TL;DR
The AI community fixates heavily on how to build models while ignoring why we build them, resulting in a damaging monoculture oriented around labor automation and raw productivity. This position paper critiques the flawed assumptions behind this mainstream mindset and proposes an alternative framework—"AI for Just Work"—that centers labor rights, systemic limits, and the fundamental right to refuse AI deployment.
What Problem Does This Paper Address?
Machine learning research frequently treats technological progress as an unquestionable good, driven by an unstated consensus: build bigger models to automate tasks, replace human labor, and maximize productivity. The authors argue that this "mainstream imagination of AI" (MIA) operates on four unexamined myths:
- AI can become omniscient and solve all complex challenges.
- Replacing workers with AI automatically liberates humanity from toil.
- Automation naturally trickles down into widespread shared prosperity.
- AI harms are merely accidental bugs or the fault of bad actors.
By accepting these premises uncritically, the field falls into techno-solutionism, obfuscates the massive human "ghost work" and material resources required to build ML systems, and creates tools that disempower workers and exacerbate inequality.
Methods
Rather than introducing a machine learning model, the authors construct an axiomatic sociotechnical framework titled "AI for Just Work", drawing from science and technology studies (STS), labor economics, and feminist epistemology:
- Foundational Assumptions (A1–A3):
- A1 (Intrinsic Limits): AI models are reductive abstractions that depend entirely on human experience and planetary resources; they cannot replace the real world.
- A2 (Valuing Labor): Workers and work must not be degraded. The definition of "work" expands to encompass reproductive, emotional, and precarious data labor.
- A3 (Ethics Over Growth): Human and ecological flourishing must take precedence over pure efficiency and economic growth.
- Derived AI Properties (P1–P4):
- P1: Problem formulation and evaluation must be grounded in domain-specific, real-world tasks rather than abstract benchmark metrics.
- P2: Thorough limitation, failure mode, and safety analyses must be default requirements in AI evaluation, on par with reporting performance gains.
- P3: AI systems and data collection pipelines must be designed to be rejectable (informed refusal as an institutional power check).
- P4: Systems must actively augment worker power and agency rather than automate them out of bargaining leverage.
The authors operationalize these principles through a case study in Medical Image Synthesis (MISyn), outlining practical criteria such as mandatory metadata labeling to prevent clinical misinformation and a non-technical checklist empowering clinicians to challenge the necessity of synthetic data.
Datasets
Not discussed in the paper. As a position paper, no new datasets were constructed or trained on.
Key Insights and Ideas
- The "Blind Spot" Worldview: Borrowing from philosophy of science, the paper highlights how ML often confuses the map with the territory—treating mathematical models as superior to messy, grounded human experience while ignoring the hidden labor (labelers, contractors) needed to keep the illusion of automation alive.
- Technological Determinism is False: Technological harms are not inevitable side effects of "progress." They stem directly from early design choices and institutional incentives.
- Productivity Does Not Automatically Lift All Boats: Citing labor economics, the authors emphasize that automation typically creates displacement effects that depress wages unless accompanied by worker bargaining power and technologies designed to create new labor-intensive tasks.
- Refusal as a Technical Feature: The right to say "no" to automated systems and data harvesting is vital to counter the default assumption that every problem requires an AI solution.
Key Results
Because this is a conceptual and methodological position paper, results are qualitative:
- A structured critique dismantling four core justifications of mainstream AI ideology using multidisciplinary evidence.
- A six-dimensional comparative audit table comparing standard research practice to the proposed "AI for Just Work" paradigm in medical image synthesis (e.g., transitioning from model-space evaluation to a five-phased clinical validation pipeline; implementing explicit synthetic labeling standards).
Limitations and Future Work
- Institutional Inertia: The dominant AI paradigm forms the "path of least resistance" in modern academia and industry, reinforced by venture funding, corporate incentives, and standard peer review practices.
- Risk of "Ethics Washing": The authors caution that ethical frameworks are frequently co-opted without altering underlying power dynamics unless paired with hard structural checks—such as labor organizing, collective bargaining, and regulatory enforcement.
- Future Work: The authors call for the AI community to expand public dialogue with marginalized stakeholders, reform ML conference review criteria to mandate limitation analyses, and formulate concrete technical sub-agendas that prioritize worker-owned, labor-augmenting applications.
Summary updated