A roadmap for building BC’s AI public policy framework

Introduction

Employers and governments are moving quickly to adopt so-called artificial intelligence in workplaces and social services, despite the many known risks the use of these systems pose to people and the environment.

While governments and employers are accepting technology companies’ projection of AI’s inevitable integration into our daily lives, this software’s future is actually heavily contested. Behind the illusion of inevitability, the software marketed as AI is an intentionally designed technology and the choices made in its design telegraph the values and goals of those who sell it: the concentration of wealth for a few by stealing from the many. 

Employers are using AI-driven automation to deskill and displace workers and “gigify” work. AI software is built on the mass appropriation of human creative and intellectual work without consent or compensation. Our personal and professional data are scraped, sold and fed into systems that profit from our stolen labour, likeness and social lives. Rather than being able to automate our labour, that work is fundamentally transformed into a more “scalable” task, where the intrinsic value humans bring to their work is deemed worth sacrificing for corporate profit.

“Because AI is sold as more efficient, more objective and less biased than the people it replaces, it is important to ask who is profiting…”

The environmental costs are staggering, as data centres consume enormous amounts of energy and drinking water, while the extraction of rare minerals devastates communities in the Global South.

And AI systems frequently replicate and deepen structural inequalities, embedding racial, gender and economic biases in work, governance and public services.

So why the rush to invest in so-called AI?

Because AI is sold as more efficient, more objective and less biased than the people it replaces, it is important to ask who is profiting from AI and at whose expense. This roadmap examines:

• how AI software actually works and why that matters, who is pushing for rapid AI adoption and what they stand to gain,
• why the economic promise of AI is not holding up,
• what harms the technology is causing,
• where Canada stands in the growing AI regulatory gap, and 
• what a community-driven approach to AI policy might look like in response. 

British Columbia is at a particularly urgent inflection point. Neither the province nor the federal government has AI-specific legislation in force. Federal and provincial AI ministers have mandates focused on commercialization, not rights protection. And new federal bills are expanding surveillance and policing powers—which increasingly rely on AI-driven facial recognition, communications scraping and risk scoring—with no coherent public framework to govern them.

We must design policy not based on the lofty promises of what AI is meant to be, nor under the coercive threat of its inevitability, but rather based on its actual design, deployment and material impact. We need an approach to AI that ensures decisions about the future of this technology are shaped by those most affected, not by corporate interests alone.

Defining AI

This section explains how the technology marketed as “artificial intelligence” works, where the term came from and how it actually depends on hidden human labour.

“Artificial intelligence” was a marketing term coined by academics to draw more funding to a field of computer science established in the 1950s that promised to mimic human cognition with automation, or what they called the science of making an intelligent machine. 

But there is no clear consensus on the boundaries of what counts as intelligence in humans, let alone in computers. If anything, there is a fairly large consensus that no technology we currently have is truly “intelligent” in any human sense—with cognitive abilities to learn, understand and apply knowledge across different domains—despite all the anthropomorphic language we use. 

This science-fiction AI is what the tech industry calls “artificial general intelligence” (AGI). Though the AI tools we encounter today, like ChatGPT, Gemini or Claude, can seem to understand us, learn and teach, these abilities are entirely illusions. 

These tools fall under a narrower category of AI: ones that can respond to instructions with a degree of autonomy to achieve specific tasks within a pre-programmed domain.

How the machine “thinks”

So how do these software systems work? How are they able to do such a convincing job (though not always) of mimicking human cognition without being intelligent at all?

The recipe has three key ingredients: algorithms, data and human labour.

Algorithms are the instructions that tell computers what to do. Although we’ve been solving mathematical problems with step-by-step procedures since antiquity, the word “algorithm” itself comes from a Latinized version of the ninth-century mathematician al-Khwarizmi, who developed a more systematic way of solving mathematical problems through a finite sequence of logical steps. Ada Lovelace is widely credited for writing the first algorithmic instructions for a computing machine in the 19th century. Essentially, all computer programming is a set of algorithms.

Now, because so much of our activity is digitized, we generate huge and complex sets of data with every click on your computer and phone, every purchase, every medical visit, every GPS-tracked movement. Many of the tools we loosely call AI are essentially software systems, which developers claim can make better sense of that data than a human could and can act on that data through automated decision-making.

Machine learning, deep learning and generative AI are ever more complex forms of what is fundamentally statistical correlation. Over millions of iterations, these systems associate certain inputs with certain outputs—which pixels tend to correspond to a face, which words tend to follow others in a sentence. In other words, they operate through pattern-matching. 

So machine learning does not “learn” in the human sense. Despite biological metaphors comparing it to a human brain, an artificial neural network is a statistical computation software wherein successive layers of calculations assign probabilistic weights to data without any inherent understanding of its meaning. 

“…what looks like creativity or intelligence in generative AI is really the predictive recombination of prior human work, stolen at scale without consent or compensation.”

And although they are called large language models (LLMs), because computers process numbers not words, a software like ChatGPT first breaks text into “tokens”—chunks of characters that might be a whole word, a word fragment or a single punctuation mark—and assigns each token a number. It then maps correlations between tokens to predict the most likely next token of characters.

Critical scholars refer to these technologies as “stochastic parrots” because they stitch together sequences of linguistic forms entirely without grasping the actual meaning of the text they generate. Others call them “synthetic text extruders,” or even simply “plagiarism machines.” 

So what looks like creativity or intelligence in generative AI is really the predictive recombination of prior human work, stolen at scale without consent or compensation. And human labour is required at every step of the process.

Artificial artificial intelligence

Machine learning is sold as the codifying of human knowledge and the automation of learning, but what it really aims to do is codify and automate human labour.

The underlying data these systems are trained on is drawn from us—from our digitized social lives, art, literature, music and research and from our online conversations. And rendering that vast material intelligible to a machine has required an immense amount of human labour.

These systems rely on what some researchers call “ghost work.” A global army of underpaid workers has had to minutely label every bit of sound, every image, every text, give feedback and correct errors so that machine learning algorithms could classify what an ear or an eye looks like, what sounds correspond to which words or what a verb does in a sentence.

Tech companies outsource these tasks to a geographically dispersed, “hyperflexible precariat” of workers usually hired through crowdwork platforms, or by intermediaries as independent contractors, and paid by the click, without unions, benefits or job security. Jeff Bezos has described the service these workers provide through platforms like Amazon’s Mechanical Turk as “artificial artificial intelligence.” The work is not only precarious but also traumatic. To make generative AI safe for public use, models must be taught what is unacceptable. This requires workers to act as guardrails, repeatedly viewing and filtering out toxic, violent and abusive content, subjecting them to severe psychological harm.

“This illusion masks the reality that the technology is wholly dependent on the hidden, exploited and uncompensated labour of millions of people.” 

By combining these elements—stealing human creative work through mass copyright infringement, stealing personal data through mass privacy violations, presenting statistical guesswork as intelligent reasoning and describing the system in anthropomorphic terms—the AI industry creates what one scholar calls a “Potemkin village of technological progress and knowledge objectivity.” This illusion masks the reality that the technology is wholly dependent on the hidden, exploited and uncompensated labour of millions of people. 

Looking at this process as a whole—data extracted from human activity, organized by human labour, transformed by human-written algorithms into statistical predictions and deployed back into society in ways that shape future behaviour—helps us see AI not as an autonomous intelligence but as an industrial process that turns human experience into a commercial resource.

Bullshit and bias

This section explains how the methodological design of artificial intelligence software reproduces bias and how AI software developers evade accountability with anthropomorphizing language.

The underlying logic of AI software directly produces and reproduces power dynamics, harms and biases in several compounding ways.

In order to calculate the statistical correlations between inputs and outputs, these software compress complex social realities into numerical distances. Because these systems are developed to manipulate the form of language without any access to its meaning, they are fundamentally disconnected from reality. And because the systems are designed to optimize for singular, measurable objectives, the social complexities and human values that resist measurement can never enter into the calculation.

“When [AIs] generate falsehoods, they are not “hallucinating” but producing what some argue can be called ‘bullshit’ in the philosophical sense: words produced with absolute structural indifference to whether they are true or false.”

The tech industry frequently labels incorrect AI outputs as “hallucinations”—an anthropomorphizing term that implies the machine has a mind trying to perceive the world and makes a mistake. But software can neither “perceive” nor “care” about truth. When they generate falsehoods, they are not “hallucinating” but producing what some argue can be called “bullshit” in the philosophical sense: words produced with absolute structural indifference to whether they are true or false. They are, in other words, bullshit machines designed to generate truthiness. Their programmed goal is to produce sequences of words that look convincing and human-like—with many systems optimized toward “affirming users’ moral and interpersonal positions even when those stances are widely judged as harmful or unethical.” 

But structural indifference is a design, and responsibility for that design sits with the companies that chose it. Had they wanted a software concerned with truth, these companies would have built a different one—with access to a discrete set of information that can simply be recalled. Their methodological choices instead serve their bottom line. False promises of mechanical objectivity sell the software as neutral, while the language of “thinking” machines manufactures public trust, obscures how the technology actually works and insulates these companies from accountability and oversight.

Whose intelligence gets encoded?

The very definition of intelligence has long been weaponized as a tool of exclusion and power. Enlightenment thinkers established a fiction of intelligence based on abstract, disembodied, rule-bound reasoning—a framework that elevated educated, white European men as the standard of rationality, while characterizing women, enslaved people and colonized subjects as irrational or physiologically incapable of mental discipline. Early AI researchers built their systems on this exact fiction, viewing the human mind as a “persistent and unimaginative beast who can follow rules blindly,” and concluding that machines executing algorithms therefore possessed human-like intelligence. The narrow, disembodied conception of intelligence that once justified colonial hierarchies is now embedded in the systems being deployed to sort, classify and make decisions about people’s lives.

Critical scholars have long rejected this definition of intelligence. True human reasoning, they argue, cannot be reduced to universal logic: it is embodied, contingent and shaped by class, race and personal experience.

Data is never neutral

Whether in the sciences, in governance or in AI systems, data is never neutral. The process of turning social life into data is itself reductive: it strips away context, reduces human experience to quantifiable fragments and carries the imprint of the inequalities from which it was collected.

Because these systems work by brute-force pattern-matching, they require unfathomable amounts of training data. To satisfy that demand, AI developers indiscriminately scrape the internet. But internet data is not an objective representation of humanity: it structurally overrepresents younger users from developed, Western countries and environments like Reddit or YouTube that are dominated by white, male demographics. The training data encodes the hegemonic worldview—white supremacist, misogynistic and ableist ideologies—and when developers compress that data into a statistical model, it amplifies those biases, cementing a “dictatorship of the past” in which old taxonomies and prejudices are endlessly regenerated.

Because this is a science built on correlation, not causation, more data does not necessarily mean more truth. As datasets grow, they accumulate more noise: misleading associations and spurious connections, taking incidental patterns in the data and using them to make automated decisions. Financial institutions, for example, make lending and insurance decisions using digital traces like social media activity or purchasing patterns that were never meant to measure creditworthiness, because “spurious correlation is fine, so long as it is profitable spurious correlation.” 

“The training data encodes the hegemonic worldview—white supremacist, misogynistic and ableist ideologies—and when developers compress that data into a statistical model, it amplifies those biases…”

These models project historical conditions into the future, and the institutions acting on their outputs lock marginalized groups into cycles of disadvantage.

Under this logic of correlation, there are consequences both to being too visible in data and to being invisible within it. Hypervisibility produces feedback loops. In so-called predictive policing, for example, algorithms trained on racially biased arrest records label certain neighbourhoods or individuals as “high risk.” Police use that label to justify increased patrols and arrests in those same areas, which reinforces the original bias. The algorithm is not predicting crime, it is automating the systemic discrimination that generated the historical data in the first place, deepening racial criminalization rather than preventing it.

Invisibility in the data can produce a different kind of harm. For example, in healthcare the underrepresentation of women—especially Indigenous, Black and racialized women—in medical research and diagnostic datasets means that AI diagnostic tools are likewise more likely to fail to match their symptoms to a diagnosis. The absence of these bodies and experiences from datasets translates directly into worse health outcomes.

The same logic of algorithmic sorting operates well beyond policing and health; it extends across welfare distribution, military targeting and other systems marketed as neutral while reproducing old hierarchies. Researchers have described the systemic encoding of racial and class hierarchies into ostensibly neutral technical systems as “automated inequality” or the “New Jim Code.”

The AI bubble

This section covers why a technology that loses money and falls far short of what its promoters claim keeps attracting billions in investments anyway.

Despite all the hype, AI is not performing commercially as well as its promoters claim. One widely publicized MIT study argues that about 95% of AI pilot projects fail to become profitable. Even as governments and investors pour billions into the field, the flagship AI companies are themselves struggling to make profits. 

Because these are private companies, assessing the financial reality of the largest AI software developers is difficult. And as these companies maneuver towards public offerings, they are filling the media with curated financial narratives that make their true profitability even more unclear. Nonetheless, we know that OpenAI recorded a staggering US$8 billion in losses in 2025, up from their $5 billion loss in 2024—though some calculations put that loss as high as US$38 billion. HSBC analysts project OpenAI will remain unprofitable through 2030 due to $1.4 trillion in committed infrastructure costs over the next eight years. Anthropic, the company behind Claude, reported losses of US$5.3 billion in 2024 with no break-even expected before 2028. 

These shortfalls are for a number of interconnected reasons.

Building AI systems requires immense resources

Building, training and running AI systems is materially, and consequently environmentally, incredibly costly—and the costs are growing faster than the technology’s returns. 

Most AI tools rely on cloud infrastructure rather than local processing, which means that the computing power needed to run AI software does not happen on a customer’s device but instead at a data centre: a sprawling facility equipped with servers operating around the clock. By the end of 2025, the five largest US cloud companies—Amazon, Meta, Alphabet, Microsoft and Oracle—had signed US$969 billion in future data centre leases, with plans to spend another $635-690 billion building more of them in 2026. Although it is difficult to calculate the exact amounts of energy, minerals and water required to run these data centres, the scale is also difficult to overstate. 

A 2019 benchmark study found that training a single natural language processing model can emit more than 600,000 pounds (270,000 kg) of carbon dioxide emissions. According to IBM, “large facilities can require between 100 and 300 megawatts of electricity to operate – enough to power hundreds of thousands of homes.”

“the AI industry is projected to withdraw between 4.2 and 6.6 billion cubic metres of drinking water by 2027—more than the 2.7 billion cubic metres used annually by all Canadian households combined.”

Because the servers running these models generate enormous heat, data centres require vast amounts of potable water to keep them cool. As Al Jazeera reports, “a 100-megawatt facility can consume about 2.5 billion litres of water a year – equivalent to the annual needs of 80,000 people.” Researchers estimate that generating a single 100-word response from a chatbot can use the equivalent of a bottle of drinking water. As the demand for data centres grows, the AI industry is projected to withdraw between 4.2 and 6.6 billion cubic metres of drinking water by 2027—more than the 2.7 billion cubic metres used annually by all Canadian households combined. Even with all this water for cooling, “researchers found that land surface temperatures around data centres rise by an average of 2C with some areas recording increases as high as 9C.”

The servers, processing units, batteries and cables that power AI software also require rare-earth elements and minerals like lithium and cobalt, with extraction processes linked to widespread environmental contamination and alarmingly high rates of congenital disorders in the communities living closest to the mines.

And the returns on all this expenditure are shrinking. Improving translation accuracy by a tenth of a point in one study required an additional US$150,000 in processing power, plus the associated emissions. Researchers now argue that the core strategy driving AI software’s rapid gains—training ever-larger models on ever-larger datasets—has hit a ceiling that more spending cannot break through. A few years ago, doubling the size of a model from 10 to 20 billion parameters produced a 10-15% jump in performance. Today, doubling from 100 to 200 billion parameters delivers only a 1-2% improvement, at vastly greater cost.

The technology does not work as advertised

Running AI systems requires constant updates, maintenance and human supervision. AI tools remain clumsy at basic tasks—from writing coherent prose to handling customer service interactions. And AI software frequently make mistakes: model “hallucinations” and logic errors are common, while limited context windows mean outputs frequently omit crucial information. In the rush to deploy AI, we are likely coding serious errors into essential services—not to mention publishing and circulating AI-generated misinformation at scale. 

While AI software outputs can appear reasoned and well-sourced, commentators are pointing to the Gell-Mann amnesia effect to describe our tendency to easily spot errors in AI-generated materials on a subject where we have expertise, yet nonetheless confidently trust its outputs on subjects we know little about.

Tech news website CNET was caught publishing AI-generated financial articles that included elementary mathematics mistakes, plagiarism and factually misleading financial advice. A public database has catalogued over 1,000 cases in the US alone of lawyers filing AI-generated briefs that cited entirely fabricated case precedents. In Canada, BC’s Civil Resolution Tribunal found Air Canada liable for the incorrect information its AI chatbot gave a customer about the airline’s bereavement discount. 

“commentators are pointing to the Gell-Mann amnesia effect to describe our tendency to easily spot errors in AI-generated materials on a subject where we have expertise, yet nonetheless confidently trust its outputs on subjects we know little about.”

Ontario’s Auditor General found that AI transcription tools used by doctors to record conversations with their patients were prone to errors and inaccuracies. In Australia, an AI medical scribe fabricated a claim that a patient was taking illegal drugs and an AI diagnostic tool recorded the wrong breast for a cancer diagnosis and falsely attributed epilepsy to a patient who did not have it. After AI was added to a surgical navigation device in the US, the FDA received reports of at least 100 malfunctions of the system misinforming surgeons about the location of their instruments inside patients’ bodies, with at least 10 patients injured between 2021 and 2025. 

A recent BBC study found that 45% of AI-generated responses “misrepresented news content.” Though there are competing ways of measuring “hallucinations,” benchmarking tools consistently find AI software to make mistakes even on closed-domain questions such as summarizing or answering questions about a given text. For open-ended questions, the rate of fabricated or incorrect responses is significantly higher. 

AI tools have been so ineffective that some companies that claimed to be using AI were actually using humans all along. Amazon’s Just Walk Out technology, for example, was marketed as a revolutionary cashierless shopping experience, using computer vision and AI software to automatically detect what customers picked up and bill them. In reality, it relied on roughly 1,000 workers in India watching the cameras, because the company could not get the AI to work. In the US, the Securities and Exchange Commission has started issuing fines for “AI washing,” the practice of making false claims about AI-labelled technology.

Other companies that were once vocal boosters of AI have quietly rehired a human labour force. Klarna went from claiming a chatbot could do the work of 700 employees to hiring humans back after service quality collapsed and customers complained about being trapped in Kafkaesque loops. Recent polls are finding that more and more employers who replaced workers with AI regret the decision and hire workers back.

However—and this is a major caveat—many of these workers are being hired back to assist AI, editing, correcting and reformatting its outputs behind the scenes. So what we’re seeing is that instead of replacing human labour outright, employers are using AI to reorganize work into more repetitive, underpaid, precarious work.

With billions in investment riding on the promise that AI systems can replace human labour, the illusion of autonomous, job-replacing AI is itself a product—one that inflates valuations, attracts subsidies and justifies the enormous capital expenditure being poured into the industry. The tech industry’s use of anthropomorphizing language—calling an AI system a tutor, assistant or coworker—deliberately telegraphs an intention to replace the human workers in those roles, whether or not the technology is capable of doing so. 

“Employers are also closing the door on entry-level work as they put senior professionals to work quality-controlling AI outputs instead of training junior staff.”

Although a significant gap remains between what AI software can do and what tech companies promise, employers are nonetheless deploying these tools to replace workers wherever they can. Workers in fields such as software engineering, journalism, editing, translation, closed captioning, animation, administration and customer support work have already been affected. Employers are also closing the door on entry-level work as they put senior professionals to work quality-controlling AI outputs instead of training junior staff. Automation cannot replicate what these workers once did, but employers can use it to restructure the process of production for scale and profit. Employers seem to be willing to accept a loss in quality for a gain in their bottom line.

The political economy of AI

This section investigates the political economy of AI: whose interests the technologies marketed as artificial intelligence actually serve. From fighting against workers’ power, to monopolizing digital infrastructure, to expanding invasive surveillance, these software services are concentrating power and control into the hands of a few corporations.

If the technology is expensive, unreliable and dependent on hidden human labour, why are governments, social services and businesses racing to adopt it? The answer is, in part, because AI’s value to capital is less about efficiency than it is about monopoly and power.

The disciplinary impact on labour

The threat of automation has always been used to reduce bargaining power—from the automated looms of the 18th century to the factory floors of the 20th, the illusion of machine autonomy has served to discipline labour. When workers are told they are replaceable by machines, they are more likely to accept worse working conditions, longer hours and lower pay. 

Microsoft, for example, announced major layoffs under the banner of “AI transformation” while simultaneously hiring new workers through the H-1B visa program—essentially replacing stable jobs with cheaper, more compliant labour. 

The modern illusion of artificial intelligence relies on the same trick as its industrial predecessors: fetishizing software as an autonomous, creative “mind” while hiding “the body” of the machine—the exploitative global supply chains, the ecological devastation and the armies of underpaid workers who constantly correct, label and sustain the algorithms. 

“The modern illusion of artificial intelligence relies on the same trick as its industrial predecessors: fetishizing software as an autonomous, creative “mind” while hiding “the body” of the machine.”

And this separation enables a very specific form of workplace control. Even when AI does not work well, it still works as discipline. Employers can use it to set new standards for what a single worker is expected to produce, especially if you ignore quality and focus on quantity. The automation of managerial functions—or algorithmic management—fragments, standardizes and quantifies every aspect of the labour process. Employers deploy an array of tracking tools—from “bossware” that tracks keyboard and screen activity to physical sensors and wearable technologies that monitor a worker’s movements, interactions and even vital signs.

This pervasive surveillance generates a massive information asymmetry: management possesses granular, real-time data on every worker, while workers are kept in the dark about how this data is collected, weighted or used. As algorithms take over scheduling and task assignment, worker autonomy is severely reduced. Workers are subjected to optimized, “just-in-time” scheduling that creates highly unpredictable and unsustainable working patterns. Stripped of their professional discretion, employees often feel compelled to optimize their behaviour to satisfy the algorithm’s metrics, and in some cases this pressure incentivizes workers to disregard safety rules, bypass professional ethics and work in increasingly dangerous ways. Forcing metrics as “performance indicators” can also change the nature of the work altogether, such as a media organization’s target for “clicks” changing how journalists make decisions about what news to cover and how to cover it.

“AI is not rendering human labour obsolete. It is functioning as a tool to discipline labour, extract maximum surplus value and extend managerial control.”

At the same time, every time workers use these tools, they generate the data that trains AI software to immitate their work. The more we use these tools in our workplaces, the more data we hand over to the companies deskilling and displacing us. We end up in an inverse situation: instead of these being tools to assist us at our jobs, our jobs become to assist these tools. Author Cory Doctorow has dubbed this the reverse centaur. And in many fields, there are deep concerns that complex, reflective expertise will be reduced to mere data processing and form-filling.

AI is not rendering human labour obsolete. It is functioning as a tool to discipline labour, extract maximum surplus value and extend managerial control—locking the global workforce into a race to the bottom on pay and working conditions.

Control over infrastructure and data

On the surface, we are watching a wave of start-ups and companies experiment with what AI can do, making it look like a robust and competitive sector. But underneath is a decades-long process of vertical integration. A handful of tech giants—Meta, Google, Microsoft, Amazon—now own everything from the deep-sea cables to the data centres, the apps and platforms we use, the AI models, the cloud computing power, and of course all the data that flows through this entire stack. They are even invested in the energy sector, because this technology is such a drain on power. The AI hype, in turn, is helping these companies expand their infrastructure at minimal cost, thanks to public subsidies and the AI (fool’s) gold rush. 

In other words, AI is an economy of scale: the production of AI requires such vast financial resources to secure specialized talent, proprietary datasets and massive computational power that it restricts the market to a few mega-corporations. They have been allowed to entrench a rentier economic model where all other businesses and social systems—from healthcare to police departments to AI start-ups—must pay these gatekeepers for access to computing power, data storage and the “foundational” AI models they control. 

Some scholars argue this marks a shift toward “technofeudalism”: rather than profiting from traditional production, tech giants act as digital landlords, extracting monopoly rents by enclosing digital spaces, controlling access to data-generating environments and transforming human social interactions into capitalized assets. As these companies cement this lock-in, they are beginning to raise the price on access to their services and infrastructure, creating what some have dubbed a “subprime AI crisis” or more plainly, “enshittification.” So, while AI pilots may rise and fall, the Big Tech oligopoly makes money either way.

The result is an unprecedented concentration of economic power that threatens democracy.

Surveillance, policing and war

While advertising revenue built the initial tech monopolies, the most profitable and strategically significant revenue streams for platform giants have shifted decisively toward surveillance, policing and military contracts. 

The AI economy is now deeply integrated with state violence, geopolitics and a highly lucrative defence industry, powering an expansive, unavoidable surveillance dragnet. Companies like Amazon, Microsoft and Google operate a multi-layered profit model: they collect data on billions of users through consumer platforms, provide the cloud infrastructure that processes surveillance operations and develop specialized AI tools for security agencies and militaries. Companies like Palantir function as a critical integration layer that runs on this infrastructure to unify fragmented government databases with data from tech platforms, creating a symbiotic ecosystem where all parties profit from the same surveillance apparatus. They then sell these infrastructures and AI tools to the Israel Defense Forces (IDF), the Pentagon, US Immigration and Customs Enforcement (ICE), the Canadian Department of National Defence, the RCMP and the Ontario Provincial Police, and these groups in turn field test and refine the technology.

In other words, AI infrastructure is deliberately dual-use. Tech companies use the exact same cloud systems and machine-learning models for both civilian and military purposes, allowing them to monetize both spheres simultaneously. 

At the state level, neoliberal governments are not exactly dupes of AI and predictive analytics—they are active participants with their own incentives to automate social control while wrapping these actions in the “objective” aura of mathematics. By automating welfare distribution, policing and even incarceration, states can replace democratic imperatives with technologically administered, punitive systems that target the poor and marginalized. AI surveillance tools are deployed by agencies like ICE, Canada Border Services Agency (CBSA), Immigration, Refugees and Citizenship Canada (IRCC) and European border forces to automate the tracking, profiling and exclusion of migrants and refugees. In occupied Palestine, Israel uses AI-enabled facial recognition databases at military checkpoints to monitor, segregate and dictate the movement of every Palestinian.

“neoliberal governments are not exactly dupes of AI and predictive analytics—they are active participants with their own incentives to automate social control while wrapping these actions in the “objective” aura of mathematics.”

Militaries and tech start-ups use wars, occupations and even genocide to test, prototype and refine AI drones, autonomous weapons and biometric surveillance on human populations in real time. The Israeli military has deployed AI systems known as The Gospel, Lavender and Where’s Daddy? in its genocide of Palestinians. These systems use machine learning to analyze cellular data, social media and facial recognition surveillance to automatically generate “kill lists,” producing targets at a speed vastly exceeding human capability. Then they use the purported sophistication of “smart” bombs to mask the use of imprecise “dumb” bombs in order to maximize civilian casualties while displacing moral culpability of these atrocities onto machines (or AI washing).

This violence then boomerangs. Companies market their technologies as “battle-tested” and export them globally. US-developed tools like ShotSpotter—originally deployed in Black American neighbourhoods—are sold to the Israeli military, and Israeli firms like NSO Group and Corsight sell their AI tools to police forces, border patrols and authoritarian regimes worldwide. Former prime minister Stephen Harper now leads Awz Ventures—whose advisory board includes former heads of CSIS, the CIA, MI5 and Mossad—an investment firm “devoted to launching security companies in Israel,” including Corsight AI, which has been used to mass-surveil and flag Palestinians for detainment and torture by the IDF.

The profits are enormous. Big Tech corporations and Silicon Valley start-ups compete fiercely for billions of dollars in government defence and policing contracts. Google and Amazon signed the US$1.2 billion Project Nimbus contract to provide cloud computing and AI services to the Israeli government and military. Microsoft secured a US$480 million contract to provide augmented-reality combat headsets to the US Army. To streamline this pipeline, the Pentagon created the Defence Innovation Unit to act as an internal venture capital fund, funnelling taxpayer money directly into tech start-ups. Venture capital firms like Peter Thiel’s Founders Fund and defence start-ups like Anduril and Palantir (also co-founded by Thiel) profit by aggressively lobbying the government and fuelling geopolitical paranoia. Tech executives actively push a narrative of an existential AI arms race to convince the state to pour unchecked billions into their technologies. Through this fearmongering, the Pentagon acts as a “speculative investor of last resort” for tech capital.

The result is that the Big Tech oligopoly now has the technical capacity and legal impunity to monitor, analyze and weaponize every dimension of human existence in the service of capital accumulation, state violence and social control.

AI as digital imperialism/colonialism

The political economy of AI is not a radical break from the past. It is a direct continuation and intensification of historical colonialism and imperialism—an ongoing global system of domination that extracts value from the global majority to enrich a concentrated minority—through the exploitation of workers, the extraction of natural resources and the deployment of surveillance and weapons on marginalized populations, all in service to the concentration of wealth in a handful of corporations.

Just as Europeans used the legal fiction of terra nullius—nobody’s land—to seize Indigenous territories, tech companies further treat human social life and cultural production as animo nullius—no person’s mind. They treat global data as an unclaimed, raw natural resource that is just there to be harvested without compensation or consent, mirroring the predatory extractive practices of historical colonialism. Through unregulated “knowledge extractivism,” corporations indiscriminately scrape datasets containing copyrighted texts, personal photos and cultural artifacts. By framing their algorithms as acts of “novel transformation or discovery,” these companies erase the diverse cultures, histories and labourers who actually produced the knowledge that AI mimics.

“Technology empires” then exercise imperial power through their monopoly control over the three core pillars of the digital ecosystem: software, hardware and network connectivity.

“Technology empires” then exercise imperial power through their monopoly control over the three core pillars of the digital ecosystem: software, hardware and network connectivity. By controlling the semiconductor supply chain and imposing export controls on advanced chips, dominant states and companies engage in “weaponized interdependence,” ensuring that smaller nations, companies and communities remain technologically dependent consumers rather than producers. Because AI models are trained predominantly on Western, English-speaking internet data, they actively exclude and devalue Indigenous and non-Western knowledge systems. This enforces a form of cognitive imperialism, embedding the cultural norms, racial biases and epistemologies of the Global North into algorithms while erasing the diverse realities of the majority of the world.

The colonial supply chain demands not only the cognitive exploitation of ghost workers paid poverty wages to train algorithms, but the physical devastation of the regions where the minerals are mined and the waste is dumped, turning them into “environmental sacrifice zones.” The benefits of AI accrue largely to corporations and privileged consumers in the Global North, while the devastating consequences—toxic pollution, droughts, the mining of cobalt and lithium, the dumping of e-waste—are disproportionately suffered by marginalized populations in the Global South. This is environmental racism operating at a global scale.

The regulatory vacuum in Canada

This section explores the gap between the more developed AI regulations of the European Union and the relative regulatory vacuum in Canada, and what the rhetoric of BC’s and Canada’s AI ministers tells us about how hard we’ll have to fight for that gap to close.

The European Union has moved AI policy forward on multiple fronts 

The European Union’s (EU) AI Act regulates AI systems by four risk categories. Applications posing an “unacceptable risk”—considered a clear threat to the safety, livelihoods and rights of people—are completely prohibited. This includes indiscriminate scraping of the internet or CCTV for facial recognition databases, emotion recognition in workplaces or educational institutions, social scoring or software that generates non-consensual sexually explicit content. Some “high risk” AI software that can negatively affect safety or fundamental rights are allowed but face strict obligations (like mandatory risk assessments, human oversight and high-quality data requirements) before they hit the market. These include AI used in critical infrastructure (like transport), hiring and managing workers (e.g., CV-sorting software), education (e.g., automated exam scoring), determining access to essential services (e.g., credit scoring for loans) and law enforcement, migration and border control. AI with “transparency risk” covers systems where users might be tricked or confused about whether they are dealing with a human or real content such as chatbots or generative AI. AI software deemed “minimal risk” remains unregulated, although the Act imposes specific transparency, copyright and risk-mitigation requirements on general-purpose AI models, the foundation models that power many downstream AI applications.

The EU’s AI Act includes a tiered penalty structure issued directly by the European Commission of up to €35 million or 7% of global annual turnover (whichever is higher) for deploying or developing any of the banned, “unacceptable risk” AI systems; up to €15 million or 3% of turnover for failing to comply with mandatory obligations for high-risk AI systems (such as missing risk assessments, poor data governance, inadequate technical documentation or lack of human oversight); up to €7.5 million or 1% of turnover for misleading authorities, such as by supplying incorrect, incomplete or misleading information; and up to €15 million or 3% of global annual turnover for general-purpose AI model providers who violate transparency, copyright or systemic-risk mitigation rules, fail to cooperate with audits or refuse the European Commission access to evaluate their models. 

In addition to the EU’s AI Act, the General Data Protection Regulation (GDPR) ensures that AI models cannot be indiscriminately trained on EU citizens’ personal data without proper consent or legal basis, and it mandates that data processing not subject users to solely automated decision-making. It also grants individuals the right to access their data, correct it or have it deleted.

The Digital Markets Act is an antitrust and competition law designed to stop large tech monopolies from maintaining unfair control over digital markets. “Gatekeepers,” like Google, Apple, Meta and Amazon, that provide “core platform services” (search engines, app stores, social networks, messaging services) cannot rank their own products higher than competitors’ products on their platforms, must allow users to easily uninstall pre-loaded software and must allow developers to use alternative in-app payment systems.

The Digital Services Act enforces a tiered platform accountability for harmful content and algorithmic transparency, wherein the larger the platform, the stricter the rules. Very Large Online Platforms (VLOPs) with over 45 million EU users must have fast, transparent mechanisms for users to flag illegal content and provide dispute mechanisms for content removal. The Act requires platforms to explain why their algorithms recommend certain content, allow users to opt out of personalized AI feeds entirely and assess how their algorithms might be amplifying illegal content or risking public security. The Act also bans targeted advertising to minors.

Finally the Platform Work Directive requires algorithmic transparency in workplace management, with protections covering an estimated 28 million gig workers such as ride-hailing drivers, food delivery couriers and freelance data encoders (microtaskers). The regulations include mandatory human oversight for termination of employment and a ban on processing certain personal data (like a worker’s emotional state or personal beliefs) using AI.

Taken together, the EU has built an interlocking regulatory architecture that governs AI systems, platform monopolies, data rights and worker protections.

Canada is operating in an AI regulatory vacuum

Despite launching the world’s first national AI strategy in 2017 and investing over C$4.4 billion in AI research infrastructure since 2016, Canada still has no enforceable regulatory framework governing how AI systems are developed, deployed or audited.

Canada’s foundational consumer privacy law, the Personal Information Protection and Electronic Documents Act (PIPEDA), dates back to 2000 (the era of websites and emails and not much else). Attempts to modernize it, like the Consumer Privacy Protection Act bundled inside Bill C-27, failed to address the core issue: that consent-based models do not protect citizens when saying no means being excluded from modern life. 

Also bundled under Bill C-27 was the Artificial Intelligence and Data Act (AIDA) introduced in 2022 as part of Canada’s Digital Charter. AIDA was meant to regulate “high-impact” AI systems, but it lacked meaningful definitions, clear enforcement mechanisms and public-interest safeguards. Civil society groups and legal experts criticized AIDA throughout its passage for being drafted with almost no public consultation.

“Canada still has no enforceable regulatory framework governing how AI systems are developed, deployed or audited.”

Ultimately, Bill C-27 languished in committee through 2024 and died on the order paper when parliament prorogued in January 2025. 

In June 2025, the federal AI minister confirmed that AIDA would not be revived, promising a “light, tight, right” approach—supposedly light enough not to “stifle innovation,” tight enough to close “real risks.” The result of this promise is two bills that were tabled in 2026: Bill C-34, the Safe Social Media Act, and Bill C-36, the Protecting Privacy and Consumer Data Act, which replaces the privacy provisions of PIPEDA. 

Bill C-34 would require AI chatbot operators to interrupt an interaction and direct users to crisis services where someone expresses suicidal ideation and to stop systems posing as human beings or as licensed professionals—but only where those systems interact with children and vulnerable users. Bill C-36 adds a transparency measure for algorithmic decision-making, but as the Canadian Civil Liberties Association (CCLA) notes, much of what the government presents as new—the right of deletion, heightened protection for children’s data, limits on surveillance pricing—restates protections that have been in federal privacy law for 25 years. Its recognition of privacy as a “fundamental right” sits alongside broad exceptions for corporations’ activities, which the CCLA notes have “the ability to ignore consent requirements and retain personal data indefinitely if superficially de-identified.”

Both bills would be administered by a new Digital Safety and Data Protection Commission, which dilutes the Privacy Commissioner’s role as steward of commercial privacy protections and folds privacy enforcement together with content moderation under one body. How that commission is built, and who is consulted while it is built, will shape AI governance in Canada.

At the provincial level, some provinces have started to introduce privacy and automated decision-making protections aimed at AI. 

Quebec’s Law 25 is an effort to modernize the province’s privacy law, closer in line with the EU’s GDPR. The regulations apply to any public or private sector organization or business that processes the personal information of individuals residing in Quebec. Tracking technologies, cookies and profiling tools must be switched off by default and require explicit consent (opt-in) before they can be switched on. Requests for consent must be written in clear, simple language and presented separately from the companies’ other terms and conditions. Individuals have the right to withdraw their consent at any time, access and correct their data, have their data de-indexed or obtain a copy of their data in a portable format.

The Quebec law also requires transparency regarding automated decision-making and profiling. If a public body or business uses automated processing by a machine to make a decision, they must explicitly inform the affected individual. Individuals are granted the right to know what personal data was used, understand how that data was used to make the decision, request corrections and submit observations to a staff member capable of reviewing the machine’s decision. Organizations must conduct a privacy impact assessment before launching major electronic service delivery systems, acquiring or developing new information systems or transferring personal data outside of Quebec, and they must name someone accountable for compliance such as a privacy officer (otherwise by default the CEO). Penalties reach C$25 million or 4% of worldwide annual turnover (whichever is higher) for the most serious violations.

In Ontario, the Working for Workers Four Act requires employers to disclose in job postings if they use AI to screen, assess or select applicants, and the Strengthening Cyber Security and Building Trust in the Public Sector Act (Bill 194) requires public sector institutions to inform the public when they use AI in any of their processes. Bill 194 also allows the government to set standards on the public sector regarding accountability frameworks,  risk management, human oversight and the outright prohibition of high-risk uses of AI.

AI ministers’ mandates

The federal and BC governments have recently appointed new AI ministers. Both ministers claim to balance innovation with ethics, but their actual mandates are overwhelmingly oriented toward commercialization and public-private partnerships rather than rights protection.

In May 2025, Prime Minister Carney appointed Evan Solomon as Canada’s first-ever Minister of Artificial Intelligence and Digital Innovation. The portfolio sits inside Innovation, Science and Economic Development Canada. Its mandate centres on AI investment, adoption, commercialization, talent retention and the building of AI infrastructure—with little substantive commitment to rights-based regulation. While the minister did not receive an AI-specific mandate letter, the prime minister’s single mandate letter to all cabinet included:

The combination of the scale of this infrastructure build and the transformative nature of artificial intelligence (AI) will create opportunities for millions of Canadians to find new rewarding careers – provided they have timely access to the education and training they need to develop the necessary skills.

Government itself must become much more productive by deploying AI at scale, by focusing on results over spending, and by using scarce tax dollars to catalyse multiples of private investment.

In BC, Premier Eby appointed Rick Glumac as Minister of State for Artificial Intelligence and New Technologies in July 2025. The provincial mandate letter directs the minister to identify investment opportunities, increase adoption of AI tools and promote BC’s AI research and business nationally and internationally. Rights protection, algorithmic accountability and harm mitigation are absent from the mandate letter. 

In September 2025, Minister Solomon launched an AI Strategy Task Force and a “national sprint” as the vehicle for developing a renewed national AI strategy. The task force comprised 28 members drawn primarily from industry, academia and select think tanks. 

More than 160 academics, civil society organizations, union representatives and legal experts signed an open letter protesting the sprint’s compressed timeline, industry-weighted composition and leading survey framing. The signatories of the open letter warned that the task force was dominated by “the exact subset that stands to profit the most from an ‘all in’ approach to AI,” and that the communities who “disproportionately bear the brunt of AI-facilitated harms” had the least representation on it. 

In direct response, a coalition of civil society organizations—including the BC Civil Liberties Association (BCCLA) and the CCLA—launched the People’s Consultation on AI, collecting submissions from individuals, communities, labour organizations, academics and impacted groups as an independent democratic counterweight to the national AI strategy. These submissions were forwarded to the federal government.

Unfortunately, although the federal strategy was delayed from its promised end-of-2025 release, with a summary of inputs published in February 2026, which used generative AI tools to “analyze” and summarize the inputs, their key concerns were not heeded. 

“Unfortunately, although the federal strategy was delayed from its promised end-of-2025 release, with a summary of inputs published in February 2026, which used generative AI tools to “analyze” and summarize the inputs, their key concerns were not heeded.”

The resulting national AI strategy released in June 2026, AI for All, is an open policy of aggressive adoption across both private and public sectors with no regard for risks and harms. Its primary measure of success is AI adoption. The threat to public confidence, in the government’s account, is that Canadians are not using AI enough: the current adoption rate, it warns, “risks undermining public trust.” The national strategy promises 250,000 new AI-related jobs with no count of the number of workers AI-branded technologies could displace over time. It commits to putting AI into businesses, classrooms, hospitals and the public service, with no mention of the errors these systems produce or the biases they reproduce. Beyond Bills C-34 and C-36, the strategy does not restrict any use or application of AI.

“The resulting national AI strategy released in June 2026, AI for All, is an open policy of aggressive adoption across both private and public sectors with no regard for risks and harms.”

Expanding surveillance without accountability

At the same time that the government fails to regulate AI, several major bills are actively expanding state surveillance and enforcement power—enabled by or intersecting with the very technologies that remain ungoverned.

Tabled in 2025, Bill C-2, the Strong Borders Act, would have allowed police, border officials and security agencies to access private information without prior judicial authorization. It empowered the government to order electronic service providers—including social media platforms, messaging apps and mobile device operators—to provide information about customers’ private activities. Civil liberties groups described it as threatening “human rights, refugee and migrant rights, and privacy of all residents of Canada.” Over 300 organizations called for it to be withdrawn. 

In response, the government split the bill. The border and immigration measures became Bill C-12, the Strengthening Canada’s Immigration System and Borders Act. Introduced in response to the privacy criticisms, it nonetheless replicates many of C-2’s anti-migrant and anti-refugee provisions. “The bill allows unrestricted information sharing about migrants across all government levels,” writes the Migrant Rights Network, and “undocumented workers asserting labour rights could face deportation when employers report them to border enforcement.” The bill received royal assent in March 2026.

“Introduced in response to the privacy criticisms, [Bill C-12] nonetheless replicates many of C-2’s anti-migrant and anti-refugee provisions. ”

The unprecedented digital-access powers returned as Bill C-22, An Act respecting lawful access, which retains the authority to compel electronic service providers to disclose customers’ information and activities. Despite what legal experts have warned about “the bill’s sweeping scope, significant constitutional and human rights risks, transparency and accountability deficits, and dangers to encryption and Canada’s cybersecurity,” it passed its third reading in the House of Commons in June 2026 and is currently under review by the Senate.

While Bill C-8, An Act respecting cyber security, introduces mandatory cybersecurity obligations for vital services, it also grants government a power to spy that critics say “empowers government officials to secretly order telecommunications companies to install backdoors inside encrypted elements in Canada’s networks.” 

Bill C-9, the Combatting Hate Act, introduces a new “intimidation offense” and removes the requirement for the Attorney General’s consent before hate propaganda charges can be laid. Civil society groups—including the CCLA, BCCLA, Canadian Muslim Public Affairs Council and over 40 other organizations—argue that its vague language will be used to criminalize peaceful protest and silence dissent. 

Together, these bills consolidate surveillance and policing powers based on digital and AI technologies—but there is no coherent legislation to govern whether or how these technologies can be used, let alone ensure human rights are protected in the process.

Digital sovereignty for whom?

The federal government has rhetorically—though not materially—made digital sovereignty a nation-building priority. 

In September 2025, the prime minister directed the new Major Projects Office to develop a Canadian Sovereign Cloud, placing it on the same tier as pipelines, ports and nuclear reactors. Budget 2025 committed nearly $1 billion over five years to large-scale “sovereign” public AI infrastructure. The stated rationale is to reduce Canadian dependence on US-owned data infrastructure and to protect against the US Clarifying Lawful Overseas Use of Data (CLOUD) Act, which enables American law enforcement to compel cloud service providers to hand over data held abroad. As things stand, under the CLOUD Act, the US government can unilaterally access data on Canadian soil held by US companies, such as a Canadian branch of Microsoft. 

But despite the Canadian government’s rhetorical promotion of digital sovereignty, Canada has been negotiating with the US since 2022 to develop a CLOUD Act agreement, which would give US law enforcement similar access to Canadian data held in Canada by entirely Canadian companies, outside of Canadian judicial oversight. Bill C-22 compels tech companies to build in the surveillance capability such an agreement requires.

“While this so-called “sovereign” cloud rhetorically prioritizes data sovereignty from foreign governments and corporations, it does not address data sovereignty for Canadian residents.”

And while this so-called “sovereign” cloud rhetorically prioritizes data sovereignty from foreign governments and corporations, it does not address data sovereignty for Canadian residents—including Indigenous Peoples, racialized communities, migrants and others disproportionately surveilled by Canadian state actors. 

This is a critical moment for intervention. 

Policy solutions

This section brings together policy recommendations from civil liberties organizations, the labour movement and academia to intervene in the false inevitability of “artificial intelligence” as it is being sold to us today.

If we accept that AI is not an inevitable force but a political and economic project, then alternative policies become possible. A social-good approach to AI would begin by asking what forms of collective ownership and democratic control could ensure that technology serves social needs rather than capital accumulation.

But first, “for effective policy, solutions must be mapped onto accurately described harms.”

Labour policy

Because the extraction of labour without adequate compensation sits at the centre of the AI business model, workplace protections are an urgent priority. This is not an exhaustive list, but an example of the type of universal protections that must be urgently implemented.

Click on a category to view details:

Broader policy directions

Beyond the workplace, a social-good approach to AI must look to collective ownership and democratic control to ensure that AI technology serves social needs rather than capital accumulation. 

The question of how to govern data ecosystems as collective resources rather than private commodities is not new. Indigenous Peoples have been developing and practising data sovereignty frameworks for decades in response to long histories of colonial data collection, extractive research and the systemic misuse of community knowledge by governments, corporations and academia. These frameworks are rooted in Indigenous Peoples’ inherent rights to govern their peoples, lands and resources and must be implemented on their own terms as a matter of Indigenous self-determination. They also represent the most rigorous models for what collective data governance looks like in practice and as such, they have much to teach any community grappling with the concentration of power in AI data infrastructures. 

“…a social-good approach to AI must look to collective ownership and democratic control to ensure that AI technology serves social needs rather than capital accumulation.”

The principles of OCAP^Ⓡ (ownership, control, access and possession), established in 1998, assert “that First Nations alone have control over data collection processes in their communities, and that they own and control how this information can be stored, interpreted, used, or shared.” The National Inuit Strategy on Research, established in 2018, calls for Inuit ownership over data gathered on Inuit populations, wildlife and environments and for Inuit representational organizations to be the rightful gatekeepers of research in Inuit Nunangat. 

The CARE principles for Indigenous data governance, drafted in 2018, extend these principles globally and ground the governance of data ecosystems in four commitments: that data ecosystems be designed so that Indigenous Peoples derive collective benefit from them; that Indigenous Peoples’ authority to control how they and their lands, resources and knowledges are represented in data be recognized and respected; that those working with Indigenous data bear a responsibility to demonstrate how that data supports Indigenous self-determination, including by building data capacity and literacy within Indigenous communities; and that ethics—minimizing harm, pursuing justice and accounting for future use—guide every stage of the data life cycle.

The provincial and federal governments should treat these frameworks as binding standards. No Indigenous data—whether collected through public institutions, research partnerships or automated systems—should be incorporated into AI training sets or algorithmic decision-making without free, prior and informed consent from the relevant Indigenous governing bodies. Public funding should be directed toward Indigenous-led data infrastructure and digital capacity, and Indigenous communities must have binding authority over how data about their peoples, territories and knowledge systems is collected, stored and used.

The recommendations that follow extend these principles to the broader AI data ecosystem—that data governance must be democratic, transparent, rooted in collective rights and accountable to the communities it affects.

Click on a category to view details:

The choice is not between embracing or rejecting AI, but between surrendering to its current capitalist-driven form or reshaping it for collective benefit. Technology has always been a terrain of struggle. The more we treat Big Tech’s AI as inevitable, the more we foreclose the possibility of building technology in service of a society where human creativity and collective capacity come first.