Beyond Bias Detection: Community Auditors and Normative Reasoning in AI Oversight

AI technologies are being rapidly adopted by government agencies across federal, state, and municipal levels. Public sector organizations now use machine learning models for a range of tasks, including predictive analytics for determining benefits eligibility, fraud detection in welfare programs, resource allocation for public housing, and predictive policing to anticipate crime hotspots [1, 9, 16]. These applications promise increased efficiency and data-driven decision-making in domains traditionally guided by human administrators.

While replacing discretionary human judgment with automated systems might streamline operations, it also risks concealing the value judgments and trade-offs inherent in policy decisions. For instance, predictive policing algorithms may disproportionately target neighborhoods of color if trained on historically biased police data, echoing longstanding patterns of over-policing [19]. Similarly, automated welfare eligibility systems might unfairly flag or sanction low-income individuals due to biased design or data, exacerbating feelings of alienation and unfairness.

A particular risk of adoption is that AI tools may not align with local cultural and contextual factors. Many AI systems are developed for broad markets and may embed assumptions or data that do not align with a particular community's needs or values. When a city or agency adopts an off-the-shelf algorithm (e.g., for policing or child protective services), it may not account for local context, leading to misaligned or unfair outcomes [41]. In practice, this could mean that an algorithm trained on data from one jurisdiction performs poorly or inappropriately in another with different demographics or policy priorities. Such misalignment can undermine the tool's effectiveness and legitimacy in the eyes of both officials and the public.

Therefore, without adequate oversight, public-sector AI systems could reinforce existing power imbalances, biases, and institutional inequities, thereby undermining democratic accountability and eroding public trust in government institutions.

In recent years, there has been considerable momentum in developing frameworks, guidelines, and tools for auditing AI systems for potential issues such as bias, unfairness, and broader societal harms (e.g., [2, 5, 8, 29]). An algorithm audit is a technique that systematically interacts with an algorithm and analyzes its outputs to infer details about its internal processes and potential external consequences [28].

Most algorithm audits are conducted by specialized groups, including industry professionals, academic researchers, activists, and regulatory bodies [28], who use a range of toolkits and frameworks to evaluate RAI principles. For example, Algorithmic Impact Assessments (AIA) (i.e., systematic evaluations of an algorithm's potential effects on fairness, privacy, and rights) are used before the system is deployed [35]. AIAs are analogous to environmental impact statements for algorithms, often involving public input and review to anticipate harms and inform mitigation strategies. In addition to process-oriented assessments, technical toolkits are available to examine and improve model fairness. For example, IBM's AI Fairness 360 toolkit and Google's What-If Tool provide practitioners with libraries and interfaces to test bias models, explore how changes in data or thresholds affect outcomes, and compare different fairness metrics. These tools help translate abstract criteria into concrete diagnostics and adjustments throughout an AI system's development lifecycle. Using such techniques, developers can proactively identify biases and adjust the model or decision-making policies accordingly.

However, expert-driven audits often miss significant issues that end users readily identify once systems are deployed in real-world environments [22, 39]. Expert auditors may lack the cultural insight or lived experience necessary to recognize certain forms of harm or to know where to look for them [14, 22, 39, 43]. Additionally, some harms only emerge, or are only perceived as problematic, within specific use contexts, which formal audits may not adequately simulate [13, 18, 20, 22, 38, 39].

As such, evaluating an AI system's impact is not purely a technical task. For instance, understanding fairness in a housing allocation algorithm might require insights from social work or urban studies to know what equity in housing means for different communities. The most effective responsible AI practices bring together data scientists, legal experts, policymakers, and community stakeholders to interpret audit findings and determine acceptable trade-offs [4, 7, 38]. This socio-technical collaboration is necessary because fairness is socially defined; what counts as a fair outcome or a sufficient explanation can vary with community values and historical context.

While traditional expert-led audits have been instrumental in identifying algorithmic harms, they also exhibit notable limitations, particularly in recognizing context-specific or culturally embedded biases. A growing body of work has begun exploring how everyday users can meaningfully contribute to the auditing process. Recent research and industry practices highlight a shift toward more user-engaged approaches, where non-experts are empowered through tools, platforms, and incentive structures to detect and report harmful algorithmic behaviors.

Numerous cases have demonstrated how everyday users, rather than experts, have detected critical flaws in deployed systems ranging from search engines [10] to review platforms [40] and machine translation tools [33]. In response to these trends, Shen et al. [39] introduced the concept of "everyday algorithm auditing," which describes how ordinary users, through routine engagement with algorithmic systems, can collaboratively uncover and investigate biased behaviors. These users often organically form hypotheses and test them to reveal systemic issues. Building on this insight, DeVos et al. [14] conducted behavioral studies to explore why users are frequently so effective at surfacing algorithmic harms that expert audits might overlook.

Recognizing users’ potential in algorithm auditing, researchers have increasingly focused on developing systems that directly engage users in uncovering harmful algorithmic behaviors [14, 25]. This line of work has produced various tools, including interfaces, interactive visualizations, and crowdsourcing pipelines, that support user participation in identifying biases and problematic outputs [3, 11, 23, 31]. These approaches span a spectrum from practitioner-led audits to user-driven initiatives where individuals take greater control. Examples include Search Atlas by Ochigame and Ye [32], which allows users to compare Google search results across different countries, and Dynabench by [23], a platform where users generate test inputs and flag problematic model behaviors. Lam et al. [25] advanced this approach with IndieLabel, empowering users to detect biases and author audit reports for decision-makers. Industry efforts have likewise begun to incorporate user engagement. Twitter's 2021 "algorithmic bias bounty" challenge invited users to surface biases in its image cropping algorithm [36] while Meta adopted Dynabench to audit natural language processing models [23]. Google's AI Test Kitchen also encourages users to experiment with LLM-powered agents and report harmful behaviors.

However, algorithmic auditing is not merely a technical process but a deeply normative, participatory practice. Community-centered approaches to AI alignment foreground whose values guide system behavior. For example, Bergman et al. [6] proposed the STELA methodology, which uses deliberative sessions with underrepresented groups to ‘elicit latent normative perspectives’ and rich contextual insights about what AI should do. Similarly, Reisman et al. [35] argues that accountability frameworks (e.g., Algorithmic Impact Assessments) exist to ensure that public agencies protect ‘basic democratic values, such as fairness and due process,’ when deploying automated systems. In these frameworks, community stakeholders play an important role in surfacing the value judgments and trade-offs inherent in different definitions of fairness. This literature suggests that audits should invite affected communities to apply their own fairness criteria, so that audits reflect pluralistic, context-sensitive values rather than a single technical metric.

User-driven auditing research further shows that non-expert auditors naturally use personal values and lived experiences when interpreting system behavior. DeVos et al. [14] found that users’ search and evaluation strategies are “heavily guided by their personal experiences with and exposures to societal bias.” That collective sensemaking among auditors is crucial for uncovering harms. In practice, audit participants form normative expectations of how an algorithm should behave, rooted in their own notions of right and wrong. These expectations function as a values-informed lens through which auditors notice biases and imagine remedies. Unlike purely quantitative audits, this value-based reasoning makes latent value tensions visible (e.g., conflicts among privacy, accuracy, and equity). In privileging diverse perspectives, non-expert auditing can reveal when technical definitions of fairness conflict with community priorities.

We argue that auditorship should be conceived as a normative reasoning practice based on these insights. Foregrounding whose values count and how they might conflict, audits surface the hidden trade-offs and equity tensions in AI systems, guiding more democratic and aligned design of algorithmic tools. Non-expert algorithmic audits differ fundamentally from technical audits in that they foreground subjective assessments of fairness rooted in lived experience. While technical audits evaluate algorithms against predefined metrics (e.g., demographic parity, equal opportunity), non-expert audits ask stakeholders to assess whether algorithmic outcomes align with their notions of right and wrong [7, 35]. This normative orientation emerges organically because: (1) affected communities possess contextual knowledge about historical inequities that shape their expectations of fairness [43], (2) participants naturally invoke moral reasoning when evaluating high-stakes decisions affecting vulnerable populations [14], and (3) the absence of technical training means auditors rely on ethical intuitions rather than formal metrics [39].

The question then becomes, what defines a good auditor or, more broadly, a good crowd of auditors in public sector AI? A good auditor is not simply a detector of errors but should be a normative reasoner, i.e., someone who engages with the underlying value trade-offs in algorithmic systems. Prior work in participatory AI and value alignment emphasizes the importance of surfacing “latent normative perspectives” [6]. This suggests that good auditors must be attuned not just to technical flaws, but to the broader social and ethical implications of automated decisions. Drawing from existing literature, several qualities emerge:

• Good auditors bring contextual sensitivity, recognizing that fairness is not universal but shaped by domain-specific factors such as the stakes of a decision or the vulnerability of affected groups. Rather than applying a single standard across all settings, effective auditors adapt their reasoning to the nuances of each scenario [14]. Research in user-centered audit studies reveals that participants draw on their own experiences of bias to identify system harms that experts might overlook, demonstrating the value of situated knowledge in algorithmic evaluation [14].
  
• Effective auditors exhibit principled subjectivity, that is, their evaluations reflect stable value commitments (e.g., equity, harm prevention), even as they accommodate situational complexity. This echoes findings in participatory design research, where stakeholders maintain consistent orientations while remaining responsive to contextual details [26]. Good auditors are neither rigidly dogmatic nor arbitrarily subjective; they balance personal commitments with situational awareness.
  
• Good auditors demonstrate discernment in their ability to identify affected groups and articulate plausible risks that stem from algorithmic deployment. This quality is particularly important when harm is diffuse or socially contingent. Research on everyday algorithm auditing shows that users often detect harms that escape formal evaluation because they possess cultural insight and lived experience necessary to recognize certain forms of bias [14, 39]. Finally, a good audit crowd reflects value pluralism by aggregating judgments from individuals with diverse lived experiences and priorities, illuminating tensions and trade-offs that might otherwise remain invisible in formal evaluation protocols [35].
  

With these principles in mind, we evaluate auditors’ performance across multiple dimensions of good auditorship. We assess whether participants demonstrate normative reasoning by articulating value-driven concerns, exhibit contextual sensitivity by adapting their evaluations to different AI deployment contexts, show principled subjectivity through consistent yet flexible fairness orientations, and display discernment by accurately identifying at-risk populations and high-stakes scenarios.

We developed an exemplar-based introduction and structured audit protocol to guide participants through the evaluation of AI systems. Our approach was designed to make algorithmic auditing accessible to non-experts while maintaining the rigor necessary to capture meaningful fairness judgments.

3.1.1 Guiding Audit Overview. Since auditing an AI system can be challenging for participants without prior exposure to the process, we created a guided introduction that illustrates key audit concepts and procedures through a worked example. In this introductory material, we presented a fictional community member, Eric, who demonstrates how one might approach auditing an algorithmic system. Eric's inclusion draws from research showing that relatable, narrative-based exemplars can reduce cognitive load in complex tasks and help participants understand nuanced decision-making processes [12, 30].

When choosing a persona, we considered facets of Eric's identity. We selected social work as Eric's profession to model non-expert reasoning. Social workers frequently navigate ethical trade-offs and advocate for marginalized populations, making this profession appropriate for demonstrating values-informed auditing. We deliberately chose a male persona despite social work being a female-dominated profession. Here, we sought to avoid reinforcing gender stereotypes about care work while still conveying empathy and a focus on community. Overall, in our characterization of Eric aimed, we sought to emphasize the value of human-centered perspectives in AI evaluation. In this, Eric is portrayed as a compassionate social worker who collaborates with community groups to address issues like housing insecurity and employment barriers.

Figure 1figure.caption.2 (left) shows Eric's profile and background, illustrating his role as a bridge between technical systems and community concerns. This persona provides participants with a concrete reference point for how real-world experience and community knowledge inform algorithmic evaluation. Through Eric's reasoning process, participants gain a deeper understanding of how context factors (in Eric's case, socio-historical) are factored into algorithmic assessment.

Fig. 1. On the left, an auditor persona (Eric) who serves as an exemplar to guide participants through the audit process by modeling how individual knowledge and social work experience inform algorithmic evaluation. On the right, a demonstration scenario showing an algorithmic hiring system used in the exemplar-based introduction. Participants reviewed this system description to understand how AI models process application data (experience, education, skills, previous jobs) and make hiring recommendations before proceeding to evaluate actual public sector scenarios.

Following Eric's introduction, participants encountered a demonstration scenario featuring an AI system designed to help companies decide which job applicants to interview. This exemplar was intended to ensure participants understood the audit logic before proceeding to the main evaluation tasks. The scenario provided background information on how the model ranks applicants based on education, work experience, and other factors, as well as the potential implications of relying on algorithmic recommendations in hiring decisions. We provided sufficient technical detail to convey the model's purpose and inputs without overwhelming participants with unnecessary complexity. This approach aligns with established guidelines for scenario-based evaluations in HCI research, which emphasize accessible contexts that elicit engaged and authentic responses [17, 41].

Figure 1figure.caption.2 (right) describes the information that the developer might provide for a hiring algorithm. The description includes data collection (e.g., résumés) and where the model might be implemented. The overview concluded with participants reviewing sample audit questions that probe AI systems for concerns related to fairness, transparency, and accountability. To scaffold this process, we displayed Eric's responses in a distinct red font alongside each question. This design served two purposes. First, it demonstrated possible reasoning for identifying overlooked or context-specific factors (e.g., training data gaps that might disadvantage certain groups). Second, it modeled how to justify fairness judgments, reinforcing the importance of reflective reasoning in algorithmic evaluation [7, 26].

Figure 2figure.caption.3 shows the interface where Eric's responses appear, accompanied by hover icons that participants can activate to reveal additional justification or commentary. This interactive feature encourages deeper engagement by allowing participants to consider Eric's reasoning before deciding whether to adopt or challenge his perspective. Consistent with participatory design principles, such elements demonstrate how non-technical stakeholders can meaningfully contribute to algorithmic evaluation when provided with accessible, context-rich scaffolding [25].

Fig. 2. Audit interface (right) displaying interactive scaffolding where Eric's sample responses appear in red text with expandable tooltips. This design allows participants to view example reasoning before formulating their own assessments, supporting the exemplar-based learning approach.

Participants were recruited through Prolific, an online platform known for its diverse participant pool and reliable data collection [15]. Prolific was selected to enable targeted recruitment criteria, including age (18 years and older) and English language proficiency, consistent with similar crowdsourced studies in social computing [24, 41].

After clicking our Prolific advertisement, respondents completed a screening questionnaire to determine if they were qualified for the study. The screening included questions about (1) knowledge of AI in decision-making contexts, (2) a short video on AI bias, and (3) two questions to evaluate what they learned from the video (i.e., “What is one reason why AI systems can develop biases?” and “What is a possible consequence of biased AI systems?”). These questions probe whether participants understand that AI models can inherit biases from training data, a key point highlighted in fairness and accountability research [4], and whether they understand the real-world implications of bias (e.g., discrimination against specific groups). Responses to these questions allowed us to evaluate a baseline understanding of AI concepts (or the ability to learn them). If participants answered either question incorrectly, they were removed from the study.

We used quantitative modeling to understand how participants reason about fairness, risk, and harm in various AI scenarios. Our analytical approach was designed to assess the four dimensions of good auditorship identified in our theoretical framework, i.e., normative reasoning, contextual sensitivity, principled subjectivity, and discernment. We discuss each in turn:

• To assess contextual sensitivity, we first examined if participants made differentiated risk assessments across scenarios rather than applying uniform judgments. We fit a linear mixed-effects model with scenario as a fixed effect and participant ID as a random intercept, allowing us to estimate systematic variation in risk perceptions by use case while accounting for repeated measures within individuals. We conducted pairwise comparisons of marginal means across scenarios using Tukey's HSD adjustment for multiple comparisons to identify specific contextual differences in perceived risk.
  
• To test for principled subjectivity (i.e., the extent to which participants maintained consistent value orientations while adapting to contextual specifics) we estimated a cumulative link mixed model (CLMM) predicting ordinal risk ratings. Fixed effects included scenario context, participant demographics (gender, age, income, education, minoritized identity status), fairness-related beliefs measured on Likert scales, and theoretically relevant values from the Portrait Values Questionnaire (PVQ-21), including benevolence, universalism, security, tradition, and power. Participant ID was included as a random intercept to capture individual-level consistency in risk sensitivity. This model structure allowed us to assess whether fairness orientations remained stable across contexts (evidenced by significant random effects) while still being responsive to scenario-specific features (evidenced by significant fixed effects for scenario).
  
• To evaluate participants’ discernment (i.e., their ability to distinguish between high-risk and low-risk AI applications), we compared participant risk ratings with expert-coded severity assessments. The research team independently rated each scenario's expected harm severity using a 5-point scale based on three criteria: (1) likelihood of disproportionate harm, (2) vulnerability of affected populations, and (3) severity of real-world consequences (See Table 3table.caption.8). We then fit a linear mixed-effects model treating expert severity codes as predictors of participant risk ratings, with participant ID as a random intercept. A significant positive relationship would indicate that participants could meaningfully differentiate scenarios based on substantive equity concerns rather than responding arbitrarily.
  
• To assess normative reasoning capabilities, we analyzed participants’ identification of vulnerable groups using precision and recall metrics. For each scenario, we compared participant-identified groups to researcher-coded expected vulnerable populations, calculating the percentage of correctly identified groups (precision) and the percentage of expected groups that were identified (recall). We also conducted chi-squared tests to examine relationships between bias recognition (groups identified as potentially harmed) and design prioritization (groups selected as most important to protect), revealing potential tensions between awareness and actionable focus.
  

We included several attitudinal measures in our analysis, reflecting how participants think about fairness in decision-making (e.g., emphasis on equity, inclusion, objectivity). Participants responded to six Likert-scale items, each reflecting a different conception of fairness in algorithmic systems. Several patterns emerged. First, there was overwhelming agreement that fairness should account for individuals’ needs and circumstances. Specifically, 90% of participants agreed with the statement “AI should consider people's needs and well-being to create fair and personalized experiences,” while only 4% disagreed. This response demonstrates a broad endorsement of personalized, human-centered fairness. Second, participants broadly supported equity- and inclusion-oriented approaches to fairness. When asked whether algorithms should “sometimes consider people's backgrounds to make fairer and more inclusive decisions,” 87.3% agreed and only 6.4% disagreed. Similarly, 79.1% agreed that systems should “take people's backgrounds into account to address past inequalities and promote fairness,” with just 9.1% disagreeing. These responses suggest that most participants recognize the need for historically informed and context-sensitive models of fairness.

Third, participants largely rejected the purely formal notions of fairness based on objectivity. For the statement “AI should base decisions only on objective data to ensure fairness and eliminate discrimination,” only 16.4% agreed, while 73.6% disagreed. This view challenges dominant narratives that equate fairness with objectivity or neutrality, highlighting a general skepticism toward technocratic definitions of fairness. Fourth, efficiency and profit motives were viewed as incompatible with fairness. In response to the statement “AI should prioritize objectives like efficiency or profitability over people's needs and well-being,” 70% disagreed, while 20% agreed. This underscores a perceived tension between organizational performance goals and moral obligations toward fairness. Fifth, opinions were more divided when participants considered goal-focused design approaches. Asked whether algorithms should “focus only on their main goals and not consider people's needs or well-being,” 49.5% agreed, 29.4% disagreed, and the rest were neutral. This ambivalence may reflect a lack of clarity about what constitutes a “main goal,” or discomfort reconciling goal-orientation with human-centered concerns.

The findings in the previous section reveal a general orientation among auditors toward contextual, equity-aware, and person-centered understandings of fairness. Most participants rejected the idea that fairness can be achieved through neutrality or strict objectivity. Instead, they endorsed fairness approaches that consider individual and group-based circumstances. These attitudes provide important scaffolding for understanding how participants later approached specific audit tasks, especially when weighing trade-offs or identifying at-risk groups. They also suggest a readiness to engage with nuanced, value-laden conceptions of fairness in algorithmic systems. To evaluate the quality and character of participants’ audit behavior, we turn to several dimensions of what we term ‘good auditor-ship.’ By this, we mean the degree to which participants engage in thoughtful, informed, and context-sensitive evaluation of algorithmic harms. Specifically, we examine whether participants align their fairness preferences with context (e.g., varying their responses based on the domain of use or the stakes involved); demonstrate stability in risk assessments (i.e., showed consistency in evaluating similar risk scenarios); exhibit discernment in assessing different types of risks (rather than responding uniformly across contexts); recognize fairness trade-offs and could prioritize among competing fairness goals when asked to make optimization choices; identified impacted or marginalized groups, thereby demonstrating sensitivity to equity and historical disadvantage.

We next examined the extent to which individual differences, e.g., demographics, fairness attitudes, and personal values, influenced participants’ assessments of algorithmic risk. As in earlier models, we focused on participants’ ordinal risk judgments at the scenario level, using a five-point Likert scale ranging from “Very Low” to “Very High.” To evaluate whether these assessments reflected stable dispositions or adaptive responses, we estimated a cumulative link mixed model (CLMM) predicting risk ratings. Fixed effects included the AI deployment context, participant demographics (gender, age, income, education, and minoritized identity status), fairness-related beliefs (e.g., support for individual needs or efficiency), and a subset of theoretically relevant values from the PVQ-21 framework (e.g., benevolence, power, and tradition). A random intercept for each participant accounted for baseline variation in risk sensitivity.

The model explained a substantial share of variance (marginal R² =.318, conditional R² =.456), reinforcing the interpretation that both contextual features and certain attitudinal differences meaningfully shaped participants’ perceptions of AI risk. As in prior analyses, scenario context remained a significant and powerful driver of risk judgments. Compared to the baseline, participants assigned significantly higher risk ratings to Immigration Fraud Detection (β = 1.18, SE = 0.53, p =.027), while both Monitor Student Performance (β = -1.37, SE = 0.49, p =.005) and Optimize Renewable Energy Production (β = -1.30, SE = 0.56, p =.021) were rated as significantly less risky.

Among the fairness attitude variables, one predictor reached significance, i.e., participants who expressed neutrality toward the importance of demographic inclusion were less likely to perceive a scenario as risky (β = 0.93, SE = 0.43, p =.031). This may reflect uncertainty or ambivalence about whether background-based considerations should shape model design, and aligns with prior findings that principled reasoning (not blanket concern) guides fairness reasoning. Other fairness predictors, including support for individual needs or efficiency, were not statistically significant in this specification.

None of the demographic variables (gender, age, income, education, or minority association) was significantly associated with risk perception. Likewise, none of the value-based predictors (benevolence, universalism, tradition, security, or power) emerged as significant, suggesting that while these orientations may inform broader social judgments, they were not reliably associated with higher or lower risk sensitivity in this task. The variance of the participant-level random intercept (τ₀₀ = 1.03) indicates some individual variation in baseline risk perception, but these effects were modest relative to the dominant role of context.

Again, we find evidence that participants acted as principled, context-sensitive evaluators rather than as biased or ideologically driven evaluators. Their judgments were shaped by the specific contours of each scenario and, to a lesser extent, by attitudinal orientations toward fairness.

One of the most significant findings of this work is the consensus among auditors regarding fairness principles, despite their demographic backgrounds. This finding both supports and extends prior literature on fairness perceptions while challenging assumptions about demographic variation in algorithmic fairness preferences. Participants in our study rejected purely technical approaches to fairness in favor of contextual, equity-aware approaches. Specifically, 90% of participants agreed that AI should consider people's needs and circumstances, 87.3% supported algorithms considering backgrounds for inclusive decisions, and 73.6% disagreed with basing decisions solely on "objective" data. This consensus aligns with Binns et al. [7] finding that citizens view fairness as fundamentally relational rather than mathematical, requiring consideration of individual circumstances and social context. Importantly, our quantitative analysis found that demographic variables including gender, age, income, education, and minority status—were not significantly associated with risk perception or fairness attitudes. This finding contrasts with some prior work suggesting demographic differences in algorithmic fairness preferences [42], but supports research indicating that context and system characteristics often matter more than individual demographics in shaping fairness judgments [22]. The consensus we observed may reflect the inherently social nature of fairness and that technical metrics must be grounded in shared social understandings of justice [38]. Our participants’ rejection of purely objective approaches (73.6% disagreement) echoes findings that citizens view algorithmic neutrality as insufficient for achieving fairness in high-stakes public decisions [41].

The patterns we observe demonstrates the need for participatory approaches that allow communities to negotiate fairness criteria for specific contexts rather than imposing universal technical standards. While there may be broad agreement on fairness principles such as the importance of considering individual circumstances and addressing historical inequities, the operationalization of these principles requires context-specific deliberation. Furthermore, the tension between consensus on principles and variation in application reflects the challenge of translating abstract fairness concepts into concrete algorithmic implementations [35]. Our results suggest that effective AI governance requires both recognition of shared democratic values and mechanisms for non-expert input on context-specific applications of these values. Rather than seeking demographic representation to capture different fairness perspectives, our results suggest that participatory processes should focus on facilitating deliberation about how shared fairness principles apply to specific deployment contexts, aligning with recent work on deliberative AI governance stressing collective reasoning over demographic sampling [6].

A major implication of this research is that community-based algorithm audits are a feasible and valuable form of participatory AI oversight. We demonstrated that ordinary citizens, when provided with an accessible scenario description, can engage in substantive evaluation of an algorithm's fairness and even identify potential biases or improvements. This finding contributes to the growing body of HCI work on end-user algorithm auditing [14, 25, 39] by extending it to the public sector context. Designing effective participatory audits will require careful attention to usability (so that non-experts understand the AI system sufficiently), representativeness (ensuring diverse participation, especially of those most impacted by the algorithm), and empowerment (giving citizens confidence that their input will have an impact). Our study offers a template for how such audits might be structured in practice, using scenario-based evaluations that resonate with participants’ experiences, and facilitating deliberation to draw out collective insights. We see an opportunity for designers and researchers to create toolkits that support non-expert audits – for example, simplified audit interfaces that allow individuals to test “what if” scenarios on an algorithm (similar to Google's What-If Tool), or visualizations that show the community how decisions are distributed across groups. By lowering the barrier to auditing, such tools could enable non-expert groups or civil society organizations to regularly audit local government algorithms and publish their findings, much like watchdog NGOs.

Importantly, although not tested in this study, participating in the audit process may increase trust in AI. As individuals and communities have a say in how algorithms are constructed, they may feel more confident in and accepting of the outputs of models even if they do not fully align with their values [26]. This suggests a need for feedback loops where participatory audits not only gather data about fairness but also build algorithmic literacy and trust among citizens, as long as the process is genuine and their voices are heard. Our work contributes to CSCW and CHI discussions by illustrating a concrete way to operationalize participatory algorithmic governance. Audits can act as a “safety net” that catches technical audits that miss and as a bridge between AI developers and the public.

Finally, our study proposes insight into how AI developers might integrate feedback throughout the lifecycle of an algorithm. One contribution is identifying that fairness issues often become apparent only when viewed through the lived realities of affected communities. The literature suggests that good auditors are reflective, situated, and morally engaged participants whose contributions extend beyond performance assessment to normative critique. Designing for this kind of auditorship requires supporting both individual reasoning and structured collective deliberation, so that audits can function not only as diagnostic tools but as mechanisms of democratic alignment in AI governance. Therefore, AI development teams (or government contractors building public algorithms) should seek input early, during problem formulation and dataset selection, not just after deployment. Techniques from participatory design could be employed by convening focus groups from target user communities to discuss what fairness means to them in the given context, which could inform the choice of model objectives or constraints. We show that individuals can engage with the concept of an algorithm and raise relevant concerns, such as transparency or undue harm to a group, so their input can be specific and actionable. We suggest establishing feedback channels, such as user advisory boards for algorithms, community testing phases (analogous to beta tests, but for policy algorithms), and ongoing surveys or reporting tools that allow citizens to flag perceived unfair decisions. The AI Incident Database is one such materialization of these recommendations. Incorporating these into the development cycle could lead to AI systems that are more robust to public scrutiny.