Who owns Workplace Data? Post-Deployment Training and Digital Public Goods
In the modern economy, data has emerged as the primary currency of control. The rise of foundation models and agentic artificial intelligence (AI) has transformed workplace data into a strategic asset, encoding both the tacit and explicit know-how of individual workers and entire organisations. Every daily task, decision, and digital interaction is captured as training data to refine systems capable of automating complex human workflows.
Operating under the legal fiction that data generated within corporate digital ecosystems is a proprietary asset, private firms and global technology providers have amassed an asymmetric collection of information. This regime leaves workers vulnerable to unilateral exploitation while stripping local firms and sovereign states of operational knowledge.
As policy analyst Farzad Shahinfard notes, workplace data in the contemporary AI era possesses a predatory 'second life.' What begins as a digital footprint returns to the workplace as an unyielding productivity target. Private corporate ownership converts a record of what a worker did into an unchallengeable dictate of what they must do next.
The harms of corporate data sovereignty extend beyond individual workplaces to threaten broader economic sovereignty. According to research by the UK’s Institute for the Future of Work (IFOW), foundation model providers and enterprise software vendors capture critical business intelligence through continuous enterprise interactions, a process known as 'post-deployment training.' Through these interfaces, raw worker inputs generate operational data captured and processed by hyperscalers and foundation model vendors to post-train general models.
In this manner, the tacit knowledge, specialised processes and operational methods of local businesses are systematically extracted and integrated into centralised AI models owned by foreign technology conglomerates. This extraction drains enterprise value from adopting firms, transferring their competitive 'moats' to global platforms and concentrating market power.
In jurisdictions like the UK, while early-stage AI startups receive domestic equity investment, late-stage capital and mergers & acquisitions (M&A) are overwhelmingly dominated by foreign investors (71% US acquisitions), funnelling local workplace intelligence out of the domestic economy.
The Evolution of Data Annotation
Hidden human labour within the value chain of frontier model development is not new. Human annotation (or labelling) of digital data traces its roots to early academic efforts to benchmark machine learning, culminating in the landmark 2009 release of ImageNet. To catalog millions of images, researchers turned to Amazon’s (now defunct) Mechanical Turk, demonstrating that distributed crowdsourcing could generate training data at an unprecedented scale. This breakthrough catalysed the deep learning revolution of the 2010s, transforming a niche academic practice into a vital commercial sector.
As enterprise investment in computer vision, natural language processing and autonomous driving surged, basic crowdsourcing proved inadequate for strict enterprise quality and data privacy standards. Instead, specialised vendors capable of handling meticulous perceptual tasks - such as drawing 3D bounding boxes on lidar point clouds, segmenting radiology scans, and transcribing multi-dialect audio - emerged to fill the void.
The generative artificial intelligence boom fundamentally reshaped the annotation landscape. Simple perceptual labelling shifted toward cognitive alignment via Reinforcement Learning from Human Feedback (RLHF), supervised fine-tuning, and adversarial ‘red-teaming’. Instead of merely identifying traffic lights, annotators now evaluated factual accuracy, stylistic nuance, safety and complex logical chains.
Historically, this unseen labour was concentrated in the Global South. As highlighted in Karen Hao’s book Empire of AI, major data annotation companies like Scale AI (through its Remotasks division) targeted regions experiencing severe economic crises but possessing high English proficiency, such as Venezuela, the Philippines, and Kenya. This created a transient, highly exploitative model where companies would abruptly shut down operations and relocate whenever local workforces demanded better terms and conditions.
However, as a recent study titled 'Surveillance at Scale' by the interdisciplinary research group DipLab points out, this dynamic is rapidly shifting. While standard labelling teams still handle foundational pipelines, companies increasingly recruit domain experts, including software engineers, clinicians, attorneys, and research scientists, to train frontier reasoning models.
Human-in-the-loop workflows have matured in parallel. Automated pre-labelling algorithms now handle baseline processing, leaving humans to resolve difficult ambiguities and validate synthetic datasets. This demand for expert feedback has pushed the data annotation industry into the wealthiest nations of the Global North.
DipLab’s study focuses on 35 data annotators across 13 European countries working for Outlier, a division of Scale AI aimed at higher-end tasks. Despite their advanced education and the complexity of their labour, the study finds that these European professionals have been thoroughly 'Uberised'; a university degree no longer shields them from gig economy exploitation.
Much like traditional platform workers in food delivery or ride-hailing, Outlier annotators face sudden, unexplained account deactivations ('robo-firings') with no transparency or meaningful opportunity to appeal. Significantly, Outlier’s workforce is employed on a ‘piece-rate’ basis. Compensation is structured per task within aggressive time limits. Workers might be offered a fixed rate for a task completed within 20 minutes, only to have their pay halved if they exceed that window. This creates relentless pressure to rush through complex evaluations.
Additionally, task allocation is deeply unequal: an elite fraction of workers receives the bulk of assignments while others receive none. Annotators also perform extensive unpaid labour, spending hours studying changing guidelines and constantly monitoring the platform to catch available tasks. While companies like Scale AI position the work as flexible contracting, workers are forced to remain constantly on call, sacrificing personal time to meet company demands.
Existing Governance Models for Workplace Data
So long as private corporations maintain exclusive sovereignty over workplace data, they are incentivised to optimise its exploitation for profit and capital accumulation. As DipLab’s research (and Salaree’s research into social care workers in England) shows, this occurs at the direct expense of employment rights and worker well-being.
Existing legal frameworks struggle to halt this value drain due to structural deficits. Data protection regimes like the EU’s GDPR treat workers as individual 'consumers' or 'data subjects,' relying on individual rights such as access and consent that fail to account for the inherent power imbalances and economic coercion present in employment relationships.
In jurisdictions like Australia, private-sector employee records are explicitly exempted from federal privacy laws, leaving millions of workers without statutory protection against data mishandling; similarly, US federal and state privacy statutes routinely exclude employment data.
Finally, legal blind spots persist because traditional frameworks treat technology adoption as an exclusive employer prerogative, leaving data governance detached from collective bargaining and social dialogue.
As technology outpaces legacy legislation, workplace regulation is entering a phase of convergence where data governance, occupational health and safety, privacy and wage integrity intersect into an integrated regulatory system. Regulatory bodies increasingly recognise that algorithmic management tools are not merely operational enhancements, but structural risks.
Because private entities are bound by fiduciary duties to maximise shareholder value, they are constrained from self-regulating their data practices. Addressing this challenge requires exploring alternative models of workplace data ownership - it seems inevitable that the state, bound by constitutional frameworks and public oversight, will need to play a more prominent role in safeguarding worker welfare in the future.
Alternative Governance Models for Workplace Data
International bodies like the International Labour Organisation (ILO) emphasise that workplace data governance must be grounded in legal principles. Treating workplace data as a public good rebalances power between employers and employees. Under public custodianship, businesses would not 'own' employee data; instead, they would lease restricted, anonymised, and context-specific access for operational necessity under strict regulatory oversight.
To protect workers, basic privacy and human dignity must be mandatory, preventing employers from forcing staff to surrender data rights in employment contracts. Regulations should strictly ban invasive practices like emotion tracking, continuous surveillance and union monitoring.
Portable public data profiles would allow workers to carry verified credentials and work history from job to job, liberating them from proprietary employer scoring systems.
What are the reference points for this framework?
One model to consider is Digital Public Infrastructure (DPI). The World Bank describes DPI as 'shared, reusable building blocks, fostering whole-of-society collaboration—including strong involvement from the private sector—and embedding core principles such as openness, interoperability, data protection, and user choice.'
This contrasts with traditional digital transformation programmes that rely on fragmented, single-vendor proprietary solutions for identity, payments, and data management, driving high costs and extractive data monopolies. Transitioning workplace data into a state managed 'digital commons' transforms personal metrics and collective operational know-how into a public utility in several key ways.
First, it counters vendor lock-in, as open-source architectures prevent corporations from trapping operational data inside closed ecosystems, reducing long-term IT costs and national technology dependencies. Second, it preserves economic sovereignty: hosting data architecture on public, sovereign servers or distributed secure trusts prevents foreign hyperscalers from harvesting domestic business intelligence.
Third, it enables democratically auditable rules, ensuring that public data architecture and the algorithms governing work are subject to public oversight, algorithmic neutrality checks, and statutory enforcement. Finally, public infrastructure ensures algorithms for hiring and scheduling are subject to mandatory audits to detect and eliminate hidden biases.
The defining political challenge of the artificial intelligence era is determining who retains the power to write the rules that govern human work. Ultimately, the debate is less about who currently owns workplace data, and more about who should.
Comments