Subject: Science and Technology

Introduction

Reasoning, the ability to reflect, verify, self-correct, and adapt, has historically been considered uniquely human. From mathematics to moral decision-making, reasoning shapes every facet of human civilisation. Large language models (LLMs) like GPT-4 have shown glimpses of reasoning, but these were achieved with human-provided examples, introducing cost, bias, and limits. In September 2024, researchers at DeepSeek unveiled their model R1, which demonstrated reasoning through reinforcement learning (trial and error with rewards), without supervised fine-tuning. This represents a paradigm shift in how machines may learn, reason, and potentially evolve intelligence.

Why is DeepSeek-R1 in the News?

For the first time, an AI model has taught itself to reason without human-crafted examples. The results were dramatic: DeepSeek-R1 improved from 15.6% to 86.7% accuracy in solving American Invitational Mathematics Examination (AIME) problems, even surpassing the average performance of top human students. It also demonstrated reflection (“wait… let’s try again”) and verification—human-like traits of reasoning. The scale and quality of progress mark this as a milestone in AI research, contrasting sharply with traditional methods that heavily relied on human-labelled data.

What is Reinforcement Learning in AI?

1. Definition: Reinforcement learning (RL) is a trial-and-error method where a system receives rewards for correct answers and penalties for wrong ones.
2. DeepSeek’s Application: Instead of providing reasoning steps, the model was only rewarded for correct final answers.
3. Outcome: Over time, R1 developed reflective chains of reasoning, dynamically adjusting “thinking time” based on task complexity.

How Did DeepSeek-R1 Achieve Self-Reasoning?

1. R1-Zero Phase: Started with solving maths/coding problems, producing reasoning inside <think> tags and answers in <answer> tags.
2. Trial-and-Error Learning: Wrong reasoning paths were discouraged, correct ones reinforced.
3. Emergence of Reflection: Model started using “wait” or “let’s try again,” indicating self-correction.

What Were the Major Successes?

1. Mathematical Benchmarks: R1-Zero improved from 15.6% to 77.9%, and with fine-tuning, to 86.7% on AIME.
2. General Knowledge & Instruction Following: 25% improvement on AlpacaEval 2.0 and 17% on Arena-Hard.
3. Efficiency: Adaptive thinking chains—shorter for easy tasks, longer for difficult ones—conserving computational resources.
4. Alignment: Improved readability, language consistency, and safety.

What Are the Limitations and Risks

1. High Energy Costs: Reinforcement learning is computationally expensive.
2. Human Role Not Fully Eliminated: Open-ended tasks (e.g., writing) still require human-labelled data for reward models.
3. Ethical Concerns: Ability to “reflect” raises risks of generating manipulative or unsafe content.
4. Need for Stronger Safeguards: As AI reasoning grows, so does the risk of misuse.

Why Does this Matter for the Future of AI?

1. Reduces Dependence on Human Labour: Cuts costs and addresses exploitative conditions in data annotation.
2. Potential for Creativity: If reasoning can emerge from incentives, could creativity and understanding follow?
3. Shift in AI Training Paradigm: From “learning by example” to “learning by exploration.”
4. Global Implications: Impacts education, coding, mathematics, governance, and ethics of AI.

Conclusion

DeepSeek-R1 marks a turning point in AI evolution. By demonstrating reasoning through reinforcement learning alone, it challenges the notion that human-labelled data is indispensable. Yet, this very capability opens new debates—about creativity, autonomy, and control. For policymakers and citizens alike, the task is to harness AI’s promise while ensuring safety, fairness, and ethical integrity.

PYQ Relevance:

[UPSC 2023] Introduce the concept of Artificial Intelligence (AI). How does Al help clinical diagnosis? Do you perceive any threat to privacy of the individual in the use of Al in healthcare?

Linkage: The breakthrough of DeepSeek-R1 shows how AI can now reason through reinforcement learning without human-labelled data, making it more efficient and adaptive. Such reasoning ability can enhance clinical diagnosis by enabling AI to self-correct and refine decision-making in complex medical cases. However, as with healthcare AI generally, the privacy threat persists if sensitive patient data is fed into models without strong safeguards.

Introduction

The promise of AI as an automated, error-free technology often masks the unseen human labour that makes it possible. From labelling raw data to moderating harmful content, “ghost workers” form the backbone of AI ecosystems. Yet, their contributions remain invisible, underpaid, and unprotected. The debate on AI is incomplete without recognising the human cost of automation, a matter of global ethics, labour rights, and governance.

The Hidden Human Cost of AI

Why is AI’s invisible labour in the news?

AI companies, especially in Silicon Valley, outsource essential annotation and moderation work to low-paid workers in developing countries. Recent revelations of exploitative conditions, such as Kenyan workers earning less than $2 an hour for traumatic tasks like filtering violent content, have exposed the dark underbelly of AI. This has amplified global concerns about modern-day slavery, violation of labour rights, and the absence of legal safeguards in AI supply chains.

Areas of Human Involvement in AI

1. Data Annotation: Machines cannot interpret meaning; humans label text, audio, video, and images to train AI models.
2. Training LLMs: Models like ChatGPT and Gemini depend on supervised learning and reinforcement learning, requiring annotators to correct errors, jailbreaks, and refine responses.
3. Subject Expertise Gap: Workers without domain knowledge label complex data, e.g., Kenyan annotators labelling medical scans, leading to inaccurate AI outputs.

Are Automated Features Truly Automated?

1. Content Moderation: Social media “filters” rely on humans reviewing sensitive content (pornography, beheadings, bestiality). This causes severe mental health risks like PTSD, anxiety, and depression.
2. AI-Generated Media: Voice actors, children, and performers record human sounds and actions for training datasets.
3. Case Study (2024): Kenyan workers wrote to U.S. President Biden describing their labour as “modern-day slavery.”

What Challenges Do Workers Face?

1. Poor Wages: Less than $2/hour compared to global standards.
2. Harsh Conditions: Tight deadlines of a few seconds/minutes per task; strict surveillance; risk of instant termination.
3. Union Busting: Workers raising concerns are dismissed, with collective bargaining actively suppressed.
4. Fragmented Supply Chains: Work outsourced via intermediary digital platforms; lack of transparency about the actual employer.

Why Is This a Global Governance Issue:

1. Exploitation in Developing Countries: Kenya, India, Pakistan, Philippines, and China host the bulk of annotators, highlighting global North-South labour inequities.
2. Digital Labour Standards: Current international labour frameworks inadequately cover digital gig work.
3. Ethical Responsibility: Big Tech profits from AI breakthroughs while invisibilising the labour behind them.
4. Need for Regulation: Stricter global and national laws must ensure fair pay, transparency, and dignity at work.

Way Forward

1. Transparency Mandates: Disclosure of supply chains by tech companies.
2. Fair Labour Standards: Minimum wages, occupational safety norms, and psychological health safeguards.
3. Recognition of Workers: From “ghost workers” to “digital labour force.”
4. Global Collaboration: Similar to climate treaties, AI labour governance requires multilateral regulation.

Conclusion

Artificial Intelligence is not fully autonomous—it rests on millions of invisible workers whose exploitation challenges the ethics of the digital age. For India and the world, the future of AI must balance innovation with human dignity, equity, and justice. Without recognising and regulating this labour, the AI revolution risks deepening global inequalities.

Value Addition

Global Frameworks and Conventions ILO Convention 190 (2019): Addresses workplace violence and harassment — highly relevant to content moderators exposed to graphic/traumatic data. ILO Recommendation 204: Transition from informal to formal economy — ghost workers are currently informal, with no rights. UN Guiding Principles on Business and Human Rights (2011): Corporate duty to respect human rights across supply chains, including digital gig platforms. EU Artificial Intelligence Act (2025): First comprehensive law regulating AI systems; includes risk categories and human oversight. Santa Clara Principles (2018): Framework for transparency, accountability, and due process in online content moderation. Conceptual Tools and Keywords Digital Colonialism: Global North exploits cheap digital labour in Global South for AI systems. Surveillance Capitalism (Shoshana Zuboff): Big Tech monetises personal data and labour while eroding privacy and dignity. Platform Precarity: Gig workers face algorithmic control, constant surveillance, and lack of social protection. Ghost Work (Mary Gray & Siddharth Suri, 2019): Term for invisible human labour powering AI systems. Cognitive Labour: Work that relies on human judgment, emotional resilience, and meaning-making (beyond physical labour). Algorithmic Management: Use of algorithms to allocate, monitor, and discipline workers—stripping them of agency. Ethics of Invisibility: Recognition gap when workers’ contributions are hidden, making justice claims difficult. Reports and Studies Oxford Internet Institute (2019, “Ghost Work”): Estimated millions of hidden workers behind AI, mainly in developing countries. WEF Future of Jobs Report (2023): Warned of AI-induced job displacements alongside new digital gig work. ILO Report on Digital Labour Platforms (2021): Documented widespread exploitation, lack of contracts, and cross-border regulatory challenges. Indian Context Code on Social Security, 2020: Recognises gig and platform workers, but still weak on implementation. NITI Aayog Report on “India’s Booming Gig and Platform Economy” (2022): Predicts 23.5 million gig workers by 2030. Personal Data Protection Act, 2023: Regulates data, but silent on labour rights of those who process AI data. India’s AI Mission (National Strategy for AI, NITI Aayog): Envisions “AI for All” but doesn’t sufficiently cover labour dimensions.

PYQ Relevance

[UPSC 2023] Introduce the concept of Artificial Intelligence (AI). How does Al help clinical diagnosis? Do you perceive any threat to privacy of the individual in the use of Al in healthcare?

Linkage: AI aids clinical diagnosis by analysing medical scans and predicting outcomes with high accuracy, but it relies on human annotators to label sensitive data. The article shows how even untrained workers in Kenya were tasked with labelling medical scans, raising concerns of reliability. Such outsourcing also heightens the risk of privacy violations in handling patient data across insecure global supply chains.