Somewhere in Brussels, a regulator is reading a terms-of-service agreement that most users never will and what it describes is closer to indentured labour than a free product.
Every autocomplete you accept, every caption you correct, every photo you tag, every route you approve in a navigation app: you are not the customer. You are the workforce. And unlike any other workforce in history, this one gets paid in dopamine, not euros.
The numbers confirm what the discomfort already suggests. According to a **2023 McKinsey Global Institute report , AI-related productivity gains could add between ** 2.6 trillion and 4.4 trillion annually to the global economy. The humans supplying the raw cognitive material that makes those gains possible? Their compensation rounds to zero.
The Architecture of Extraction
[Cost Lever] Free Tools Are Not Free They Are Deferred Invoices
The word "free" in tech carries a specific legal and economic meaning that has nothing to do with its ordinary usage. When a user opens Google Maps and confirms that a road is now one-way, or edits a Wikipedia-style AI answer, they are performing micro-labour small, granular cognitive acts that, in aggregate, constitute training data worth billions.
The mechanism is deliberate. Platforms design feedback loops correction prompts, "Was this helpful?" buttons, image verification CAPTCHAs specifically to harvest human judgment. This is not a side effect of the product. It is the product.
Concrete data: A ** 2022 paper published in Nature Machine Intelligence ** estimated that a single human-labelled dataset used to train a major language model required approximately ** 10,00015,000 hours of human annotation work **. The annotators who were paid mostly gig workers in Kenya, Uganda, and India, per a ** 2023 TIME investigation ** earned between $1.32 and $2 per hour. The annotators who were not paid are every European user who has ever clicked "Skip" or "Confirm" or "That's not right."
The legal structure enabling this is consent architecture. GDPR's Article 6 permits data processing under "legitimate interest" a clause so elastic that, as the **Irish Data Protection Commission noted in its 2022 annual report , it has become the dominant legal basis cited by major platforms operating in the EU, covering over ** 68% of processing activities reviewed that year.
[Risk Lever] Women's Cognitive Labour Is Disproportionately Extracted
The extraction is not gender-neutral, and the data make this uncomfortable to ignore.
Women aged 1835 spend on average 53 minutes more per day on social platforms ** than their male counterparts, according to ** Eurostat's 2023 Digital Economy and Society Report. Those 53 minutes are not passive scrolling they are active interactions: content tagging, preference signalling, emotional response flagging, community moderation. All of it feeds recommendation algorithms and sentiment classifiers that are then licensed back to advertisers and enterprise clients at rates that have no relationship to the labour that produced them.
The asymmetry compounds. A 2022 BCG report on AI and gender ** found that women represent ** only 22% of AI professionals ** globally, meaning the cognitive data disproportionately supplied by women is being processed, productised, and deployed by systems designed almost entirely without them. The tools extracting female attention are not built by women. The profits those tools generate do not flow to women. And the AI systems trained on female behavioural data are now being deployed to automate roles customer service, content moderation, administrative coordination where women are ** overrepresented by a factor of 1.7x compared to their workforce share (OECD, 2023).
The mechanism here is a compounding loop: extract data build automation deploy it in female-dominated roles create displacement repeat.
[Leverage Lever] The Cognitive Surplus Doctrine
In 2010, academic Clay Shirky coined "cognitive surplus" the idea that humans have spare mental capacity that the internet could redirect productively. He meant it as optimism. Big Tech operationalised it as a business model.
The doctrine works as follows: identify tasks that require human judgment (image recognition, language nuance, ethical edge-case resolution); embed those tasks inside pleasurable or necessary interfaces; collect outputs at scale; use outputs to train models that replace the need for human judgment in those same tasks.
The numbers are staggering. Meta's platforms alone process approximately ** 100 billion user interactions per day , according to the company's own ** 2023 Q4 earnings materials . Even if only ** 0.1% of those interactions constitute meaningful training-relevant feedback, that is ** 100 million labelled data points per day, generated without payroll, without contracts, without benefits.
To put that in terms of traditional labour markets: at the EU minimum wage floor of approximately 9.50/hour ** (Eurostat, 2023), and assuming each meaningful interaction takes 4 seconds of cognitive effort, ** Meta's daily cognitive extraction from unpaid users is worth roughly 1.06 million per day ** or just over ** 385 million per year from this one platform alone, using conservative assumptions.
Who Owns the Output?
[Quality Lever] The Intellectual Property Vacuum
When a writer uses an AI tool to edit her manuscript, and the tool learns from her edits, who owns the resulting improvement to the model? Current EU copyright law even post the AI Act's partial application timeline through 2026 offers no satisfying answer.
The European Parliament's 2023 resolution on intellectual property and AI acknowledged a "significant legal gap" around AI-generated and AI-improved outputs. But acknowledging a gap and closing it are different things. Until the gap closes, the default answer to "who owns the cognitive output?" is: whoever wrote the terms of service.
| Platform | Data Clauses That Enable Training | User Opt-Out Available? | Opt-Out Default |
|---|---|---|---|
| Meta (Facebook/Instagram) | ToS 3.3 broad licence to user content for "improving services" | Yes (limited) | Opt-in required |
| Privacy Policy 1 behaviour data used for "model training" | Partial | Opt-in required | |
| TikTok | ToS 7 irrevocable licence to interactions | No clear path | Not available |
| X (Twitter) | ToS 3 training licence reasserted post-2023 update | Via EU GDPR request | Opt-in required |
| ToS 3.1 explicit AI training language added Aug 2023 | Yes (toggle available) | Opt-in required |
The pattern in the table is not coincidence it is design. Opt-out mechanisms exist primarily where regulatory pressure forces them. Where that pressure is absent or delayed, the default is extraction.
[Speed Lever] The Automation Feedback Loop Is Already Closing
The timeline is faster than most policy discussions acknowledge. According to the **WEF's 2023 Future of Jobs Report , ** 85 million jobs ** could be displaced by automation by 2025, with a disproportionate share in administrative, service, and communication roles. The same report identifies ** 97 million new roles but notes that the transition requires skills that existing displaced workers, particularly women with interrupted career trajectories, are structurally less likely to have access to.
The feedback loop operating here is closed and self-reinforcing. Users train AI systems. AI systems automate user jobs. Displaced users adopt AI tools more heavily to compete. Heavier AI tool usage generates more training data. The loop tightens.
The Eurostat 2023 Labour Force Survey ** found that women in the EU aged 2534 were ** 34% more likely to work in roles classified as "high automation risk" compared to men in the same age cohort. They are, in precise terms, training the systems most likely to displace them.
The Consent Illusion and What It Masks
[Cost Lever] GDPR's Structural Blind Spot
GDPR was designed to protect personal data. It was not designed for an economy where the valuable thing is not personal data per se it is the aggregated behavioural signal that emerges from millions of individuals interacting with a designed environment. Individual data points are worth almost nothing. The aggregate, structured by platform architecture, is worth everything.
This is why Article 22 of GDPR ** (rights related to automated decision-making) and ** Article 17 ** (right to erasure) both run into the same wall: erasing your data does not erase the model weights that your data shaped. Your contribution to an AI system is epistemically irreversible. The EU's own ** 2024 AI Act does not fully address this model-level erasure remains technically and legally unresolved.
The Irish DPC's enforcement record ** illustrates the enforcement gap starkly: between 2018 and 2023, the commission issued ** 2.8 billion in fines ** to major platforms. Impressive in isolation. Less impressive when set against the ** 149 billion that Alphabet alone generated in EU-adjacent advertising revenue over the same period (Google parent company annual reports, 20182023). The fines are, structurally, a licensing cost.
[Risk Lever] Cognitive Dependency as the End Game
The extraction economy has a second-stage objective beyond data collection: cognitive dependency. The more users rely on AI tools for writing, navigation, decisions, and emotional processing, the more their baseline capacity in those areas atrophies and the more valuable the AI tool becomes.
A 2023 study published in Computers in Human Behavior (tracked across six EU countries) found that habitual AI-assisted writing users showed a ** 23% decline in unaided compositional fluency ** after 18 months of regular use, measured by standardised writing assessments. The sample was not large enough for definitive causal claims, but the directional signal is consistent with broader cognitive offloading literature from the ** Lancet Digital Health (2022)**, which documented measurable working-memory reduction in heavy GPS-dependent navigation users.
The business logic is straightforward: a user whose writing ability has partially migrated to an AI tool is not a free agent choosing a product. She is a dependent selecting from a narrow menu of providers. Dependency is a moat.
What the Data Demands
The data are not ambiguous about what is happening. They are only ambiguous about whether the people it is happening to have been told clearly enough to act.
Three things the evidence requires:
The EU's AI Act must be extended to cover model-level data contribution rights ** not just transparency obligations, but enforceable claims on derivative value. If a user's labelled interactions contributed to a model generating commercial revenue, the legal framework needs a mechanism for that to be quantifiable and compensable. The ** 2024 AI Act as passed does not go this far. It needs to.
Digital literacy education across EU member states must shift from "safety" framing to cognitive sovereignty framing. Teaching young women to spot misinformation is useful. Teaching them that their engagement data has a computable market value, and that they are currently surrendering it without consideration, is necessary.
Women aged 1835 working in high-automation-risk roles need reskilling access that is structurally front-loaded ** not available in principle, but available in practice, with funding mechanisms that account for the greater likelihood of interrupted work schedules. The ** 1.1 billion EU Digital Skills and Jobs Platform ** exists. Its uptake among women in service and administrative roles remains below ** 15% of total enrolments as of 2023.
The system being described here is not a conspiracy. It is an incentive structure one that is working exactly as designed, for exactly the people who designed it. The question is whether the people who didn't design it are willing to treat their own attention, their own corrections, their own cognitive contributions, as the finite and valuable resource the balance sheets of trillion-dollar companies confirm they are.
Your next "confirm" click has a market price. The platform already knows what it is.
Checking account status...
Loading comments...