Kenya’s proposed protections for AI workers could become a defining test of whether the global artificial intelligence boom can continue to rely on human labour without leaving the people behind the models exposed to low pay, opaque contracts, relentless monitoring, and psychological harm. The draft Kenya Artificial Intelligence and Other Emerging Technologies Policy places data annotators, content moderators, AI evaluators, and other workers at the centre of a national AI strategy, recognising that the systems marketed as autonomous still depend heavily on people performing difficult and often unseen work.
For Windows users, this development matters more than it may initially appear. Modern AI products, including assistants embedded in productivity software, cloud services, search experiences, and developer tools, are trained, tested, moderated, and refined by distributed workforces across the world. Kenya’s approach raises a direct question for the technology industry: should the companies that sell polished AI services also carry clearer responsibility for the human working conditions behind them?
The answer in Kenya’s draft is increasingly clear. AI governance is not only about model safety, copyright, privacy, or hallucinations. It is also about labour.
Kenya has opened public consultation on a draft national policy covering AI and other emerging technologies. Among its most significant provisions are planned occupational protections for people working throughout the AI value chain.
The framework calls for the development and enforcement of standards covering:
The policy is still a draft, not a completed labour law or immediately enforceable wage regulation. That distinction is crucial. It lays out the government’s intended direction and proposes standards that may later need detailed regulations, guidance, enforcement mechanisms, and legal authority before they change everyday employment practice.
Still, the language is more consequential than a broad statement of principles. Kenya is proposing that AI companies and the outsourcing providers they use should be measured against locally defined standards of care rather than leaving conditions entirely to commercial contracts between global technology buyers and business-process outsourcing firms.
That could reshape one of the world’s most important pools of AI support labour.
That view is misleading.
These judgements make training data more useful and help AI systems learn patterns. They also support testing and evaluation after a model is deployed.
In generative AI, annotators and evaluators can be asked to compare multiple answers and determine which is more accurate, safer, clearer, less biased, or better aligned with a user’s instruction. That work is central to the quality of AI assistants used in commercial products.
This is not simply routine clerical work. Repeated exposure to traumatic material can carry serious mental-health risks, especially when productivity targets reward speed, discourage breaks, or make workers fear dismissal for seeking reassignment.
Kenya’s draft policy acknowledges this core reality: AI and platform work can create unique occupational harms that general employment rules may not adequately address. That recognition is one of the proposal’s strongest elements.
For services such as Microsoft Copilot, quality work may involve assessing whether an answer is grounded, whether an AI feature follows a prompt correctly, whether it handles sensitive data appropriately, and whether safety controls respond reliably to harmful requests.
This means the AI workforce is not shrinking into irrelevance as automation advances. In many areas, human labour is becoming more specialised, more difficult to audit, and more tightly connected to product reliability.
That status has created opportunity. AI-related work can provide entry points into the digital economy, generate export revenue, build operational experience, and support a wider technology ecosystem. For many workers, these roles may offer an alternative to unemployment or informal work.
But the same conditions that make Kenya attractive to global technology firms can create a power imbalance.
However, the model also complicates accountability.
A technology company may set project requirements, quality targets, tools, content categories, and deadlines. An outsourcing provider may formally employ the workers. A further subcontractor may manage a portion of the workflow. If conditions fail, workers can face a maze of separate corporate entities, contracts, and legal jurisdictions.
The Kenyan policy’s emphasis on duty of care is therefore notable. It signals that responsibility should not disappear simply because a high-risk task is placed inside a vendor arrangement.
A fair-pay framework does not necessarily mean that every worker in every country must be paid an identical salary. Costs of living, local labour law, qualifications, tax systems, and job structures differ.
But there is a meaningful difference between localised compensation and a system where workers have no clear way to understand how pay was calculated, how task rates change, whether required training is paid, or how their wages compare with the market value of comparable work.
Kenya’s proposal aims to move the discussion beyond secrecy.
The policy direction suggests that pay should be benchmarked transparently against international rates for equivalent work, with companies expected to disclose how their compensation structures compare.
This proposal has clear strengths, but its design will determine whether it delivers meaningful improvements or becomes a reporting exercise.
The benefits could include:
A workable system would need to account for:
Workers exposed to disturbing material can experience stress, sleep disruption, anxiety, burnout, and symptoms associated with secondary trauma. A one-off wellbeing seminar or an employee-assistance number is not enough where a role routinely requires exposure to extreme content.
Monitoring may have legitimate purposes. Employers need to protect customer data, prevent fraud, secure systems, manage quality, and understand workloads. But surveillance becomes harmful when it is excessive, undisclosed, or used as an automated disciplinary tool without context.
For AI workers, monitoring can create a difficult contradiction. Employees are asked to make careful human judgments about nuanced and distressing material, yet their own performance may be assessed through rigid metrics that reward speed above accuracy, safety, or recovery time.
A proportionate approach should require companies to explain:
The public consultation process is a necessary first step. Yet the ultimate impact will depend on what happens after consultation ends.
Compliance reporting could be useful, but self-reported disclosures are only as reliable as the methodology, audit rights, and consequences behind them. Companies must not be allowed to publish reassuring statements while workers lack access to the same underlying records.
If a global AI company contracts with a Kenyan provider, specifies the work, controls the tools, dictates productivity expectations, and benefits from the output, there is a strong argument that it should not be able to disclaim all responsibility for working conditions.
At minimum, meaningful accountability would require lead companies to conduct due diligence, assess vendor practices, fund appropriate safety measures, support independent audits, and remedy violations discovered in their supply chain.
A weaker framework could let major technology brands point to vendor contracts while vendors argue that they lack the budget or authority to improve conditions. That outcome would preserve the very accountability gap the draft policy seeks to address.
AI is increasingly embedded in everyday work. Copilot features can summarise meetings, draft text, analyse information, assist with code, and support business workflows. Behind the user experience are complicated systems that need training data, model evaluation, safety testing, and ongoing content controls.
An organisation choosing AI services can ask providers whether they maintain robust standards for outsourced human labour. That does not require businesses to investigate every contractor individually. It does mean treating workforce practices as part of responsible vendor governance.
Useful procurement questions include:
A model based primarily on keeping labour costs low is vulnerable. It can lead to high turnover, reputational damage, legal disputes, poor-quality output, worker illness, and abrupt client exits. It also creates a race to the bottom in which jurisdictions compete by accepting weaker protections.
A standards-based model offers a more durable advantage.
Kenya could position itself not merely as a lower-cost source of digital labour, but as a jurisdiction known for high-quality, accountable AI operations.
That approach aligns economic development with worker dignity. It also gives international customers a clearer basis for choosing Kenyan partners when they need reliable, ethical AI support services.
The risk, however, is displacement. If standards are introduced without cross-border coordination, some firms may attempt to move work to countries with lower costs and fewer protections. Kenya will need to combine enforcement with investment in higher-value AI skills, local innovation, education, and infrastructure so that the country is not dependent on the most easily relocatable tasks.
They are not peripheral. They are part of the product.
The proposed framework correctly identifies the principal areas where AI labour protections have lagged behind the industry’s rapid growth: pay transparency, mental wellbeing, contracts, monitoring, grievances, and responsibility across outsourcing chains. It also recognises that workers who make AI safer should not be asked to absorb the human cost of that safety alone.
The next stage will determine whether the policy becomes a meaningful standard or a well-intentioned statement. Kenya will need enforceable rules, practical benchmarks, independent oversight, confidential reporting channels, and obligations that reach beyond the immediate employer to the companies benefiting from the work.
For the wider technology sector, the message is harder to ignore. The future of AI will not be judged solely by how capable models become or how seamlessly assistants integrate into Windows and cloud services. It will also be judged by whether the human workforce behind those systems is paid fairly, protected properly, heard clearly, and treated as indispensable rather than invisible.
For Windows users, this development matters more than it may initially appear. Modern AI products, including assistants embedded in productivity software, cloud services, search experiences, and developer tools, are trained, tested, moderated, and refined by distributed workforces across the world. Kenya’s approach raises a direct question for the technology industry: should the companies that sell polished AI services also carry clearer responsibility for the human working conditions behind them?
The answer in Kenya’s draft is increasingly clear. AI governance is not only about model safety, copyright, privacy, or hallucinations. It is also about labour.
Overview: A Draft Policy With Global Consequences
Kenya has opened public consultation on a draft national policy covering AI and other emerging technologies. Among its most significant provisions are planned occupational protections for people working throughout the AI value chain.The framework calls for the development and enforcement of standards covering:
- Fair and transparent pay
- Written and understandable employment contracts
- Psychosocial and mental-health support
- Safeguards against harmful-content exposure
- Accessible grievance and redress procedures
- Proportionate workplace surveillance
- Clearer employer accountability
- Compliance expectations for domestic and international firms operating in Kenya
The policy is still a draft, not a completed labour law or immediately enforceable wage regulation. That distinction is crucial. It lays out the government’s intended direction and proposes standards that may later need detailed regulations, guidance, enforcement mechanisms, and legal authority before they change everyday employment practice.
Still, the language is more consequential than a broad statement of principles. Kenya is proposing that AI companies and the outsourcing providers they use should be measured against locally defined standards of care rather than leaving conditions entirely to commercial contracts between global technology buyers and business-process outsourcing firms.
That could reshape one of the world’s most important pools of AI support labour.
The Invisible Workforce Behind AI
Public discussion of AI often focuses on large language models, data centres, semiconductor supply chains, advanced chips, and billion-dollar investments. The work performed by humans is frequently treated as a temporary implementation detail rather than a permanent part of the technology stack.That view is misleading.
Data annotation is foundational work
Data annotation is the process of assigning labels, descriptions, classifications, rankings, or other structured information to raw data. A worker might identify objects in an image, transcribe audio, judge whether a chatbot answer follows instructions, classify a sentence by sentiment, or determine whether a user prompt contains harmful material.These judgements make training data more useful and help AI systems learn patterns. They also support testing and evaluation after a model is deployed.
In generative AI, annotators and evaluators can be asked to compare multiple answers and determine which is more accurate, safer, clearer, less biased, or better aligned with a user’s instruction. That work is central to the quality of AI assistants used in commercial products.
Content moderation is a high-risk occupation
Content moderation is different but closely related. Moderators review material uploaded to digital services and decide whether it violates platform rules or legal requirements. They may encounter images, video, text, and audio involving extreme violence, sexual exploitation, self-harm, abuse, hate speech, fraud, or other distressing subject matter.This is not simply routine clerical work. Repeated exposure to traumatic material can carry serious mental-health risks, especially when productivity targets reward speed, discourage breaks, or make workers fear dismissal for seeking reassignment.
Kenya’s draft policy acknowledges this core reality: AI and platform work can create unique occupational harms that general employment rules may not adequately address. That recognition is one of the proposal’s strongest elements.
AI quality evaluation is becoming more important
The third category—AI quality evaluation—is likely to grow rapidly. As companies release increasingly capable assistants, they need people to test outputs across languages, cultures, domains, and safety scenarios.For services such as Microsoft Copilot, quality work may involve assessing whether an answer is grounded, whether an AI feature follows a prompt correctly, whether it handles sensitive data appropriately, and whether safety controls respond reliably to harmful requests.
This means the AI workforce is not shrinking into irrelevance as automation advances. In many areas, human labour is becoming more specialised, more difficult to audit, and more tightly connected to product reliability.
Why Kenya Has Become a Critical AI Labour Hub
Kenya has developed a major role in data annotation, content moderation, digital work, and business-process outsourcing. Its large English-speaking workforce, established technology sector, growing digital infrastructure, and strategic position in Africa have made it attractive to international contractors and technology companies.That status has created opportunity. AI-related work can provide entry points into the digital economy, generate export revenue, build operational experience, and support a wider technology ecosystem. For many workers, these roles may offer an alternative to unemployment or informal work.
But the same conditions that make Kenya attractive to global technology firms can create a power imbalance.
Outsourcing can obscure responsibility
AI companies often procure data labelling and moderation services through specialist vendors rather than employing workers directly. This model can make administrative sense: an outsourcing partner may handle recruitment, facilities, payroll, local operations, and project management.However, the model also complicates accountability.
A technology company may set project requirements, quality targets, tools, content categories, and deadlines. An outsourcing provider may formally employ the workers. A further subcontractor may manage a portion of the workflow. If conditions fail, workers can face a maze of separate corporate entities, contracts, and legal jurisdictions.
The Kenyan policy’s emphasis on duty of care is therefore notable. It signals that responsibility should not disappear simply because a high-risk task is placed inside a vendor arrangement.
Low cost should not mean low standards
The draft policy also confronts an uncomfortable economic reality. Global AI companies can benefit from significant differences in labour costs between countries, while workers doing essential work may have limited visibility into the value their efforts create.A fair-pay framework does not necessarily mean that every worker in every country must be paid an identical salary. Costs of living, local labour law, qualifications, tax systems, and job structures differ.
But there is a meaningful difference between localised compensation and a system where workers have no clear way to understand how pay was calculated, how task rates change, whether required training is paid, or how their wages compare with the market value of comparable work.
Kenya’s proposal aims to move the discussion beyond secrecy.
The Fair-Pay Reference Framework: Ambitious but Complex
One of the most ambitious provisions is a proposed fair-pay reference framework for AI-related roles, including data annotation, content moderation, and AI quality evaluation.The policy direction suggests that pay should be benchmarked transparently against international rates for equivalent work, with companies expected to disclose how their compensation structures compare.
This proposal has clear strengths, but its design will determine whether it delivers meaningful improvements or becomes a reporting exercise.
What the framework could achieve
A properly designed reference system could give workers, regulators, and employers a shared basis for evaluating compensation. It could expose extreme gaps, encourage better wage records, and make it harder for firms to market work as highly skilled while paying rates that do not reflect the demands or risks involved.The benefits could include:
- Better clarity around hourly pay, task rates, bonuses, deductions, and overtime
- More consistent treatment across contractors and subcontractors
- A clearer distinction between entry-level annotation and specialist evaluation work
- Stronger incentives to compensate workers for harmful-content exposure
- More reliable documentation for labour inspections and worker grievances
- Greater visibility into unpaid tasks such as onboarding, testing, training, and compliance checks
The central question: comparable to what?
The phrase international rates for equivalent work sounds straightforward but will be difficult to implement fairly. A content moderator reviewing graphic material, a multilingual evaluator grading complex model responses, and a basic image-labelling worker may all sit under the broad label of “AI worker,” but the nature, risk, and required expertise of their jobs can differ substantially.A workable system would need to account for:
- Skill level and domain expertise
Coding, law, medicine, cybersecurity, language expertise, and advanced reasoning tasks should not be priced like basic repetitive classification. - Exposure to harmful material
A worker handling traumatic content requires different safeguards and potentially different compensation from someone labelling benign consumer-product images. - Employment status
Full-time employees, temporary staff, agency workers, freelancers, and remote contractors may all have different benefit structures. A nominal hourly rate cannot be compared honestly without considering paid leave, health coverage, social contributions, equipment, and job security. - Hours actually worked
Pay transparency must include unpaid waiting time, mandatory training, quality reviews, appeals, testing, and administrative tasks. - Local living conditions
International comparison should not become a simplistic currency conversion. It must take account of local costs, legal wage floors, and the goal of decent work.
Mental Health Support Must Be More Than a Hotline
The proposed policy places mental health and psychosocial support alongside compensation and contracts. That is an essential decision.Workers exposed to disturbing material can experience stress, sleep disruption, anxiety, burnout, and symptoms associated with secondary trauma. A one-off wellbeing seminar or an employee-assistance number is not enough where a role routinely requires exposure to extreme content.
What effective duty of care should look like
The practical standard should include a full system of prevention, not merely support after harm has occurred. Strong occupational protection guidelines would be expected to include:- Honest recruitment disclosures about the nature of content workers may review
- Pre-assignment risk assessments for high-exposure projects
- Rotation away from distressing task categories
- Reasonable limits on cumulative exposure
- Paid breaks that are not treated as productivity failures
- Confidential, trauma-informed counselling
- Clinical referral pathways for workers who need additional care
- The right to request reassignment without retaliation
- Training for managers on trauma, stress, and respectful escalation
- Productivity targets that reflect task complexity and worker wellbeing
The surveillance problem
Kenya’s draft also addresses workplace surveillance, calling for measures that are proportionate and respectful of worker rights. This is especially relevant to digital labour, where every click, pause, correction, mouse movement, and completed task can be tracked.Monitoring may have legitimate purposes. Employers need to protect customer data, prevent fraud, secure systems, manage quality, and understand workloads. But surveillance becomes harmful when it is excessive, undisclosed, or used as an automated disciplinary tool without context.
For AI workers, monitoring can create a difficult contradiction. Employees are asked to make careful human judgments about nuanced and distressing material, yet their own performance may be assessed through rigid metrics that reward speed above accuracy, safety, or recovery time.
A proportionate approach should require companies to explain:
- What data is collected about workers
- Why it is needed
- How long it is retained
- Whether it affects pay, scheduling, promotion, or dismissal
- Whether automated systems make or recommend employment decisions
- How a worker can correct inaccurate data or appeal an outcome
A Policy Is Not Yet an Enforcement Regime
The proposed protections deserve attention, but the enthusiasm around them should be tempered by an important legal and operational fact: a draft policy does not automatically create enforceable rights.The public consultation process is a necessary first step. Yet the ultimate impact will depend on what happens after consultation ends.
The implementation gap
The policy will need answers to several difficult questions:- Which government body will enforce compliance?
- Will rules apply to every firm in the AI supply chain or only formally registered employers?
- Will international technology buyers be jointly responsible with local vendors?
- What penalties will apply for non-compliance?
- How will workers report violations safely?
- Will anonymous reporting be available?
- Can regulators inspect vendor facilities and audit digital records?
- How often will pay and wellbeing reports be submitted?
- Will workers and unions help define the benchmarks?
- How will the standards apply to remote contractors and gig workers?
Compliance reporting could be useful, but self-reported disclosures are only as reliable as the methodology, audit rights, and consequences behind them. Companies must not be allowed to publish reassuring statements while workers lack access to the same underlying records.
The outsourcing liability test
The largest unresolved issue is likely to be lead-company accountability.If a global AI company contracts with a Kenyan provider, specifies the work, controls the tools, dictates productivity expectations, and benefits from the output, there is a strong argument that it should not be able to disclaim all responsibility for working conditions.
At minimum, meaningful accountability would require lead companies to conduct due diligence, assess vendor practices, fund appropriate safety measures, support independent audits, and remedy violations discovered in their supply chain.
A weaker framework could let major technology brands point to vendor contracts while vendors argue that they lack the budget or authority to improve conditions. That outcome would preserve the very accountability gap the draft policy seeks to address.
What This Means for Microsoft, Copilot, and Enterprise AI
The policy is relevant to the entire AI industry, including companies building tools for Windows PCs, Microsoft 365, Azure, enterprise security, developer platforms, and cloud-based productivity services.AI is increasingly embedded in everyday work. Copilot features can summarise meetings, draft text, analyse information, assist with code, and support business workflows. Behind the user experience are complicated systems that need training data, model evaluation, safety testing, and ongoing content controls.
Responsible AI cannot stop at the model boundary
Enterprise buyers rightly ask questions about data privacy, security controls, compliance, transparency, model accuracy, and vendor lock-in. They should also recognise that responsible AI procurement has a labour dimension.An organisation choosing AI services can ask providers whether they maintain robust standards for outsourced human labour. That does not require businesses to investigate every contractor individually. It does mean treating workforce practices as part of responsible vendor governance.
Useful procurement questions include:
- Does the provider have enforceable standards for data annotators and content moderators?
- Are outsourced workers covered by written contracts and fair grievance procedures?
- Are workers paid for all mandatory training and evaluation time?
- How are harmful-content projects assessed and managed?
- Are mental-health services confidential and trauma-informed?
- Does the company audit contractors and subcontractors?
- Can workers report concerns without retaliation?
- Are automated performance-management systems transparent and appealable?
Kenya’s Competitive Advantage Could Become Better, Not Weaker
Some technology firms may argue that stronger labour requirements make Kenya less competitive as an outsourcing destination. That interpretation is too narrow.A model based primarily on keeping labour costs low is vulnerable. It can lead to high turnover, reputational damage, legal disputes, poor-quality output, worker illness, and abrupt client exits. It also creates a race to the bottom in which jurisdictions compete by accepting weaker protections.
A standards-based model offers a more durable advantage.
Better conditions can support better AI
When workers receive clear contracts, predictable pay, proper training, manageable workloads, and support for high-risk tasks, firms can build a more skilled and stable workforce. That matters as AI work shifts from simple labelling toward complex evaluation, specialised domain review, multilingual testing, and safety assurance.Kenya could position itself not merely as a lower-cost source of digital labour, but as a jurisdiction known for high-quality, accountable AI operations.
That approach aligns economic development with worker dignity. It also gives international customers a clearer basis for choosing Kenyan partners when they need reliable, ethical AI support services.
The risk, however, is displacement. If standards are introduced without cross-border coordination, some firms may attempt to move work to countries with lower costs and fewer protections. Kenya will need to combine enforcement with investment in higher-value AI skills, local innovation, education, and infrastructure so that the country is not dependent on the most easily relocatable tasks.
The Real Measure of Success
Kenya’s draft AI worker protection framework is significant because it challenges an assumption that has shaped much of the AI economy: that the people who train, test, and moderate systems are peripheral to innovation.They are not peripheral. They are part of the product.
The proposed framework correctly identifies the principal areas where AI labour protections have lagged behind the industry’s rapid growth: pay transparency, mental wellbeing, contracts, monitoring, grievances, and responsibility across outsourcing chains. It also recognises that workers who make AI safer should not be asked to absorb the human cost of that safety alone.
The next stage will determine whether the policy becomes a meaningful standard or a well-intentioned statement. Kenya will need enforceable rules, practical benchmarks, independent oversight, confidential reporting channels, and obligations that reach beyond the immediate employer to the companies benefiting from the work.
For the wider technology sector, the message is harder to ignore. The future of AI will not be judged solely by how capable models become or how seamlessly assistants integrate into Windows and cloud services. It will also be judged by whether the human workforce behind those systems is paid fairly, protected properly, heard clearly, and treated as indispensable rather than invisible.
References
- Primary source: TechTrendsKE
Published: 2026-07-25T08:33:53+00:00
Kenya Unveils Draft AI Worker Protection Policy - TechTrendsKE
Kenya's proposed AI workers policy introduces pay benchmarks, mental health support and labour protections for content moderators and data annotators.techtrendske.co.ke - Related coverage: kenyanews.go.ke
- Related coverage: business-humanrights.org
Kenyan platform workers demand better conditions to address job insecurity, unrealistic performance targets, and constant digital surveillance, while calling for transparent ratings and reliable social welfare
Check out this page via the Business and Human Rights Centre
www.business-humanrights.org
- Related coverage: ilo.org
Kenya Validates National Strategy to Support Sustainable Reintegration of Returning Migrant Workers
Validating Kenya’s draft National Return and Reintegration Strategy to strengthen coordinated support for returning migrant workers and help them access livelihoods, services, and sustainable reintegration pathways.www.ilo.org
