US interest costs are soaring and debt is mounting This US federal debt crisis squeezes education and future investment The article urges early fiscal reform that shields human capital
In fiscal year 2025, the United States is set to
AI risk lists miss real hiring patterns Job data shows specific roles shrinking and new AI roles growing Education and policy must follow real vacancy data, not myths
Every week, new headlines warn that artificial intelligence could disrupt our careers. One Microsoft-backed analysis of workplace chats indicates that for specific roles, up to 90% of daily tasks could potentially be performed by AI tools, with interpreters, historians, and writers at the highest risk. However, AI job market data presents a different picture. Analyzing 180 million recent job postings, overall demand decreased by a modest 8% year over year, while postings for machine learning engineers increased by 40% following a 78% rise the previous year. Meanwhile, fewer than 1% of advertised roles explicitly require generative AI skills, despite those skills offering higher pay and appearing more frequently on resumes. For educators and policymakers, the clear message is that AI job market data already reveals where work is changing, and it often does not align neatly with the commonly circulated "40 jobs at risk" lists.
The risk lists currently circulating are based on clever but abstract models. The Microsoft study that supports one "40 jobs most at risk" article examined how well AI could perform tasks in different occupations. It produced scores that placed translators, writers, data scientists, and customer service representatives at the top, with some roles deemed 90% automatable. In contrast, manual or relational jobs such as roofing, cleaning, and nursing support rank much lower, exposing only a small portion of tasks. A separate economic study, using two centuries of patent and job data, reaches a similar conclusion: unlike past automation waves that primarily affected low-paid, low-education jobs, AI is expected to affect more well-paid, degree-intensive, often female-dominated fields.
These models are significant as they help identify which types of thinking and communication overlap most with AI's current strengths in pattern recognition, text generation, and number analysis. They also reveal a crucial structural risk: AI may increase pressure on knowledge workers rather than on routine manual staff, reversing the usual trend of technological disruption. However, these models are fundamentally thought experiments. They treat each occupation as an average set of tasks, assume rapid adoption of tools, and do not fully account for budgets, regulations, and managerial decisions that shape hiring in real companies. They indicate where automation could occur, while AI job market data, derived from tens or hundreds of millions of postings, illustrates where employers are actually focusing their efforts today.
The 8% decrease in job postings identified in a recent analysis of 180 million global vacancies establishes the context: labor demand is cooling, but not collapsing. Within this context, specific roles at risk of being replaced by AI stand out. Postings for computer graphic artists dropped by 33% in 2025, while those for photographers and writers fell by 28%, following similar declines the year prior. Jobs for journalists and reporters decreased by 22%, and public relations specialists saw a 21% decline. These are not vague categories; they represent the execution end of creative work, where tasks often involve producing large volumes of images or text based on a brief. The data indicates that when AI reduces the cost of that output, employers require fewer entry-level workers, even as they continue to hire designers and product teams engaged in research, client interaction, and strategic decisions.
The same dataset reveals a second, less expected cluster of losses in regulatory and sustainability roles. Postings for corporate compliance specialists fell by around 29%, while sustainability specialists saw a 28% decline, closely followed by environmental technicians. The drop occurs across all job grades. Sustainability managers and directors decreased by more than 30%, and chief compliance officer roles fell by over a third. Here, AI is not the main factor. Instead, political backlash against environmental, social, and governance rules, along with changing enforcement priorities, appears to be influencing this pullback. The lesson is uncomfortable yet crucial: some of the most significant job losses in the current AI era arise not from automation but from policy decisions that make entire areas of expertise optional. No exposure index based solely on technical abilities can reflect that reality.
Healthcare provides a clearer view of AI automation in practice. Jobs for medical scribes, who listen to clinical encounters and create structured notes, have dropped by 20%, even as similar healthcare administrative roles remain relatively stable. Medical coders show no significant change, while medical assistants are slightly below the broader market. This trend aligns with the use of AI-powered documentation tools in consultation rooms to transcribe doctor-patient conversations and generate clinical notes almost instantly. Nevertheless, even in this case, the data indicate a limited scope of substitution rather than a widespread wave of job losses. A specific documentation task has become more cost-effective, while the broader team supporting patients remains intact.
On the demand side, AI job market data presents a contrary trend to the alarming headlines. Postings for machine learning engineers surged by 40% in 2025 after a 78% increase in 2024, making it the fastest-growing job title in the 180-million-job dataset. Robotics engineers, research scientists, and data center engineers also experienced growth in the high single digits. Senior leadership jobs decreased only 1.7% compared to the 8% market baseline, while middle management roles fell 5.7% and individual contributor jobs declined 9%. In marketing, most roles mirrored the overall market. Still, postings for influencer marketing specialists increased by 18.3%, on top of a 10% rise the previous year, signaling significant demand for trusted human figures in a landscape filled with AI-generated content.
New research from Indeed, summarized in an AI skill transformation index, further supports this view. By examining nearly 3,000 work skills in job postings from May 2024 to April 2025, the team estimates that 26% of jobs will undergo significant changes due to AI, 54% will see moderate changes, and 20% will have low exposure to AI. Yet only 0.7% of skills are considered fully replaceable today. Software development, data and analytics, accounting, marketing, and administrative assistance rank among the most affected groups, with two-thirds to four-fifths of their tasks potentially being supported by AI. Still, job postings in many of these categories remain stable or evolve, indicating that employers are redesigning roles instead of eliminating them. When AI job market data and exposure models are analyzed together, the consistent message leans more toward "significant task reshuffling" rather than "mass unemployment."
For educators, the main risk is not that AI will render entire degrees obsolete overnight. The real danger lies in curricula continuing to follow outdated job titles while the underlying tasks evolve. AI job market data already illustrates that within the same broad field, some roles are declining while others are expanding. In creative industries, execution roles centered on producing content to a brief face pressure. At the same time, strategy, research, and client-facing design work tend to be more robust. In data and software, routine coding and reporting tasks are increasingly performed by tools, while higher-level architecture, problem framing, and governance gain value. Educators who still define careers solely as "graphic designer" or "software developer" risk steering students toward aspects of those jobs that are already being automated.
A more practical approach begins with the signals of skills in job postings. Although ads explicitly seeking generative AI skills remain a small portion of the total, demand for AI-related skills has surged. A study of job vacancies in the UK finds that jobs requiring AI skills pay about 23% more than similar positions without those skills. LinkedIn also reports a 65% year-on-year increase in members listing AI skills. Furthermore, global surveys indicate that AI competence is now a key expectation for nearly half of the positions employers are hiring for. For universities, colleges, and online providers, this implies two responsibilities. First, they must incorporate practical AI literacy—how to use tools to draft, analyze, and prototype—into most degree programs, not just those in computer science. Second, they need to teach students how to interpret AI job market data on their own, helping them parse postings, understand which skills are bundled, and identify where tasks are shifting within their chosen fields.
Administrators and quality assurance teams can also leverage AI job market data to rethink how programs are evaluated. Instead of primarily relying on long-term employment statistics linked to broad occupational codes, they can monitor live vacancy data in collaboration with job platforms and labor market analytics firms. When postings for execution-focused creative roles decline significantly over two years, while postings for creative directors and product designers remain stable, this should prompt a review of how design courses allocate time between production software and client-facing research or strategy. When machine learning engineers, data center engineers, and robotics specialists all experience substantial growth, technical programs should adjust prerequisites, lab time, and final projects to enable students to practice directing and debugging AI systems rather than just manual coding.
For policymakers, the contrast between theoretical exposure lists and AI job market data serves as a caution against broad, one-size-fits-all narratives. Suppose regulatory and sustainability roles are declining due to political decisions instead of technical obsolescence. In that case, workers in those fields require a different support package than those in areas where AI is clearly reducing demand for specific tasks. Career-change grants, regional transition funds, and public-sector hiring guidelines should align with real vacancy trends rather than abstract rankings of which jobs are "exposed" to AI. At the same time, growth in high-skill AI infrastructure roles suggests that investments in advanced training must complement industrial policy regarding data centers, cloud infrastructure, and robotics so local education systems can fill the jobs created by capital spending.
The education system also plays a key role in helping new workers interpret the confusing landscape of alarming forecasts and optimistic stories. Students now encounter lists of "40 jobs at risk," projections that a quarter of jobs will be "highly transformed," and examples of medical scribes or junior copywriters losing work to AI tools. Without proper direction, the instinctive response can lead to paralysis. Programs that present AI job market data to learners—showing which roles in their fields are shrinking, which are growing, and which skills command higher pay—can help ground those fears in reality. They can also highlight a recurring pattern in the data: the jobs that withstand these changes are those where humans provide direction, exercise judgment, build trust, and interpret outputs that AI alone cannot safely manage.
In this context, the most helpful question for educators and policymakers is no longer "Which 40 jobs are most at risk?" but rather "Which human-centered tasks does AI increase in value, and how do we train for those?" The current wave of AI job market data provides early answers. It indicates declining demand for repetitive execution, uncertain and politically influenced signals in specific regulatory areas, and significant growth in roles related to design, governance, and infrastructure for building and supervising AI systems. It shows that many jobs will be redefined rather than eliminated, and that skills in directing, critiquing, and contextualizing AI outputs are already commanding higher wages. If institutions treat this evidence as a living syllabus—reviewed each year, openly discussed, and translated into course design—they can move past hypothetical lists and support learners in navigating the jobs AI is actively reshaping today, rather than those it might replace in the future.
The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.
Bloomberry / Chiu, H. W. (2025, November 14). I analyzed 180M jobs to see what jobs AI is actually replacing today. Bloomberry.
Brookings Institution / Eberly, J. C., Kinder, M., Papanikolaou, D., Schmidt, L. D. W., & Steinsson, J. (2025, November 20). What jobs will be most affected by AI? Brookings.
FinalRoundAI / Saini, K. (2025, September 23). 30 Jobs Most Impacted by AI in 2025, According to Indeed Research. FinalRoundAI.
Indeed Hiring Lab / Recruit Holdings. (2023, December 15). Webinar: Indeed Hiring Lab – Labor Market Insights [Transcript].
Sky News. (2025). The 40 jobs "most at risk" from AI – and 40 it can't touch. Sky News Money.
Tech23 / Dave. (2025, October 13). The 40 Jobs Most at Risk from AI…and What That Means for the Rest of Us. Tech23 Recruitment Blog.
Korea’s strong healthcare still leaves families exposed to severe income loss from serious illness Health shocks cut wages, pushing seniors and students into long-term financial strain Korea must pair universal care with robust income protection to prevent poverty after illness
Deepfake image rights alone cannot stop fast, zero-cost copying of abusive content We need layered protection that combines law, platform duties, and strong school-level responses Education systems must train students and staff, act quickly on reports, and treat synthetic abuse as a shared responsibility
Generative AI has made image abuse a widespread issue. A 2024 analysis of nearly 15,000 deepfake videos found that 96 percent were non-consensual intimate content, particularly targeting women. In 2023, there were about 95,820 deepfake videos indexed online, with 98 percent being pornographic and mostly showing women and girls. Most of these are pornographic materials that the victims never agreed to create or share. At the same time, the cost of copying these files is nearly zero. Once a clip is uploaded, it can be mirrored, edited, and reposted across platforms in minutes. In response, lawmakers are swiftly working to strengthen protections against deepfakes, introducing new criminal offenses and proposing that each person’s face and voice be treated as intellectual property. Yet a crucial question remains: can rights that depend on slow, individual enforcement really protect people in schools and universities from abuse that spreads in seconds?
Discussions about the rights to deepfake images often begin with a hopeful idea. If everyone owns their likeness, victims gain a legal tool to demand removal, fines, or damages. Granting individuals copyright-like control over their faces, voices, and bodies seems empowering. It aligns with existing ideas in privacy, data protection, and intellectual property law, delivering a straightforward message to anxious parents and educators: your students' faces will belong to them, not to machines. In countries with strong legal compliance and resources, this model might deter some abuse and simplify negotiations with platforms. Denmark’s initiative to grant copyright over one’s likeness and to fine platforms that do not remove deepfakes exemplifies this approach.
The challenge lies in scale and speed. Non-consensual deepfake pornography already dominates the landscape. A study of 14,678 deepfake videos found that 96 percent involved intimate content created without consent. A 2025 survey on synthetic media use concluded that non-consensual pornography represents the vast majority of deepfake videos online. For targeted individuals, the initial leak is often just the beginning. Copies circulate rapidly between sites, countries, and languages, faster than any legal notice can reach them. They are remixed, cut into shorter clips, and reposted under new titles. Granting victims greater rights over deepfake images does not reduce copying; it merely adds more paperwork for them to deal with after the fact, usually at their own cost, with no guarantee that overseas hosts or anonymous users will comply.
The impact on young people is especially concerning. Research for the European Parliament shows that half of children aged 8 to 15 in the United Kingdom have seen at least one deepfake in the last six months, with a significant portion of synthetic sexual content involving minors. A separate police-commissioned survey in 2025 found that about a quarter of adults felt neutral or unconcerned about creating and sharing non-consensual sexual deepfakes, with 7 percent admitting to being targeted. Only half of those victims ever reported the abuse. For a teenager who sees a fake explicit video of themselves spreading through their school overnight, the promise of future compensation will likely seem hollow. Deepfake image rights do not address the shame, fear, and social isolation that follow, nor do they restore trust in institutions that failed to act promptly.
In theory, deepfake image rights empower victims. In practice, they shift the burden of responsibility. Someone who finds a sexualized deepfake must first recognize the abuse, gather evidence, seek legal help, and start a case that could take months or years to resolve. Meanwhile, the clip can be shared on anonymous forums, in private chats, and through overseas hosting services. Reports from law enforcement agencies, including the FBI in 2023, indicate a sharp rise in sextortion cases using AI-generated nudes created from innocent social media photos or video chats. Victims face threats of exposure unless they pay, yet even when police intervene, images already shared often remain online—rights based on individual enforcement lag behind automated, targeted, and global harms.
The emerging patchwork of laws highlights this gap. Denmark is working to give individuals copyright-like control over their images and voices, with penalties for platforms that fail to remove deepfakes. Spain plans heavy fines for companies that fail to label AI-generated content, following the European Union's broader requirement to mark synthetic audio, images, and video clearly. The 2025 Take It Down Act in the United States criminalizes non-consensual intimate deepfakes and requires major platforms to remove reported content within 48 hours and prevent duplicates from appearing again. The United Kingdom’s Online Safety Act makes sharing non-consensual deepfake pornography an offense. It is being bolstered by plans to criminalize its creation. On paper, victims now have more legal protection than ever before.
However, these gains are uneven and fragile. Many low- and middle-income countries lack specific deepfake regulations and have limited capacity to enforce existing cybercrime laws. Even where offenses exist, policing is often slow, under-resourced, and insufficiently sensitive to gender-based digital violence. A 2024 survey by researchers in the United Kingdom found that while about 90 percent of adults are concerned about deepfakes, around 8 percent admit to having created them. An international study of synthetic intimate imagery suggests that about 2.2 percent of respondents have been targeted and 1.8 percent have participated in production or sharing. The gap between concern and behavior indicates that social norms are still developing. Legal rights without swift, collective enforcement risk becoming a luxury for those who can afford legal representation in wealthy democracies, rather than a baseline for school and university safety worldwide.
The legal blind spots also show up in countries with major commercial pornography industries. Japan’s adult video sector is estimated to produce thousands of titles each month, generating tens of billions of yen in revenue. However, deepfake pornography targeting actresses, idols, and everyday women has only recently led to arrests for creating AI-generated obscene images. Regional legal scholars note that non-consensual deepfake creation often falls into grey areas, with no precise commercial distribution, making remedies slow and complicated. When deepfake abuse exploits popular, legal adult content markets, the harm can be dismissed as "just another clip." Deepfake image rights may appear robust in theory. Still, without quick investigative pathways and strong platform responsibilities, they rarely provide timely relief.
If rights to deepfake images aren't enough, what would proper protection look like, especially in educational settings? The first step is to consider deepfakes as an infrastructure issue, not just a speech or copyright concern. Technical standards are essential. Content authenticity frameworks that embed secure metadata at the point of capture can help verify genuine material and flag manipulated media before it spreads. Transparency rules, such as those requiring labeling of AI-generated images and videos, can be strengthened by requiring platforms to automatically apply visible markers rather than relying on users to declare them. National laws that already demand quick removal of intimate deepfakes, penalties for repeated non-compliance, and risk assessments for online harms can be aligned. This way, companies face clear, consistent responsibilities across different jurisdictions, rather than a confusing system that they can exploit.
Education systems can use this legal foundation to create a layered response. At the curriculum level, media literacy must go beyond general warnings about "fake news" and include practical lessons on synthetic media and deepfake image rights. Students should understand how deepfakes are made, how to check authenticity signals, and how to respond if a peer is targeted. Systematic reviews of deepfake pornography indicate that women and girls are disproportionately affected, with cases involving minors rising rapidly, including in school environments. Surveys show that many victims do not report incidents due to fear of disbelief, ridicule, or retaliation. Training for teachers, counselors, and university staff can help them respond quickly and compassionately, collect evidence, and utilize swift escalation channels with platforms and law enforcement. Institutions can establish agreements with major platforms to prioritize the review of deepfake abuse affecting their students and staff, including cases that cross borders.
Policymakers also need to revise procedures so that victims do not bear the full burden of enforcing their rights against deepfakes. Standardized takedown forms, centralized reporting portals, and legally mandated support services can lighten the load during a crisis. Rules requiring platforms to find and remove exact copies and easy variations of a reported deepfake, rather than putting the burden on the victim to track each upload, are crucial. The Take It Down Act's requirement that platforms remove flagged content within 48 hours and prevent its reappearance reflects this need, even while civil liberties groups express concerns about overzealous filtering. Substantial penalties for platforms that ignore credible complaints, along with independent audits of their response times, can enhance compliance beyond mere box-ticking. Internationally, regional human rights organizations and development agencies can fund specialized digital rights groups in the global South to help victims navigate foreign platforms and legal systems.
For educators, administrators, and policymakers, the goal is to connect these layers. Rights over one’s image should be seen as a minimum standard, not a maximum limit. Schools and universities should adopt clear codes of conduct that treat non-consensual synthetic imagery as a serious disciplinary offense, independent of any criminal proceedings. Professional training programs for teachers and social workers can cover both the technical basics of generative AI and the psychological aspects of shame, coercion, and harassment. Public awareness campaigns should focus less on miraculous detection tools and more on simple norms: do not share, do not joke, report, and support. Educational institutions should not passively accept regulations; they are vital environments where the social significance of deepfake abuse is shaped and where deepfake image rights can take on real meaning or remain empty promises.
Deepfake image rights are still necessary. They signal that a person’s face, voice, and body should not be raw material for anyone’s experiments or fantasies. However, rights that exist mainly on paper cannot keep pace with systems that replicate abuse at nearly no cost. The statistics on non-consensual deepfakes, on children's exposure, and on victimization highlight that the problem is real and not limited to celebrities or election campaigns. It is part of daily digital life. Protecting students and educators requires a shift from individual lawsuits to shared infrastructure: technical verification, strong platform responsibilities, fast takedown pathways, and campus cultures that treat synthetic sexual abuse as a joint concern. Suppose we do not build that layered protection now. In that case, the next generation will inherit a world where their deepfake image rights appear strong on paper but are essentially meaningless in practice.
The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.
Crest Advisory. (2025). Police-commissioned survey on public attitudes to sexual deepfakes. The Guardian coverage, 24 November 2025.
European Parliament. (2025). Children and deepfakes. European Parliamentary Research Service Briefing.
Japan News Outlets. (2025). Japan grapples with deepfake pornography as laws struggle to keep up. AsiaNews Network and The Straits Times reports, October 2025.
Kira, B. (2024). When non-consensual intimate deepfakes go viral. Computer Law & Security Review.
National Association of Attorneys General. (2025). Congress’s attempt to criminalize non-consensual intimate imagery: The benefits and potential shortcomings of the Take It Down Act.
New South Wales Parliamentary Research Service. (2025). Sexually explicit deepfakes and the criminal law in NSW.
Romero Moreno, F. (2024). Generative AI and deepfakes: A human rights approach to regulation. International Journal of Law and Information Technology.
Sippya, T., et al. (2024). Behind the deepfake: Public exposure, creation, and concern. The Alan Turing Institute.
Umbach, R., et al. (2024). Non-consensual synthetic intimate imagery: Prevalence, victimization, and perpetration. Proceedings of the ACM Conference on Human Factors in Computing Systems.
World Economic Forum. (2025). Deepfake legislation: Denmark moves to protect digital identity.
Russia–North Korea ties and Arctic shipping are reshaping the US–South Korea alliance The Northern Sea Route turns the Arctic into a new strategic front Education and policy must integrate Arctic routes into Korean security planning
Thailand is pivoting to green-powered growth Its green shift loosens Asia’s oil dependence Education must fast-track clean energy skills
For once, the most critical number in global energy comes from a country w
China quietly rations rare earths through strict export paperwork, not formal bans Its dominance in mining and magnet production turns bureaucracy into global leverage Schools and policymakers must plan for lasting supply risk and build alternative sources
AI data colonialism exploits hidden Global South data workers Education can resist by demanding fair, learning-centered AI work Institutions must expose this labor and push for just AI
Currently, between 154 million and 435 million people perform the “invisible” tasks that train and manage modern AI systems. They clean up harmful content, label images, and rank model outputs, often earning just a few dollars a day. Meanwhile, private investment in AI reached about $252 billion in 2024, with $33.9 billion directed toward generative AI alone. The global generative AI market is already valued at over $21 billion. One data-labeling company, Scale AI, is worth around $29 billion following a single deal with Meta. This disparity isn’t an unfortunate side effect of innovation; it is the business model. This is AI data colonialism: a digital plantation economy in which workers in the Global South provide inexpensive human intelligence, enabling institutions in the Global North to claim “smart” systems. Education systems are right in the middle of this exchange. The crucial question is whether they will continue to support it or help dismantle it.
AI data colonialism is more than just a catchphrase. It describes a clear pattern. Profits and decision-making power gather in a few companies and universities that own models, chips, and cloud services. The “raw material” of AI—labeled data and safety work—comes from a scattered workforce of contractors, students, and graduates, many from lower-income countries. Recent studies suggest that this data-labor force is already as large as the entire national labor market, yet most workers earn poverty-level wages. In Kenya and Argentina, data workers who moderate and label content for major AI companies earn about $1.70 to $2 per hour. In contrast, similar jobs in the United States start at roughly $18 per hour. Countries are effectively exporting human judgment and emotional resilience rather than crops or minerals. They import AI systems built on that labor at a premium.
Working conditions in this digital plantation mirror those of the old ones, though they are hidden behind screens. Reports detail workers in Colombia and across Africa handling 700 to 1,000 pieces of violent or sexual content per shift, with only 7 to 12 seconds to decide on each case. Some workers report putting in 18 to 20-hour days, facing intense surveillance and unpaid overtime. A recent study by Equidem and IHRB reveals widespread violations of international labor standards, from low wages to blocked union organizing. A TIME investigation found that outsourced Kenyan workers were earning under $2 per hour to detoxify a major chatbot. New safety protocols introduced in 2024 and 2025 aim to address these issues. However, 81% of content moderators surveyed say their mental health support is still lacking. This situation is not an error in the AI economy; it is how value is extracted. Education systems—ranging from universities to boot camps—are quietly funneling students and graduates directly into this pipeline.
If we accept the current model, AI data colonialism becomes the default “entry-level job” for a generation of learners in the Global South. Many of these workers are overqualified. Scale AI’s own Outlier subsidiary reports that 87% of its workforce holds college degrees, while 12% have PhDs. These jobs are not casual side gigs; they are survival jobs for graduates unable to find stable work in their fields. Yet their tasks are treated as disposable piecework rather than as part of a structured learning journey into AI engineering, policy, or research. For education, this is a massive waste. It parallels training agronomy students, only to have them pick cotton indefinitely for export.
There is evidence that a different model can work. In India, Karya provides data services that pay workers around $5 per hour—nearly 20 times the minimum wage—and offers royalties when datasets are resold. Karya reports positive impacts for over 100,000 workers and aims to reach millions more. Projects like this demonstrate that data work can connect to income mobility and skill development instead of relying on exploitative contracts. Simultaneously, the market for data annotation tools is growing rapidly, from about $1 billion in 2023 to a projected $5.3 billion by 2030, while the global generative AI market could increase eightfold by 2034. The funding exists. The challenge is whether educational institutions will ensure that data work linked to their programs is structured as meaningful employment focused on learning rather than as digital labor.
Education systems have more power over AI data colonialism than they realize. Universities, vocational colleges, and online education platforms already collaborate with tech companies for internships, research grants, and practical projects. Many inadvertently direct students into annotation, evaluation, and safety tasks framed as valuable experience, often lacking transparency about working conditions. Instead of viewing these roles as a cost-effective way to demonstrate industry relevance, institutions can turn the model on its head. Any student or graduate engaged in data work through an educational program deserves clear rights: a living wage in their local context, limits on exposure to harmful content, mental health support, and a guaranteed path to more skilled roles. This work should count toward formal credit and recognized qualifications, not just as “extra practice.”
Some institutions in the Global South are already heading in this direction through AI equity labs, community data trusts, and cooperative platforms. They design projects where local workers create datasets in their languages and based on their priorities—such as agricultural support, local climate risks, or indigenous knowledge—while gaining technical and governance skills. A cooperative approach, as advocated by proponents of platform cooperativism in Africa, would allow workers to share in the surplus generated by the data they produce. For education providers, this means shifting the concept of “industry partnership” away from outsourcing student moderation and toward building locally owned shared AI infrastructure. That change transforms the dynamic: instead of training students to serve distant platforms, institutions can empower them to become stewards of their own data economies.
Policymakers often claim that data-intensive jobs are short-term, since automation will soon eliminate the need for human labelers. However, current investment trends indicate otherwise. Corporate AI spending hit $252.3 billion in 2024, with private investment in generative AI reaching $33.9 billion—more than eight times the levels of 2022. The generative AI market, worth $21.3 billion in 2024, is expected to grow to around $177 billion by 2034. The demand for high-quality, culturally relevant data is rising alongside it. Annotation tools and services are also on a similar growth trajectory. Even if some tasks become automated, human labor will remain essential for safety, evaluation, and localization. Assuming that AI data colonialism will disappear on its own simply allows the current model another decade to solidify.
Regulators are beginning to respond, but education policy has not kept up. New frameworks, such as the EU Platform Work Directive, create a presumption of employment for many platform workers and require algorithmic transparency. A global trade union alliance for content moderators has established international safety protocols, advocating for living wages, capped exposure to graphic content, and strong mental health protections. Lawsuits from Kenya to Ghana demonstrate that courts recognize psychological injury as genuine workplace harm. Education ministries and accreditation bodies can build on this momentum by requiring any AI vendor used in classrooms or campuses to disclose how its training data was labeled, the conditions under which it was done, and whether protections for workers meet emerging global standards. Public funding and partnerships should then be linked to clear labor criteria, similar to requirements many already impose for environmental or data privacy compliance.
For educators and administrators, tackling AI data colonialism also means incorporating the labor that supports AI into the curriculum. Students studying computer science, education, business, and social sciences should learn how data supply chains operate, who handles the labeling, and what conditions those workers face. Case studies can showcase both abusive practices and fairer models, like Karya’s earn–learn–grow pathways. Teacher training programs should address how to discuss AI tools with students honestly—not as magical solutions, but as systems built on hidden human labor. This transparency prepares graduates to design, procure, and regulate AI with labor justice in focus. It also empowers students in the Global South to see themselves not just as workers in the digital realm, but as future builders of alternative infrastructures.
The plantation metaphor is uncomfortable, but it effectively highlights the scale of imbalance. On one side, hundreds of millions of workers, often young and well-educated, perform repetitive, psychologically demanding tasks for just a few dollars an hour and without real career advancement. On the other side, a small group of firms and elite institutions accumulate multibillion-dollar valuations and control the direction of AI. This represents the essence of AI data colonialism. Education systems can either normalize it by quietly channeling students into digital piecework and celebrating any AI partnership as progress, or they can challenge it. This requires implementing strict labor standards, investing in cooperative and locally governed data projects, and teaching students to view themselves as rights-bearing professionals rather than anonymous annotators. If we fail, the history of AI will resemble a new chapter of familiar exploitation. If we succeed, education can transform today's digital plantations into tomorrow's laboratories for a fairer, decolonized AI economy.
The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.
Abedin, E. (2025). Content moderation is a new factory floor of exploitation – labour protections must catch up. Institute for Human Rights and Business.
Du, M., & Okolo, C. T. (2025). Reimagining the future of data and AI labor in the Global South. Brookings Institution.
Grand View Research. (2025). Data annotation tools market size, share & trends report, 2024–2030.
Okolo, C. T., & Tano, M. (2024). Moving toward truly responsible AI development in the global AI market. Brookings Institution.
Reset.org (L. O’Sullivan). (2025). “Magic” AI is exploiting data labour in the Global South – but resistance is happening.
Startup Booted. (2025). Scale AI valuation explained: From startup to $29B giant.
Stanford HAI. (2025). 2025 AI Index report: Economy chapter.
TIME. (2023). Exclusive: OpenAI used Kenyan workers on less than $2 per hour to make ChatGPT less toxic.
UNI Global Union / TIME. (2025). Exclusive: New global safety standards aim to protect AI’s most traumatized workers.
Karya. (2023–2024). Karya impact overview and worker compensation.
EU digital regulation simplification will shape Europe’s tech power Fragmented rules now act as an uncertainty tax on digital innovators The article urges clear, one-stop rules that turn high standards into an advantage.
The article expands actuarial neutrality from pensions to the whole public budget It shows how early retirement weakens tax revenues, pension capital, and debt sustainability It urges linking retirement-age rules to public-finance neutrality while protecting vulnerable workers
