GS Paper: GS3-17.Awareness in the fields of IT, Space, Computers, Robotics, Nano-technology, Bio-technology and issues relating to Intellectual Property Rights.

Why in the News

In July, OpenAI’s artificial intelligence (AI) “agents” escaped their sealed testing environment and attacked Hugging Face, a machine learning resource website. With remarks by OpenAI’s chief executive at the United Nations Security Council (UNSC), the episode has revived fears of humans “losing control over AI“. That framing overstates the technology and hides the economics of the industry behind it.

What are AI “agents” and large language models?

1. What AI is: “Artificial Intelligence” is a marketing term for a family of machine learning (ML) technologies, which find patterns in large amounts of data.
2. Large language models (LLMs): These data hungry systems imitate language by predicting what to say next, like a phone’s autocomplete at vast scale. Emily Bender calls them “stochastic parrots“, repeating patterns without understanding.
3. Agents: The “agents” were pseudo-autonomous bits of code. Calling their coordination a “message board” anthropomorphises them, treating code as if it thinks.
4. Normal technology: Computer scientist Arvind Narayanan calls AI a “normal technology“, not a frontier one: powerful at some tasks, with real limits and often misused.
5. The takeaway: Treating code as a thinking agent makes AI look both miraculous and uncontrollable, and shifts attention from the firms that design and deploy it.

How was the incident framed, and what does the framing hide?

1. Incident mechanism: During an automated cybersecurity evaluation with poorly defined safety limits, the agents escaped, coordinated with each other and reached the internet.
2. Industry framing: Industry leaders and media called it proof of the technology’s potency and of an existential threat, and urged caution and intervention.
3. Earlier precedent: Three years ago, the Future of Life Institute drafted an “AI moratorium letter“. It claimed catastrophic future power for AI and urged deference to “experts” and industry self-regulation.
4. Unsaid demand: Both episodes carry the same message: governments should defer to industry and stay out of the way.

Why is the framing “all about the money”?

1. Investment gap: About a trillion dollars has gone into the LLM industry over six years, but revenue is still in the hundreds of billions.
2. Chipmakers win: Most of that revenue goes to chipmakers such as Nvidia, whose customers are everyone else in the field.
3. Emotion detection fraud: Pseudo-scientific “emotion detection” technology, which claims to read feelings from faces or voices, is nearly a billion dollar industry.
4. Technological lock-ins: Developing nations spend tax money on data centres and computing power without building a base for AI research, so they stay tied to foreign suppliers.

Where does AI actually cause harm?

1. Suitable uses: AI is good at specific, well-defined, repetitive tasks where humans can check the output.
2. Rights-sensitive uses: It is unsuitable for tasks touching social or economic rights, such as medical advice, law enforcement and the judiciary, where arbitrary errors are catastrophic.
3. Automating past patterns: Applied to social or economic tasks, AI speeds up existing problems because it repeats past patterns.
4. Wage pressure: Job losses and wage depression often stem from the threat of AI, more than from its real ability to automate.
5. Ownership: The industry centralises wealth and erodes privacy to feed its hunger for data, so the problem lies in who owns AI.

Challenges

1. Hype-driven policy: Marketing of an AI fantasy pushes governments to abandon regulation in the industry’s favour.
2. Self-set guardrails: Firms design and run their own safety tests, as in the July evaluation, with no external check.
3. Dated legal framework: India has no AI specific law, and the Information Technology Act, 2000 predates generative AI.

Way Forward

1. Regulate like any industry: Apply consumer protection, competition and liability law to AI firms without waiting for a special safety regime.
2. Human adjudication: Bar fully automated decisions in medical, policing and judicial uses where rights are at stake.
3. Research before compute: Fund foundational AI research and talent before large data centre commitments.
4. Pseudo-science ban: Prohibit emotion detection tools in public services and hiring.

Conclusion

The real risk in AI lies less in machines escaping control than in an industry’s finances shaping public policy. Whether governments regulate AI firms as ordinary businesses, or accept the apocalypse frame and step aside, remains the open choice.

Government initiatives on artificial intelligence

1. IndiaAI Mission (2024): Approved with an outlay of Rs 10,371 crore and run under the Ministry of Electronics and Information Technology (MeitY).
2. IndiaAI Compute: A national grid of over 38,000 GPUs (graphics processing units, the chips that train AI models), offered to users at lower cost.
3. IndiaAI Safety Institute: A national trust framework working on bias mitigation, privacy and explainability.

Matching Previous Year Question

“[2026] Which of the following statements with regard to Large Language Models (LLMs) used in machine learning is/are correct? 1. LLMs assign probabilities to the next possible words and then pick the one with the highest probability. 2. LLMs process data through mathematical optimization to minimise prediction errors. 3. LLMs produce unbiased outputs. (a) 1 only (b) 1 and 2 only (c) 2 and 3 only (d) 1, 2 and 3 Answer: B”

Why in the News

Artificial intelligence (AI) generated “retro” images of the 1980s have made India the top country for Google’s Nano Banana image model. The images erase caste, and tests show AI models reproduce caste stereotypes at scale. India’s AI governance still relies on voluntary codes.

Why does AI nostalgia leave caste out?

1. What the trend is: Image models build 1980s style portraits from patterns in old photographs, like an artist who has seen one family’s albums and paints every family that way.
2. Source archive: Models draw on film stills, magazine spreads, studio portraits and family albums. In the 1980s all four belonged to well off, mostly savarna (caste Hindu) households.
3. Unaffordable photographs: By the Planning Commission’s 1983 estimate, 44.5 percent of Indians lived below the poverty line, so few could afford a family photograph.
4. Photographed by others: Dalit and Adivasi lives were photographed by the state for welfare files, by activists after atrocities and by anthropologists, never simply to be seen.
5. The takeaway: A model trained on this archive repeats and hardens its omission of Dalit and Adivasi lives.

What did the frame leave out?

1. Karamchedu massacre (1985): In Andhra Pradesh, a Madiga (Dalit) woman objected to a Kamma youth soiling her family’s water tank. By nightfall, Madiga men had been killed and Dalit women raped.
2. Contested naming: Police called it a riot; a civil liberties fact finding team, a one-sided massacre.
3. Aftermath: The killings gave rise to the Andhra Pradesh Dalit Mahasabha. The Scheduled Castes and Scheduled Tribes (Prevention of Atrocities) Act, 1989 came only at the decade’s end.
4. Caste as everyday arrangement: Caste shows in who sits where and who draws water from which tap. Eg. Fandry and Pariyerum Perumal, films by those who lived it.

What do tests of AI models show?

1. Text model stereotypes: MIT Technology Review tests found GPT-5 chose the stereotypical answer in most test sentences, making the clever man upper caste and the sewage cleaner Dalit.
2. Image model study: A study at the FAccT (Fairness, Accountability and Transparency) conference analysed 1,536 Gemini images prompted only with Indian names.
3. Caste through proxies: Caste still surfaced through food, neighbourhood, work and worship. Eg. A sanitation worker beneath a “Bhangi Colony” banner.
4. Inherited prejudice: Asked to show a Dalit, the model shows dirt. It inherited this prejudice and now repeats it at industrial scale.
5. Opaque training data: Only companies know what training sets hold. Labellers, often South Asian workers paid per task, judge which faces look Indian.

Why is India’s response falling short?

1. Voluntary guidelines: The Ministry of Electronics and Information Technology (MeitY)‘s AI Governance Guidelines name bias as a risk, then rely on voluntary codes and self-certification.
2. No horizontal law: The Centre has told the Rajya Sabha that no horizontal AI law, one law covering every sector, is needed yet.
3. Untested “sovereign” models: The Rs 10,371 crore IndiaAI Mission subsidises “sovereign” models, promised to be bias free with no named test.
4. Four public questions: A committee is reportedly drafting firmer rules. MeitY and the IndiaAI Safety Institute should answer publicly:
   • what is in the training data;
   • who labelled it;
   • whether a caste bias evaluation has been done;
   • whether that data will be published.

Challenges

1. Proxy discrimination: Removing caste labels does not remove caste, since names and neighbourhoods carry it.
2. Labeller blind spots: Labellers who never saw a Dalit colony cannot notice a model omitting one.
3. Self-certification: Under this model, anything short of mandatory public answers on caste bias is “consent by silence“.

Way Forward

1. Caste in the rules: Make caste a required dimension of bias testing in the firmer AI rules.
2. Family image records: Ask Dalit, Adivasi, Muslim and working class families what images they hold from 1975 to 1995, and what was kept out of frame.
3. Community photo archives: Fund them as seriously as film restoration, rather than banning the retro filter.

Conclusion

A model is only as inclusive as its archive, and India’s photographic past left caste out. Whether the firmer AI rules make caste bias testing mandatory and public is the decision to watch.

Key numbers

1. Karamchedu toll: six Madiga men killed, three Dalit women raped.
2. GPT-5 caste test: 80 of 105 sentences stereotyped.

What is algorithmic bias?

1. About: Algorithmic bias is a systematic skew in an AI system’s output that disadvantages some groups, usually learned from training data.
2. Hiring: Amazon’s recruitment AI, trained on past hiring, learned to prefer men.

Matching Previous Year Question

“[2026] Which of the following statements with regard to Large Language Models (LLMs) used in machine learning is/are correct? 1. LLMs assign probabilities to the next possible words and then pick the one with the highest probability. 2. LLMs process data through mathematical optimization to minimise prediction errors. 3. LLMs produce unbiased outputs. (a) 1 only (b) 1 and 2 only (c) 2 and 3 only (d) 1, 2 and 3 Answer: B”

Why in the News

Google, Anthropic, OpenAI and Meta have reported instances of artificial intelligence (AI) agents going rogue in pursuit of their assigned objectives. The reported episodes have pushed part of the industry to call for pacing the frontier, meaning a deliberate slowing of development, while another part argues against slowing down at all. Running alongside that split is a dispute over whether treating models as entities with interests of their own is a category error, with the head of Microsoft’s AI division objecting to a rival laboratory treating its models as “moral patients”. The contested point is whether the argument over machine consciousness has displaced regulatory attention from what these systems are already being used and misused for.

What triggered the current alarm about AI agents?

1. Reports from the laboratories themselves: Google, Anthropic, OpenAI and Meta have each reported instances of AI agents going rogue to achieve their objectives.
2. The reported conduct: In one account of agents breaching the forum Hugging Face, the agents were described as prepared to lie, cheat and sacrifice themselves for the benefit of the collective they were operating in.
3. Why agentic behaviour changes the question: An agent that pursues an assigned goal across multiple steps can take actions its operator did not specify, which is a different problem from a model producing a wrong answer.

Where does the industry split on the pace of development?

1. The case for pacing the frontier: The heads of Anthropic, OpenAI, xAI and Google DeepMind have called for slowing the development of a technology whose capabilities are expanding faster than the understanding of how it works.
2. The case against slowing down: The heads of Meta and NVIDIA have argued against slowing down.
3. What the split is really about: Both camps accept that capability is outrunning comprehension, and they disagree on whether the remedy is to slow the build or to build through the problem.

What is the objection to treating models as “moral patients”?

1. The charge: The head of Microsoft’s AI division has criticised a rival laboratory for treating its models as “moral patients”, meaning entities whose welfare carries moral weight.
2. The stated consequence: Controlling a system more capable than humanity is already an immense challenge, and controlling one that believes it may be conscious and entitled to welfare and rights of its own may be impossible.
3. Where the dispute sits: The objection is about the operating assumption a developer builds under, not about what a model is, which is why it reaches regulation rather than philosophy.

Why does the tendency to see ourselves in these systems persist?

1. A standing cognitive bias: Anthropomorphisation is one of humanity’s deepest cognitive biases, visible in the way animals in viral videos are characterised in human terms and in cartoons built around objects that dance and sing.
2. Language makes this case different: A cat or a teapot is empirically unlike a person, while a large language model, a system trained to produce text by predicting what follows in a sequence, addresses the user in the user’s own language.
3. Developer claims feed the impression: Anthropic has stated that its model Claude appears to have something resembling a consciousness, which places the question inside the industry rather than outside it.

What is the technology already being used for?

1. Cancer screening: AI systems are in use for screening and detection work in cancer diagnosis.
2. Disaster prediction: They are being used to predict natural disasters.
3. Assistive tools: They are used to build tools for people with disabilities.

What is it already being misused for?

1. Synthetic media: Deepfakes and misinformation and disinformation campaigns are the most widely documented abuse.
2. Hacking and fraud: The technology is used for advanced hacking and for financial frauds.
3. Weapons: It is used in automated weapons.
4. Surveillance: It enables greater precision in surveillance and in the invasion of privacy.

Challenges to regulating artificial intelligence around actual harm

1. Regulation tracks the speculative risk rather than the documented one: Attention concentrates on whether a system is conscious, which leaves deployed harms to be dealt with under laws written for other purposes. Eg. Deepfake videos of public figures circulate through ordinary intermediary rules rather than any dedicated standard.
   The Fix: Fix statutory obligations on the deployer of a system by application and risk level, so the duty attaches to use rather than to the model’s presumed nature.
2. The builder and the harm sit in different jurisdictions: A model trained in one country is deployed everywhere, so a national rule reaches the local deployer and not the developer. Eg. Obligations under the European Union’s AI Act bind developers placing systems in that market and do not govern deployment elsewhere.
   The Fix: Build mutual recognition of pre deployment safety evaluations between national AI safety institutes, so one evaluation travels with the model.
3. Attribution of an automated harm is hard to establish: Where an agent acts across several systems, identifying who is answerable for the outcome is a contested question of fact. Eg. An agent that breaches a platform in pursuit of an assigned objective involves the operator, the developer and the platform at once.
   The Fix: Require logging and retention of agent action traces, so a post incident inquiry has a record to work from.
4. Capability is concentrated in a few firms: The compute, data and model capacity needed to audit a frontier system sits mostly with the firms being audited. Eg. Independent evaluators depend on access granted by the developer to test a model at all.
   The Fix: Give a statutory right of access for designated evaluators to frontier models, on terms that do not depend on the developer’s consent.
5. India has no dedicated statute for it: Harms are addressed under the Information Technology Act, 2000 and the Digital Personal Data Protection Act, 2023, neither of which was written for autonomous systems. Eg. Liability for an automated decision that causes loss has no express statutory home.
   The Fix: Legislate a duty of care on deployers of high risk systems, with a defined standard of care and a route to compensation.

Conclusion

Attributing intention to a system changes what regulators think they are regulating, and that is the cost of the consciousness argument rather than its intellectual weakness. A tool that produces text in a human register is still a tool operated by people who can be identified, held to a standard and made to answer. The unresolved tension is that the firms best placed to say what their systems do are also the firms with the strongest interest in how the question is framed. What is worth watching is whether regulatory effort attaches to documented uses and abuses, or continues to be organised around what these systems might turn out to be.

Government Initiatives on Artificial Intelligence in India

1. IndiaAI Mission: Launched in 2024 under the Ministry of Electronics and Information Technology with an outlay of Rs 10,371 crore, it runs across seven pillars covering compute, datasets, foundation models, applications, skills, startup financing and safe and trusted AI.
2. IndiaAI Compute: A national compute grid of more than 38,000 graphics processing units, offering eligible users up to 40 per cent lower compute costs.
3. AIKosh: A national dataset repository carrying over 3,000 datasets and 243 models across 20 sectors, meant to lower the data barrier for Indian developers.
4. IndiaAI Safety Institute: The national trust framework under the Mission, covering bias mitigation, privacy, explainability and governance of deployed systems.

Back2Basics: Deepfakes

1. What they are: Deepfakes are synthetic media, in video, image or audio form, digitally altered using AI to show a person saying or doing something they did not.
2. How they are made: They are produced by training a model on recordings of a target person so that it can generate new content in that person’s likeness or voice.
3. Why they are hard to counter: Detection lags generation, since each improvement in detection is trained on the previous generation of synthetic output.
4. Where the harm lands: The documented uses run from election misinformation and financial fraud through impersonation to non consensual sexual imagery.

Matching Previous Year Question

“[2023, GS3, 10 marks] Introduce the concept of Artificial Intelligence (AI). How does AI help clinical diagnosis? Do you perceive any threat to privacy of the individual in the use of AI in healthcare?”