Building on our earlier PRISMA review (Udoh, Yuan, & Rorissa, 2022), we revised our initial search query and expanded our search to include additional databases, including those that have become key venues for publishing on information access and algorithmic fairness. We used predetermined keywords with the query: Information AND (retriev* OR system*) AND (search* OR engine) AND Fair* AND “Information access.”

We searched eight databases in three rounds, limiting our inclusion to peer-reviewed full-text articles published from 2015 to 2025, written in English, and relevant to algorithmic fairness and information access. Titles and abstracts were screened using the predetermined inclusion/exclusion criteria, and the full texts of selected articles were reviewed to ensure they met these criteria.

From the initial collection of 635 articles, we removed 164 duplicates, resulting in 471 unique sources. The authors independently screened the titles and abstracts of these articles, removing 359 that were irrelevant to the review, selecting 112 articles for further screening, and including 103 in the final sample. Figure 1 summarises the search results and workflow for the literature search.

Algorithmic fairness in information access has emerged as a central concern for information research, while remaining conceptually fragmented across technical, legal, and socio-cultural domains (Udoh et al., 2024). This fragmentation undermines the field’s ability to evaluate and design systems that equitably mediate access to information, opportunities, and social participation (Ekstrand et al., 2019; Udoh et al., 2022). We synthesise recent PRISMA-based scholarship on algorithmic fairness and information access to examine key concepts, stakeholders, metrics, solutions, and their application in core technologies. These strands reveal a pressing need for a comprehensive, information-access–specific definition of algorithmic fairness that can support cumulative research and practice.

The literature converges on the view that AI/ML fairness is a multidimensional construct, though the dimensions are treated in isolation. A foundational distinction is between individual (that seeks to ensure that similar individuals are treated similarly) and group (focuses on comparable outcomes for socially salient groups defined by protected attributes such as gender, race, or disability) fairness (Barocas & Selbst, 2016; Dwork et al., 2012). Another recurring distinction is between fairness in treatment (intentional differentiation) and fairness in impact (which captures unintended but systematically unequal outcomes) (Barocas & Selbst, 2016). In algorithmic contexts, these are typically operationalised via statistical parity and related measures, which assess whether decisions or exposure to information are independent of group membership (Friedler et al., 2021). These concepts highlight that fairness cannot be reduced to the absence of explicit discrimination; instead, the distributional effects of ranking and recommendation policies should be considered.

The literature also identifies several further dimensions relevant to information research. Internet or ‘net’ neutrality concerns equitable access to online content and services, requiring that providers avoid blocking, throttling, or prioritising specific sites or applications (Conn, 2010). Network fairness examines whether users or applications receive a fair share of system resources and how network position can confer self-reinforcing advantages (Venkatasubramanian et al., 2021). Fairness in ranking and exposure focuses on how search and recommendation systems allocate visibility and attention, explicitly recognising that exposure is a resource (Balagopalan et al., 2023; Diaz et al., 2020). Finally, legal and policy dimensions of fairness are anchored in accessibility protections, such as Section 508 in the United States (U.S.), and cross-national regulatory comparisons (Sonnenberg, 2020; Yang & Chen, 2015).

These dimensions point to fairness as a layered construct encompassing equal treatment and impact, exposure and resource allocation, infrastructural conditions (net and network neutrality), and compliance with accessibility and anti-discrimination regimes. However, these layers rarely are integrated into a unified conceptualisation specific to information access.

Algorithmic fairness spans multiple, interrelated issues across information systems, including bias, inclusion, network neutrality, and the broader harms of automated decision-making such as stereotyping, misrepresentation, and disenfranchisement. These concerns arise in contexts like internet access, search and recommendation systems, information quality, digital literacy, and metadata practices.

Inclusion remains a central challenge, particularly in equitable access to digital resources. Algorithmic filtering and prioritisation can undermine network neutrality and restrict access (Conn, 2010; Yang & Chen, 2015). Persistent disparities affect older adults, people with disabilities, and marginalised populations, reflecting gaps in both access and digital literacy, especially in developing regions (Gebremichael & Jackson, 2006; Yang & Chen, 2015).

Result diversity highlights how systems often reproduce dominant perspectives due to biased data and design. For instance, although women comprise 28% of U.S. chief executive officers (CEOs), they appear in only 10% of Google Image results for “CEO” (Kay et al., 2015). Efforts such as diversity-aware summarisation aim to mitigate this (Celis & Keswani, 2020), yet disparities persist, including in multilingual retrieval, where Bantu-language speakers may experience worse performance than English speakers (Chavula & Hussein, 2016).

Network fairness and net neutrality address how infrastructure and policy shape equitable access. Unequal network structures can amplify disparities through feedback loops, while policy debates, such as U.S. net neutrality, demonstrate how access to essential information, including healthcare, can be constrained (Conn, 2010; Venkatasubramanian et al., 2021).

Algorithmic systems also shape visibility and opportunity through ranking and recommendation. Biases may stem from training data (exogenous) or broader systemic conditions (endogenous), producing allocational and representational harms (Blodgett et al., 2020; Rekabsaz et al., 2021). These dynamics align with concerns about algorithmic oppression, where systems reproduce inequality through mechanisms like stereotyping and redlining.

Bias and stereotyping persist due to unrealistic assumptions, such as completely protected–attribute data, and flawed modelling choices that skew exposure (Gao & Shah, 2021; Joachims, 2021; Kirnap et al., 2021 ). Inequities are compounded by differences in readability and information quality, disproportionately affecting users with lower digital literacy, disabilities, or age-related barriers (Hamwela et al., 2018; McKearney et al., 2018; Yang & Chen, 2015). Privacypreserving behaviour may also lead to less dependable or more extreme recommendations (Spinelli & Crovella, 2020).

Accountability and transparency remain difficult to standardise. Fairness is measured through computational metrics, while accountability and transparency rely on audits and user studies, prompting efforts to develop clearer evaluative frameworks (Bernard & Balog, 2023). Interpretability adds another dimension: users frequently disagree with system outputs, for example, 81% disagreed with Alexa’s top result in one study, highlighting the gap between system logic and user expectations (Dash et al., 2022).

Overall, fairness in information systems extends beyond accuracy to include equitable access, representation, interpretability, and the distribution of informational benefits and harms.

The reviewed literature increasingly adopts stakeholder and multi-stakeholder perspectives, emphasising that those most affected by algorithmic systems should participate in their governance (Costa et al., 2020; Webb et al., 2018). Studies of patient-facing health information highlight misalignments between the content and presentation of online resources and the preferences and capacities of people with chronic conditions, such as low back pain, signalling the need for participatory design and evaluation (Costa et al., 2020; McKearney et al., 2018).

Stakeholders identified include technology firms, regulators, researchers, educators, librarians, civil society organisations, and end users, particularly marginalised communities such as people with disabilities, older adults, and users with lower literacy (Webb et al., 2018; Yang & Chen, 2015). In recommender systems, Sonboli et al. (2022) articulate a multi-stakeholder framework that distinguishes between consumer fairness (users), item fairness (content providers), and platform fairness (system operators), and argue that fairness objectives must be negotiated among these groups rather than specified unilaterally.

Within the information research community, librarianship and media ecosystems are positioned as active stakeholders in fairness. Critical librarianship initiatives call for exposing subject-heading bias, teaching search engine bias, and curating social-justice-oriented guides, while credibility models for news communities leverage interaction graphs to identify trustworthy sources and ‘citizen journalists’ (Bains, 2020; Mukherjee & Weikum, 2015). These strands suggest that algorithmic fairness in information access is not a purely technical property of systems, but also involves professional practice, governance, and collective responsibility.

Fairness metrics in information access have evolved from simple decision-level measures to more nuanced, exposure–oriented, and context-sensitive frameworks. In ranking contexts, Diaz et al. (2020) define fairness as equal expected exposure for items, (or groups, of equal relevance, explicitly accounting for position bias and user attention. Kırnap et al. (2021) introduce unbiased estimators for fair ranking metrics that address incomplete judgments and logging bias, enabling more accurate evaluation of exposure fairness under realistic data constraints.

Beyond specific metrics, researchers propose frameworks for evaluating fairness in ranked lists. Sakai and colleagues advance group fairness and relevance (GFR) measures that manage nominal and ordinal groups, as well as soft group membership, and quantify how far exposure distributions depart from fairness targets (Sakai et al., 2023). Complementing this, Bernard and Balog (2023) provide a systematic review of fairness, accountability, transparency, and ethics in information retrieval and conclude that the field lacks standard definitions and calls for taxonomies of requirements that can be operationalised in evaluation (Bernard & Balog, 2023).

Where explicit protected attributes are unavailable, estimation and auditing approaches have been proposed. Quantification-based estimators approximate group composition and adjust outputs to promote fairer distributions, while counterfactual auditing evaluates whether individual-level outcomes change under hypothetical modifications to sensitive attributes (Cornacchia et al., 2023; Fabris et al., 2023). In recommender systems, multi-stakeholder fairness frameworks define distinct metrics for consumer, item, and platform fairness, and axiomatic models specify desired fairness properties for ranking systems, including search and job recommendations (Giner, 2023; Sonboli et al., 2022).

These developments underscore that fairness is increasingly measured at the level of exposure and attention, but that metric design is shaped by assumptions about context, group definitions, and stakeholder priorities. This reinforces the need for an overarching conceptualisation that clarifies how different metric families relate to each other in information access settings.

The identified algorithmic harms centre on inclusion and equity across protected characteristics such as race, gender, age, religion, national origin, marital status, socioeconomic status, and proxy variables like location.

Proposed solutions address fairness in search, retrieval, and recommendation systems through agnostic ranking models (Ai et al., 2023; Cornacchia et al., 2023; Rekabsaz et al., 2021; Rieger et al., 2024), distribution-aware recommender systems (Ekstrand et al., 2024; Diaz et al., 2020), diversity and multi-sided exposure methods (Balagopalan et al., 2023; Dhaliwal et al., 2021; Gao et al., 2022; Gao & Shah, 2020; Sonboli et al., 2022; Wu et al., 2022 ), image summarisation diversity (Celis & Keswani, 2020), and pairwise fairness measures (Fabris et al., 2023).

Model-based approaches address bias directly. For example, AdvBERT uses adversarial learning to remove protected attributes while preserving relevance, though proxy variables, such as location for race, can still reintroduce bias (Rekabsaz et al., 2021). Counterfactual auditing methods, such as PRET, instead assess fairness by testing outcome changes under altered sensitive attributes (Cornacchia et al., 2023).

Several fairness interventions engage librarianship and media ecosystems to preserve fairness and facts in the face of misinformation and disinformation. Critical librarianship calls for naming subject-heading bias, teaching search-engine bias, and curating social-justice guides, while credibility models for news communities infer credible articles, trustworthy sources, and expert ‘citizen journalists’ from interaction graphs (Bains, 2020; Mukherjee & Weikum, 2015).

Preprocessing decisions and evaluation strategies further shape fairness outcomes. Metadata decisions significantly affect information retrieval (IR) performance, while randomised evaluation and optimisation frameworks promote equal expected exposure and parity of attention (Diaz et al., 2020; Roy et al., 2018). Several solutions focus on preserving fairness at the resource ranking level in search engines, including uncertainty-aware retrieval and unbiased estimation of fairness metrics from sparse labels, and improved tag recommendation using topic models (Guiver & Snelson, 2008; Kirnap et al., 2021; Krestel et al., 2009).

Accountability and transparency are advanced through system design and auditing. For instance, IBM’s CLEVER (Continuous Learning for Evolving and Verifiable Embodied Representations) search system provides visibility into ranking pipelines, building on Kleinberg’s (1999) HITS (or Hypertext Induced Topic Search) algorithm framework, while broader efforts combine audits and user studies to evaluate fairness (Kumar et al., 2006).

Another group of solutions that aim to ensure the fairness of algorithms within search engines includes cluster-based retrieval for improved access (Lamprier et al., 2010), spam-resilient ranking (Du et al., 2007), and fairness in advertising mechanisms such as keyword bidding (Rey & Kannan, 2010).

Context-aware and user-centred approaches further refine fairness, including models that incorporate task context, can better infer user intent to retrieve relevant and valuable information (Koskela et al., 2018), while search behaviour (e.g., literature search and data seeking) is more closely intertwined than previous literature suggests (Kramer et al., 2021).

Several sources frame fairness in terms of inclusion and accessibility to resources for people with disabilities. These accessibility solutions range from early communications stations for blind users (Rahimi & Eulenberg, 1974) to modern legal and policy audits of government compliance with Section 508 (Sonnenberg, 2020), and cross-national comparisons of the rules and laws of Internet accessibility protections, and calls for stronger private-sector accessibility mandates (Yang & Chen, 2015). Mobile and geospatial systems also enhance equitable access through integrated, location-based services (Sun et al., 2017).

Sociotechnical interventions also address misinformation and bias through critical librarianship and credibility modelling in news ecosystems (Bains, 2020; Mukherjee & Weikum, 2015). Fairness–aware rank fusion methods, such as WISE Fusion, aim to ensure proportional or equal representation in high-stakes contexts like hiring (Cachel & Rundensteiner, 2024), much more than the fairness and utility performance of state-of-the-art methods.

These interventions show that achieving fairness in information access in practice requires coordinated technical, evaluative, and institutional strategies.

Search engines, information retrieval systems, recommender systems, ranking algorithms and mechanisms, and websites are among information access technologies relevant to the discussion of algorithmic fairness.

In information access systems, fairness concerns whether benefits and resources are distributed equitably across separate groups (Ekstrand, Burke, & Diaz, 2019; Jeong, 2006). Fairness concerns could be raised regarding the algorithm, the data they utilise to produce results, the output, or the search results (Makri et al., 2021), and interfaces such as graphical displays used by visually impaired or blind users (Jeong, 2006). The potential for bias in search engines is significant, from the introduction or embedding of ‘social, political, and moral values in their ranking functions’ (Ekstrand, Burke, & Diaz, 2019, p. 7) to biased crawling and indexing, where webpages in specific languages could be under-indexed. Therefore, any solutions to algorithmic fairness issues in information access must directly address factors relevant to these technologies, including developing and applying fairness metrics that assess web usability, readability, and quality (Hamwela, Ahmed, & Bath, 2018; McKearney et al., 2018), and are pivotal in objectively evaluating the fairness of search system ranking decisions (Kirnap et al., 2021; Rekabsaz et al., 2021).

The two major sectors directly impacted are online retail and streaming services, where the influence of issues such as biased training data on the fairness of these technologies is substantial (Joachims, 2021). The heightened scrutiny around search engines and fairness issues has prompted some to propose fairness evaluation metrics that consider protected attributes like gender or ethnicity (Kirnap et al., 2021), while others (Shah & Bender, 2024) advocate for reassessing the core information retrieval research focus to adequately address the tremendous impact algorithmic decision systems have on information access, individuals, and society. It suggests that algorithmic fairness cannot be fully understood without considering the interplay of data, algorithms, interfaces, and user characteristics across the entire information access pipeline.

The literature explicitly connects algorithmic fairness to broader notions of information access as a human right rooted in equality, justice, and non-discrimination (Udoh et al., 2022, 2024). As AI and ML systems increasingly shape how individuals and communities search for, encounter, and interpret information, fairness in these systems becomes a prerequisite for equitable participation in knowledge production, public discourse, and everyday life (Ekstrand et al., 2019; Li et al., 2014).

However, fairness is currently conceptualised through multiple, partially overlapping lenses: statistical parity and disparate impact; exposure equality and diversity; accessibility and disability rights; stakeholder participation and commonsense expectations; and socio-technical accountability and transparency (Bauer & Schiele, 2024; Bernard & Balog, 2023). While this diversity reflects the multidimensional nature of fairness, it also creates a fragmented landscape in which definitions, metrics, and interventions are difficult to compare and aggregate.

Thus, while information research has developed rich work on specific fairness dimensions, metrics, and solutions, it lacks a comprehensive, field–specific definition of or a conceptual framework for algorithmic fairness in the context of information access (Udoh et al., 2024). Developing and debating such a framework within the Information Research community would support more coherent evaluation, design, and governance of information access systems in an increasingly algorithmic world.

Several overlapping and intersecting themes emerged from the previous section, and we triangulated those findings with current themes not captured in the review. Again, the overlapping/emerging themes form the foundation for FAIRS (‘Fair Algorithms for Information Retrieval and Seeking’) that conceptualises algorithmic fairness in the context of information access (Figure 2). We present the framework before discussing crucial concepts that should constitute any conceptualisation of algorithmic fairness in the context of information access.

Bias appears in the literature as a pervasive, overlapping theme that links data practices, modelling choices, and power relations in information access systems. It manifests in stereotyping, redlining, disenfranchisement, and ‘dirty’ or biased data, often through imbalances in user and item data, content production, and interaction logs that feed retrieval, ranking, and evaluation models (Ferguson, 2017; Noble, 2018; Richardson et al., 2019). Scholars argue that algorithmic systems rely on flawed assumptions, such as that data accurately reflects reality, that the future mirrors the past, and that algorithms are neutral, despite significant human judgment shaping their design and outcomes (Barocas & Selbst, 2016; Moses & Chan, 2016; Selbst et al., 2019).

This work converges on the view of bias as socio-technical: preexisting in institutions and hierarchies, technical in design and implementation, and emergent in use (Crawford, 2016; Friedman & Nissenbaum, 1996; Kroll et al., 2017). Moreover, algorithms align with and advance ideological worldviews (Beer, 2009, 2013; Danaher, 2016; Fuller & Goffey, 2012; Mackenzie, 2015).

Algorithmic systems both personalise and sort while also enabling discrimination, exclusion, and the reinforcement of racialised and gendered orders, as documented in search engines, social media feeds, and recommender systems (Beer, 2009; Gillespie, 2014, 2019; Noble, 2018; Pasquale, 2015; Tufekci, 2015). In information access specifically, this includes representational harms such as misrepresentation and stereotyping, as well as allocative harms tied to unequal exposure and opportunity, including in criminal justice and credit scoring (Ferguson, 2017; Pariser, 2012; Richardson et al., 2019; Wu et al., 2022). For example, the same search query entered by three users, each from a different continent at the same time, tends to produce different sets or orders of results.

Across these strands, concerns about ‘algocracy’ foreground how hiddenness and opacity in data collection, modelling, and deployment undermine accountability, enabling algorithmic systems to act as de facto arbiters of public relevance and legitimacy (Danaher, 2016; Kroll et al., 2017; Peter, 2017; Selbst et al., 2019). The resulting picture, which FAIRS adopts, treats bias not as a purely statistical artifact but as a structuring condition of contemporary information ecosystems that must be confronted through contextual, multi-level interventions in data, models, interfaces, and institutions (Crawford, 2016; Friedman & Nissenbaum, 1996; Noble, 2018; Pasquale, 2015).

The user is a key stakeholder in evaluating algorithmic fairness. In the context of information access, the question is ‘Fair to whom?’ To whom should the algorithmic decision system be fair? And how much is fair enough?

In different application domains of algorithmic systems, users face different fairness challenges. In job and candidate seeking search, whether the user receives a fair set of job opportunities in the recommendations or ads in their feed, whether the match or fit score is a fair score, does not redline or overvalue some candidates, whether the candidates have fair visibility when recruiters filter for top candidates (Geyik & Kenthapadi, 2018), and what fairness concerns come from regulatory requirements or procedures. Similarly, fairness in discoverability also directly impacts users’ news and music discoverability. For example, do the music and news items receive fair exposure and visibility across topics and genres, are users exposed to diverse content regardless of demographics or location, and are recommendations and search results free from biases affecting both users and creators (Epps-Darling, Bouyer, & Cramer, 2020)? The point is, ‘News search and recommendation influences user exposure to news articles on social media, news aggregation applications, and search engines’ (Ekstrand et al., 2022, p. 12). This tends to place users in filter bubbles, exposing them primarily to news items that reinforce their beliefs and deepen ideological polarisation (Alstyne & Brynjolfsson, 2005; Pariser, 2011). The bottom line is that information access systems ought to meet various objectives for multiple stakeholders.

Fairness discussions in information access also need to consider the various technologies involved, including search engines, information retrieval systems, recommender systems, ranking algorithms, and web–based interfaces as the core infrastructures through which algorithmic fairness issues materialise (Ekstrand et al., 2019; Makri et al., 2021). These systems are consistently framed as socio-technical pipelines, spanning crawling and indexing, learning-to-rank, personalisation, and interface design, whose design choices embed social, political, and moral values and can systematically skew exposure and opportunity across groups (Ekstrand et al., 2019; Joachims, 2021; Kirnap et al., 2021; Rekabsaz et al., 2021). A substantial subset of studies highlights modality- and context–specific access technologies, including assistive interfaces, mobile platforms, and large language model (LLM)-based systems, demonstrating how accessibility constraints, biased training data, and under-indexing of particular languages or regions exacerbate digital divides and representational harms (Chavula & Suleman, 2016; Jeong, 2006; Schmidt et al., 2010; Shah & Bender, 2024).

Collectively, this work motivates FAIRS’ treatment of information access technologies as a unifying axis in the literature, linking concrete systems and interfaces to fairness metrics, evaluation methods, and governance interventions that seek to ensure equitable exposure, usability, and information quality for diverse users and content providers (Diaz et al., 2020; Hamwela et al., 2018; McKearney et al., 2018; Sonboli et al., 2022; Udoh et al., 2022, 2024).

Algorithmic fairness debates still hinge on which conception of fairness is assumed and what counts as ‘fair enough’ in applied settings, especially given the dominance of Euro-American legal and statistical traditions (Sambasivan et al., 2020; Udoh et al., 2024). Recent work argues for treating fairness as a realistic sufficiency threshold rather than an ideal of perfection, emphasising context-sensitive standards against which complex information access systems can be evaluated and governed (Ekstrand, 2022, 2024; Sambasivan et al., 2020). Drawing on FAIRS, we conceptualise algorithmic fairness as a multidimensional responsibility structure in which accountability, transparency, safety, privacy, and ethics specify interrelated but analytically distinct conditions for responsible information access (Bernard & Balog, 2023; Udoh et al., 2022, 2024).

Fairness as accountability links measured disparities in exposure, error, or relevance to concrete mechanisms for traceability, audit, and redress, making fairness metrics and audits actionable governance tools rather than purely diagnostic instruments (Bernard & Balog, 2023; Ekstrand, 2022; Selbst et al., 2019). Fairness as transparency concerns the scrutability of system purposes, data flows, and optimisation logics for affected stakeholders, including intelligible explanations of decisions, objectives, and trade-offs; without such transparency, neither informed consent nor meaningful contestation of unfair outcomes is possible (Bernard & Balog, 2023; Guo et al., 2023; Udoh et al., 2024).

Fairness as safety foregrounds obligations to anticipate, mitigate, and monitor informational harms, such as exposure to misinformation, discriminatory content, or unsafe recommendations, by integrating harm-sensitive objectives and pessimistic or worst-case evaluation methods that emphasise the most vulnerable users and providers rather than aggregate utility (Diaz, 2024; White & Hassan, 2014). Fairness as privacy highlights how unequal exposure to data collection, profiling, and inference constitutes an unfair burden, positioning privacy both as a constraint on fairness interventions and as a dimension of fairness concerned with whose data, queries, and interaction histories are most heavily exploited (Akter et al., 2020; Udoh et al., 2024).

Finally, fairness as ethics provides the normative grounding that connects these sub concepts to broader commitments to non-discrimination, equity, and respect for persons, extending beyond mere compliance toward critical reflection on how information access systems reproduce or resist structural injustices and representational harms (Gotterbarn et al., 2018; Noble, 2018; Shah & Bender, 2024; Sonboli et al., 2022).

Synthesising these strands, we define algorithmic fairness in the information access domain as:

Some barriers to algorithmic fairness include historical and persistent discrimination, censorship, content divide, disproportionate access to infrastructure, deliberate digital misrepresentation, disparate distribution of content, insufficient legislation and regulation, and lack of net neutrality. Discussions on algorithmic fairness in information access must pay special attention to the fact that some groups are more susceptible to these barriers than others. In the United States, there is a long history of discrimination against black and minority groups dating back to the county poorhouses of the 19th century (Eubanks, 2018). These surreptitious redlining activities persist online through the opaque algorithmic ranking, sorting, and recommendation systems.

The metrics used to determine what and how much is fair in the context of information access also need to be considered. There needs to be greater transparency about the sampling strategies used to determine who is included in the dataset, accompanied by detailed codebooks that not only define the variables but also specify how each variable is recorded, particularly for categorical variables. The selection and definition of variables are crucial, as observations or proxies must be valid and unbiased measures of the target construct to avoid measurement bias. A codebook reflects specific perspectives on how observations should be recorded or coded, which is extremely important in contexts where protected and personal variables are used.

Considering group fairness, group size, or the amount of representation can also contribute to unfairness (Rolf et al., 2021), because naive modelling is likely to be more accurate at predicting the majority group than the minority groups. This is often the root of representational or distributional harms online, where more data and examples from dominant geographical and cultural environments are used to train algorithms for filtering, recommendation, and ranking, and to determine how much exposure a resource gets and who has more access to it.

Technologies do not exist in a vacuum; they are deployed in domains and environments coloured by culture, politics, economics, legal and policy framework, and social norms, among others. Therefore, fairness discourses in the context of information access need to focus on implementation contexts as well. Despite the pervasiveness of the Internet, net neutrality is not a common feature across all geopolitical zones, as policies, laws, and guidelines govern Internet use, digital information searches, and access to information. With censorship, for instance, entire groups and countries are denied unfettered access to Internet resources. In contrast, some geopolitical regions have no say in how digital resources are distributed and must contend not only with a digital divide but also a content divide.

Similarly, input data and the algorithmic decision results need to be collected and interpreted, considering the data’s social and cultural context to avoid unfairly disregarding local knowledge, perspectives, societal norms, mores, and ethos, as well as bias. A vital direction would be to focus on the entire data pipeline, as each point in the pipeline is a potential vulnerability to unfairness and unfair access practices.

Most work on algorithmic fairness in information access has evolved around classification and rank-based personalisation, typically operationalised through group and individual fairness metrics, exposure constraints, and multi-stakeholder trade-offs (Barocas & Selbst, 2016; Diaz et al., 2020; Dwork et al., 2012; Giner, 2023; Sonboli et al., 2022). FAIRS reframes fairness not merely as a distributional outcome, but as a responsibility structure that spans data, models, interfaces, and governance (Bernard & Balog, 2023; Ekstrand et al., 2022; Shah & Bender, 2024). This shift echoes philosophical accounts that treat algorithmic fairness as an institutional obligation to avoid unjustified differential treatment and impact, rather than as a purely statistical criterion (Wachter & Mittelstadt, 2019).

First, FAIRS integrates bias, user roles, technologies, barriers, metrics, and implementation context into a single conceptual map, thereby countering the fragmentation that separates ‘fairness in ranking,’ ‘accessibility,’ and ‘policy/regulation’ into parallel but rarely connected discourses (Conn, 2010; Ekstrand et al., 2019; Sonnenberg, 2020; Yang & Chen, 2015). Ekstrand et al. (2022) emphasise stakeholder and exposure trade-offs in information access; Selbst et al. (2019) foreground sociotechnical abstraction; and Burke (2017) distinguishes consumer– and provider-side fairness in recommender systems. FAIRS extends these contributions by insisting that any serious account of fairness in information access must treat bias, stakeholders, access technologies, barriers, metrics, and context as jointly constitutive, not optional add-ons (Friedman & Nissenbaum, 1996; Nissenbaum, 2001; Noble, 2018; Udoh et al., 2022, 2024).

Second, FAIRS makes explicit that accountability, transparency, safety, privacy, and ethics are analytically distinct but interdependent fairness sub concepts that define what responsible information access requires (Bernard & Balog, 2023; Makri et al., 2021; Rekabsaz et al., 2021). This aligns with recent work that maps fairness, accountability, transparency, and ethics (FATE) in information retrieval as overlapping but under-specified notions in need of clearer taxonomies and operational requirements (Bernard & Balog, 2023; Visier, 2023). By embedding these sub concepts at the core of the framework, FAIRS links fairness metrics and audits to concrete governance demands, such as traceability of decisions, meaningful contestation, harm mitigation, and ethical orientation, rather than treating them as external constraints (Danaher, 2016; Gao & Shah, 2021; Joachims, 2021).

Third, FAIRS foregrounds the user not as a passive recipient of rankings and recommendations, but as a central stakeholder whose informational autonomy, exposure, and opportunities are directly shaped by algorithmic decisions across search, recommendation, and filtering systems (Ekstrand et al., 2022; Geyik & Kenthapadi, 2018; Kramer et al., 2021). This perspective resonates with work on filter bubbles, confirmation bias, and the homogenisation of recommendations, which documents how iterative feedback loops can entrench exposure patterns and worldviews (Alstyne & Brynjolfsson, 2005; Pariser, 2011; Tufekci, 2015). FAIRS situates these phenomena within a broader picture of distributive and representational harms, thus connecting experiential user harms to structural issues such as digital divides, accessibility failures, and network or net neutrality constraints (Gebremichael & Jackson, 2006; Hamwela et al., 2018; Venkatasubramanian et al., 2021; Yang & Chen, 2015).

These elements motivate a more formal, comprehensive definition of algorithmic fairness tailored to information access:

This definition builds on and extends existing accounts that emphasise fair opportunities for exposure and quality for both consumers and providers in information access systems (Burke, 2017; Ekstrand et al., 2019; Ekstrand et al., 2022).

While FAIRS provides a more holistic lens, it also foregrounds tensions that are not easily resolved within existing technical or policy toolkits and reveal why a responsibility-based conception of fairness is normatively attractive but practically demanding (Barocas & Selbst, 2016; Blodgett et al., 2020; Sambasivan et al., 2021).

The multi-stakeholder nature of information access systems means that ‘Fair for whom?’ and ‘Fair how?’ cannot be answered independently (Costa et al., 2020; Sonboli et al., 2022; Webb et al., 2018). Provider-side fairness, consumer-side fairness, and platform-side objectives (e.g., engagement, revenue, regulatory compliance) often come into conflict (Burke, 2017; Ekstrand et al., 2019). For instance, equalising exposure for underrepresented artists or news outlets may reduce short-term click–based relevance for some users, while optimising for maximum utility may exacerbate existing inequalities in visibility and opportunity (Ekstrand et al., 2022; Gao et al., 2022; Wu et al., 2022).

FAIRS do not resolve these conflicts; instead, it requires that stakeholders explicitly surface and govern them (Costa et al., 2020; Webb et al., 2018). The framework positions fairness metrics (e.g., equal expected exposure, pairwise fairness, demographic parity, calibration) as tools for articulating trade-offs rather than as stand-alone solutions (Diaz et al., 2020; Fabris et al., 2023; Giner, 2023; Kırnap et al., 2021). This stance directly acknowledges impossibility results showing that different fairness criteria cannot simultaneously be satisfied under realistic conditions and that value-laden choices among these criteria are unavoidable (Friedler et al., 2021; Wachter & Mittelstadt, 2019). In doing so, FAIRS invites designers and regulators to treat these conflicts as sites of negotiation shaped by legal, cultural, and ethical commitments, not purely technical optimisation problems (MacCarthy, 2019; Sonnenberg, 2020).

Another central tension lies between fairness constraints and traditional information retrieval goals of relevance and utility (Guiver & Snelson, 2008; Lewandowski & Spree, 2011; Li et al., 2014; Menczer et al., 2004). Fair ranking approaches that constrain exposure or prioritise diversity can degrade conventional effectiveness metrics, while ranking systems optimised solely for relevance risk reinforcing dominant perspectives and undeserving marginalised users or content (Celis & Keswani, 2020; Chavula & Suleman, 2016; Kay et al., 2015). Studies of search engines and scientific information show that highly ranked results often privilege specific framings or communities, with lower-ranked results offering systematically different perspectives that remain effectively invisible to most users (Ekstrand et al., 2019; Li et al., 2014).

Within FAIRS, this tension is reframed as a normative design choice: whose utility, over what time horizon, and under which constraints count as relevant (Bauer & Schiele, 2024; Ekstrand et al., 2019). Distributionally informed evaluation and stochastic ranking approaches already move in this direction by considering exposure distributions and worst-case group performance, but they still leave open fundamental questions about how to integrate long-term social welfare, epistemic diversity, and harm reduction into “relevance” (Diaz et al., 2020; Ekstrand et al., 2024; Gao & Shah, 2020). FAIRS thus highlights the need to reconceive relevance itself as a fairness–aware construct, especially in high-stakes domains such as health, employment, and civic information, where unfair access can have material consequences (Joachims, 2021; McKearney et al., 2018).

FAIRS also illuminates a persistent friction between fairness interventions and privacy protections (Cornacchia et al., 2023; Fabris et al., 2023; Spinelli & Crovella, 2020). Group fairness metrics often require knowledge of protected attributes, yet many legal and ethical regimes encourage or mandate minimisation, obfuscation, or non-collection of such data (Barocas & Selbst, 2016; Joachims, 2021). Proxy-based or quantification-based approaches attempt to approximate group composition without direct identifiers, but they introduce their own risks of misclassification, opacity, and potential reidentification (Cornacchia et al., 2023; Fabris et al., 2023).

By treating privacy as both a constraint on and a dimension of fairness, FAIRS underscores that fairness interventions can themselves produce unfair burdens when they require invasive or asymmetric data collection from specific groups (Research Agenda for Algorithmic Fairness, 2022; Wachter & Mittelstadt, 2019). At the same time, privacy-preserving practices that indiscriminately suppress sensitive information may obscure systematic harms and impede targeted remediation (Blodgett et al., 2020; Noble, 2018). This creates a genuine governance dilemma: How to design accountability and audit mechanisms that can detect and address unfairness without deepening surveillance or eroding informational autonomy for already vulnerable communities (Danaher, 2016; Tufekci, 2015; Udoh et al., 2022, 2024).

A further tension arises from the non-portability of fairness notions across cultural, legal, and infrastructural contexts. Much of the fairness discourse is grounded in Western, Euro-American understandings of discrimination, legal protections, and institutional capacities, which do not straightforwardly apply to contexts with different histories, regulatory environments, or social norms (Gebremichael & Jackson, 2006; Sambasivan et al., 2021). Work on the non-portability of algorithmic fairness in India, as well as cross–national comparisons of accessibility laws and net neutrality regimes, illustrates how seemingly neutral technical criteria can fail or even backfire when transplanted across domains (Conn, 2010; Sambasivan et al., 2021; Yang & Chen, 2015).

FAIRS directly incorporates implementation context as a constitutive element, raising an open question: How far can a single framework travel before it becomes an instrument of epistemic or regulatory colonialism (Noble, 2018; Udoh et al., 2022, 2024)? The emphasis on contextsensitivity, sociotechnical analysis, and multi-stakeholder governance is intended to mitigate this risk, yet it also implies that instantiations of FAIRS will, and must, diverge across settings (Costa et al., 2020; Webb et al., 2018). This variability complicates efforts to standardise fairness metrics or benchmarks globally and demands further work on reconciling local adaptation with crosscontext comparability and accountability (Ekstrand et al., 2019).

The FAIRS framework, while more comprehensive than prevailing models, has several limitations that must be acknowledged.

First, FAIRS is primarily conceptual and synthesising: It identifies the necessary components and relations for thinking about fairness in information access, but does not specify operational metrics, algorithms, or governance procedures (Bernard & Balog, 2023; Ekstrand et al., 2019). It offers guiding questions rather than prescriptive answers, and its efficacy depends on how practitioners, policymakers, and communities translate its categories into concrete practices and constraints (Costa et al., 2020; MacCarthy, 2019). This leaves a gap between high-level responsibility claims and the everyday realities of engineering and product decisions, where competing incentives and resource constraints are acute (Gao & Shah, 2021; Shah & Bender, 2024).

Second, the framework centres algorithmic fairness in the information access domain, which may limit its immediate applicability to adjacent but distinct contexts such as credit scoring, criminal justice, or hiring platforms that are less obviously ’information access’ systems yet still rely heavily on ranking and classification (Barocas & Selbst, 2016; Wachter & Mittelstadt, 2019). While many FAIRS components (e.g., bias, stakeholders, metrics, context) generalise, the specific emphasis on exposure, visibility, and access as primary goods may underemphasise other forms of harm or benefit in those domains (Pasquale, 2015; Zarsky, 2016).

Third, FAIRS is grounded in a particular scholarly canon and methodological approach, a PRISMA-based review of largely peer-reviewed, English-language literature from 2015 to 2025, plus additional triangulated sources (Page et al., 2021; Udoh et al., 2022, 2024). This corpus inevitably reflects selection biases, including the relative underrepresentation of non-English scholarship, practitioner know-how, and community-based or activist perspectives on algorithmic harms (Crawford, 2016; Noble, 2018). As such, FAIRS risks systematically underweighting certain experiences of unfairness (e.g., those documented primarily in grey literature or local advocacy) and overweighting concerns prominent in well-resourced academic and industrial research communities (Kontiainen et al., 2022; Sambasivan et al., 2021).

Finally, the framework presupposes the feasibility and desirability of embedding fairness considerations at the design stage and throughout the system life cycle (Ekstrand et al., 2019; Shah & Bender, 2024). In practice, information access systems are often legacy, opaque, and tightly coupled with business models or infrastructural dependencies that resist substantial redesign (Danaher, 2016; Joachims, 2021). Retrofitting FAIRS into such environments may require political and economic changes that extend well beyond the control of technical teams or even individual organisations (Ekstrand et al., 2022; MacCarthy, 2019). This limitation underscores that FAIRS is as much a normative agenda for governance and institutional reform as it is a conceptual framework for system design (Costa et al., 2020; Webb et al., 2018).

The FAIRS framework raises a set of open questions that define a research and practice agenda for algorithmic fairness in information access.

In sum, FAIRS advances the field by offering a domain-specific, responsibility-oriented definition and framework for algorithmic fairness in information access, while openly acknowledging that any serious pursuit of fairness must grapple with deep trade-offs, context dependencies, and structural constraints (Ekstrand et al., 2019; Sambasivan et al., 2021; Udoh et al., 2022, 2024). The framework is intended to provide shared language and structure for more rigorous, context–aware, and ethically grounded conversations among researchers, practitioners, regulators, and affected communities about what fairness in information access should mean and how it might be pursued in practice (Costa et al., 2020; Research Agenda for Algorithmic Fairness, 2022; Webb et al., 2018).

FAIRS answers the concern raised by Selbst et al. (2019) of the lack of a universally accepted and contextually sensitive definition of algorithmic fairness by giving a context-sensitive definition of algorithmic fairness for information access and by specifying the substantive elements that must be considered together: Bias, users, technologies, barriers, metrics, and implementation context. Unlike Burke (2017), which focuses on multisided fairness in recommender systems and especially provider- and consumer-side fairness, FAIRS expands the unit of analysis to the broader information access domain, including search, retrieval, ranking, recommender systems, accessibility, and governance. Most importantly, FAIRS defines fairness not only as exposure or allocation across stakeholders but as an interrelated responsibility structure spanning accountability, transparency, safety, privacy, and ethics across data, models, interfaces, and governance.

Our Framework aims to establish the core constitutive concepts in the algorithmic fairness discourse, enriching and balancing the discourse and providing a better foundation for other research efforts, especially in the information access domain. With this foundational work, new fairness models applicable to and appropriate for the information access context become a viable research agenda. It also provides a reference point for AI, ML, and information technology practitioners, vendors, and researchers on the ethical implications and challenges of integrating AI and ML in information access spaces.