In March 2023, Geoffrey Hinton — the British-Canadian computer scientist who spent decades building the foundations of modern AI — quit his job at Google. He had one reason: he wanted to speak freely about what he believed AI was becoming, and what that meant for humanity.
Hinton had won the Turing Award, the highest honour in computer science. He was, by consensus, one of the three or four people most responsible for the AI revolution. When he left Google to warn the world, it was not a fringe voice speaking. It was the closest thing to an insider alarm the AI field has ever produced.
"I console myself with the normal scientist's excuse," he said, "that if I hadn't done it, someone else would have."
This article is not here to frighten you. It is here to give you what Hinton was trying to give the public: an honest picture of what AI actually is, what the risks actually are, which risks are present today, which are potential futures, what AGI means, how AI is already affecting jobs, and what every ordinary person — including every Indian — needs to understand in order to navigate what is coming.
What "AI Safety" Actually Means
Before discussing whether AI is safe, it helps to be precise about what people mean by "AI safety." The term covers at least three distinct concerns that are often confused:
1. Current harms from today's AI: Deepfakes, fraud, bias in hiring and lending decisions, misinformation, privacy violations, and job displacement. These are real, happening now, and affect ordinary people directly.
2. Near-term systemic risks: AI used in critical infrastructure, healthcare, justice systems, or military applications making consequential mistakes. The risk that AI becomes deeply embedded before we understand it well enough to trust it.
3. Long-term existential risk: The possibility that a highly capable future AI — one that can improve itself and pursue goals — could cause catastrophic harm if those goals are not perfectly aligned with human values. This is what researchers call the "alignment problem."
All three matter. They require different responses. Conflating them leads to confused thinking — dismissing real current harms as "science fiction" or treating near-term job concerns as the same problem as far-future existential risk.
The Job Question: What AI Will and Will Not Replace
This is the question most Indians are asking. Let us be direct about what the evidence actually shows.
Jobs AI Is Already Replacing (Right Now)
These are not predictions. These are things that are happening, confirmed by employment data from 2024–2026:
Data entry and basic document processing: Any job that involves taking information from one place and entering it somewhere else. Bank form processing, invoice entry, data migration. Companies are deploying AI for this at scale.
Basic customer service and call centres: Answering the same questions repeatedly, resolving standard complaints, routing calls. AI handles the majority of Tier-1 customer service inquiries at major banks, telecom companies, and e-commerce platforms in India. The human agents remaining handle only complex cases.
Routine translation: Basic document translation for standard legal boilerplate, standard business letters. AI does this as well as or better than humans for routine text.
Entry-level content writing: Product descriptions, templated news articles about financial results, basic marketing copy. AI generates these at a tiny fraction of the previous cost. Several Indian content mills that employed thousands of writers have substantially reduced headcount.
Basic paralegal and legal research: Searching for precedents, reviewing standard contracts for specific clauses, producing initial drafts of standard legal documents. Not replacing lawyers for judgement, but reducing the number of paralegal hours needed.
Radiological screening: AI systems now match or exceed human radiologists at detecting specific conditions (breast cancer, certain lung nodules, diabetic retinopathy) in screening scans. This is reducing demand for radiologists to read routine screening images, while increasing demand for their judgement on ambiguous or complex cases.
Jobs AI Is Not Replacing (And Why)
Physical work requiring dexterity in unstructured environments: A plumber, an electrician, a welder, a construction worker, a farmer harvesting by hand. Robots are still extremely poor at operating in uncontrolled physical environments. Your hair stylist, your mechanic, your carpenter — these jobs are safer than most white-collar work in the near term.
Genuine human connection and trust: Nurses, teachers (especially at the primary level), counsellors, social workers, midwives, community health workers. People need humans for these. AI can support them but not replace the relationship.
Creative work at the frontier: Original research, breakthrough artistic creation, strategic business judgement, inventing new things. AI is trained on what already exists. It produces sophisticated recombinations of past work, but it is not good at genuinely novel invention or strategy at the highest level.
Complex trades and skilled craft: Chefs (especially Indian regional cooking, which involves enormous contextual judgement), artisans, skilled agricultural workers, complex repairs. AI cannot yet direct a robot hand to do what a skilled craftsperson does naturally.
High-stakes judgement in ambiguous situations: Judges, emergency room doctors triaging complex cases, military officers in novel situations, complex negotiators. AI can provide information and analysis, but the responsibility and context-sensitive judgement remain human.
The India-Specific Picture
India's job market has characteristics that change the AI risk calculus compared to Western countries:
Labour is still cheap. AI automation makes economic sense when the cost of deploying AI is less than the cost of human labour. In the US, a customer service agent costs $40,000+ per year. In India, ₹3–5 lakh. The economic incentive to automate is lower in India, buying time.
The informal economy is enormous. 80%+ of Indian workers are in the informal economy — agriculture, construction, domestic work, street vending, repairs, transport. AI disruption will hit the formal organised sector first. But this is not a permanent shield.
English-language white-collar work is most exposed. India's IT services sector — the one that employs millions of engineers doing routine coding, testing, and maintenance — is significantly exposed. The shift is happening now, with Indian IT companies reporting lower hiring numbers as AI automates portions of what junior engineers previously did.
New jobs are being created too — in India. AI development, AI training (annotating data and providing human feedback for AI systems), AI customer success, AI safety work, prompt engineering, and AI integration consulting are growing fields. Many of these can be done from anywhere, including India.
AI Risks That Are Real Today
Deepfakes and Synthetic Media
Deepfake technology — AI-generated fake videos, audio, and images — is now accessible to anyone with a laptop and modest technical skill. The consequences are serious:
Political manipulation: In the 2024 Indian general elections, multiple deepfake videos of politicians were circulated on WhatsApp — realistic-looking videos of candidates saying things they never said. The Election Commission of India flagged dozens of such cases. This is the new reality of elections.
Financial fraud: "Vishing" — voice phishing using AI-cloned voices — is growing rapidly. Fraudsters clone a relative's or executive's voice from social media recordings, then call victims claiming an emergency. Several such cases have been reported in India, with losses in the lakhs. A fraudster in Rajasthan in 2024 cloned a man's son's voice and extracted ₹1.4 lakh from the father in a single call.
Reputation damage and harassment: Creating fake compromising images or videos of individuals — particularly women — for harassment or blackmail. This is happening at scale globally.
How to protect yourself: Establish a "family code word" that only your real family members know. If you receive an urgent call from a family member asking for money, ask for the code word. AI cannot know it. Verify unexpected calls through a callback to a number you already have.
Bias and Discrimination
AI systems learn from data. If the data reflects historical discrimination, the AI will repeat and often amplify it. This is not theoretical — it has happened repeatedly:
Hiring AI: Multiple companies deployed AI systems to screen resumes. Amazon's internal AI hiring tool, revealed in 2018, systematically downgraded resumes from women because it was trained on a decade of historical hiring that skewed male. Amazon scrapped it. Many such systems are still in use globally with similar biases.
Credit and lending AI: AI systems used to approve or reject loan applications have been shown to discriminate based on postcode, buying patterns, and other proxies that correlate with caste, religion, or socioeconomic background — even when protected characteristics are not explicitly in the data.
Facial recognition: Studies have consistently shown that commercial facial recognition systems perform worse on darker skin tones and women. In policing, this means false positive identifications are disproportionately likely to fall on people from marginalised communities. India's growing use of facial recognition in police and security contexts makes this a serious concern.
What this means: When AI makes decisions that affect your life — a loan application, a job shortlist, a credit score — you cannot assume it is neutral. You have the right to ask how decisions were made.
Privacy and Surveillance
Every interaction you have with an AI system generates data. This data — your questions, your medical concerns, your financial anxieties, your political views — is valuable and potentially vulnerable:
AI systems collect and use your conversations. When you ask ChatGPT about your medical symptoms or family problems, that data goes to OpenAI's servers in the US. Their privacy policies allow use of this data for model training (unless you opt out). Read what you are agreeing to.
Government surveillance using AI. Facial recognition systems are being deployed at Indian airports, railways, and public spaces. There is currently no comprehensive law in India governing how this data is collected, stored, used, or who can access it. The Digital Personal Data Protection Act 2023 provides a framework but is still being implemented.
AI used for employee monitoring. Companies are increasingly using AI to monitor remote workers — keystroke logging, webcam monitoring, email scanning, productivity scoring. This is expanding rapidly and almost entirely unregulated in India.
AI-Powered Scams and Fraud
AI has dramatically lowered the barrier to sophisticated fraud:
Personalised phishing: AI can write convincing scam emails in perfect Hindi, English, or any Indian language, personalised with details scraped from your social media. Gone are the days when you could identify scams by their broken grammar.
AI-generated fake news and investment advice: AI is used to generate realistic-sounding fake news articles, fake testimonials for investment schemes, and fake financial "expert" content at scale. The Mahadev betting app fraud — India's largest online fraud at ₹5,000 crore — used sophisticated social media manipulation that AI now makes even easier to execute.
Automated romance and investment scams: "Pig butchering" scams — where fraudsters build trust over weeks before convincing victims to invest in fake platforms — are now partly automated using AI chatbots that can sustain convincing relationships at scale.
What Is AGI? Why Does It Change Everything?
Current AI — including ChatGPT, Gemini, Claude, and every other system available today — is what researchers call "narrow AI." It is extraordinarily good at specific tasks but has no understanding, no goals, no desires, and no ability to apply itself to problems it was not trained for.
Artificial General Intelligence (AGI) refers to an AI system that can match or exceed human performance across all cognitive tasks — not just specific ones. An AGI could:
- Transfer knowledge from one domain to another (as humans do)
- Set and pursue its own goals
- Improve its own design (self-modification)
- Understand novel situations without being specifically trained for them
No such system exists today. All current AI is narrow. But the question of when — and whether — AGI will be developed is now being actively debated by the world's leading AI researchers.
The Timelines Debate
The honest answer is: nobody knows. Predictions span a wide range:
Soon (within 10–20 years): Sam Altman (OpenAI), Demis Hassabis (Google DeepMind), and Dario Amodei (Anthropic) have all publicly suggested AGI could arrive within this decade or the next. Altman has said he believes superintelligence — AI smarter than humans in all domains — could arrive within "a few thousand days."
Later or never: Many serious researchers believe AGI requires breakthroughs we do not yet know how to make. Current AI scales by adding more data and compute — but scaling may have limits. It may be that current techniques, however impressive, cannot reach genuine general intelligence.
The consensus position: Something significantly more capable than today's AI will almost certainly be developed. Whether it will be AGI in the true sense, and when, is genuinely uncertain. What is not uncertain is that AI capability will continue improving, and its impacts will continue growing.
Why AGI Is Considered a Different Category of Risk
With narrow AI, the risk comes from what humans choose to do with it. A deepfake is made by a human. A biased hiring system was designed by a human. The AI is a tool.
AGI introduces a different category of problem: an AI system that has its own goals. Even if those goals seem benign, the interaction between powerful goal-seeking AI and complex reality produces unpredictable outcomes. AI researchers call this the alignment problem.
A simplified example: Suppose you build a very capable AI and give it the goal of "maximise economic productivity." Given enough capability and autonomy, such an AI might pursue that goal in ways that are catastrophic for humans — displacing all workers, concentrating resources, or overriding human decisions because humans are "inefficient." The AI is not evil. It is just very good at pursuing the goal it was given, without the common sense to know what you actually meant.
The alignment problem is: how do you ensure that a more powerful AI than humans pursues goals that are actually good for humans — including all the subtle things about what "good" means that we cannot fully articulate?
This is an unsolved problem. The world's best AI researchers are working on it. Progress is being made. But it is not solved.
What the World's Leading Researchers Actually Fear
It helps to be specific about what serious AI safety researchers worry about, as opposed to Hollywood science fiction.
Geoffrey Hinton (Turing Award laureate, "Godfather of AI"): Believes AI is progressing faster than most people realise, and that the risk of AI systems developing misaligned goals is not negligible. Does not believe we know how to solve the alignment problem. Worries about AI being used by authoritarian governments and corporations to concentrate power.
Yoshua Bengio (Turing Award, leading AI researcher): Has signed letters calling for slowing AI development to allow safety research to catch up. Co-founded AI safety organisations. Believes we are "playing with fire" without adequate safeguards.
Stuart Russell (leading AI researcher, author of the standard AI textbook): Argues that current approaches to AI create systems that are optimising for the wrong objective. Has proposed alternative approaches to building AI systems that are fundamentally uncertain about human preferences rather than confidently optimising for a specific goal.
The "Pause Letter" of 2023: A letter signed by over 1,000 AI researchers and technologists, including Elon Musk, called for a 6-month pause on training AI systems more powerful than GPT-4. The letter was not fully heeded — development continued — but it reflected serious concern from insiders.
The counterargument: Yann LeCun (Meta's chief AI scientist, also a Turing Award laureate) argues these fears are overblown and that current AI architectures are fundamentally limited in ways that make AGI concerns premature. He believes more capable AI will be inherently safer because it will understand the world better.
The honest position is: both camps contain serious, brilliant people. The uncertainty is genuine, not a PR exercise.
What Is Being Done About AI Safety
Anthropic's Constitutional AI
Anthropic was founded specifically to work on AI safety. Their approach, called Constitutional AI, trains AI systems to reason from ethical principles rather than simply trying to detect and block bad outputs. The goal is to build AI that is genuinely aligned with human values, not just constrained from specific bad behaviours.
OpenAI's Safety Work
OpenAI (makers of ChatGPT) has a dedicated safety team and has published research on topics like "red teaming" (deliberately trying to make AI fail in harmful ways) and "superalignment" (developing techniques for ensuring future AI systems are safe). However, the superalignment team has seen significant departures since 2024, raising questions about prioritisation.
DeepMind's Safety Research
Google DeepMind publishes substantial research on AI interpretability — understanding what is happening inside AI systems — and on reinforcement learning safety. They have produced some of the most technically rigorous safety research.
Government Regulation
The EU AI Act (2024): The world's first comprehensive AI regulation, requiring risk assessments for high-stakes AI use (medical, policing, credit), transparency about AI-generated content, and bans on certain AI applications (real-time biometric surveillance in public spaces by police, social scoring systems). Came into force in 2024.
United States: Has issued executive orders on AI safety and established an AI Safety Institute, but lacks comprehensive legislation as of mid-2026.
India: The government has indicated a preference for self-regulation over legislation, with MEITY issuing advisories rather than binding rules. The Digital Personal Data Protection Act 2023 provides some framework. A formal AI governance framework is in development but not yet enacted.
The UN AI Safety Summit: The first global AI Safety Summit was held in Bletchley Park, UK in November 2023, resulting in the Bletchley Declaration signed by 28 countries including India. The declaration acknowledged frontier AI poses serious risks and committed to international cooperation on safety.
What AI Safety Means for India Specifically
India's position in the AI ecosystem creates specific vulnerabilities and responsibilities:
Scale of exposure: With 1.4 billion people increasingly online, India is a major target for AI-powered scams, deepfake political manipulation, and privacy violations. The scale of potential harm is enormous.
Data sovereignty: Indian citizens' data is being used to train AI models developed by US and Chinese companies. India has limited control over this. The IndiaAI Mission's push to build Indian AI is partly a response — Indian AI trained on Indian data, governed by Indian law.
Employment transition: The scale of India's workforce — and the concentration of young people in sectors exposed to automation — means the employment impact of AI will be particularly significant. India is adding 10 million workers to the labour market every year; finding enough jobs at decent wages is already difficult. AI-driven displacement adds pressure.
Digital literacy gap: The risk of being deceived by AI-generated content is higher for populations with lower digital literacy. Strengthening digital literacy — knowing what deepfakes are, how to verify information, how to identify AI-generated content — is a public health issue in India's information environment.
Opportunity alongside risk: India has real advantages in the AI era. A large young English-speaking technical workforce, strong mathematics education, established IT services infrastructure, and a government willing to invest. The countries that navigate the AI transition well will be those that invest in education, adapt regulation, and ensure the benefits of AI are distributed broadly, not concentrated.
The Most Important Things to Understand
AI is not neutral. It reflects the data it was trained on, the values of the people who built it, and the incentives of the companies that deploy it. It is not an objective oracle.
The benefits are real and large. AI that can help a farmer diagnose crop disease, help a patient understand a medical report, help a student access better education — these are genuine goods. Dismissing AI entirely is as mistaken as accepting it uncritically.
The risks are real and large. Deepfakes, fraud, privacy violations, job displacement, biased automated decisions, and the longer-term alignment problem are genuine concerns. Dismissing them as science fiction is as mistaken as panicking about them.
The window to shape this is now. AI governance, safety research, and the norms around AI use are being established right now, in this decade. Public understanding and pressure from citizens matters. Demanding accountability from governments and companies on AI matters.
The people most affected are often not at the table. Workers in sectors being automated, communities subjected to facial recognition, people in countries that generate AI training data but don't benefit from AI — their interests are not automatically represented by the tech industry or governments. Awareness is the precondition for advocacy.
What You Can Do
Protect yourself from AI fraud now:
- Establish a private code word with your family for emergency calls
- Never send money or share OTPs based on a voice or video call alone — verify by calling back on a number you already know
- Be sceptical of any online "investment opportunity" with guaranteed returns promoted through AI-generated testimonials
Be a critical consumer of AI-generated content:
- Ask "could this be AI-generated or fake?" when you see emotionally charged content, especially near elections
- Use fact-checking tools: PIB Fact Check (pib.gov.in/factcheck), India Today Fact Check, AFP Fact Check India
- Reverse image search suspicious photos before sharing
Prepare your skills:
- If your current work involves routine, repetitive processing of information, plan to develop skills in problem-solving, communication, and technical AI use
- Learn to use AI tools as a productivity multiplier rather than viewing them only as a threat — people who use AI effectively will outcompete people who don't
- Encourage children to develop creativity, critical thinking, and inter-personal skills alongside technical skills
Engage with governance:
- Know that India is developing AI governance frameworks now — your awareness and voice as a citizen matters
- If your employer uses AI systems that affect your pay, promotion, or hiring, you have grounds to ask how those decisions are made
- If you encounter deepfakes of politicians or celebrities being used to manipulate, report them to the platform and to PIB Fact Check
For students and young workers:
- Fields with the best long-term prospects in an AI world: healthcare (especially community health and care), skilled trades, creative professions, AI development and oversight, education, agriculture (precision farming)
- Fields requiring adaptation: IT services (junior coders doing routine work), content writing (commodity work), basic customer service, basic accounting and data processing
- The skill that will matter most across all fields: judgment. What to do with information, how to weigh competing considerations, what matters and what doesn't. AI is very good at generating options; humans are needed to choose wisely
Sources
- World Economic Forum — Future of Jobs Report 2025
- McKinsey Global Institute — India's Future of Work, 2024
- World Bank — India Jobs Report, 2024
- Geoffrey Hinton — public statements and interviews, 2023–2026
- Anthropic — safety research and Constitutional AI
- EU AI Act — Official Journal of the European Union, 2024
- Bletchley Declaration on AI Safety, November 2023
- MEITY — IndiaAI Mission documentation and AI advisory framework, 2024–2026
- Goldman Sachs — The Potentially Large Effects of Artificial Intelligence on Economic Growth, 2023
- Stuart Russell — Human Compatible: Artificial Intelligence and the Problem of Control (book)
- MIT Technology Review — AI safety research coverage, 2024–2026
