Abstract
People with disabilities 1 in Asia are increasingly accessing new employment opportunities through artificial intelligence (AI)-driven platforms, such as food delivery and data annotation, which are often portrayed as pathways to economic empowerment. However, this article interrogates two key questions: what are the stakeholders within the platformised data annotation ecosystem, and how do their interactions shape the work–life experiences of individuals with disabilities? Drawing on a 3-year empirical study initiated in July 2019 – including 107 days of participant observation at seven data annotation bases across China and 155 semi-structured interviews with diverse stakeholders – this article applies Van Dijck’s ‘platformisation tree’ framework to analyse a multifaceted ecosystem. In this metaphorical tree, key actors including central government, local authorities, tech capital and Disabled People's Organisations (DPOs 2 ), assume roles akin to roots, trunk and branches. Meanwhile, annotators with disabilities are depicted as the leaves, positioned at the periphery where they experience both innovative forms of inclusion and heightened risks of marginalization. This analysis reveals the complex dynamics of coordination, negotiation and conflict among these actors as they collectively build and regulate the data annotation workforce. Ultimately, this study contributes to our understanding of labour precarity, platformisation and disability inclusion in AI-driven economies.
Keywords
This article is a part of special theme on Infrastructure, Labour, & Social Change. To see a full list of all articles in this special theme, please click here: https://journals.sagepub.com/page/bds/collections/infrastructure_labour_social_change?pbEditor=true
Introduction
The 21st century has witnessed a remarkable expansion of artificial intelligence (AI) industries, prompting extensive scholarly inquiry into the human labour underpinning AI-enabling tasks such as data annotation (Gray and Suri, 2019; Ross et al., 2010; Schmidt, 2022). Data annotation – also referred to as data labelling – entails the systematic tagging of raw data (e.g. photos, texts, videos and audio) so that machine learning algorithms can interpret real-world contexts and generate accurate predictions. This painstaking process underlies key technological applications in computer vision, natural language processing, audio-based speech recognition and autonomous vehicles.
Within the burgeoning global data annotation market, platformised labour processes predominate, encompassing work allocation, wage payment and labour-process monitoring (Gray and Suri, 2019; Le Ludec and Casilli, 2023; Posada, 2024; Wu, 2023; Wu and Xia, 2023; Xia and Wu, 2025). In China, tech giants such as Baidu, Alibaba and Tencent – collectively known as the BAT trio – have aggressively pursued AI initiatives, particularly in the wake of ChatGPT's successes. Rather than examining platforms as standalone entities, this study focusses on the tech capital (Reft, 2020) operating behind these platforms.
As data annotation work proliferates, economically under-resourced regions like Henan, Shanxi and Guizhou have emerged as hubs for outsourcing (Kshetri, 2021). This expansion also draws attention to the participation of marginalised groups – such as working mothers (Wu and Xia, 2023) and people with disabilities (Wu, 2023) – in the data annotation industry. Despite facing substantial barriers to traditional employment (e.g. stigmatization, inadequate accommodations, limited educational opportunities) (Kohrman, 2015), individuals with disabilities are increasingly finding roles as ‘AI trainers’ in data annotation (CCTV News, 2024). As previous research states (Gray and Suri, 2019; Posada, 2024; Tubaro et al., 2020), these remote and flexible work arrangements offer potentially inclusive employment options, the platformised labour model also poses risks of exploitation and marginalization.
The unique characteristics of the Chinese context further distinguish these dynamics. As Lieberthal (2003) demonstrates, the central government sets overarching policies that are variably implemented by local governments, whose distinct incentives and strategies drive regional development. In line with this, Chen et al. (2024) illustrate how divergent local approaches to policy implementation shape broader governance models, while Bo (2020) provides evidence that political hierarchy reforms contribute to regional specialization. Together, these studies underscore that the interplay between central directives and localised governmental incentives is a critical factor influencing both the economic landscape and labour conditions within China's platformised data annotation ecosystem.
The distinctive characteristics of China's governance structure further shape the dynamics of its data annotation industry. As Lieberthal (2003) demonstrates, while the central government issues broad policy directives, their implementation is filtered through local governments whose varying incentives and strategies drive region-specific development. Building on this, Chen et al. (2024) show how diverse local policy interpretations generate divergent governance outcomes, and Bo (2020) provides evidence that political hierarchy reforms have contributed to regional specialization. These insights help differentiate China's data annotation industry from other national contexts studied in existing platform labour research. Together, these studies highlight the importance of central-local interactions in shaping both economic development and labour conditions – an essential but underexplored dimension in China's platformised data annotation ecosystem.
This article analyzes that ecosystem based on 3 years of fieldwork and draws on Van Dijck's (2021) concept of the ‘platformisation tree’ – a multi-layered framework that captures the interplay between state actors, market forces and social organisations. We conceptualise the central government, local authorities, tech capital and Disabled Persons’ Organisations (DPOs) as key actors co-producing and regulating the conditions of platform labour. While previous studies have used the platformisation tree metaphor to map China's global platform strategies (Zheng, 2025) or to examine how platform logics reshape urban governance through historical and sociotechnical mediations (Renzi and Frenzel, 2025), this article shows that the metaphor is particularly effective in illuminating the platformised labour model in AI data annotation (Le Ludec and Casilli, 2023; Posada, 2024). It helps uncover how state policy, corporate capital and advocacy networks intersect in ways that both enable and constrain labour.
In doing so, this article contributes to critical discussions on platformised labour by centreing the experiences of workers with impairments – a group often marginalised in labour and technology debates – and situates their experiences within a broader political economy analysis. While existing literature has tended to focus on data work at the level of firms, platforms and supply chains (Chandhiramowuli et al., 2024; Le Ludec and Casilli, 2023; Miceli et al., 2020; Posada, 2022; Wang et al., 2022), we extend this scholarship by examining the multi-scalar dynamics among central and local governments, tech capital and civil society actors. This political economy perspective reveals the complex and sometimes contradictory forces that structure labour in China's data annotation industry and highlights the importance of contextual political structures in shaping platformised work.
We begin by detailing our theoretical framework of the platformisation tree and its relevance to the data annotation ecosystem, followed by a discussion of the working lives of people with disabilities within this framework. Our methodological approach is then outlined and our empirical findings are presented by conceptualizing the central government as the root, the negotiation between local authorities and tech capital as the trunk, DPOs as the branches, and data workers’ negotiation and resistance as the leaves. In the final section, we summarise our contributions to both data work research and platformisation studies.
Platformisation and the data annotation ecosystem
Platformisation has emerged as a critical concept in understanding the integration of digital platforms into societal and industrial domains. Poell et al. (2019) define platformisation as the penetration of platforms into infrastructures, economic processes and governmental frameworks across diverse sectors of society. Previous discussion on platformisation often conceptualises the layered architecture of platforms as a collection of ‘stacks’, emphasizing their modular and accumulative features (Bratton, 2016; Schwarz, 2017; Tiwanna, 2014; Walton, 2017). However, platforms are not equal, nor are they ‘stacked’ arbitrarily; rather, platform ecosystems are hierarchically structured and interdependent, granting some platforms disproportionate power. As Donovan (2019) suggests, the ‘stack’ metaphor may no longer sufficiently capture the intricate and evolving dynamics of the system.
Based on a combined Science and Technology Studies (STS) and political economy approach, Van Dijck (2021) adopts the metaphor of a tree to illustrate the structure of the platformisation ecosystem. In this metaphor, the roots represent the foundational digital infrastructure – cables, satellites and data centres – that underpin the ecosystem. The trunk symbolises intermediary platforms such as social networks, search engines and payment systems that mediate between infrastructure and users. The branches extend into societal and industrial sectors, shaping specific domains such as education, healthcare and commerce. This metaphor emphasises the layered and hierarchical nature of platform ecosystems, where entities consolidating control over the trunk layer hold disproportionate power over the system. Van Dijck's assertion that the platformisation tree may assume different forms in varying contexts further motivates this study's examination of China's data annotation work.
The platformised nature of data annotation work and its impact on workers’ conditions has been widely examined (Chandhiramowuli et al., 2024; Gray and Suri, 2019; Posada, 2024; Tubaro et al., 2020). For example, Posada (2024) investigates data workers in Venezuela, highlighting how the platformisation of wages diminishes worker autonomy through power dynamics favoring platforms that set transaction rules and prices. His findings offer a multifaceted perspective on platform labour, illustrating that the financial rewards of intermediation often come at a significant cost to workers in precarious situations. Similarly, Chandhiramowuli et al. (2024) develop the concept of a ‘regime of counting’ to examine how control and authority are exercised by AI requesters over annotation workers, revealing the unequal employment relations and power asymmetries embedded in data annotation work.
Following Van Dijck's argument, we position the central government as the ‘root’ that both enables and regulates data annotation work in China. However, we refine this dual role by highlighting how negotiations between local authorities and tech capital constitute the ‘trunk’ of the data annotation tree – encompassing both regulatory initiatives and restrictions. The forces emerging from the root and transmitted through the trunk can either nurture the flourishing of the ‘branches’ (in this case, DPOs) or precipitate their decline. Within this framework, workers with disabilities are conceptualised as the ‘leaves’ of the data annotation tree. Their negotiations and interactions during the work process are nourished by the regulatory and economic forces emanating from the roots, trunk and branches, illustrating how their labour conditions are deeply shaped by the broader ecosystem.
Our study bridges the platformisation tree research with existing studies on data annotation labour. By adopting Van Dijck's tree metaphor, we reveal the driving forces behind the data production chain. Moreover, as Van Dijck emphasises, the trunk – embodied by platform companies – expands continuously through vertical integration, the infrastructuralization of intermediary platforms, and cross-sector expansion, thereby consolidating control over the entire ecosystem. Likewise, our findings show that the dynamic negotiations between local authorities and tech capital, representing the trunk in our study, evolve and reinforce control over the data annotation ecosystem, ultimately strengthening governance over the AI industry. This analytical lens clarifies how the interweaving of digital and social infrastructures shapes the labour conditions of marginalised communities, and exposes the complex interactions among the central government, local authorities, tech capital and DPOs within the platformised data annotation ecosystem.
People with disabilities and work–life experiences
There is a growing body of literature on the quality of working life among people with disabilities (Hall and Wilton, 2010; Hao and Xiao, 2022; Jones, 2016; Randle and Hardy, 2017; Williams et al., 2017). The central argument is that workers with impairments, much like other marginalised groups such as women, immigrants and non-white working classes, face socially structured discrimination in the workplace. For instance, Randle and Hardy (2017) document how people with impairments are excluded from the UK film and television labour market, paralleling the systemic barriers experienced by other marginalised groups due to factors such as age, gender, ethnicity, or migration, as embedded in the production model. Some studies, however, also highlight the ways in which workers with impairments resist discrimination and forge distinct work identities (Hall and Wilton, 2010; Jammaers and Williams, 2021). Jammaers and Williams (2021), for example, demonstrate that these workers develop adaptive strategies – such as self-medicating and energy conservation – to navigate the intersections of work and life.
Recent research has further interrogated the complex dynamics between people with disabilities and technology within platform work contexts (Hong, 2024; Wu, 2023). Wu (2023) engages with concepts such as ‘crip technoscience’ (Hamraie and Fritsch, 2019) and ‘disability expertise’ (Hartblay, 2020) to explore the practices of workers with disabilities in AI data annotation. She critiques the ‘epistemic violence’ (Ymous et al., 2020) present in technoscientific imaginaries, where individuals with impairments are often reduced to marginal cases or mere objects in thought experiments (Shew, 2020). Rather than simply being marginalised, however, Wu argues that workers with disabilities in data annotation can harness their expertise to renegotiate the terms of techno-capitalism. Complementing this perspective, Kim (2017) examines the dual nature of cure – while it negates disability as a valid lived experience, it simultaneously legitimises practices that inflict harm by framing pain as necessary for achieving an idealised, abled norm. Expanding on these ideas, Hong (2024) introduces the notion of ‘curative platforms’, which, though they promote inclusion and rehabilitation, can also reinforce exclusion and normalization. These narratives suggest that while platform work can potentially empower marginalised groups, it may also constrain their transformative potential.
Building on these lines of inquiry, our study investigates how workers with impairments engage in AI development through data annotation, specifically examining how they negotiate and adapt to the platformised annotation ecosystem. By drawing on Van Dijck's ‘platformisation tree’ metaphor, our analysis links macro-level structural dynamics to the lived realities of data workers – conceptualised as the ‘leaves’ of the tree. In this ecosystem, the central government (the ‘root’) provides overarching regulatory frameworks, while local authorities and tech capital (the ‘trunk’) negotiate and implement these policies. The ‘branches’ of the tree, exemplified by DPOs, further mediate these interactions. Within this framework, we explore two interrelated dimensions: (a) how data workers with disabilities negotiate and adapt to the platformed annotation work, and (b) the impacts of this hierarchical structure on their everyday labour experiences. Based on our empirical data, we focusses on three focal points: (a) the ways in which this macro-level structural dynamic frames workers with disabilities as ideal annotators; (b) the tensions and negotiations arising from crip time versus industrial time; and (c) the challenges and negotiations posed by machine-driven, computerised annotation rules. In linking these focal points back to the platformisation tree, we aim to demonstrate how the interplay of central government policies, localised negotiations and market-driven forces shape the conditions under which data workers operate, thereby offering a comprehensive view of the opportunities and constraints within China's data annotation ecosystem.
Research methods
Our empirical data collection spanned a 3-year fieldwork period starting in July 2019, covering seven data annotation bases across China, including Shanghai and five other provinces representing diverse geographical regions. The disability cases only cover a 2-year period from January 2020 to July 2021 (see Table 1). Access to the field sites was facilitated through our connection with Kai, a data annotation worker who introduced us to Boss Zhang, the leader of a DPO and our key informant. After we explained the purpose and objectives of our study to Boss Zhang, he was supportive of our research, seeing it as an opportunity to enhance his DPO's efforts in empowering people with disabilities. With his assistance, we conducted participant observation at Bases D, E and F, which were operated by his organisation.
Fieldwork details.
While Boss Zhang played a central role in facilitating access, his involvement did not limit or shape the scope of our interactions with annotators. We were able to speak with nearly all workers at these bases – specifically, all but two individuals who were either absent or declined to participate. Boss Zhang maintained an open and non-interventionist stance throughout our fieldwork, consistent with his advocacy-driven role in promoting the rights of people with disabilities. Nevertheless, we remained attentive to possible power dynamics and made efforts to ensure that participants felt free to express themselves candidly, independent of organisational leadership.
Access to Base C was obtained via an administrative staff member from the China Foundation for Disabled Persons (CFDP). He welcomed our research as a means to investigate how data annotation projects can be organised for people with disabilities across different provinces.
We adhered to principles of informed consent, confidentiality, and voluntary participation. Participants were provided with detailed information about the study's aims and methods, and they were assured that their identities would remain confidential. To acknowledge their time and contributions, we compensated interviewees with 200 RMB per interview, ensuring that this did not exert undue influence on their willingness to participate.
During our presence at these bases totalling 107 days, we maintained an observation journal exceeding 200,000 words. Additionally, we conducted 155 semi-structured interviews with various stakeholders, including project managers, algorithm engineers, administrative staff from local government big data departments, DPO leaders, managers, workers and their families affiliated with data annotation companies. These interviews, averaging 1 to 2 hours each, included 16 conducted via video calls due to one of us being outside of China at the time of writing. Additionally, we organised ten focus groups as part of our fieldwork. We list the details below in Table 1.
Open coding was our initial step in the analysis, during which we read the fieldnotes and interview transcripts line by line to identify and formulate any and all ideas, themes, or issues they suggested, without relying on preconceived categories (Emerson, Fretz and Shaw, 2011; Rubin and Rubin, 2012). This process allowed us to remain open to diverse and sometimes unexpected aspects of the data, which we then further refined in subsequent phases of coding and analysis. Madison (2012) claims that ‘coding and logging data is the process of grouping together themes and categories that you have accumulated in the field’ (p. 43). We therefore assign descriptive codes, such as professional workplace, ghost employment, hukou, 3 etc., to all relevant data segments and group similar codes together. Madison (2012) notes that grouping themes create points of views. Throughout constant comparison and iterative refinement, we identified emerging themes and developed a coding scheme to organise the data. We also maintained memos to document insights and reflections throughout the analysis process.
As a result, we classified the bases into three groups based on operational characteristics: Bases A and B are operated by local government-led organisations; Bases C, D, E and F are managed by DPOs; and Bases A and G are overseen by college-led organisations. For this study, we focus on Bases C, D, E and F. Our dataset comprises 80 in-depth interviews, 4 focus groups and an observation journal (totalling 114,745 words). Although our focus is on these selected cases, data from other bases were used as complementary information.
Base C, established in October 2019 by Wellbeing – a tech company owned by the CFDP – and operated by a couple (with Sun, the male partner, having a leg disability). Located in a local industrial incubation park, it occupies half of the park's 2000-square-metre area with 3 years of waived rent. Sun financed the renovation costs, which were promised to be reimbursed by the local DPF 4 later. The base employs 23 call centre workers (9 disabled) and 28 data annotators (26 disabled), whose primary responsibilities include tagging pictures and text on a piecework basis (earning between a few hundred and 2000 RMB per month).
Bases D, E and F are overseen by our key informant, Boss Zhang, who leads a team of over 40 workers with disabilities across these sites. Base D, directly managed by Boss Zhang's DPO, employs blind workers for semantic annotation tasks. Base E, jointly operated by Boss Zhang's DPO and a local DPO named Cow, employs physically disabled workers for audio transcription and text annotation. Base F, in collaboration with Boss Zhang's DPO and another local DPO named Swallow, employs blind and physically disabled workers for semantic annotation and Cantonese audio transcription. Workers at these bases receive a stable monthly salary ranging from 3000 to 3600 RMB. Detailed interview and focus group information is provided in the appendices.
Findings
The root: Central government as both enabler and regulator
In Van Dijck's (2021) conceptualization of platformisation tree, she positions state as the roots that shape and support the growth of platforms, by its twofold roles of enabler and regulator. In details, as the enabler, the state creates legal systems and political decisions within which platforms operate. As the regulator, the state addresses issues regarding to public interests, including users’ privacy, monopoly power and labour rights, in order to regulate actions of platforms. In other words, the state is not only foundational to the growth of the platform economy, but also a key factor in governing its impacts. Hong (2024: 2598–2599) agrees with Van Dijck's emphasis on the position of state in the platformisation process, as she takes Grab's service in Singapore as a case to explain how Singapore state bound up ensuring regulation, public/private interest and stakeholder interests. For example, she shows Grab was financially supported by a subsidiary of Singapore's state-owned Venture. Furthermore, Grab also collaborate with Singapore government's policies to support the public interests, by setting up its mission as creating economic empowerment for every local resident.
Likewise, in Chinese context, state also plays roles of both enabler and regulator. For example, during our business trip to Chengdu, Boss Zhang explained that during consultations with major disability organisations for the 14th Five-Year Plan, a proposal was included to establish a disability entrepreneurship incubation base in each province, although it was ultimately rejected. The policy aimed to leverage central government approval to enable local governments to allocate funds – typically for flexible expenses like rent and renovations – and to collect disability insurance contributions from businesses, thereby facilitating significant resource allocation for establishing these enterprises. This demonstrates that state decisions, such as policy formulation and financial support, provide the infrastructure needed for platform initiatives to thrive. By offering policy backing, the state creates a framework within which local governments can invest in and manage platform-related projects.
Furthermore, the Chinese state also plays the role of regulator. In mid-2021, the Chinese government launched the ‘Common Prosperity Policy’ to address growing inequality, aiming to redistribute wealth to disadvantaged groups through measures such as tighter regulation of technology companies. As part of their response to the new anti-monopoly regulations, major tech companies increasingly engaged in charitable and public welfare initiatives to align with government objectives. For instance, Fox, one of the leading tech firms, outsourced its in-house data annotation services to Boss Zhang's organisation, framing the project as a way to empower individuals with disabilities and promote social inclusion, thereby demonstrating compliance with the policy. According to Boss Zhang, four different departments within Fox contacted him to initiate projects involving workers with disabilities during that period (Fieldnotes, Chengdu Business Trip). This demonstrates how the state's regulatory interventions shape tech companies’ behaviours, ensuring they operate within parameters that prioritise public interests, including social inclusion of people with disabilities.
Both Van Dijck (2021) and Hong (2024) emphasise the central government's dual roles in platformisation but overlook the pivotal mediating role of local governments. On one hand, local authorities serve as direct implementers of central government policies; on the other hand, through a top-down effect, local government resources track central policies by supplementing and enriching the support available to tech capital within the central policy framework (Chen et al., 2024; Lieberthal, 2003). Base E, a three-story building, houses an exhibition space on the first floor adorned with photographs of Wusi's engagements with prominent figures, including officials from various levels of the DPF and a notable image of her shaking hands with President Xi Jinping in 2018.
Governance of Base E involves fifteen board members, predominantly appointed by the local DPF. Although these members rarely attend meetings, they ensure the coverage of annual utility and network expenses and allocate a budget of 500,000 RMB. The building's renovation, costing 3 million RMB, was funded by the Ministry of Civil Affairs. These financial and infrastructural supports commenced in 2018 following Wusi's meeting with President Xi as a representative of Shenzhen's disabled community. According to Boss Zhang, this encounter led to an influx of resources seeking collabouration with her (Fieldnotes, Chengdu Business Trip). Consequently, Wusi was appointed as a standing committee member of the Chinese People's Political Consultative Conference (CPPCC) and received prestigious honors such as the National May Fourth Youth Medal. These titles attracted numerous projects from organisations including the local DPF, Women's Federation and trade unions, resulting in sustained financial and organisational support for Base E. ‘Money kept coming in, and they couldn’t not do it…As she's someone who's already shaken hands with President Xi’ (Fieldnotes, Chengdu Business Trip, Boss Zhang).
By following the central government's symbolic endorsement – exemplified by the handshake photo with President Xi – local authorities enable the establishment and operational success of regional bases, such as Base E. They do so by providing supplementary policies and resources from various local departments. However, this process is not straightforward; local governments do not simply implement central government policies and decisions regarding tech capital. Instead, as a part of the trunk, they negotiate with technology capital entities, sometimes regulating or even restricting them. The following example illustrates the complex dynamics between local governments and tech capital.
The growing trunk: Negotiations between local governments and tech capital
While the roots provide a tree's foundational support, the branch – signifying negotiations between tech giants and local governments – facilitates the expansion of data annotation work. Local governments frequently interpret and implement central policies in distinct ways, resulting in a complex and intertwined relationship with technology companies. For example, Tiger, an internet giant joining the data annotation as a significant player, chose carefully its in-house data annotation space by measuring various benefits provided by different local governments: “We prefer a low-cost space in a high-tech park… Local governments want to have a large amount of tax revenue and the new industry can employ local people. We prefer a long-term cooperation, in which the local government can provide free spaces, subsidies, and benefits for our talents, such as the education of their children… The first-tier city is too expensive for an in-house annotation space. Data annotation is not a high-skilled work that needs sophisticated talents, but it still has to gather a group of people in a physical space. Put bluntly, we can't even recruit enough workers in a third- or fourth-tier city. So, we prefer second-tier cities. We must still compare different benefits from local governments in the second-tier cities…” (Focus group: FF20210102, a team leader)
After comparing the offers from different local governments, Tiger set up its in-house data annotation service in a second-tier city in southwestern China. The local government promised to help Tiger develop its digital games, the most profitable product of Tiger, once the local government was authorised to examine and publish digital games. The deal was decided not only by the department requiring the data annotation service; rather, it was a ‘hands shaking’ (Focus group: FF20210102) deal, negotiated among different departments, including the PR and the government-relations departments.
This example highlights the intricate and multi-dimensional nature of negotiations between local governments and tech giants, where economic incentives and strategic partnerships play a central role. Such interactions reveal the broader dynamics of platformisation at the regional level, where local governments leverage their policies and resources to attract platform-based industries, while tech companies strategically evaluate and select locations that maximise their operational and financial advantages. These negotiations are not merely transactional but involve a complex interplay of interests, including long-term commitments, workforce availability, infrastructure support and additional benefits, such as regulatory assistance for other profitable ventures.
As previously mentioned, Base C was established in October 2019 by Wellbeing, relying on financial support pledged by the local DPF. However, tensions soon emerged between Wellbeing and the local DPF. One prominent incident involved the Wellbeing's plaque-unveiling ceremony, initially scheduled for early January 2020, and later postponed to February due to a provincial leader's commitment at a national conference. Crucially, the revised date was communicated directly to Sun, bypassing the local DPF and causing dissatisfaction. In response, the local DPF, disregarding higher-level directives, contacted Sun via WeChat to handle the event independently. Yet, because the CFDP and a provincial leader were expected to attend, the ceremony required a vice-mayoral reception. Additionally, when signing its contract with Wellbeing, the local DPF unilaterally modified the original template. This combination of reduced-level reception and contract alterations fuelled ongoing friction between Wellbeing and the local DPF, ultimately impeding Base C's subsequent operations. (Fieldnotes, Base C).
This underscores that local governments, though frequently instrumental in advancing tech initiatives, can also exercise discretionary authority in ways that impede such ventures. Conflicts over scheduling, contractual terms and ceremonial protocol, as seen in the Base C example, highlight how local governments can withhold or reconfigure crucial resources – be they financial, institutional, or symbolic. Such tensions reveal that the relationship between local governments and tech capital is neither uniformly cooperative nor strictly directive but is instead shaped by ongoing negotiation, mutual dependency, and, at times, resistance.
These tensions reveal that the success or failure of data annotation initiatives is not solely determined by local-level implementation or central directives. Rather, a delicate interplay of political authority, policy priorities and technological investments shapes the outcomes for the organisations engaged in data annotation work – here referred to as ‘branches’, often DPOs supplementing platformised data annotation ecosystems. These branches must navigate the forces stemming from the ‘root’ (central government) and passing through the ‘trunk’ (negotiation between local authorities and tech capital). In the next section, we will explore how these branches either flourish or fail within this multi-layered ecosystem, highlighting the diverse trajectories that emerge from the negotiation between policy frameworks, resource allocation and grassroots practices.
The branches: Some thrive, others are abandoned
When the flow of energy from the tree's roots to its branches proceeds smoothly, the branches thrive. As previously discussed, the platformised data annotation work has given rise to various organisations that are closely integrated with platform ecosystems, including the DPOs in this article (Wu and Xia, 2023; Xia and Wu, 2025). Base F offers a positive illustration of this process. By aligning with the central government's supportive stance – symbolically represented by a handshake with President Xi – the local government introduced a wide range of policies and resources that fostered the base's establishment and operations, as well as the recruitment and stability of workers with disabilities.
For example, in Shenzhen – where Base F is located – the coveted household registration status provides a range of disability-related benefits. These include housing subsidies, 5 access to free assistive equipment such as wheelchairs, and a monthly allowance of 250 to 1000 RMB, 6 depending on the subdistrict 7 (Fieldnotes, Base F). However, obtaining a Shenzhen hukou is challenging for individuals with disabilities. One key obstacle is the educational requirement, which mandates at least a college diploma – an attainment level that many people with disabilities do not meet. Additionally, the physical examination can disqualify those with certain disabilities, such as visual impairments, by deeming them medically unfit. As a result, illegal intermediaries exploit this situation, charging 30,000 to 40,000 RMB to facilitate Shenzhen hukou applications for people with disabilities. For instance, Linda, an annotator with visual impairments at Base D, obtained her Shenzhen hukou by paying such an intermediary fee. Similarly, Zhang from Base E and Hao from Base D, connected through the organisation, sought Linda's assistance in contacting the same intermediary (Fieldnotes, Base D, E and F). Later, Wusi leveraged her local networks across various departments to help most workers in Base D, including Boss Zhang, relocate their household registration to Shenzhen (Fieldnotes, Chengdu business trip).
As a result, workers with disabilities who obtain a Shenzhen hukou can access additional financial subsidies from the local government. Moreover, among people with disabilities, Shenzhen is widely recognised for its superior inclusiveness compared to other first-tier cities like Shanghai and Beijing. For instance, metro staff in Shenzhen frequently offer assistance to wheelchairs users, checking whether they need help navigating stairs or installing a ramp. Lin, a wheelchair user at Base F, took exams where staff members proactively installed a ramp whenever she encountered classrooms with steps, and several exam venues were already equipped with elevators. (Interview: F20210107). This support stabilises the employment of people with disabilities at Base F and encourages further recruitment.
Yet, not all ‘branches’ receive ongoing nourishment: some are effectively abandoned so that the ‘trunk’ may grow taller. Base C illustrates such a case. Although the local government initially assisted Base C with recruiting people with disabilities, challenges soon arose. In Heilongjiang province, disability-related affairs are closely regulated, requiring county-level DPF offices to issue acceptance letters and cover lodging and meal stipends for trainees. Sun and his spouse relied on these subsidies by charging a daily 1000 RMB training fee for up to 18 days – the allowable maximum – even though participants typically received only a few days of instruction before being assigned to data annotation. Consequently, many workers lacked both adequate knowledge and awareness of tax requirements under the labour dispatch contracts, which mandate a 20% income withholding that is returned at year's end. Believing their wages were being withheld, some workers lodged complaints against Sun's organisation. The firm Wellbeing grew dissatisfied with these disputes and the local DPF's failure to resolve them. A corruption scandal subsequently prevented the local DPF from reimbursing the upfront renovation costs as promised, causing Base C to collapse under unmet financial obligations. (Fieldnotes, Base C)
However, these ‘branches’ are not merely passive recipients of external support. DPOs exercise considerable agency in recruiting, training and retaining workers, while also allocating resources from tech capital and local governments (Wu and Xia, 2023). For instance, when two employees at Base D applied for data annotation positions at Apple, Boss Zhang learned of it through his personal connections. Sensing a negotiation opportunity, Boss Zhang warned Fox that Apple might be ‘poaching’ their employees. Ultimately, the two employees did join Apple. Nevertheless, rather than penalise Boss Zhang for the attrition, Fox increased salaries for Base D's remaining employees, as it underscored the caliber of Base D's workforce and services. When the two workers left, Peng – the supervisor overseeing Fox's annotation business – organised an online farewell event. He invited Boss Zhang to bring in additional recruits, and included Base D's employees in Fox's biennial pay-adjustment roster. This leverage enabled Boss Zhang to negotiate favorable terms with Fox, such as higher salaries and expanded business opportunities.
These ‘branches’ – DPOs engaged in data annotation here – demonstrate the complex interplay between central policy frameworks, local government discretion and tech capital interests. While some branches, like Base F, flourish through coordinated support, others, like Base C, may be abandoned due to funding constraints or fraught relationships with local authorities. Simultaneously, DPOs exhibit agency in brokering resources for workers with disabilities, at times aligning with governmental or corporate objectives, and at other times advocating for employees’ interests. In the following sections, we shift focus from organisational dynamics to the lived experiences of workers with disabilities – the ‘leaves’ in this study – examining how these broader structural forces shape individuals’ day-to-day work and challenges.
The leaves: Worker with disabilities framed as ideal annotators
Our interviews with engineers (Interviews: O20220101; O20220103; O20220105; O20220115) reveal that a low turnover rate among data annotation workers significantly enhances algorithm accuracy. Both the objective and subjective judgments of algorithms rely heavily on skilled annotators. In this context, DPOs leverage the stability of workers with disabilities – who tend to be more reliable than other annotators – to attract tech capital. Due to their historical exclusion from the mainstream job market (Randle and Hardy, 2017), workers with disabilities highly value stable employment, which they associate with work dignity and social inclusion (Xia and Wu, 2025). For instance, Hui, a worker with a second-level physical disability who uses crutches, sustained an injury in the bathroom requiring four stitches. Despite this, he chose not to report the incident to the DPO, fearing he might lose his job (Fieldnotes, Base E). At Base D, most workers have been employed for over 4 years since the organisation began its data annotation work, a remarkable tenure in an industry typically characterised by high turnover (Gray and Suri, 2019).
DPOs also promote workers with disabilities as exceptionally skilled annotators. We found out that blind individuals utilise their vivid imagination to compensate for their lack of sight, making them adept at understanding and interpreting others’ intentions. These skills are particularly suited to semantic annotation, which requires imagination and the ability to tag metadata documents with relevant concepts (Interviews: D20200101; D20200102; D20200103; D20200104). Another characteristic leveraged by DPOs is the workers’ unwavering focus. For example, individuals with visual impairments rely on screen-reading software to access digital content, preventing them from engaging in distractions like phone entertainment during work (Fieldnotes, Base D).
By portraying workers with disabilities as creative, stable, focussed and eager to work, DPOs align their initiatives with the interests of tech companies. At the same time, these companies frame their support as compliance with the central government's ‘Common Prosperity Policy’, using the empowerment of workers with disabilities to enhance their public image. In this way, workers with disabilities are constructed as the ideal data annotators, meeting the needs of both DPOs and tech capital.
However, these negotiations often prioritise productivity and adherence to corporate or policy agendas, sometimes at the expense of workers’ well-being. Hong (2024) illustrates that in Singapore, people with disabilities are largely left to contend with free-market forces and minimal legal protections, relying on workfare to incentivise employers. Similarly, while DPOs help bridge gaps left by the market, they may inadvertently contribute to worker exploitation. Hui's injury, for instance, underscores how the pressure to conform to ‘normate’ expectations (Garland-Thomson, 1997) can lead to self-suppression and neglect of personal well-being. This analysis highlights the need to critically assess the terms of inclusion offered by both DPOs and tech capital to ensure that the pursuit of inclusion does not come at the cost of workers’ health and rights. These tensions set the stage for the next section, where we explore the conflict between crip time and industrial time – a conceptual lens that sharpens our understanding of workers with disabilities’ position as ‘leaves’ in the platformisation tree. While the metaphor captures the multi-layered governance structure (roots, trunks, branches, leaves), this section interrogates how temporal regimes reinforce that hierarchical logic, often relegating workers with disabilities to the most exposed and precarious roles in the ecosystem.
Tensions and negotiations: Crip time vs. industrial time
As previously noted, data annotation work is highly platformised, governed by algorithmic controls that reinforce an ‘industrial time’ often at odds with ‘crip time’ (Hendren, 2020; Kafer, 2021), which describes how disabled bodies and minds adjust or disrupt industrial temporal norms. This tension disproportionately impacts workers with disabilities. Fox employs a ‘half-hour slicing management’ system that combines algorithmic surveillance with human oversight (Lei, 2021; Veen et al., 2019). Every thirty minutes, a data quality specialist – coordinating with a base-level team leader – issues warnings to annotators who fall below performance thresholds. In January 2021, tensions arose at Base E when physically disabled workers required extended bathroom breaks due to mobility constraints, exacerbated by constipation from prolonged wheelchair use. Managers at Fox dismissed these breaks as ‘excuses’ for unmet performance targets, triggering significant conflict. To address the issue, local DPOs invited Fox's data quality specialists to visit Base E in May. After observing firsthand how long workers needed for restroom use, Fox rescinded the policy.
This case not only illustrates the clash between industrial and crip temporalities, but also speaks to the broader ‘regime of counting’ (Chandhiramowuli et al., 2024) – a system in which performance, productivity and compliance are continuously quantified. The annotation process becomes a site of algorithmic discipline, where disabled annotators’ lived temporalities are rendered invisible unless actively defended. By situating this case within the platformisation tree, we show that the logic of counting and time control disproportionately manifests at the ‘leaves’ – the most precarious point in the ecosystem – while upstream actors remain insulated from its consequences. This highlights the analytical utility of our framework: it makes visible not only structural hierarchies, but also how temporal and quantification regimes operate unevenly across them.
Recent scholarship, however, emphasises the emergence of non-normative management strategies devised by DPOs (Wu, 2023; Wu and Xia, 2023; Xia and Wu, 2025). Drawing on the concept of ‘crip time’, Wu (2023: 10) illustrates how a DPO negotiates a fixed-salary contract for collective performance, thereby accommodating individual crip time within a corporate schedule. Expanding on this perspective through our broader fieldwork with disabled annotators, we find that such negotiations are undertaken not only by DPOs but also by workers themselves. For example, Base D is divided into six teams, with the Content Playback Team handling the largest number of daily annotations – often unable to finish in a single day. Consequently, once other teams complete their daily quotas, they ‘actively support’ the Playback Team (Fieldnotes, Base D). As one employee noted: “We receive our quota on Friday. Under normal circumstances, the Playback Team won’t finish that week's quota by Thursday, so on Friday they still have to work on the previous week's quota. The other teams will put aside any newly assigned tasks and help the Playback Team first. Only when the Playback Team is completely done does everyone move on to the new data for each team.” (Interview: D20200111)
At the same time, both new and experienced employees employ various strategies to maintain similar annotation volumes. For instance, the Chat Team, consisting primarily of newer workers, learns about veteran employees’ workloads from daily end-of-day reports. In their private WeChat group called ‘Serious Chat’, they discuss reducing their own daily quota. As a result, they lowered their targets from 750 to 650 annotations, ultimately stabilizing at 550 to match the senior employees’ workload (Fieldnotes, Base D). As one participant explained: “Anyway, in the end, everyone just talks it over and decides roughly how many tasks to do each day, and that's pretty much it.” (Interview: D20200108)
The same ‘cooperative understanding’ is also evident across different bases. Because of the random distribution of ‘double annotation’, the same data packet may occasionally be assigned to two annotators in separate locations – such as Base D and Base F – resulting in collabourative pairs tasked with double annotating. Although visually impaired annotators can accelerate their screen-reading software to 40 or 50 times the normal speed, they still cannot match the pace of physically impaired annotators who can ‘read multiple lines at a glance’. Visually impaired workers also rely primarily on keyboard operations, which is slower than using a mouse, and they must spend additional time confirming typos. Consequently, for the same data packet, physically impaired annotators typically label faster than their visually impaired counterparts.
Nevertheless, physically impaired workers at Base F deliberately slow their pace so that their daily output exceeds that of their visually impaired partners at Base D by only 100 to 300 tasks. Setting this gap involves careful calculation: “If we [physically impaired annotators] produced the same output as they [visually impaired annotators], the higher-ups at Fox would think we’re not capable. So we need to produce more, but not by too much. If the gap is too big, then the people [visually impaired annotators] over in Shanghai wouldn’t be happy… [All of] this is something we worked out among ourselves.” (Interview: F20210102)
In an effort to ‘catch up’ to industrial time, blind annotators in Wu's (2023: 10) study adopt what she describes as ‘disrupt time’, shortcutting the linear progression of tasks through high-speed reading. While these techniques challenge normative assumptions of inefficiency, we also find they can cause physical harm, primarily to the annotators’ ears. Because they rely on screen-reading software all day, they must wear headphones for extended periods, resulting in ear pain and fatigue. Furthermore, the digital voice used by these tools is emotionless, leaving workers craving more natural human interaction after work: “(The voice of the screen-reading software is) very robotic and emotionless. Personally, I really want to take off my work headphones and listen to real people talking outside.” (Interview: F20200101)
Challenges and negotiations: computerised annotation rules
The rules of data annotation are typically articulated through algorithmic, computer-oriented language. For instance, Picture 1 shows interview questions used at Bases D, E and F, with all potential answers expressed in a programming-like format. For those unfamiliar with data annotation, even interpreting what these answers mean can be time-consuming, let alone determining which one is correct. The HR manager at Base D remarked that the primary criterion for hiring annotators is not simply identifying the right answer but assessing whether a candidate's reasoning ‘aligns with machine logics’ (Fieldnotes, Chengdu Business Trip).
Picture 1: Job interview questions
Moreover, these annotation rules frequently change, often without notice from upstream developers: “After [the developers] change the rules, they don’t tell us, which causes a lot of trouble… They keep changing their minds… Sometimes they even forget what the original rules were…” (Interview: D20200104)
Fox, for example, provided a 2000-page document detailing its annotation guidelines. However, through ongoing communication between developers and annotators, nearly 1600 pages were eventually discarded, indicating the fluid nature of these regulations. One initial rule divided user requests into three categories – alarm, calendar reminder, and countdown – based on the instruction's duration (‘longer than 60 seconds’ as alarm and ‘60 seconds or shorter’ as countdown). Before long, this approach was superseded by a keyword-based system that prioritised one keyword over another: for instance, ‘countdown’ outranked ‘alarm’, which in turn superseded ‘calendar reminder’, while ‘shout’ took precedence over ‘call’, and ‘call’ outranked ‘tell me’. Instructions containing an event were prioritised over those without. Consequently, ‘Wake me up at 8 a.m. tomorrow’ was categorised as ‘alarm’, whereas ‘Remind me to cook at 8 a.m. tomorrow’ was classified as ‘reminder’, due to the keyword ‘cook’. If the user said, ‘Wake me up at 8 a.m. tomorrow to cook’, the keyword ‘wake’ was considered higher priority, so the command remained ‘alarm’ despite containing ‘cook’ (Fieldnotes, Base D). Such frequent revisions illustrate how everyday thinking is repeatedly adapted to machine-oriented logics.
To cope with this monotonous, machine-driven work, Zhao, Sun's spouse at Base C, deliberately fostered a familial atmosphere to retain annotators with disabilities. In the early days of the base, she visited the dormitories of more than twenty new recruits from different counties almost every evening, chatting with them to encourage them to stay on (Interview: C20200103). During our field work, Zhao was seen moving among different workstations daily, chatting informally with each annotator. (Fieldnotes, Base C).
Annotators themselves also employ strategies to mitigate the monotony, such as working and chatting simultaneously: “Working while chatting makes the job go more smoothly… It's not like customer service, where you have to talk or not talk. Sometimes, you just shout for someone to come over and check how to frame something. After they look at it, we chat a bit, then go back to work. Before you know it, the numbers of finished tasks add up.” (Interview: C20200110)
This relaxed, family-like environment prompts many annotators to ‘never take time off – because they’re afraid of getting laid off if they do’, while others ‘quietly work overtime every day’ (Interview: C20200103).
At Base D, annotators also chat while working, discussing both job-related matters and personal life. Occasionally, visually impaired workers wearing headphones to use screen-reading software will protest loudly against excessive chatter (Fieldnotes, Base D). Among themselves, visually impaired employees adopt nicknames such as ‘Mosquito’, ‘Puppy’ and ‘Chick’, jokingly calling their workplace a ‘zoo’ (Interview: D20200104).
By cultivating this supportive, lighthearted environment, disabled annotators actively reduce the difficulty imposed by complex and mechanistic annotation rules, making their work more manageable overall.
Discussions and conclusion
Drawing on Van Dijck's (2021) conceptualization of the ‘platformisation tree’, we have illustrated how the central government in China simultaneously acts as an enabler and a regulator in the data annotation industry. In a similar vein, Hong (2024) argues that Grab's success in Southeast Asia – credited to its social mission and willingness to collabourate with governments (Tam et al., 2018) – demonstrates how ‘platformisation’ can integrate state-directed public infrastructure ideals into an evolving superapp model. Here, however, we emphasise that local governments in China also operate within the central government's dual roles as both enabler and regulator. Rather than merely executing top-down policies, they supply and enrich resources for tech capital under the central policy framework. The case of ‘Wusi's handshake with President Xi’ serves as a potent symbol: once local authorities recognised central government endorsement, they exercised considerable agency in supporting data annotation work.
Local governments thus become the ‘trunk’ of this data annotation tree, negotiating extensively with tech capital on issues such as long-term commitments, workforce availability, infrastructure support and additional benefits. This negotiation is not always smooth; tensions and conflicts can lead local authorities to withhold or reconfigure financial, institutional, or symbolic resources. As the negotiation forces travel from the ‘root’ (central government) through the ‘trunk’ (local government and tech capital), they ultimately determine whether ‘branches’ (DPOs) flourish or fail. Base C's collapse underscores how conflicts at the trunk level can decisively shape the fate of a data annotation base.
Unlike conventional brokers – often characterised by their tendency to commodify or exploit workers (Burt, 1992; Rodriguez, 2010; Stinchcombe, 1990) – DPOs in this study occupy a more ambiguous position. Sometimes, they advocate for workers’ interests and leverage crises as opportunities, as seen when Apple recruited annotators from Base D and the DPO negotiated favourable terms. At other times, they align with governmental or corporate agendas.
Within this platformisation tree (as picture 2 shows), workers with disabilities represent the ‘leaves’. Their portrayal as ideal annotators is not merely symbolic; it encapsulates their continuous negotiation with the tensions and challenges that arise from the dynamics flowing from the root to the trunk and branches. For example, DPOs facilitate inclusion initiatives by portraying workers with disabilities as ideal data annotators. However, critical disability studies caution that such initiatives may inadvertently reinforce oppressive structures by enforcing conformity to able-bodied norms and stifling diverse lived experiences (Mitchell and Snyder, 2015; Titchkosky, 2011). Hong (2024: 2598) similarly contends that access framed in inclusive or rehabilitative terms can simultaneously exclude certain bodies and normalise others, thereby eroding disability's radical potential. Depicting workers with disabilities as creative, stable, focussed and eager to work underscores an aspiration towards normative standards and may encourage workers to fit predefined, able-bodied expectations, potentially at the expense of individual identities and needs.
Picture 2: China's data annotation tree
According to Hong (2024: 2606), this tension arises when workers experience a ‘misfit’ between their embodied realities and the structure of work, resulting in emotional distress. She interprets this misfit as a form of violence that operates alongside the platform's rhetoric of ‘being your own boss’. Similar conflicts emerge in our study's accounts of ‘crip time’ versus ‘industrial time’. Fox's ‘half-hour slicing management’ system, for instance, is designed for maximum efficiency and productivity, thereby imposing algorithmic controls that disregard the embodied differences of workers with disabilities (Mitchell and Snyder, 2015; Titchkosky, 2011).
These norms of productivity and efficiency are also reflected in the standardised and computerised annotation rules, which not only change frequently but are written in machine-oriented language. In response, DPOs recruit workers with disabilities who exhibit a ‘gossiping spirit’ and cultivate a family-like environment to mitigate the monotony of the work. Within this context, ‘normate’ (Garland-Thomson, 1997) becomes embodied in the ‘platform's rhetoric’ (Hong, 2024: 2606) of thinking in a machine-like manner, reinforcing able-bodied ideals in surveillance-intensive workplaces (Castaneda et al., 2019; Hong, 2024). Although such inclusion initiatives aim to incorporate workers with disabilities into the labour force, they can inadvertently perpetuate oppression by compelling adherence to standard notions of time (Harlan and Robert, 1998; Randle and Hardy, 2017). This process directly affects workers’ experiences, as demonstrated by physical strain – ear pain and fatigue from wearing headphones for extended periods – and by individuals like Hui, who fear reporting injuries due to concerns about job security. Taken together, these cases illustrate how striving for the ‘normate’ can lead to self-suppression and overlooked well-being, raising important questions about the terms of inclusion that ‘data annotation tree’ provide.
In sum, this article extends the ‘platformisation tree’ metaphor to illuminate the multi-layered dynamics shaping data annotation work for workers with disabilities in China – an empirical context that has been largely overlooked in platform and data labour studies. Unlike previous applications of the metaphor, which often emphasise macro-level platform governance or urban infrastructural developments, our study highlights how coordination, negotiation and conflict among central government, local authorities, tech capital and DPOs co-produce labour conditions on the ground. In doing so, we contribute to the theoretical repertoire of platform studies by critically reassessing the utility and limits of the platformisation tree in capturing uneven power relations and actor interdependencies in a context shaped by China's unique governance structures. Moreover, by engaging with political economy approaches to data work, we foreground how platformised labour is governed not only by algorithms or corporate logic, but also by state-led developmental imperatives and local implementation strategies. These findings underscore the complexities of achieving meaningful inclusion and suggest that balancing state, corporate and worker interests requires sustained and critical attention to both policy design and workplace practice.
Supplemental Material
sj-pdf-1-bds-10.1177_20539517251386046 - Supplemental material for The platformisation tree in China's AI data annotation ecosystem
Supplemental material, sj-pdf-1-bds-10.1177_20539517251386046 for The platformisation tree in China's AI data annotation ecosystem by Bingqing Xia and Tongyu Wu in Big Data & Society
Supplemental Material
sj-pdf-2-bds-10.1177_20539517251386046 - Supplemental material for The platformisation tree in China's AI data annotation ecosystem
Supplemental material, sj-pdf-2-bds-10.1177_20539517251386046 for The platformisation tree in China's AI data annotation ecosystem by Bingqing Xia and Tongyu Wu in Big Data & Society
Footnotes
Acknowledgements
The authors are deeply grateful to the reviewers for their careful and patient readings, as well as for their insightful suggestions that have greatly strengthened this article. They sincerely thank all participants and interviewees for generously sharing their experiences and for their kindness in supporting the research. In particular, they would like to express their special thanks to Boss Zhang for his invaluable assistance in facilitating access to his organisations and for providing professional and detailed insights into the working lives of people with disabilities.
Ethical approval
The study was approved by East China Normal University (Reference Number: 41300-20101-222421) on 15 February 2019.
Informed consent
Informed consent was obtained verbally before participation. The consent was audio-recorded in the presence of an independent witness.
Funding
The authors received no financial support for the research, authorship and/or publication of this article.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.
Data availability statement
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
Supplemental material
Supplemental material for this article is available online.
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
