Perceptions of voluntary horizontal confidential food safety data sharing: an exploratory interview study with food industry leadership Authors: Linda Kalunga, Katherine Koebel, Carrie Alexander, Martin Wiedmann, Aaron Smith, Aaron Adalja, Renata Ivanek Published: 2026 ABSTRACT Voluntary horizontal sharing of confidential food safety data among companies for joint analysis can improve food safety, efficiency, and decision-making, especially for rare events. Despite its potential benefits, horizontal data sharing in the food industry has lagged, with limited research exploring the reasons for hesitation. To address this gap, we conducted in-depth semi-structured interviews with 27 food industry leaders. Four themes emerged: (1) benefits of data sharing, (2) technical barriers, (3) trust as a determinant of data sharing decisions, and (4) data governance as a solution. Across these themes, we found that companies face trade-offs when deciding to share food safety data, weighing risks against benefits. Data sharing decisions were strongly influenced by trust in stakeholders (e.g., industry peers, regulatory bodies, customers) and in data protection measures. Data governance emerged as a solution to concerns stemming from trust, such as loss of control once data is shared. Taken together, the findings reveal underlying tensions between individual firm incentives and collective benefits, including uneven cost-benefit distributions, opportunism concerns, and participation cost asymmetries. These findings offer timely insights to guide data sharing initiatives and prioritize areas for future research. Introduction The rise of a data-driven economy has prompted companies across industries to recognize the strategic value of voluntary horizontal data sharing, enabling collective insights that no single organization can achieve alone. In this study, we focus on such strategic data sharing in the critical domain of food safety across the food industry (defined here as companies involved in the primary production, processing, packaging, and distribution of food). Data sharing is a process of providing third parties access to others’ original observations or measurements, such as numbers, text, or images, which may be sensitive1, to enable joint analysis and extract useful information. We define voluntary horizontal food safety data sharing as the sharing of food safety data among organizations operating at the same level of the value chain, including firms that may otherwise compete. Horizontal data sharing can involve both intra-sectoral sharing (e.g., among companies that produce the same commodity, such as packaged fresh-cut lettuce) or, less frequently, inter-sectoral sharing (e.g., among companies that produce different products, such as ready-to-eat meats and dairy, but share common interests, such as validation of sanitation procedures for slicers). Horizontal sharing does not encompass industry sharing data with government agencies (even though companies may be concerned that their data could become discoverable to government agencies, e.g., in a legal action) or industry sharing data vertically within the supply chain (e.g., with consumers, even though companies may be concerned about consumers gaining access through data leaks). Prior research suggests that sharing data could improve efficiency by reducing private costs and increasing productivity2,3. Despite these potential gains, food safety data sharing initiatives remain surprisingly uncommon in the food industry. A publicly known example of voluntary horizontal confidential data sharing for food safety improvement through data-driven insights and collaboration is the Western Growers GreenLink initiative. This initiative involves approximately 20 leafy green operators sharing food safety data4, allowing for aggregation and analysis to generate actionable insights5. Operators’ participation in the Western Growers GreenLink initiative may be motivated by the produce industry’s reliance on collective reputation6, whereby a single foodborne outbreak can affect all operators due to limited product branding. While this shared vulnerability illustrates one factor shaping data sharing decisions in the produce industry, it remains unclear what other factors motivate broader participation, both within the produce industry and in other food industries that do not face the same collective risk. Prior literature indicates that concerns surrounding data confidentiality, security, bad publicity, and lack of standardized data formats shape data sharing decisions7–9. Additionally, the risk of legal liability is a barrier to participation in data sharing initiatives10, due to concerns that confidentially shared data may be exposed. Du et al. further note that limited collaboration and weak partnerships function as relational barriers to sharing2, suggesting that these factors could affect implementation of horizontal food safety data sharing within the food industry. Notably, the food industry has voluntarily coordinated on food safety through initiatives such as the Global Food Safety Initiative (GFSI), which aligns standards and certification requirements to reduce audit redundancies11. This demonstrates the industry’s capacity to work collectively towards shared food safety goals. However, voluntary food safety coordination efforts have largely remained in this standards-alignment domain and have not extended to food safety data sharing. This also raises the question as to why the industry does not engage in broader food safety data sharing. One useful approach for understanding this challenge is through the collective action problem lens, which we introduce here to situate our study conceptually. Collective action problems arise when multiple actors stand to benefit from a shared resource or goal, but participation is difficult because individuals must bear private costs or risks while the benefits depend on others also contributing12. For example, while the food industry as a whole would benefit from pooled insights, individual actors must share sensitive data with potential legal liability implications. We draw on the collective action diagnostic framework of Weldon et al.13. Rather than enumerating all elements of the framework, we focus on its initial diagnostic logic, which emphasizes clarifying the specific problem of (in)action, the nature of the shared resource, and whether the challenge primarily concerns provision or use13. In the food safety context, the specific problem is limited participation in voluntary horizontal data sharing, despite potential benefits, such as identifying patterns relevant to rare food safety failures, supporting advanced predictive modeling, and promoting the adoption of best practices. Understanding this challenge also involves clarifying the nature of the resource being created through data sharing. In this case, the resource has a dual (club goods vs. public goods) structure. The shared data itself and aggregated analytical insights can be thought of as club goods, meaning they are excludable (people are prevented from benefiting if they don’t contribute) and non-rivalrous (one person’s use does not reduce availability for others)14. This creates a direct link between contribution and benefit, because only the firms that share data gain access to the pooled insights. However, the downstream outcomes of data sharing, such as improved food safety practices and public health, function as public goods, which are non-rivalrous and non-excludible15. Therefore, the broader benefits extend to the entire industry, regardless of who participated. This dual resource structure shapes both the provision challenge (whether firms decide to contribute their sensitive data despite costs and risks) and the use challenge (whether participants trust that shared data will be protected and used appropriately). Weldon et al. also note that collective action problems usually develop around challenges related to resource allocation, trust, externalities, and free-riding13. In data sharing, companies could potentially face all four of these challenges: they must (i) decide whether to invest resources such as time, labor, and money to prepare data to meet the requirements for participation, (ii) trust that sharing won’t backfire by advantaging competitors, (iii) also believe that others will contribute rather than free-ride on pooled insights while withholding their own data, and (iv) potentially confront a disconnect between who bears the costs and who benefits. Similarly, co-opetition dynamics, the simultaneous pursuit of cooperation and competition16,17, amplify this collective action problem, as firms must balance potential gains from collective insights against the risk of advantaging competitors. However, empirical evidence examining how food industry leaders experience and articulate these tensions in the context of horizontal food safety data sharing remains sparse. Existing empirical research has advanced our understanding of data sharing in general, as well as in agriculture (of relevance to the current study), by examining barriers, costs, and benefits, and willingness to share through quantitative approaches (surveys, choice experiments)18–21, and qualitative work (case studies, expert interviews)22–24. However, the quantitative studies are limited in their ability to capture the nuanced perceptions and contextual concerns that shape food industry leaders’ attitudes towards data sharing. Additionally, none of these studies address the unique dynamics of horizontal food safety data sharing, where firms must simultaneously navigate competitive positioning, liability exposure, and collective food safety improvement. This contextual gap necessitates an exploratory qualitative approach to capture industry leaders’ perceptions of barriers, benefits, and conditions that shape participation decisions in horizontal food safety data sharing25. Semi-structured interviews with information-rich individuals represent a method of choice for this initial exploration, allowing us to probe nuanced experiences, concerns, and perceptions while allowing the researcher flexibility to pursue emergent ideas across participants26. In qualitative research, information-rich individuals are those who possess relevant knowledge or experience about the topic of interest27. In the context of this study, information-rich individuals are food industry leaders, defined as individuals tasked with strategic decision-making. Strategy has been defined by Fisher et al. as “a diagnosis that defines or explains a business challenge or opportunity, a decision or set of decisions for dealing with the challenge or opportunity, and a coherent set of actions to deliver on decisions so as to create an advantage and superior returns”28. Further, we adopt the distinction between three levels of strategy defined by Fisher et al.: corporate-level (e.g., CEO), business-level (e.g., vice presidents), or managerial-level (e.g., director, manager) within their organizations29. Therefore, we conducted semi-structured interviews with food industry leaders to investigate (i) the perceived benefits motivating companies to voluntarily engage in horizontal confidential food safety data sharing, (ii) the barriers that impede such sharing, and (iii) the conditions and strategies that can enable broader adoption of such practices from the perspective of food industry leaders. In the analysis of interviews, coding and theme development were conducted inductively, without reference to collective action and co-opetition theories, but these theoretical frameworks were subsequently used to interpret the findings and contextualize and deepen understanding of the themes identified in the interviews. By interpreting findings through collective action and co-opetition theories, our exploratory research offers insights into how horizontal food safety data sharing can be incentivized and implemented more effectively and more broadly. Results We present results from in-depth semi-structured interviews with 27 participants recruited through purposive and snowball sampling of food industry leaders between June 2023 and January 2024. To ensure sufficient expertise and organizational insight, eligible participants were adult food‑industry professionals with at least two years of corporate, business, or managerial leadership experience29 in U.S. food companies, specifically in the dairy, meat, produce, and food manufacturing sectors, as well as in food safety laboratory services. Interviews explored food safety data collection and sharing practices, and circumstances where data was collected but not shared, including underlying reasons. A semi-structured interview guide is provided in Supplementary Text S1. The transcribed interviews were analyzed using an inductive thematic approach, creating a hierarchical coding structure of parent codes (broad code categories) and child codes (specific dimensions). This allowed themes to emerge from the data rather than being determined a priori. Participants Of the 87 invited food industry leaders, 27 completed interviews (31% response rate). The interview lengths ranged from 19 to 52 min. Participant attributes are presented in Table 1, which provides an overview of represented food industry sectors and the participants’ years of experience and leadership positions within their companies.Table 1Summary of participant attributesAttributeAttribute valueNumber participants (n)Industry sectorDairy6Food Manufacturing7Meata4Produce7Laboratory Services3Years of experience2–586–101311–154≥162Position (leadership experience)CorporateC-suite and/or president3BusinessVP of food safety & quality12ManagerialDirector of food safety4ManagerialManager (plant, food safety, quality)8Total interviews27a Includes beef and poultry. Table 1Summary of participant attributesAttributeAttribute valueNumber participants (n)Industry sectorDairy6Food Manufacturing7Meata4Produce7Laboratory Services3Years of experience2–586–101311–154≥162Position (leadership experience)CorporateC-suite and/or president3BusinessVP of food safety & quality12ManagerialDirector of food safety4ManagerialManager (plant, food safety, quality)8Total interviews27a Includes beef and poultry. Summary of participant attributes a Includes beef and poultry. Summary of identified themes Inductive, codebook thematic analysis of 27 transcribed interviews yielded 65 unique codes and four themes. Coding and theme development were conducted iteratively and collaboratively, drawing on inductive semantic coding to allow themes to emerge from participants’ stated experiences. Table 2 presents the four identified themes and the overarching themes highlighting the benefits, barriers, and solution aspects of voluntary horizontal confidential food safety data sharing. A binary-coded matrix of all codes and interviews is provided on GitHub (see “Data availability” section).Table 2Themes identified in semi-structured interviews of 27 food industry leaders about voluntary horizontal food safety data sharingThemeOverarching theme(1) Horizontal data sharing can enhance food safety practices and decision-making due to access to more data pointsBenefits(2) Technical barriers, such as limited digitalization, standardization of data collection and management practices, limit the feasibility of data sharingBarriers(3) Trust, or lack thereof, is a key factor influencing data sharing decisions, often underlying specific concerns such as confidentiality, liability risks, and data misuseBarriers(4) The need exists for sustained collaboration between academia, government, and industry to identify trade-offs in data sharing, establish data governance, and enable improved food safety through information pooling.Solutions Table 2Themes identified in semi-structured interviews of 27 food industry leaders about voluntary horizontal food safety data sharingThemeOverarching theme(1) Horizontal data sharing can enhance food safety practices and decision-making due to access to more data pointsBenefits(2) Technical barriers, such as limited digitalization, standardization of data collection and management practices, limit the feasibility of data sharingBarriers(3) Trust, or lack thereof, is a key factor influencing data sharing decisions, often underlying specific concerns such as confidentiality, liability risks, and data misuseBarriers(4) The need exists for sustained collaboration between academia, government, and industry to identify trade-offs in data sharing, establish data governance, and enable improved food safety through information pooling.Solutions Themes identified in semi-structured interviews of 27 food industry leaders about voluntary horizontal food safety data sharing Code and meaning saturation analysis We examined code saturation for parent codes using an inductive thematic saturation model, defined as the criterion for continuing or stopping data collection based on whether new codes or themes emerge, with an emphasis on the quantity of new codes identified30. We assessed patterns of code saturation using sequential saturation analysis, order randomization, and bootstrap resampling. We found that no new codes were identified after interview #12 (Supplementary Fig. S1). The comparison of code emergence in the actual vs randomized interview order suggested that the saturation observed in Fig. S1 is relatively robust, though modestly influenced by interview order (Fig. S2). As a robustness check, we repeated the code saturation analysis for the business and managerial-level subgroups; all detailed results are provided in Figs. S3 – S8. The bootstrap analysis showed that a median of 8 interviews reached 80% code saturation (with the 5th and 95th percentiles of 5 and 12 interviews), and a median of 13 interviews reached 90% saturation (5th and 95th percentiles: 8 and 23) (Fig. 1).Fig. 1Cumulative percentage of simulations reaching 80% and 90% code saturation by number of interviews.Each curve is based on 10,000 bootstrap iterations, where interviews were randomly sampled with replacement. The x-axis shows the number of interviews, and the y-axis shows the percentage of iterations in which 80% and 90% of all codes were identified by that count. For example, the 90% curve reaches 100% at the 25th interview, meaning that in every simulation, 90% of the total codes had been identified by that point. Cumulative percentage of simulations reaching 80% and 90% code saturation by number of interviews. Each curve is based on 10,000 bootstrap iterations, where interviews were randomly sampled with replacement. The x-axis shows the number of interviews, and the y-axis shows the percentage of iterations in which 80% and 90% of all codes were identified by that count. For example, the 90% curve reaches 100% at the 25th interview, meaning that in every simulation, 90% of the total codes had been identified by that point. In addition to examining code saturation, we conducted meaning saturation to confirm we had captured the full depth and nuance within our parent codes by tracking the emergence of child codes across all interviews. From our 65 parent codes, we selected 31 parent codes that are central to our research and that represented variation in prevalence (number of interviews in which the code appeared)31. Parent codes with only one child code, as well as parent codes whose child codes all emerged within a single interview, were excluded from the meaning saturation analysis because they did not exhibit development of dimensions across interviews. Because the average prevalence across our dataset was 7.9 interviews, we categorized codes appearing in more than 7.9 interviews as high prevalence and those appearing in fewer as low-prevalence31. We also included both (i) concrete codes, which capture the explicit, definitive issues in data, and (ii) conceptual codes, which capture abstract constructs31. We found all but 2 code dimensions were identified by interview #6, while the rest were identified by #9 (Fig. 2). We achieved meaning saturation by interview #12 for 14 of the 31 selected codes. Seventeen codes had not reached meaning saturation by interview #12, indicating that although dimensions for most codes emerged early, some required additional interviews to fully understand their meaning. Thus, while 12 interviews were sufficient for code saturation, more were needed to achieve meaning saturation. These results support the sufficiency of our sample size.Fig. 2Timing of code development versus meaning saturation.For each parent code (y-axis), the open diamond indicates the interview in which the first dimension (child code) emerged, and the shaded square marks the interview in which the final dimensions were identified; the line between them represents the span of meaning development. Vertical dashed lines (blue, orange, black) at interviews 6, 9, and 12 represent milestones used to assess the progression of meaning development across interviews. Timing of code development versus meaning saturation. For each parent code (y-axis), the open diamond indicates the interview in which the first dimension (child code) emerged, and the shaded square marks the interview in which the final dimensions were identified; the line between them represents the span of meaning development. Vertical dashed lines (blue, orange, black) at interviews 6, 9, and 12 represent milestones used to assess the progression of meaning development across interviews. Descriptive analysis We conducted a descriptive analysis to examine how thematic emphasis varied across participant attributes, including industry sector, years of experience, and position, as shown in Fig. 3. We normalized the aggregate theme frequencies, allowing for proportional comparison across attributes. In Fig. 3, we present a heatmap of the normalized thematic emphasis for the four identified themes across participant attributes. Participants in the produce sector (n = 7; 0.36), those with ≥16 years (n = 2, 0.38), and those in C-suite or president positions (n = 3; 0.39) contributed the highest proportions to Theme 1 (benefit: more data points), indicating that these groups most often emphasized the value of expanded data access. For Theme 2 (technical barriers), participants with 2–5 years of experience (n = 8; 0.41) contributed notably more than other groups, suggesting a potential existence of a generational divide where emerging leaders were better equipped or more interested in talking about the operational challenges. The C-suite participants (n = 3; 0.42) contributed the highest proportions to Theme 3 (Trust), reflecting a stronger emphasis on trust-related concerns among corporate-level leaders. Contributions to Theme 4 (collaboration) appeared more evenly distributed, though C-suite participants (n = 3; 0.32) contributed slightly more than others, indicating that they frequently referenced the collaboration conditions that support data sharing.Fig. 3Heatmap of proportional distribution of code references by theme, normalized by participant numbers among participant attribute values.Color intensity reflects relative emphasis, with lighter yellow indicating lower proportional representation and darker green indicating higher proportional representation. Heatmap of proportional distribution of code references by theme, normalized by participant numbers among participant attribute values. Color intensity reflects relative emphasis, with lighter yellow indicating lower proportional representation and darker green indicating higher proportional representation. To examine whether themes were driven by any single participant subgroup, we conducted an additional analysis in which we identified each participant’s contribution to each theme across leadership positions. This analysis showed that all four themes were supported by participants from multiple subgroups, indicating that thematic structure was not attributable to a single leadership position (Table S1). Analysis and interdependence of themes shaping food safety data sharing decisions Next, we analyze the four themes and illustrate them with direct quotes from participants. Additional quotes, organized by theme and theme aspect, are provided in Supplementary Text S2. Theme 1: Horizontal data sharing can enhance food safety practices and decision-making due to access to more data points Participants consistently emphasized that sharing food safety data could strengthen industry-wide efforts to prevent foodborne outbreaks. A primary benefit was access to a larger and more diverse dataset, which participants believed would enable better trend analysis and root cause identification. As one participant explained, “… an individual company might not be big enough to have enough data points to build and understand trends. … if we could share our data collectively … we might be able to build a trend or understand what’s going on from an industry perspective ….” This perception reflects an important collective action dynamic of individual firms recognizing that the value of shared data depends on widespread participation to create the “critical mass” of data necessary for meaningful trend analysis. Several participants also highlighted the potential for enhancing predictive modeling and enabling earlier risk identification. They envisioned using large data sets and predictive modeling to anticipate food safety events and respond proactively: “The power of interrogating data really comes from as rich and diverse a dataset as possible, and bringing together different parts of the industry and different perspectives, … because the more we do that, the earlier we are likely to notice any early signals…” Finally, some participants stressed that data sharing could level the playing field for smaller companies, giving them access to insights that larger firms typically gain through costly research/investment. As one participant explained, “I think it will be very beneficial, especially for small companies [as they] don’t have those resources and sometimes they struggle to find what is best because they don’t have that experience.” This collective approach was seen as critical for industries like produce, where one participant highlighted that food safety risks extend beyond the individual companies and have industry-wide consequences. They noted that in the event of an outbreak, consumers tend to avoid the entire product commodity rather than distinguishing among brands: “… When an outbreak in that commodity happens, people avoid all of that product. … it’s in the entire industry’s best interest to understand what happened and to be able to kind of collectively raise that bar of food safety.” This shared vulnerability illustrates a unique aspect of collective action in the produce industry, as noted in the introduction, because the externality of one firm’s food safety failure extends to all competitors, creating both collective action motivation and a co-opetition tension (i.e., the need to cooperate with those who could gain a competitive advantage from shared data). Theme 2: Technical barriers limit the feasibility of food safety data sharing within the industry Participants identified two major technical barriers that hinder the adoption of food safety data sharing initiatives: varying levels of digitalization and a lack of standardization. While some companies are transitioning from paper-based systems, others still rely heavily on paper. One participant described their organization as “… evolving in that space of digitalization”, noting that paper documentation persists while acknowledging ongoing efforts to digitalize systems. Digitalization was perceived as strongly tied to company size and profitability/resources. Larger companies were seen as better positioned to invest in digital infrastructure, while smaller companies often faced constraints that slowed their transition away from paper-based systems. As one participant explained, “The big trend would be larger companies that can buy software programs generally will have digital, the midsize will overwhelmingly have digital, the smaller ones will use more like an Excel sheet or a Google sheet, and then the very smallest, they may use those sheets, but they also may just write things down. It really is tied to the size of the company and how much profit, I guess, they make.” This resource disparity creates a classic collective action challenge of achieving the critical mass of participation necessary for meaningful data sharing, which requires investment from actors with different capacities to bear the costs. This disparity was compounded by a lack of motivation to digitalize in an environment resistant to data sharing, as supported by one participant who remarked: “… there’s such a resistance to openly sharing data that there’s no excitement or line of sight to digitalizing”. Here, the collective action problem becomes self-reinforcing, as uncertainty about whether others will participate reduces individuals’ willingness to invest in enabling infrastructure. Beyond digitalization, one participant emphasized that standardizing data practices is the real challenge. They argued that aligning nomenclature, formatting, and protocols is essential for meaningful inference from shared data: “… the technology part is easy, it’s the getting the data into the state it needs to be, is the heavy lifting and what takes resources and funding to do so. It’s really more about setting the standards and the ability to meet those standards, especially in places where you have legacy systems in place.” Others highlighted inconsistencies in data formatting even within their own organizations, describing data standardization as “a minefield” requiring clear protocols to avoid constant cleanup, stating “…we even have a challenge within our business around how we standardize data, let alone between different businesses, units, decimal places…” This standardization challenge reflects a provision problem in collective action as firms must invest time and resources to align their data with shared specifications, bearing upfront costs without guaranteed reciprocal contributions from competitors. Finally, participants expressed concerns about potential discrepancies in food safety programs, such as environmental monitoring programs (EMPs). Specifically, they worried that positive test results could be misjudged, as it would be unclear whether they reflected actual safety issues or simply more rigorous testing. As one participant noted: “ Generally, people that do test and do get positives, do the right actions, but sharing that information becomes problematic because if there’s one entity who never finds a positive, does that mean that they’re good or that they’re not looking in the right manner? And one who finds a bunch of positives, are they a bad entity or are they doing the right thing and they’re testing in the right manner and they’re actually then thus keeping out the right amount of product from commerce.” This concern highlights a co-opetition tension: firms must share data that could be misinterpreted by competitors as evidence of inferior practices, potentially creating competitive disadvantage even when robust testing reflects better food safety practices. This risk of misinterpretation introduces a trust dimension, as firms must believe that competitors will interpret shared data appropriately rather than using it strategically. Theme 3: Trust, or lack thereof, is a key factor influencing data sharing decisions, often underlying specific concerns such as confidentiality, liability risks, and data misuse Across interviews, trust emerged as an overarching determinant of data sharing decisions within the food industry, closely intertwined with technical considerations. As one participant put it, addressing the limited sharing of food safety data will “take another issue which is not scientific, which is not data, which is trust.” While participants acknowledged the existence of technological limitations to data sharing, they stressed that trust shapes whether data sharing is embraced, noting that choices are often influenced by emotion rather than pure logic:“… We always have to remember there’s a human factor to making these [data sharing] decisions. And it becomes emotional at times, not necessarily rational either, or logical, but the human component to making decisions about this has to be considered, because the technology exists to do everything that you want to do right now, and to have it blinded or everything else. But I think we have to think about what’s the human reaction that keeps us from taking that step, especially when we know that, like I said, most people want to protect public health.” “… We always have to remember there’s a human factor to making these [data sharing] decisions. And it becomes emotional at times, not necessarily rational either, or logical, but the human component to making decisions about this has to be considered, because the technology exists to do everything that you want to do right now, and to have it blinded or everything else. But I think we have to think about what’s the human reaction that keeps us from taking that step, especially when we know that, like I said, most people want to protect public health.” This articulation of trust as a “human factor” suggests that even when the technical infrastructure exists and collective benefits are recognized, participation hinges on individual actors’ willingness to accept the vulnerability, which is fundamentally a social barrier. As trust emerged as a key factor in data sharing practices, participants described interrelated concerns that reflect both economic and social barriers to food safety data sharing. A recurring mechanism undermining trust was the perceived loss of control once data leave the company, creating vulnerability to unintended consequences such as data misuse, misinterpretation, and identifiability; one participant noted, “Once I give that data away, unless I’m absolutely confident that it’s protected, it can be used as a weapon against me. So as long as I don’t give it away, they may have a gun, but I haven’t put the bullets in it, too.” Others worried that partial data could easily be taken out of context, noting that “if all the information isn’t shared, then the appearance of what the data may be representing could be wrong and detrimental to a company,” even if corrective actions were properly implemented. This fear was not limited to technical safeguards; it also reflected deeper concerns about how people would react and how organizations weigh risks, such as reputational damage or legal risks, with some admitting, “I’d rather not share in case I’m implicated.” This loss of control represents a provision challenge in collective action because contributing firms to some extent surrender ownership of their data while remaining exposed to how others, including competitors, might use or interpret the data, creating asymmetric risk. Participants consistently linked loss of control to broader risks of external exposures, which heightened their reluctance to share data. Once data leave the company, participants feared that misuse or misinterpretation could trigger increased regulatory scrutiny, damage customer relationships, or erode competitive advantage. These concerns were reflected across regulatory, customer, and competitive exposure channels. Regarding regulatory exposures, participants expressed that their prior efforts to voluntarily share data were often overshadowed by how government entities perceived shared data, reinforcing caution: “I’ve had a couple of experiences where when you offer up those voluntary information and next thing you know, sure enough, you get a phone call inquiring about it. It’s not a positive phone call …. So, a company is opening themselves up for scrutiny.” Customer exposures presented an additional risk of sharing. Participants worried that purchasing decisions could shift rapidly in response to perceived risks, sometimes also resulting in blanket exclusions, even for localized or resolved issues: “if they [customers] learn about a risky area or specific subsection of a region, they can quickly say, “I don’t want anything from that area,” and that becomes problematic. So, of course, people want to protect their own data because they would rather fix something versus being eliminated from a marketplace… Why would I share my data if you’re just going to use it against me?” Finally, the risk of competitive exposures raised concerns about fairness and ownership. Participants feared that competitors could convert insights from shared data into proprietary tools or use shared data to gain a competitive advantage, therefore undermining collaboration: “…we could unintentionally try to sabotage each other. I know that’s an awful thing to say, but it’s very competitive. …. So, how do we protect each other from using that information against each other to have a competitive edge in a very competitive landscape?” These competitive exposure concerns epitomize the co-opetition paradox: firms recognize the collective value of pooled data for food safety yet fear that sharing creates opportunities for competitors to extract benefits, whether through proprietary tool development, strategic market positioning, or even free riding on others’ data while contributing themselves minimally. This represents both a use problem (how shared data will be distributed) and a trust problem (whether competitors will cooperate rather than act opportunistically). Additionally, several participants noted that some in the food industry may view food safety as a competitive advantage rather than a pre-competitive requirement, stating that “… not everyone’s on the same page with food safety not being a competitive advantage.” While concerns about data sharing were often framed as industry-wide hesitations, some participants noted that the resistance did not stem from individuals in food safety or quality roles, but rather from legal teams and executive leadership who worry about exposure of proprietary information: “They are technically competitors; we shouldn’t share more than we need to with industry.” If food safety is viewed as precompetitive, data sharing becomes a logical collective venture; if viewed as competitive, sharing becomes strategically irrational, creating a fundamental coordination challenge. Finally, interviewees described a technology confidence gap. Some doubted that current technological systems could fully prevent data discoverability in litigation. Maintaining control over data release was important to interviewees:“If there were technological lockdowns that gave sufficient confidence that those concerns would not happen, we would be more inclined to participate in a data sharing event. But in reality, if our company is involved in a lawsuit, that information is going to be discoverable. But I can control how that information is released to plaintiff’s counsel. I can control what is released. But so having that control is important to us. It’s our data. We own it.” “If there were technological lockdowns that gave sufficient confidence that those concerns would not happen, we would be more inclined to participate in a data sharing event. But in reality, if our company is involved in a lawsuit, that information is going to be discoverable. But I can control how that information is released to plaintiff’s counsel. I can control what is released. But so having that control is important to us. It’s our data. We own it.” For some, limited familiarity itself with existing technology for privacy protected data sharing was a barrier: “No, I don’t. I don’t have that knowledge. That’s too much technology for me.” Taken together, participants portrayed trust as a moderating factor: when trust is low, perceived risks (loss of control, data exposure, misuse, and misinterpretation concerns) dominate decisions; when trust is credibly built through clear governance, firms become more willing to share. This finding motivates the next theme on data governance. Theme 4: The need exists for sustained collaboration between academia, government, and industry to identify trade-offs in data sharing, establish data governance, and enable improved food safety through information pooling. Participants acknowledged existing food safety collaboration efforts (e.g., via participation in forums and conferences) within the industry but emphasized that horizontal data sharing remains limited. As one interviewee noted, “… industry to industry [data sharing], is very minimal, if at all.” This observation reinforces that voluntary collective action for data sharing has not organically emerged despite recognized benefits, highlighting the need for mechanisms to overcome coordination failures. Across interviews, proposed solutions focused on the importance of third-party governance, problem-specific objectives, and fostering collaboration. These elements are interrelated and operate collectively to build trust. A common solution proposed was to involve third parties, such as academic institutions, to facilitate data sharing as well as spearheading data governance with related privacy-preserving technologies “that would generate some comfort on those that are a little bit paranoid about, … what they’re sharing.” Third-party governance addresses an important collective action and co-opetition challenge of neutral intermediaries who can credibly enforce rules, protect confidentiality, and prevent opportunistic behavior in ways that competitors cannot trust each other to do directly. This solution reduces both provision risks (firms trust their data will be protected) and use risks (competitors cannot directly access or misuse contributions). Several participants also valued “developed relationships with third-party legal folks” with whom companies could consult on food safety and regulatory-specific questions as a safe space to calibrate what can be shared. Another key solution discussed was the need for problem-specific data sharing objectives. Participants stressed that “we shouldn’t be sharing data just to share data. If we’re sharing data to solve a problem, that is the right context.” This framing supports that defining clear and focused goals would make the data sharing process more practical and transparent. While many participants acknowledged the broader public health benefits, they emphasized the lack of firm-level motivators to participate in data sharing initiatives, noting that unless someone can clearly articulate the benefit to the organization contributing the data, beyond a vague appeal to public health, participation will remain limited: “ I’m feeding people, that’s public health, I’m doing lots of things that are public health. The downside of this public health contribution is, I could get sued for billions of dollars and put myself out of business. I’m not going to do it. There’s no incentive right now…” This statement articulates a collective action dilemma where public health represents a public good (non-rivalrous and non-excludable benefits) while litigation risk represents private costs. Firms rationally decline to contribute when collective benefits are diffuse but private costs are concentrated. This underscores the need for governance mechanisms that either reduce private costs or create excludable benefits (club goods) that directly reward contributors. Participants also expressed that effective data sharing within the food industry is possible, but significant efforts are required to develop guidelines for these initiatives in a way that fosters trust: “I fundamentally believe that that [horizontal data sharing] can happen. You just need to put some more structure around it such that people have more trust in it.” Participants also emphasized that a critical part of making data sharing feasible involves ensuring robust confidentiality and security measures. They acknowledged the potential risks of data misuse and the fear that once shared, information could be accessed by unauthorized parties, including regulatory bodies or customers, despite intentions for it to remain within the industry. One participant noted: “It would take significant work to ensure the confidentiality of the data and the security of the data so it couldn’t be misappropriated.” They also noted the importance of engaging with entities that the industry often views as adversarial, such as regulatory bodies that, despite protections, may gain access to the confidential data shared within the industry, stating the need for “more collaborative conversations with FDA on how would they use that data and what regulatory risk might exist.” These accounts emphasize clarity of expectations as part of building trust. These insights reflect a belief that data sharing is possible if the right frameworks are in place to address data governance, security, and trust concerns. Participants’ concerns and suggested solutions spanned multiple data governance domains. In our analysis, we mapped the risks and suggested approaches to data governance: data principles, quality, metadata, access, and lifecycle (Table 3). For example, participants’ concerns about misinterpretation connect to data quality and metadata; calls for clarity on who sees what or concerns about data misuse, reflect data access; and references to discoverability and retention relate to data access and lifecycle. This mapping illustrates that participants viewed trustworthy sharing as dependent on coherent governance across domains, rather than a single technical fix.Table 3Risks associated with data sharing raised by participants mapped against the data governance domainsData governance domainDefinitionRisks Identified by ParticipantsOwnershipLoss of controlMisinterpretationMisuseBad publicityPoor data qualityTaken out of contextUnfairly implicatedPrivacySecurity (breaches)Punitive actionLoss of salesData principlesEstablish the link with business operations, assign ownership, treat data as an enterprise-wide asset, guide policies, promote data sharing, and define roles٭٭٭Data qualityEnsures data meets specific usage requirements, including accuracy, timeliness, completeness, and credibility, for its use to be effective٭٭٭٭MetadataMetadata relates to maintaining concise and consistent data description to aid in the interpretation of data٭٭٭٭٭٭٭Data accessInvolves assigning value to different data and ensuring confidentiality, integrity, and availability through risk analysis and compliance with legal standards٭٭٭٭٭٭٭٭٭Data lifecycleUnderstanding how data is used, usage duration, retention, and storage٭٭٭٭* Indicates where a concern or risk has been mapped to a given data governance domain. Table 3Risks associated with data sharing raised by participants mapped against the data governance domainsData governance domainDefinitionRisks Identified by ParticipantsOwnershipLoss of controlMisinterpretationMisuseBad publicityPoor data qualityTaken out of contextUnfairly implicatedPrivacySecurity (breaches)Punitive actionLoss of salesData principlesEstablish the link with business operations, assign ownership, treat data as an enterprise-wide asset, guide policies, promote data sharing, and define roles٭٭٭Data qualityEnsures data meets specific usage requirements, including accuracy, timeliness, completeness, and credibility, for its use to be effective٭٭٭٭MetadataMetadata relates to maintaining concise and consistent data description to aid in the interpretation of data٭٭٭٭٭٭٭Data accessInvolves assigning value to different data and ensuring confidentiality, integrity, and availability through risk analysis and compliance with legal standards٭٭٭٭٭٭٭٭٭Data lifecycleUnderstanding how data is used, usage duration, retention, and storage٭٭٭٭* Indicates where a concern or risk has been mapped to a given data governance domain. Risks associated with data sharing raised by participants mapped against the data governance domains * Indicates where a concern or risk has been mapped to a given data governance domain. Together, our findings indicate that the identified themes function as an interconnected mechanism that shapes participation in voluntary horizontal data sharing. Specifically, data sharing is perceived as risky, in part due to low trust, whereas data governance acts as a moderator, helping build trust and reduce perceived risk, thereby making participation more feasible and unlocking the benefits of sharing. Figure 4 visualizes this mechanism.Fig. 4Conceptual model depicting the proposed moderated mediation structure in voluntary horizontal food safety data sharing.The model posits that the perceived risk of data sharing influences the collective realization of benefits indirectly through trust. Additionally, the path from perceived risk to trust is conditioned by the presence of data governance, indicating that data governance moderates the first stage of the mediated relationship. Conceptual model depicting the proposed moderated mediation structure in voluntary horizontal food safety data sharing. The model posits that the perceived risk of data sharing influences the collective realization of benefits indirectly through trust. Additionally, the path from perceived risk to trust is conditioned by the presence of data governance, indicating that data governance moderates the first stage of the mediated relationship. Discussion The attitudes, perceptions, and practices of interviewed food industry leaders regarding the voluntary horizontal sharing of confidential food safety data reflect a complex interplay of personal beliefs, industry norms, legal landscape, and lived experiences. The implications of the four identified themes (Table 2) are wide-ranging and intersect across themes (Fig. 4). For example, access to more data points hinges not only on overcoming technical barriers but also on building trust in how shared data will be used. Similarly, trust is closely linked to collaboration, as stronger collaborative relationships can increase the willingness to share sensitive data. These intersections suggest that progress in one area may depend on coordinated advancements in others. We summarize these intersections into four key topics. First, companies face trade-offs when deciding to share data, weighing the risks against the benefits. Second, the level of trust that the data sharer has in the stakeholders (e.g., among the industry peers, regulatory bodies, and customers) and privacy protection technologies strongly influence voluntary horizontal data sharing decisions. Third, participants highlighted data governance as a solution to concerns that stem from trust, such as loss of control once data is shared (Fig. 4). Fourth, although data sharing could benefit everyone, individuals may avoid or undermine it due to misaligned costs, risks, and incentives, reflecting well-documented collective action and co-opetition dynamics. We next examine these topics in relation to existing literature, discussing their implications for data-sharing practices. The trade-offs companies face in data-sharing decisions were evident in how participants described the efforts required, the potential benefits, and the risks involved in engaging in horizontal food safety data sharing. Our descriptive analysis (Fig. 3) emphasizes how these elements varied across participant groups. For example, senior leaders most frequently referenced the value of access to more data points, while also raising trust-related concerns and emphasizing collaboration. This co-occurrence suggests a tension between the perceived benefits of data sharing and concerns about associated risks, alongside cautious optimism about collaboration. This finding aligns with prior work showing that organizations weigh the collaborative benefits of data sharing against perceived risks, including trust concerns and uncertainty about how shared data may be used, creating an inherent tension in participation decisions32. This simultaneous awareness of costs and benefits suggests that the challenge is not a lack of recognition of the value of data sharing, but rather the absence of mechanisms that address the perceived risks. Building on this tension, participants identified the efforts and investments needed for companies to engage in data sharing. The first is the need to digitize existing records of food safety data for sharing purposes and to transition to a digital approach for data collection, which participants noted is still largely paper-based. An interview study with food industry stakeholders conducted by Alexander et al. supports this statement and shows that some companies within the food industry delay digitizing their data due to limited time and financial resources to convert paper to digital records33. Although Alexander’s study was about barriers in adopting new AI technologies, we believe that similar constraints may apply to data-sharing efforts. Therefore, efforts to offset the financial constraints experienced by companies could help accelerate the digitalization process within the food industry. Additionally, participants emphasized the need for data standardization (defined as the establishment of standards across the data value chain to ensure consistency and interoperability)34 for effective data sharing. Qian et al. demonstrate the utility of data standardization by developing a public repository that uses three food safety datasets from different sources35. Authors had to align and preprocess these data formats to create benchmark datasets, facilitating the development of machine learning models for food safety. Mahmood describes how the lack of data standardization and the confinement of biomedical data behind institutional firewalls hinder progress in biomedical machine learning, leading to variability in how models are evaluated36. Thus, concerns about the lack of standardization hindering data sharing are well documented both within and beyond the food industry. To illustrate how similar challenges and solutions apply within the food safety domain, we draw on publicly available examples in foodborne pathogen surveillance. In this context, swab site metadata is often recorded as unstructured free text, leading to inconsistent descriptions and poor data interoperability. To address this, researchers developed a structured, machine-readable schema that standardizes swab site metadata37. This schema has been incorporated into the One Health Enteric Package at NCBI BioSample, a public metadata standard supporting pathogen surveillance across human, animal, and environmental health38. Similarly, FoodOn, an open source ontology, provides a standardized vocabulary for describing food products, their sources, processing methods and packaging to enhance data sharing and traceability39. While such efforts demonstrate progress, standardization challenges remain prevalent at both the intra- and intercompany levels, as stated by our interview participants. Our findings suggest that without broader development and adoption of standardized practices, inconsistent data collection and nomenclature will continue to hinder data sharing efforts in the food industry and compromise the reliability of any collaborative analyses. Implementing standardized data practices across the food industry is therefore essential to ensure the reliability and interoperability of shared data. A recent review highlighted that the successful application of AI in food safety relies heavily on the availability of high-quality, standardized data40, emphasizing the need to address these foundational challenges to unlock its full potential in transforming food safety. These findings suggest that collective action and co-opetition frameworks, while useful for understanding motivational dynamics, may underplay the role of technical infrastructure as a barrier to data sharing, whereby data sharing capability may need to precede willingness. Efforts aside, several participants noted that the potential benefits of data sharing include improved public health outcomes due to access to more data points. They explained that this access could enhance predictive modeling capabilities, enabling early identification and mitigation of rare food safety incidents that could lead to foodborne outbreaks. Similarly, voluntary horizontal data sharing in the swine industry, led by academic institutions, enables predictive models for forecasting disease outbreaks and informing prevention strategies, highlighting the value of data sharing in protecting animal health41. Therefore, an implicit benefit of data sharing is the creation of larger, more comprehensive datasets that support the application of AI and machine learning to advance food safety goals, provided aggregated data are accurate, unbiased, and representative of the intended use cases42. This underscores the importance of understanding what the aggregated data represent before drawing inferences to realize the benefits of data sharing fully. While data sharing has many recognized benefits, participants highlighted that it has inherent risks that could expose companies to negative consequences if sensitive data were leaked or used for an unintended purpose. Specifically, participants were concerned about the possibility that leaked data could be misinterpreted by customers, leading to reputational harm. Bernal discusses how companies face reputational risks even when they share data for socially beneficial purposes, a practice referred to as data philanthropy43,44. Bernal argues that companies may still suffer unintended consequences if the public misinterprets the shared data. Although Bernal’s paper focuses on business-to-government (B2G) data sharing, similar concerns apply in the horizontal industry-to-industry data sharing context. Another concern expressed by participants was the potential for retrospective or prospective liability. For example, participants talked about the risks associated with whole genome sequencing (WGS) data, which could implicate companies in past events or outbreaks yet to be detected, exposing them to legal liability and increased regulatory scrutiny. Although empirical research on this remains limited, a trade journal article by Klijin et al. suggests the food industry is hesitant to adopt WGS due to legal and regulatory uncertainties, particularly concerns about regulatory data sharing obligations45. These concerns indicate that without clear legal or regulatory assurances, companies might prioritize complete avoidance of risks associated with data sharing. Participants were concerned about losing their competitive advantage if they engaged in data sharing. This concern exemplifies the core tension in co-opetition theory, which is that firms must simultaneously cooperate in certain spheres to generate collective insights/value while competing in others to protect their competitive advantage. Similarly, a survey and interview of fish farmers in Norway reveals hesitancy to share data out of fear that sharing would result in a loss of competitive advantage, especially in the absence of reciprocity46. Overall, our findings highlight a broader tension between data sharing to improve public health and business survival concerns. The efforts, investments, benefits, and risks presented above suggest a misalignment between public health priorities, individual company interests, and broader industry needs. While the benefits of improved data sharing could be shared by all stakeholders, including consumers, companies, and the food industry as a whole, the associated risks and investments are often expected to be borne primarily by individual companies. As one participant said, “… I’m feeding people, that’s public health, I’m doing lots of things that are public health. The downside of this public health contribution is I could get sued for billions of dollars and put myself out of business…” This statement underscores the extent to which participants perceive data sharing as a potentially existential risk, capturing what they mean when they speak of economic harms. An interview study of experts from industries, such as agriculture and automotive, reports that organizations tend to focus more on the risks rather than the potential benefits of data sharing22. The authors highlighted that this attitude is due to uncertainties about the added value, both monetary and non-monetary (e.g., customer rewards and strategic market position), from data sharing22. Using the privacy calculus theory (PCT) framework, a study with experts from European data sharing initiatives demonstrates how data sharing decisions are influenced by perceived benefits (e.g., access to more data) and barriers (e.g., lack of trust between actors), resulting in tensions between the perceived gains and risks23. Our findings suggest the PCT framework could provide an additional useful lens for understanding how food companies navigate the complex decisions of data sharing in a food safety context. Future research should aim to develop a clear and compelling value proposition for horizontal data sharing that demonstrates how the benefits can offset the perceived and actual risks at both the company level and across the broader industry. Lefebvre et al. emphasize the importance of creating a joint value proposition for companies involved in horizontal data sharing and provide examples of existing initiatives, such as Skywise in the aviation sector, whose >300 members come together to gain insights on plane component improvement and maintenance anticipation47. These efforts, when accompanied by trust-building initiatives, can facilitate food industry-wide voluntary data sharing to support better-informed decision-making. Trust plays a pivotal role in shaping voluntary food safety data-sharing practices within the food industry. Participant concerns about horizontal data sharing, rooted in a trust deficit, can be categorized into two groups: lack of trust among stakeholders (regulatory, customers, industry peers) and lack of trust in technology. While participants expressed that the government could play a role in facilitating data sharing, many participants (i) expressed concerns about how leaked data could be used against them by regulatory bodies and (ii) shared negative past experiences about how shared data was used against them. As one participant noted, “I’ve had a couple of experiences where when you offer up those voluntary information and next thing you know, sure enough, you get a phone call inquiring about it. It’s not a positive phone call when inquiring about what’s going on…” The participant further noted that this negative experience occurred despite assurances from the government that the information would be used to improve food safety. Such experiences highlight breaches of trust and discourage future engagement in voluntary data sharing initiatives. While our participants highlighted that food safety is not a competitive advantage, the unregulated data-sharing space provides room for potential unintended consequences if industry peers access data. One participant said, “…We could unintentionally try to sabotage each other. I know that’s an awful thing to say, but it’s very competitive…” highlighting fears that competitors might lose revenue or business as a side effect of data sharing. It is often stated that food safety should not be considered a competitive advantage because it is a non-negotiable condition, invisible to consumers, and yet expected. Also, food safety incidents often create shockwaves reverberating through the entire food industry sectors, meaning any one weak link weakens the whole industry. However, horizontal sharing of food safety data may result in a loss of business advantage relative to competitors10 because shared data may allow inferences about proprietary production processes or performance metrics48. This highlights the importance of carefully regulated accessibility49 of voluntarily shared food safety data48. Otherwise, even well-intentioned transparency could compromise competitive positioning by exposing sensitive operational details. Our participants also expressed concerns about the potential for the loss of customer trust. Specifically, participants were concerned that if food safety data were leaked or made public without proper context, customers might misinterpret the data, particularly given the technical nature of such data. While participants did not explicitly question customers’ ability to interpret such data, their concerns suggest an underlying fear that technical food safety data could be misinterpreted when viewed outside its intended context. Such misinterpretation could lead to significant economic consequences, particularly if the information were perceived as evidence of negligence or risk, even when that wasn’t the case. Using the protection motivation theory, which explains consumer behavioral changes in response to health risks such as food safety risk, Liao et al. show the impact of food recall events on consumer trust and company reputation, and the role that trust plays in food safety and how violations of this trust can lead to significant reputational damage and loss in sales50. Similarly, misinterpreted leaked data can lead to a loss in sales as customers may choose to take their business elsewhere due to perceived risks. Both our findings and those of Liao et al. highlight the need for robust data privacy and security measures to prevent unauthorized access and release of shared data. Trust in technology also emerged as a critical factor limiting data sharing. Participants questioned whether the existing data systems are secure, reliable, and capable of safeguarding their shared data at all costs. One participant said that they do not believe that the current technological lockdowns provide enough confidence, and the consequences of poor data security protection would become a reality. Most study participants were unaware of technological and AI advancements in data security, such as differential privacy, federated learning, or blockchain, tools that researchers have identified as critical for protecting privacy in shared data51. However, given the rapid development of AI and privacy-preserving technologies, we acknowledge that participants’ limited awareness of emerging technologies may have evolved since data collection. Overall, our findings highlight the need to increase awareness among the food industry stakeholders about emerging data sharing technologies and AI that ensure data security and privacy. Participants expressed that barriers to voluntary data sharing can be addressed through the development of a robust data governance framework. Participants mentioned that having some governance measures around data sharing could help reassure those who feel uneasy about what they share and how that information is used. Cue et al. similarly highlight the need for data governance to foster trust and propose the Farmers Bill of Rights to address data sharing concerns in the dairy industry52. The bill stipulates that farmers retain ownership of their data and mandates transparent communication of intended use by data users (industry)52. This aligns with the broader conversation on data governance, which the Data Governance Institute (DGI) defines as “…a system of decision rights and accountabilities for information-related processes, executed according to agreed-upon models which describe who can take what actions with what information, and when, under what circumstances, using what methods.”53 This definition underscores the importance of having clear and structured frameworks that ensure that all stakeholders understand the data sharing rules of engagement. Khatri and Brown outlined five interrelated data governance domains: data principles, quality, metadata, access, and lifecycle, each addressing key aspects of data management54. Our data reported here suggest that those domains would address a number of risks associated with data sharing identified by participants (Table 3). If data governance has the potential to address the risks associated with data sharing, the question is why such frameworks have not yet been implemented. We drew insights from a systematic literature review that identified several barriers to the development of effective data governance55. Barriers relevant to the horizontal industry-to-industry food safety data sharing initiatives include a lack of knowledge about how to implement a governance framework, organizational barriers such as a lack of priority compared to other projects, and a lack of financial resources55. To gain an understanding of how a data governance framework specific to industry-owned food safety data could be developed, we can analyze existing frameworks. One of them is the Health Insurance Portability and Accountability Act (HIPAA), which offers several parallels to sharing food safety data (such as environmental monitoring test results). Like health data, food safety data contains sensitive company safety status, has significant regulatory implications, and balances competitive interests with public health benefits. HIPAA’s governance approach ensures that health data is stored, processed, and shared in compliance with federal laws, and according to clearly defined data use agreements and limitations, standards for data de-identification, and authorization requirements that enable provider/patient control. An interview study highlights the importance of comprehensive policies, procedures, and compliance with HIPAA regulations in fostering trust56. Comprehensive rules of engagement, protections, and policies are similarly necessary for a food safety data governance framework. The U.S. federal government played a central role in the development of HIPAA, and similar government involvement is necessary to offset the cost of the food safety data governance framework. However, unlike HIPAA, the participants in our study emphasized the need for oversight by a neutral actor who does not stand to benefit or be harmed either way from shared food safety data, and they suggested that academic institutions play that role. Similarly, Sandhaus et al. found that academic institutions can act as data intermediaries in autonomous vehicle safety data sharing, with the responsibility of ensuring that data meet open-access and reproducibility standards48. Our interviews also revealed that participants are concerned about losing control over their data once shared. This suggests that maintaining control would be an important motivator for data sharing and increasing participation in data sharing initiatives. As one participant stated, “… But I can control how that information is released to plaintiff’s counsel. I can control what is released. But so having that control is important to us. It’s our data. We own it…” Based on our findings, we recommend piloting a data governance framework for voluntary horizontal food safety data sharing, which includes collaboration between industry, regulatory bodies and academic institutions. However, given the heterogeneous nature of the food industry, industry-wide governance approaches may face coordination challenges. A promising alternative is the use of a location or cohort-based arrangement in which firms with similar characteristics, such as geographic proximity or product risk profiles, share data within more defined groups. Consistent with Olson’s collective action theory, smaller and more homogenous groups are better able to sustain cooperation12. In parallel, future research should examine the conditions under which these governance arrangements are most effective, as well as their scalability and equity implications. The collective action and co-opetition dynamics reflected across our four themes suggest that the current state of limited data sharing is a product of misalignment of risks, benefits, and incentives. Our themes reveal three specific problems that together explain this state: a cost-benefit asymmetry, an opportunism risk, and a participation cost asymmetry. The perceptions of benefits in Theme 1 and risks in Themes 3 reveal a cost-benefit asymmetry as companies face immediate and individualized legal, reputational, and competitive risks, while the benefits remain collective, diffuse, and dependent on other’s participation. This dynamic is consistent with the social dilemma structure that Kollock identified as central to public goods problems57. Because food industry participants view the potential risks as existential, the risks will tend to outweigh any perceived benefits in the data sharing decision calculus. Therefore, future research should focus on mitigating the downside risk faced by individual companies. Such mitigation requires broadening the risk burden. One possible scenario is that the private costs of, for example, a lawsuit resulting from a harmful pathogen being discovered in shared data could be covered by the fund to which all participants contribute. An alternative would be that the government guarantees to cover legal costs or damages using public funds (i.e., taxpayer money) or creates a safe haven that provides protection from legal action. As with any insurance setting, the risk-sharing mechanism would need to address moral hazard, i.e., the fact that firms insured against the worst downside risks of data sharing would have less motivation to follow best practices. The potential joint liability could be reduced with appropriate legal and contractual terms. The risk of opportunism in data sharing contexts is most evident in Theme 3, as participants expressed concerns that shared data could be misused or misinterpreted. As one participant noted, “… we could unintentionally sabotage each other…” Such concerns align with prior research identifying opportunism as an inherent risk in collaboration among competitors, particularly as the scope and duration of cooperation increase58. Prior work further highlights knowledge leakage as a key mechanism through which opportunism arises in information-sharing collaborations, including unintended disclosure or strategic appropriation of shared knowledge59. The participants proposed data governance (Theme 4) as a mechanism to control the risk of opportunism. A participation cost asymmetry, reflected in Themes 1 and 2, highlights that the upfront investment required to participate in data sharing is unevenly distributed across firms. Participants noted that smaller companies tend to be less digitalized, resulting in higher entry costs, as they must first invest in digital capabilities before they can meaningfully contribute. In contrast, larger and more digitalized firms face comparatively low barriers to participation. This asymmetry produces a misalignment between a firm’s interests and resources, consistent with Oliver et al., who identify this as a key barrier to critical mass formation60. Smaller firms, with high interest but low resources, may be more likely to free ride on larger firms’ contributions, while large firms, despite greater resources, may have weaker incentives to participate. This dynamic extends Olson’s concept of the “exploitation of the great by the small”12 to the data sharing context. Importantly, Theme 4 (data governance) provides a solution for how firms can share data in co-opetition settings. By structuring access and usage rights, data governance shifts the insights gained from data sharing from a public good toward a more restricted club good, thereby correcting the misalignment in the underlying incentive structures that gave rise to collective action problems such as opportunism and free riding. While we took steps to minimize the potential introduction of biases, this study has some limitations. Our sample size (27 participants) was sufficient to reach code saturation; however, recruitment relied on professional networks, which may have introduced selection bias or resulted in faster saturation due to shared experiences or perspectives within this group, overall limiting the generalizability of our findings. Given the reluctance of agri-food professionals to discuss food safety and quality control issues, trust was essential for participant recruitment, necessitating the use of professional connections. Other safety-related data-sharing studies have faced the same constraint, as recruitment through professional networks has also been necessary in an interview-based study on autonomous vehicle safety data sharing48. While the participants were not directly asked which step in the value chain their company or their position represents, the authors interpret their responses as primarily reflecting the production and processing steps of the value chain, as well as supporting laboratory services. Considering that the sample included participants from a wide range of sectors in the food industry, we believe the findings are representative of the producers and processors in the target population. The semi-structured interview format allowed for in-depth exploration of participant responses, but this also meant that prompts were sometimes given in different orders, and time constraints prevented the exploration of all scripted questions. With interviews, there is a potential for social desirability bias, as participants may have provided responses they perceived as more acceptable or favorable, especially given their leadership roles within their company. A graduate student, rather than an established food safety authority, conducted all interviews in an attempt to minimize this potential source of bias. While we emphasized confidentiality, this bias may still have influenced the data. There is also the potential for confirmation bias in the data analysis process, which could result in biased interpretation of responses and an emphasis on evidence that confirms our preexisting beliefs over contradictory evidence. To mitigate this, a four-person team conducted the analysis, including the data familiarization, ensuring thorough discussion and validation of emerging themes against interview transcripts. Finally, because the interviews were conducted between June 2023 and January 2024, the findings reflect the participants’ knowledge and experiences, the legal landscape, and the regulatory and industry contexts during that specific timeframe. Our findings should therefore be interpreted with consideration of potential changes in practices or policies since data collection. While our descriptive analysis (Fig. 3) provides valuable insights into how participant characteristics influence data sharing decisions, it reflects thematic emphasis rather than statistical significance. Future research should use a mixed-methods design, with surveys to quantitatively validate the qualitative findings, and include broader sampling across company sizes, sectors, and regions to strengthen generalizability and enable comparative insights, such as analyses of differences between large and small companies. Despite these limitations, this exploratory study provides valuable novel insight into horizontal food safety data sharing in the food industry and highlights key considerations for future research and policy. To conclude, our main findings reveal an interesting paradox: our participants’ enthusiasm for collaborative safety improvements coexists with systemic barriers that prevent the realization of horizontal sharing of confidential food safety data. We show that this phenomenon is not simply a matter of unwillingness; it is consistent with theories of collective action and co-opetition and reflects a rational response to a governance gap in which private costs outweigh benefits. By identifying the specific barriers, benefits, and governance solutions that shape participation, this study provides an evidence-based roadmap for unlocking the potential of collaborative data-driven food safety decision-making and prioritizing future research. This work lays the foundation for horizontal data sharing initiatives that could reshape the industry’s approach to preventing foodborne illness through the collaborative analysis of aggregated food safety data. Methods Sampling and recruitment We recruited participants using non-probability sampling, specifically purposive sampling, to select individuals most likely to provide relevant insights for our research question14. We also employed snowball sampling, wherein interviewed participants suggested other potential participants (to consider) to expand the original sample25,61. We first obtained a list of potential participants from the professional networks of MW and RI. This list included the potential participant’s name, company, current position, and years in current position. MW and RI reviewed this list, and individuals were selected for invitation based on the availability of contact information (email addresses). During the interview phase, we used snowball sampling to identify additional participants, from among those suggested by interviewed participants, and ensured they met the inclusion criteria before inviting them to participate in the study. We employed code saturation analysis as a quality assurance mechanism to assess whether our sample size was sufficient to capture the relevant themes of the issue being explored62. Participant eligibility included individuals in corporate, business or managerial leadership positions (e.g., president, director or manager)29 within the U.S. food industry (focusing on food sectors: dairy, meat, produce, food manufacturing) or those representing companies supporting the food industry in a food safety capacity (e.g., laboratory services). Additionally, we only considered individuals who had at least 2 years’ experience working in the food industry, fluency in English, and who were over the age of 18. These criteria ensured participants possessed the necessary expertise and organizational perspective to provide meaningful insights. MW sent the initial and follow-up emails (two weeks apart), while LK was responsible for providing interested individuals with an informed consent document, which outlined study procedures, confidentiality measures, and potential risks and benefits to those who expressed interest in the study. All procedures involving human participants were conducted in accordance with the ethical standards of the Cornell University Institutional Review Board for Human Participants (Cornell IRB), and in adherence to the principles of the Declaration of Helsinki. This study (IRB Protocol IRB0147486) was granted an exemption determination by the Cornell IRB office on June 12, 2023, and amended on September 09, 2024. Informed consent was obtained from participants prior to each interview and documented through verbal consent, as approved by the IRB. To protect participant privacy and confidentiality, identifying information, including participant ID numbers, was removed from transcripts and quotes. As a result, quotes are presented without attribution to specific participants. Data collection We designed an interview guide that focused on two key areas: (i) how food safety data is collected and shared within the industry and (ii) instances where food safety data is collected but not shared, along with the reasons behind these decisions. The guide was piloted with three food safety professionals (two representing food industry companies and one with relevant experience in the industry and government sector) and conducted by LK and RI to ensure clarity and relevance before full implementation. Interviews were conducted between June 2023 and January 2024 via teleconference, allowing for flexibility while maintaining consistency in data collection. Before each interview, we obtained informed consent to participate in the study (consent form available in Supplementary Text S3). Each interview began with an initial inquiry into the participants’ employer, role, and tenure contextualization. Next, we asked participants, “What do you understand the term ‘industry-to-industry data sharing’ to mean?” This question was asked to ensure that both the interviewer and interviewee had a shared understanding of the type of (horizontal) data sharing examined in our study. The interviewer also provided examples to clarify the concept, such as asking about the potential benefits of data being shared among multiple companies, including competitors. After these introductory questions, the interview began. The interview guide, which outlines the full set of questions and prompts used in the study, is available in Text S1. All interviews were recorded and later transcribed for analysis. Data analysis Interview recordings were transcribed using an automated transcription service provided by NVivo Transcription (Lumivero, Denver, CO). Following this, each transcript underwent a manual validation process by LK to ensure that transcripts matched the recordings. This involved listening to the audio recordings while reading the transcripts, allowing for the correction of any discrepancies between the spoken word and the transcribed text. We used an inductive, codebook thematic analysis approach, consistent with Braun and Clarke’s typology of thematic analysis approaches63. This approach provided a structured, data-driven coding process through the development and iterative refinement of a coding frame. We conducted coding in NVivo 15, which facilitated this process by allowing us to add and edit code definitions both at and after the point of creation and thereafter, supporting consistent application of codes across the dataset. We used Braun and Clarke’s six phase process as a guide to conduct our iterative thematic analysis. These included: (i) familiarization with data, (ii) generating initial codes, (iii) generating themes, (iv) reviewing potential themes, (v) defining and naming themes, and (vi) producing the report64. During each phase, L.K., C.A., K.K., and R.I., individually familiarized themselves with the data (transcripts), and collaboratively reflected on and discussed emerging codes and themes, revisiting the data and themes throughout all six phases of thematic analysis. During phase ii, we used a combination of inductive and semantic coding approaches. Inductive coding allowed codes to emerge directly from the data without relying on a pre-existing framework, while semantic coding focused on the explicit content of participants’ responses rather than interpreting underlying meanings64. While our sample was purposively selected, we employed code saturation analysis to determine the sufficiency of our sample size65. To operationalize this model, we used a binary-coded matrix of our codes from NVivo and applied a three-step approach for enhanced transparency: (i) plotting a sequential saturation curve by tracking the cumulative number of new codes after each interview to assess when saturation occurred, following a method described by Hennink et al.;31 (ii) comparing the actual interview sequence with a randomly ordered set to assess whether saturation was influenced by interview order, also following Hennink et al.’s approach. This comparison helped determine whether the observed saturation curve in step 1 was an artifact of interview order or reflected genuine convergence, (iii) performing bootstrap analysis with 10,000 iterations using random sampling with replacement. For each iteration, we recorded the number of interviews required to reach 80% and 90% saturation66,67. We then calculated median saturation points and the 5th and 95th percentiles to characterize the variability in the number of interviews needed to reach each threshold. We used the entire dataset as the base size, with the total number of unique codes across all interviews acting as the denominator in saturation calculations62. This approach assumes the full dataset represents the universe of possible codes and each interview contributes equally to code discovery62. We defined a run length of one interview, meaning each interview was assessed individually for new codes. We defined saturation using a 0% new information threshold, consistent with existing studies62,68. Cumulative saturation was calculated using Eq. 166:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{k\,}=\frac{{C}_{k}}{{C}_{N}}$$\end{document}Sk=CkCNwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{k}$$\end{document}Ck is the cumulative number of unique codes identified up to the kth interview, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{N}$$\end{document}CN is the total number of unique codes in the dataset66, meaning saturation occurred when subsequent interviews introduced no new codes (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{k}=100 \% )$$\end{document}Sk=100%). All analyses and visualizations were conducted using Jupyter notebook via Google Colaboratory (Colab). We repeated this three-step saturation procedure for the business-level and managerial-level subgroups to assess whether patterns differed from the overall dataset. We did not conduct subgroup saturation analysis for corporate-level leaders due to the small number of interviews representing that subgroup. Together, these procedures allowed us to assess whether saturation reflected convergence across a diverse set of participants rather than redundancy driven by interview order or sampling variability. We assessed meaning saturation following Hennink et al. to determine whether we had captured the full range of nuances within our codes31. First, we exported a binary codes matrix from NVivo and classified our codes along two dimensions: (i) conceptual codes versus concrete codes (explicit, definitive issues), and (ii) high-prevalence versus low-prevalence codes. We determined prevalence by calculating how frequently parent codes appeared across interviews, then classified parent codes as high- or low-prevalence based on whether they fell above or below the mean. We then selected parent codes central to our research question for meaning saturation analysis. Lastly, we developed a Python script in Google Colab for tracking the emergence of various code dimensions across our interviews. We further conducted a descriptive analysis to explore variation in thematic emphasis across participant attributes using NVivo’s matrix coding query function. In NVivo, we first categorized all interview transcripts as cases. Using the case classification feature in NVivo, we created three participant attributes: industry sector, years of experience, and position within the company. Each attribute included predefined values (levels): (i) industry sector: dairy, produce, food manufacturing, meat, and laboratory services; (ii) years of experience: 2–5, 6–10, 11–15, and ≥16 years; and (iii) position: CEO, VP of food safety, Director of food safety, and Managers. Each case was assigned corresponding attribute values. We then generated a matrix displaying the number of coded references for each aggregated theme across the attribute values. Finally, we used Eqs. 2 and 3, to adjust for differences in the number of participants per attribute value, allowing fair comparison across groups. For each attribute value, we calculated the average number of code references as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A}_{i}=\frac{C{R}_{i}}{{n}_{i}}$$\end{document}Ai=CRiniwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{CR}}_{i}$$\end{document}CRi is the total number of code references in attribute value ii and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n}_{i}$$\end{document}ni is the total number of participants in that attribute value. We then normalized these averages to obtain proportional values by dividing the average of each attribute value by the sum of the averages for all attribute values per attribute:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{i}=\frac{{A}_{i}}{{\sum }_{i=1}^{k}{A}_{i}}$$\end{document}Pi=Ai∑i=1kAiwhere kk is the total number of attribute values per attribute. This approach allowed for a balanced comparison of thematic emphasis within attributes, minimizing bias from unequal numbers of participants across attribute values. The code references indicate thematic representation and do not imply statistical significance. Variations in calculated proportions may reflect differences in participants’ expressiveness, coding granularity, or researcher interpretation69. We also assessed whether themes were supported across the strategic leadership subgroups. We exported a binary code-presence matrix from NVivo, indicating each participant’s contribution to our identified themes, grouped participants by their corresponding leadership position (corporate, business, managerial), and summed theme presence within each subgroup (Table S2). One of the preliminary themes that emerged was on barriers to data sharing. We characterized the identified barriers into three groups, expanding on the categories defined by Panhuis et al.: technical (issues relating to data formats, data interoperability, and standardization), economic (possible economic impact as a result of data sharing), and social (relating to the perceptions, attitudes, and trust of stakeholders towards data sharing)70. While Panhuis et al. originally characterized barriers in the context of public health, we are repurposing them to analyze the challenges related to food safety data sharing in the food industry. When talking about barriers to data sharing, several participants commented on the need for data governance. We mapped the barriers raised by participants against the data governance domains54 that would address them, adapting a previously introduced framework71. This study followed the consolidated criteria for reporting qualitative research 32-item checklist (Supplementary Table S1). Supplementary information Supplementary Material Supplementary Material