputs, outputs, and other formal requirements. This stepwise refinement leads to the creation of a formal IFM model. Studying each channel obtained from the previous step one at a time helps to some degree in reducing modeler bias on presumed understanding of the whole flow. If the step by step refinement runs into problems this might prompt the need for additional interviews or reflect actual inconsistencies in need of exploration. 3. Semantic typing and actor assignment. Each site and channel was annotated with semantic types (e.g., list, sublist, unstructured description, etc.) and assigned responsible actors (client, recruiter, AI system). Chanels are similarly typed into filtering, ranking, and abstraction and assigned to sub-networks with semantic labels. Depending on what information is present it might be useful to describe it further, such as what features are actually present in a dataset. In other types a more abstract definition can be used to capture unstructured information. This lfexibility is used to facilitate a holistic perspective which might be dificult on a fixed abstraction level. 4. Identification of potential bias. After understanding types and channel semantics, the model was again reviewed, channel for channel, for transformations that could introduce problems or bias. This step also involved taking existing mitigation mechanisms into account and noting them down. Ambiguities and inconsistencies from the previous modeling process can be marked here as potential sources of bias. Key on this step is that there is no detection of actual bias possible here, from a structural account alone, instead it is the envelope of potential bias which is mapped out - which is structurally dependent. 5. Analysis and tracing of Impact paths. Finally, in Section 6 we analyze whether and how identified biases can propagate downstream to afect candidate outcomes. This is done by establishing if there is a path between a potential source of discrimination or bias through the information flow all the way to a stakeholder outcome. If such a path can be described, then there is structural risk for this impact to materialize.
For a more in depth description on IFM, it’s parts and the methodology, see the IFM Methodological Guide [25]. The results of the first four of these steps are presented in detail together with figures in the following section, followed by step five in the following analysis and discussion section.
This section presents the results from the interview study and following structural analysis.
As per Steps 1 and 2 of the IF methodology three semi-structured interviews were held with use-case stakeholders. Two with recruiters and one with an internal software developer.
A summary of the recruitment process: From these interviews, a general picture of a workflow emerged. With the help of outsourced recruiters, the client with vacancies to fill describes the vacancy to recruiters who create a job description and technical requirements. The recruitment company then in a Sourcing phase employs diferent ways to collect lists of potentially interested and suitable candidates based on requirements. Recruiters search platforms such as LinkedIn for candidates and advertise the job description on various websites. The recruitment company owns a database where candidates register themselves if they are interested in being recruited, or get registered if they apply for a position from another website. To search the database of candidates, the recruitment firm has an AI Matching tool that matches the characteristics of candidates with the job requirements to provide a ranked list of candidates. This list improves the eficiency of the recruitment assignment since there can be too many candidates in the database and searching through filters does not provide holistic comparisons.
After the Sourcing phase begins the Screening phase. The Recruiter reviews the AI list to ensure the results are valid. Depending on the number of open positions for the role, the recruiter selects top-ranked candidates from the AI matching results and additional candidates from other recruitment sources to proceed to the next step. The selected candidates are contacted by the recruiter, and those who pass this initial screening advance to a more thorough interview to assess their skills. The final list of candidates are presented to the client who selects the final candidates to get job ofers.
The information flow model was created following the methodology which is summarized in Section 4. Step 1 of the methodology is about charting out the information flow which underlies the decision process. This sketching was performed independently by two interviewers during the interview process. These sketches — from each interviewer and from each interview — are then compiled together info a collection of channels. Step 2 is about refining these channels so that they conform to the formal definitions of the IFM, thereby creating the formal model. The models created are presented as two ifgures and two tables, together describing three nested levels of decision making. Figure 1 describes the decision process of the client, and embedded in this the decision flow of the outsourced recruitment process. Table 1 gives more detail on each channel of the recruitment process marked in the figure as () − (ℎ).
Several AI tools are used at various points of the recruitment process, most prominently creating of job-descriptions in channel () and automatized candidate matching in channel (). In the interviews we explored the AI Matching system, channel () in more depth and we therefore present it in detail in Figure 2 and Table 2 covering channels (1) − (6). The bare formal model is contextualized in Step 3 by assigning semantic types, roles and actors to sites, channels and sub-networks. The most relevant of this information is shown in the tables and and figures.
As highlighted by the multi-abstraction-level representations in Figures 1 and 2 the information lfow model can be made very detailed or very abstract. The relevant level of analysis depends on the questions which are to be explored. In this case these three levels adequately describe the most relevant for our analysis of the recruitment process and their embedded context.
By considering each channel one by one, we, as Step 4 of the methodology, consider what possible biases could take part of the type of transformation the channel represents. Some channels, particularly within the AI Matching system are designed with mitigation of bias in mind, and these are also noted. We collect the information on potential biases in Table 1 and Table 2. In addition, we note some of these biases below.
In each channel where unstructured or semi-structured information is to be abstracted and represented there is a potential for information loss and distortion through abstraction. This type of abstraction occurs at several stages of the process starting with a) when the Client’s vacancy is to be re-expressed as a job description with appropriate technical requirements. While abstraction by itself can cause systematic biases by missing features or systematic disregard of potentially important features, this becomes of particular concern when interpretation is involved. Since this decision process is to a large degree one of Recruiter judgment such interpretation can not be, and should not be, entirely avoided. However, with interpretation comes potential for bias. Biased interpretation can concern on one side information of candidates, and on another — more systemic — level information from the client. We are highlighting the interpretation involved by tracking the comparison to 2, the recruiters personal understanding of the client requirements as this is the main contrast in each of the more informal ifltering steps.
At several steps of the process, it is likely that the way information, otherwise similar, is presented of included in the process afects how it is received and treated. These are marked as occasions of presentation bias in Table 1 and 2. Of particular mention is that candidates sourced from diferent pathways (AI Matching, DB-search, LinkedIn, personal contact, etc.) might be given diferent weight. This is in particular an issue since the sourcing paths themselves are mainly dependent on diferent information and might thereby shift the recruitment process towards or against more biased grounds of selection. Another type of presentation bias exist as the handof between the AI Match system and the user — the Recruiter — where the user is presented a ranked list of results. The presentation of the ranked list hides the distance in score between the highest and lowest scored possibly creating a false impression of diference.
Recruiters lack access to the internal criteria and weighting used by the AI system, which may lead to overconfidence or misinterpretation of rankings. As mentioned above, the absence of numerical scores hides potential similarity between candidates. It also hides the working mechanism of the matching from the user and prevents them to form an accurate mental model of it’s behavior. In interviews, recruiters expressed that they primarily saw the tool as a search tool — while the developers primarily saw it as a ranking tool. This diference ultimately reflects a diferent view on how it operates. Beyond merely the results, the recruiters were also unaware that the tool discards their inputed job description 0 if there is a job description available from the client, here captured as Vacancy. Efectively this means aspects of the recruiters own understanding 2 which influences 0 is prevented from influencing the outputs of the AI Matching. We can see this as a opacity bias as it is a property causing an alternative information flow (excluding 0 and 2) contrary to user expectations — thereby afecting the output. This is marked as ?0 in Table 2 channel 2.
Taking the opposite role of bias — reducing potential alternative information channels — there are mitigative eforts incorporated in some channels. Just like with potential biases we cannot assume these are efective without analyzing data, but we can note their presence and how their efects change downstream efects, just like with bias. Most importantly we can use the presence of potential bias and mitigation to analyze structural efects , which can inform design, assessment and addition of additional mitigative eforts. There are three channels with explicit mitigative purpose in the modeled information lfow. Firstly we have the AI assisted creation of the job description in channel (). Here the tool includes a bias detection tool working on the language of the job description. Further, with the aim of reducing gender biases both candidates and job descriptions go through a normalization process in channels (1) and (2) to remove gendered language from the comparison elements and . From the perspective of IFM, these normalization channels can be seen as targeted information loss — to the degree they are efective. This means a gendered-language sub-feature(s) × (, ) are lost from downstream processing after the normalization to through (1). This does not mean all potential proxies of gender is removed, but in order for them to remain, it must be possible to show an information path from a feature to a proxy which is present in . We might also keep in mind that key downstream channels, such as (3) operate particularly on language (through semantic distance comparison between features of and ).
This concludes the construction of the information flow models for the use-case, and have covered Steps 1-4 in the IFM methodology. In the next section we are going to show how this assist further analysis of stakeholder impact.
In the previous section an information flow model spanning the use case and it’s context was presented. Here we will explore how this structure can be used to analyze not just the properties of the decision making itself, but also downstream efects on stakeholders, following Step 5 of the IFM methodology.
While Biases are properties of the decision process, Impacts must be considered in relation to the afected stakeholders or subgroups thereof. In this analysis, we will primarily consider the candidates and to what degree they might be unfairly treated or discriminated on protected characteristics or otherwise. In order to do so, we begin with a list of possible discriminatory outcomes. and consider if there are paths through the information flow model which contribute to these outcomes. • O1: Gender-discriminatory hiring outcomes — An outcome in which one gender is systematically underrepresented in the final candidate list 4, despite equal underlying qualifications. • O2: Exclusion of candidates with non-standard CVs or career paths — Candidates with informal, nonlinear, or atypical backgrounds might be disfavored due to mismatches with formal criteria. • O3: Penalty for non-native language users — Candidates whose CVs or online profiles difer linguistically from standard corpora are mismatched or undervalued during semantic comparison or keyword extraction. Less recognized educational institutions or past employers might not be recognized or valued by scoring mechanisms. • O4: Location-based favoritism — Favoring of candidates from specific regions possibly linked to protected status, i.e ‘redlining’ through biased scoring heuristics.
We now analyze whether the information flow model allows for such outcomes to emerge, and if so, through which pathways. For each outcome, we identify a potential path from specific biases in the system to discriminatory efects at 3 or 4. These traces represent feasible chains of information distortion or exclusion that are not corrected before reaching decision points that afect candidates.
We will in this analysis take a deeper look at gender-based discriminatory impacts as an illustrative example and then present the result of a similar approach to 2 − 4.
Gender-based discriminatory efects often originate in pre-existing societal or historical biases getting
reinforced by embedding into datasets, models or practices [
Using the IFM, we can see that there are several possible pathways from these potential sources towards the end result. Each of these paths form a potential scenario for impact (where the bias changes the outcome). Below we consider them in turn: • Client bias. If the Client holds gender preferences, these may influence requirements 0 or the Recruiter’s understanding 2. While 0 is subject to mitigation through a bias-detector at (), the Client’s final decision (ℎ) remains outside of control. Thus, even with a neutral process (–), discrimination may still arise in the last step. • Recruiter bias. If the recruiter is biased in their assessment of vacancy requirements, this will manifest as a biased 2. This might for instance be caused by a recruiter trying to meet assumed biased requirements of the Client, real or not. A biased 2 afects all downstream steps (), ( ), (), since each evaluation of the candidate is conditioned on 2. • Semantic model bias. Societal or historical gender associations may be embedded or enhanced by semantic distance models. Here, however, in extraction for use in the AI Matching both candidate information and vacancy-requirements ( → , 0 → ) undergo a normalization of gendered terms and language. Conditional on the efectiveness of this mitigation, the path from semantic bias to outcomes is closed. • Rule-based scoring bias. Even after a language based normalization, proxies such as gaps in employment history could encode gender diferences (e.g., due to parental leave). In this case however, interviews revealed that the AI matching does not evaluate or capture dates or durations of employment history. This means information on employment gaps is not present. We therefore treat this path as closed.
Conclusion: Conditional on the mitigation, since there are no unmitigated paths, we see no reason to believe the AI Matching is discriminatory on gender. There are, however, several direct routes to gender-based discrimination through the Recruiter and, externally, the Client — and these paths are entirely unafected by the three mitigative eforts.
Nonlinear or atypical career paths could in principle disadvantage candidates if the matching system penalizes gaps or deviations from standard trajectories. However, in this case the AI matching does not evaluate time or continuity of employment. Instead, non-standard paths simply provide more opportunities for positive matches across skills and experiences. Thus, no structural path from career irregularities to discriminatory impact is identified within the AI matching. As before, possible bias remains through the Recruiter’s interpretation (2) or the Client’s evaluation (ℎ).
Candidates using non-standard linguistic forms or with CVs from less recognized institutions might be disadvantaged if the semantic model fails to normalize input or if employers or educational institutions are weighted against each other. Here, however, the extraction → do not capture employer or institution names, merely job-titles. No active path of discrimination can therefore be found within the AI matching itself. Bias could still enter through the Recruiter’s interpretation (2) or at the Client’s ifnal decision stage.
Location information plays a special role in the rule-based scoring, where distance to the employer directly modifies the weighted score. This favors candidates from metropolitan or industrial regions over those from less populated areas. While not discriminatory on protected features in itself, the structural efect of distance based scoring can intersect with socio-economic or demographic patterns, leading to geographically correlated disadvantage e.g redlining-like efects. Unlike 2 and 3, this path remains open within the AI matching and may contribute to location-based discriminatory impacts. Of special note is that this location based scoring is by design, even though potential structural interaction with demographic patterns are not.
Three fundamental impact pathways are detected within the IFM, which are not afected by mitigating eforts. Overlapping paths relating to several outcomes have been combined. 1, 2, 3 4 I1: 3 : Present →a−− t−−i− o−− n−− − Biased Client (h) 4 I2: Vac− →a−− n−− c−− y−− − −
Biased Recruiter (a) [Biased R2] 1, [2→]− − () 2, [2 →]−− (ℎ) →3−− − Client 4 1, 2, 3 I3: EC, Weighted →−s−c−− o−− r−−e− − −
Location Multiplier →−− 0−− − ()− (ℎ) 4 These downstream impacts I1-I3 are noted in Table 1 and 2 for each channel.
The above argument around outcomes 1 − 3 and 1 − 4 does not mean that there are no other harmful or discriminatory outcomes.
The analysis highlights, in line with [
With regards to the AI system the analysis revealed only one - unmitigated - direct path from the AI Matching to discriminatory outcome: the location based scoring modifier. This heuristic strongly favors candidates living close to employers, which can correlate with socio-economic or demographic segregation. Our recommendation from the study was therefore to re-examine this rule by either removing it, tailoring it to vacancy type and location, or testing its efects through simulations on synthetic or real data. This would help avoid unjustified forms of localization bias.
However, the analysis points to risks of an explainability mismatch. If the AI tool only provides a list of candidates, recruiters may treat it as authoritative without questioning its heuristics. Interviews with recruiters show several misunderstandings on how the tool actually operates. A more transparent strategy would be to provide recruiters with the tool’s sub-scores alongside the answers. This would allow them to, if desired, critically evaluate AI results, spot errors, and align their understanding of the automated outputs with their own professional judgment. It would also reduce opacity in accountability pathways by making explicit how particular features contribute to candidate evaluation.
While the analysis found most pathways of discrimination in the AI Matching blocked by mitigation, it was also showed that these mitigative mechanisms within the AI system, such as normalization steps, remain largely inefective to address potential social sources of bias. This includes the bias check on 0 which was explicitly designed to capture social biases, but is largely sidestepped by recruiters’ intuitive understanding of requirements (2).
2 can still carry and reproduce biased assumptions, regardless of the AI Matching or mitigation. To counteract this, socio-technical mitigative measures are necessary. Guidelines on gender equality in the selection process, supported by follow-up monitoring or logging, could strengthen accountability and embed mitigation within organizational practice. Structuring recruiter reasoning and tracking decision patterns over time may further support fairness monitoring and, ultimately, more equitable outcomes.
While this was outside the scope of this paper and not explored in the above analysis, the IFM model can also be used to show that not all decision steps carry equal weight for the final outcome. From the information available at each step we can strongly suspect that the channel (e) Recruiter selection and the channel (h) Client selection will carry stronger weight than the other steps, just because these channels are less structured with a larger variation and with many alternatives - detected and not detected. These steps will therefore combine high decision weight and low accountability. It might be hard to address the behavior of the client, but it is possible to put additional structure or follow-up mechanisms around channel (e).
The case study also provided validation of the IFM framework. When presented back to use case-holders, the model was perceived as accurate and illuminating, clarifying their own processes better than existing documentation. Stakeholders reported that they intend to use the material for internal education and for highlighting socio-technical origins of bias. In post-study reflections, the model was also seen as a useful pedagogic tool, capable of raising awareness and structuring internal discussions about fairness and accountability. In our validation interviews, it has also surfaced that stakeholders felt the participatory method itself was both helpful and educational in that it created reflection and overview, even before seeing the final model. These reactions confirm that IFM can support not only technical redesign but also organizational learning and governance.
At the same time, the limitations of the study must be acknowledged. Most notably, candidates were not interviewed, meaning that the analysis of impacts is necessarily more speculative than the modeling of decision structure, bias, and mitigation. The IFM model also remains interpretive, depending on the ifdelity of stakeholder accounts and the perspective of the modeler. Feedback from external stakeholders highlighted the importance of explicitly attending to intersectional discrimination and modeler bias, suggesting that additional methodological scafolding is needed to ensure robustness and inclusivity in future applications.
A recurring critique of AI ethics is the tendency to produce abstract taxonomies of fairness problems
without suficient connection to practice [
The participatory process of building the IFM model is key to this situatedness. Each stakeholder brings a partial and localized perspective — developers know technical functions, recruiters know their practices, external representatives know broader risks — and these fragments are connected through the modeling process into a coherent structural account. Thus, the resulting model embodies both the situated perspectives of stakeholders and the structural relationships between decision steps. This dual role allows the model to function as a shared artifact for reflection, bridging the gap between abstract fairness discourse and concrete socio-technical practice. Taken together, the case study shows how IFM bridges the abstraction gap through its structural representation of decision-making, its holistic coverage of technical and social components, and its grounding in situated stakeholder perspectives.
This case study demonstrates how information flow modeling can be used to identify, trace, and contextualize fairness risks in socio-technical decision processes. Applied to a real-world recruitment setting, the framework made visible how biases may emerge and propagate across both automated and human decision steps. By explicitly modeling information structure and transformation, it enabled a unified analysis of technical mechanisms and organizational practices—ofering a practical tool for increased socio-technical transparency and contextually grounded fairness assessment.
Beyond the case itself, IFM complements traditional algorithmic performance and fairness tools by providing the contextual layer often missing in real-world applications. Its structural, holistic, and situated approach helps connect abstract fairness concerns to concrete system behaviors and stakeholder impacts.
Looking ahead, IFM lends itself to further development in several directions. Scenario modeling with alternatives could support exploration of what-if configurations and their consequences. In addition, formal representation of bias and mitigation as structural substitutions would enable systematic comparison of system designs and mitigation strategies. Together, these extensions would strengthen IFM’s role in comparative audits and compliance evaluations, and link fairness analysis more directly to risk- and safety-oriented assessments under frameworks such as the EU AI Act.
This work was supported by the “AEQUITAS” project funded by European Union’s Horizon Europe research and innovation programme under grant number 101070363.
During the preparation of this work, the authors used GPT-5 in order to: Paraphrase and Reword & Improve Writing Style. After using these tools, the authors reviewed and edited the content as needed and takes full responsibility for the publication’s content. 3551624.3555290. doi:10.1145/3551624.3555290. [21] F. Delgado, S. Yang, M. Madaio, Q. Yang, The participatory turn in ai design: Theoretical foundations and the current state of practice, in: Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, EAAMO ’23, Association for Computing Machinery, New York, NY, USA, 2023. URL: https://doi.org/10.1145/3617694.3623261. doi:10.1145/3617694. 3623261. [22] M. Katell, M. Young, D. Dailey, B. Herman, V. Guetler, A. Tam, C. Bintz, D. Raz, P. M. Kraft, Toward situated interventions for algorithmic equity: lessons from the field, in: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* ’20, Association for Computing Machinery, New York, NY, USA, 2020, p. 45–55. URL: https://doi.org/10.1145/3351095.3372874. doi:10.1145/3351095.3372874. [23] J. Barwise, D. Gabbay, C. Hartonas, On the logic of information flow, Logic Journal of IGPL 3 (1995) 7–49. [24] J. Barwise, J. Seligman, et al., Information flow: the logic of distributed systems, Cambridge
University Press, 1997. [25] IFM Research Team, Information flow model (ifm) – methodological guide (v0.4 validation version), 2025. URL: https://doi.org/10.5281/zenodo.16252144. doi:10.5281/zenodo.16252144.