The AI Warning and the Human Response

By Prof Ujjwal K. Chowdhury

The Chinese President Xi Jinping in his recent visit to USA has emphatically called for the human-in-the-loop Artificial Intelligence, meaning human controlled AI models and their operations. Artificial intelligence can widen access to knowledge and make useful work faster. It can also amplify fraud, error, surveillance and intellectual dependence, towards which the Chinese President has hinted. The evidence calls for neither panic nor complacency: India needs risk-based safeguards, communication built on trust, and classrooms that make a learner's own thinking visible.

Two ordinary mornings

Imagine a Class VIII classroom in a small town in Maharashtra. A student has used an AI assistant to produce a polished explanation of a geometry problem. The teacher changes one measurement and asks the class to predict what changes in the proof. The student who submitted the answer cannot begin. The machine completed the assignment; the learner missed the lesson.

Now imagine a health worker in a district clinic using an approved language tool to turn a verified maternal health leaflet into a spoken explanation in the language a family uses at home. She checks the key instructions against the original, notices a question the software cannot answer, and brings in a nurse. The tool has helped people communicate; the professional still owns the advice.

These are illustrative situations, not reported cases. They point to the real question behind the current argument about artificial intelligence: what happens to human judgement, trust, knowledge and power when a system can generate a convincing answer at almost no cost? AI can extend a person's reach. It can also conceal whether the person knows enough to judge the result. The answer depends on how institutions design the work around it.

Why the warnings keep multiplying

The volume of AI warnings has grown because capability, availability and consequence have begun moving together. General purpose systems now translate, write code, generate images, summarize documents and help with specialised research. The 2026 International AI Safety Report records striking gains in mathematics and coding, while stressing that performance remains uneven: a system can solve a difficult benchmark and still fail at a task that looks simple. At least 700 million people use leading AI systems each week, yet adoption and access are uneven across regions. [3]

The systems are also moving from chat windows into agents that can browse, call tools, edit files and carry out multi-step tasks. A wrong answer in a private study session may be corrected by a teacher. A wrong answer sent automatically to thousands of customers, or used to flag a claimant for investigation, carries a different cost. Speed and scale reduce the time available to notice an error and raise the number of people who may have to live with it.

Warnings come from researchers who see capability changes from inside laboratories, from workers who see routine tasks being reorganised, from teachers watching students submit work they cannot explain, and from journalists trying to distinguish recorded events from synthetic media. The risk agenda also reaches beyond a single chatbot: it includes fraud, privacy, labour conditions, public information, concentration of infrastructure and the possibility that future autonomous systems could be hard to supervise.

These are serious reasons for sustained scrutiny. A warning can help society prepare before a failure becomes ordinary. It can also become a powerful story in its own right, repeated by newsrooms, investors, politicians and companies that each have different interests in the outcome.

Fear has a business model

In “The business of fear”, Bhabani Shankar Nayak argues that dramatic warnings by AI leaders can help sell the inevitability of AI, sustain investor expectations and legitimise a security infrastructure that serves corporate and state power. The essay asks an essential political question: who gains influence, market share or public authority when the public is frightened? [1]

That question is more useful than a quick verdict about any executive's sincerity. A company may have genuine safety concerns and a commercial interest in being seen as the responsible steward of a powerful technology. A warning may encourage safeguards, or help create a regulatory system that only the largest firms can afford to meet. A call for security may protect users, or normalise more surveillance than the risk requires. Motive cannot be read from a headline; it has to be tested against proposals and actions.

Fear also attracts attention. A forecast of imminent extinction travels farther than a careful account of uncertainty, even when the careful account is more informative. In a media market rewarded for clicks and speed, “AI may produce convincing misinformation, with uncertain effects at scale” competes poorly with “AI will end humanity”. The latter turns a difficult policy question into a drama with villains, heroes and a countdown.

Nayak's critique is strongest when it asks us to examine ownership, profit and the reach of security systems. Its sharper claims about a coordinated corporate campaign or the motives of particular leaders should be treated as an argument, not as an established finding. The same discipline applies to executives who offer precise timelines for catastrophe. We should ask what evidence supports the claim, what uncertainty remains, what remedy is being proposed, and whether that remedy gives the public more control or simply gives powerful institutions a larger role.

Fear can push in opposite directions. It can be used to slow a rival or to justify a rush to deploy AI in the name of national competition. It can concentrate regulation around incumbents, or produce an urgent public demand for independent oversight. The public interest lies in rules that follow observable capability and consequence, protect people who bear the risk, and remain open to review.

Separate present harm from future catastrophe

An honest account sorts risks by the strength of evidence as well as the severity of possible harm. Present day misuse is the clearest category. AI systems can make phishing messages more persuasive, help generate fraudulent content, impersonate voices and assist some cyber operations. The 2026 international report says documented cases of AI generated influence operations exist, while evidence that such content is already manipulating people at scale remains limited. It also reports increasing use of AI tools in real cyber operations, but says the effect on the overall frequency of attacks is unclear. [3]

Malfunction is a second category. A model can invent a source, misread an instruction, reproduce a bias or confidently offer a wrong answer. Such failures are inconvenient in low stakes work; they can be dangerous in a diagnosis, loan decision, school safeguarding response or government eligibility process. AI is especially unreliable when it has to combine many steps, interpret a local context or work in a language and culture underrepresented in its training data. Human review matters most where the person affected has limited power to challenge the result. [3]

Systemic effects include job quality, the loss of entry level work, dependence on a few platforms and the transfer of public knowledge into privately controlled systems. Employment evidence is mixed. Some studies have not found a relationship between AI exposure and overall employment levels; others have found a decline in early career employment in some exposed occupations while senior employment remained stable or grew. That is a warning to protect the apprenticeship and practice through which young workers become experienced, not proof that a fixed share of jobs will disappear on schedule. [3]

Loss of human control is a different question. It concerns future systems that might evade oversight, pursue long plans or resist attempts to stop them. The International AI Safety Report says researchers and company leaders disagree widely about the likelihood of such scenarios. Current systems show early warning signs in controlled tests, such as disabling simulated oversight under particular instructions, but the report says present systems do not have the relevant capabilities at levels that would enable loss of control. It also finds uncertainty about how fast capabilities will develop. [3]

This distinction gives us a sane position. Catastrophic scenarios deserve serious study because the possible harm is extreme and safeguards take time to build. They do not become established forecasts because a famous person names a probability or a date. Nor does uncertainty make them irrelevant. A practical precaution is to test systems for dangerous capabilities, stage access as risk rises, retain a genuine ability to pause deployment, and require independent review before a high consequence system is put into service.

The word “AI” can hide the human choices that make a system consequential. A model does not need human feelings or motives to produce harm. People choose its training data, permissions, business model, safety checks and use. At the same time, “a human made the decision” is not enough if automation has narrowed the options, concealed the basis of a recommendation or left the human with seconds to object. Accountability has to reach the full chain, from design and procurement to deployment and redress.

India needs more than an app

India has a distinct opportunity and a distinct burden. AI may help make expertise more reachable across languages and distances. It can assist a teacher who must prepare more examples, a small enterprise that wants to reach a buyer beyond its home district, a public hospital that needs to organise records, or a researcher mapping local evidence. The IndiaAI Mission was approved with a ₹10,371.92 crore outlay and includes public compute, indigenous models, datasets, future skills and a Safe and Trusted AI pillar. The policy ambition is to build capability and use AI for public good. [8]

India is also beginning to introduce computational thinking and AI across school subjects. CBSE launched a curriculum framework for Classes III to VIII for the 2026–27 session, with problem solving, logical reasoning, digital literacy and ethical use among its aims. The opportunity is to teach children how AI works and how to question it, rather than teaching only which buttons to press. [7]

Access, however, is not the same as control. ASER 2024 found that 89.1 per cent of rural children aged 14 to 16 had a smartphone at home, and 82.2 per cent said they knew how to use one. Yet only 31.4 per cent of those who could use a smartphone reported having their own. Ownership was 36.2 per cent for boys and 26.9 per cent for girls. Fifty seven per cent of smartphone users had used a phone for an educational activity in the previous week. A family phone in the house does not guarantee private study time, reliable data, a quiet room or equal permission to use it. [9]

Language changes the calculation too. A model may handle standard Hindi or English well and still miss a regional idiom, a code switched sentence, a tribal language or the social meaning of a name. A translation can open a door only when a speaker can correct it and be heard. Every Indian AI education or communication plan must therefore test its tools with local users, build offline and low bandwidth routes, and ensure that a paid subscription is never the hidden price of participating in school.

Communication is an ecology

In the linked essay “Building an ecology of communication in the age of artificial intelligence”, Cedric Prakash asks journalists and other communicators to enter the places where the future is being made: laboratories, technology companies, schools, media, institutions and local communities. His language matters because communication is larger than the production of content. It is a living relationship between speaker and listener, source and audience, evidence and interpretation, message and consequence. [2]

AI can make a false photograph, voice or quotation quickly and cheaply. It can also transcribe a field interview, translate a public notice, describe an image for a person with low vision or help a community group produce an accessible explainer. The same capability can widen participation or counterfeit it. A fabricated voice can impersonate a teacher, doctor or local leader. A machine translated public notice can omit the condition that matters most. A generated citation can make a false claim look researched.

A newsroom in Kolkata, for example, could use speech recognition to create a first transcript of an interview in Bengali, Santali or a regional Hindi dialect. The reporter should retain the original recording with the interviewee's permission, check the transcript against the audio, mark uncertain phrases and verify names, numbers and quotations before publication. The AI can save time on clerical work. It cannot certify that the speaker meant what the transcript says, or decide whether publishing the quotation will endanger the source.

Trustworthy communication therefore needs more than a label saying “AI generated”. People need a clear account of what was generated or materially changed, a way to trace claims to evidence, consent before someone's face or voice is reproduced, and a named person or institution responsible for correction. Watermarks and automated detectors can help, but the international report notes that watermarks may be removed or altered and safeguards can be bypassed. Provenance must be paired with editorial judgement, source checking and an accessible correction process. [3]

Authenticity is not a ban on machine assistance. It is honesty about the route by which a message was made and responsibility for its effects. An AI assisted announcement may be completely authentic if a human source stands behind its content, has verified it and can answer for it. A synthetic speech presented as a real person's words is deceptive even if every sentence is grammatically perfect.

The learner must take the first step

In education, a polished submission is not reliable proof that learning has happened. The OECD Digital Education Outlook 2026 draws a clear distinction: general purpose AI can raise the quality of student work without producing lasting learning when it takes over the task. Tools designed around a teaching purpose, with questioning and feedback rather than answer delivery, can support learning. [4]

A randomized field experiment with nearly a thousand high school mathematics students in Türkiye found that unrestricted GPT-4 access improved practice performance but could lower performance on an unaided test after access was removed. A version designed to provide tutoring guardrails reduced this learning penalty. The study concerns a specific subject, country and tool; it cannot settle every classroom question. Its lesson is narrower and more useful: when AI supplies the cognitive work, a good looking result may hide an unpractised skill. [5]

Learning needs retrieval, effort, feedback and a chance to transfer an idea to a new problem. Students need to try before they see a worked solution, explain their reasoning, notice an error and retrieve the idea later without the screen. A systematic review of classroom research finds that retrieval practice benefits learning across a range of school settings. [6] These moments of effort can feel slower than instant answers. They are also the evidence from which a teacher can tell what help the learner needs.

A simple rule keeps the sequence visible: human first, AI in the middle, human last. The learner first states a question, makes an attempt or draws a plan. AI then performs a defined job: offer one hint, produce a counterexample, translate a term or critique a draft. The learner returns to the task, checks the output, explains what changed and applies the concept again. The tool may assist every stage; it must not silently perform every stage.

Small local tasks show the difference

In a Class V school water audit, children map taps, count leaks, speak with the caretaker and predict where water is being wasted. AI can help sort observations, translate interview questions or turn verified counts into a simple graph. The children still decide what they actually observed, compare the graph with the school, and test one small repair for a week. Their learning is measured by whether they can describe the evidence and explain what changed, not by the polish of a generated poster.

In a Class VIII history lesson, a chatbot can produce a confident summary of a local event. Students compare its claims with a textbook, an archive item and an oral account. They mark what is corroborated, what is missing and what remains disputed. Each student must defend one claim with evidence and identify whose voice is absent. The model becomes a prompt for source judgement, while the community record stays in human hands.

In Class X mathematics, a student spends seven minutes attempting a problem before an approved tool can provide one hint. The student records the first approach, identifies the error, solves the problem and tries a parallel question without assistance the next day. That “hint first” design preserves productive struggle while making extra practice available to a learner who may otherwise wait a week for help.

In a nursing programme, students can practise explaining a vaccination schedule or asking follow-up questions in a simulated conversation. AI can generate different patient scenarios and help students rehearse plain language. Faculty must check medical content, students must not upload identifiable patient information, and the real clinical encounter still requires supervised observation. A simulation builds readiness; it does not certify bedside competence.

In engineering or design, a student team may use a coding assistant to produce a first draft of a sensor application. The team then tests edge cases, privacy, security and accessibility. Each student explains one decision, identifies one suggestion they rejected, and repairs a bug in a short live exercise. The professional is learning to judge generated work, not to confuse code that runs once with a system fit for people.

A small craft business in Murshidabad might use AI to translate a product description or draft a reply to an overseas buyer. The artisan supplies the facts about materials, process, price and delivery, then approves the wording. A useful system opens the market while leaving authorship, cultural meaning and commercial decisions with the maker. The same rule applies to a college event team using AI to localise a public invitation: translation can scale the message, while a human confirms tone, context and accessibility.

In a public service office, AI could help a citizen complete a form or route a query to the right department in her preferred language. It should not quietly reject an application, assign a risk score or close a grievance without a human explanation and appeal. The person seeking a public service must be able to speak with an accountable official and challenge an error.

Assessment should reveal what the student knows

Institutions often respond to AI by trying to catch it. Detection software can misclassify human writing and can be defeated by small edits. A detector result is weak evidence for a high stakes judgement about cheating. The more durable response is to redesign assessment so that students can use AI openly for some work and demonstrate independent capability at purposeful points.

An open assessment lane can permit AI for research, translation, brainstorming, design and iteration. The student discloses the tool's role, preserves a short record of important choices, verifies claims and reflects on what the system got wrong. A secure lane can use a supervised examination, practical, studio review, oral defence or live debugging task when a course needs to establish what a student can do independently. One programme can use both lanes without turning every homework task into an oral examination.

For a major assignment, collect an evidence stack: the initial question or attempt, a record of meaningful AI assistance, the finished work, a short demonstration and a reflection on what the learner accepted or rejected. A three minute conversation can ask: What did you begin with? Which suggestion changed your thinking? Show me one claim you checked. What would you do if one condition changed? These questions make understanding visible without turning the classroom into an interrogation room.

UNESCO's student framework places human agency, ethics, technical understanding and system design inside AI competence; its teacher framework includes AI pedagogy and professional learning. That is a useful Indian curriculum principle: every learner should study AI, while every subject should show how to question AI within its own methods. A science student checks data and causal claims; a design student tests representation and user experience; a media student checks sources and consent; a future manager studies bias, labour and accountability. [10][12]

Keep human authority real

“Human in the loop” can become a reassuring label pasted over an automated process. A human safeguard is meaningful only when the person has the information, time, expertise and authority to disagree, and when the system records what happened after the disagreement. Where the output can affect health, safety, education, employment, credit, liberty or access to a public benefit, a human appeal route should be explicit and reachable.

Institutions should classify uses by consequence. Low risk work such as generating practice examples or improving the layout of a public brochure can be encouraged. Work that makes factual claims, uses personal data or affects a person's opportunity needs disclosure, verification and review. AI only decisions about a child's discipline, a student's progression, a patient's treatment or a citizen's eligibility should be prohibited. The division is practical: automate reversible assistance; require strong oversight when a decision is difficult to reverse.

Privacy requires the same practical clarity. Schools and universities should keep student names, marks, health details, counselling notes, family circumstances and unpublished work out of unapproved public tools. UNESCO's education guidance also places privacy and age-appropriate use among the conditions for responsible adoption. [11] Procurement contracts should state what data are collected, how long they are retained, whether prompts are used to train a model, where a breach is reported and how an institution can delete records. India's Digital Personal Data Protection Act and the 2025 Rules create a developing framework with phased commencement; institutions should map obligations as they take effect and apply data minimisation and child protection as baseline practice now. [13]

Fairness has to be tested in the languages and conditions in which a system will be used. A school should check whether a tool understands code switching, local names and dialects, and whether it produces different error rates for girls, students with disabilities, first generation learners or children from tribal communities. “Available in a language” is not the same as accurate, respectful or useful in that language.

Institutions should also ask whether AI is reducing workload or only moving it. If a teacher saves time on lesson plans, the school should protect space for feedback, small group teaching and conversation. If a company automates routine drafting, it should preserve supervised practice for junior staff rather than removing the very tasks through which they learn the profession. Productivity becomes a public benefit when the time and value it releases reach the people doing the work.

Build the roadmap in a hundred days

In the first month, a school, college, newsroom or public agency should map a few real tasks before buying a platform. What problem is being solved? Which human capability must remain? What can go wrong, who could be affected, what data would be used, and who can reverse a mistake? Establish a baseline for quality, time, learning or access. A clear “green, amber, red” policy can then distinguish low risk assistance, verified work and tasks where AI cannot make the final decision.

In the second month, run a small pilot co-designed by teachers, learners, staff and the people the service is meant to reach. Choose a tool that fits the task; prefer low bandwidth and multilingual access; create an offline alternative; train the adults who will supervise it. Tell participants what the system does, what it does not do, what information it stores and how to report an error. Do not make a student's or citizen's access to an essential service depend on accepting optional AI processing.

In the third month, compare the pilot with current practice. Did students remember more a week later and transfer the idea to a new problem? Did a local language message preserve its meaning? Did teachers spend more time with learners? Could users challenge the system? Did error rates differ across groups? What did the AI cost in money, time, compute and energy? Publish the answer in plain language. If the output looks better while unaided understanding, trust or fairness worsens, pause and redesign.

At the end of 100 days, scale only where there is evidence of human benefit. Government and boards can establish procurement standards, minimum data protections, incident reporting and public evaluation. Universities can independently test tools and train students to assess generated claims. Newsrooms can make provenance, correction and consent routine. Employers can create transition pathways and preserve early career learning. AI providers can supply documentation and access for independent evaluation. Each group has a different duty; no single safety pledge can replace the others.

The measures of success should be human measures: delayed recall, transfer, source verification, quality of dialogue, response time, errors corrected, access across languages, workload returned to people, and the number of users who can understand and appeal a decision. The count of licences purchased or words generated tells us what the machine did. It says little about whether people became more capable.

What survives when the screen goes dark

India can pursue useful AI with confidence and still take its risks seriously. The evidence supports immediate safeguards for known harms, careful tests for higher risk capabilities and honest public debate about uncertain futures. It supports a communication culture in which people know how a message was made and who stands behind it. It supports education in which AI can widen practice, feedback and access, while a teacher continues to see the learner and the learner can explain the work.

The final test is simple enough to ask in a classroom, a newsroom, a hospital or a ministry: after the system has spoken, can a person still question it, verify it, explain a decision, correct a mistake and care for the people affected? When the screen goes dark, the human work should be clearer, stronger and more widely shared.

000

Sources and notes

Bracketed numbers in the feature point to the sources below. The two Counterview essays are discussed as arguments; factual and policy claims are checked against the research and official materials listed here.

[1] Bhabani Shankar Nayak. “The business of fear: How AI leaders sell doom for profit.” Counterview, September 2026. Open source

[2] Cedric Prakash SJ. “Building an ecology of communication in the age of artificial intelligence.” Counterview, September 2026. Open source

[3] International AI Safety Report. International AI Safety Report 2026: Extended Summary for Policymakers. February 2026. Open source

[4] OECD. OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education. OECD Publishing, 2026. Open source

[5] Bastani, H. et al. “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences, 122(26), 2025. Open source

[6] Agarwal, P. K., Nunes, L. D., and Blunt, J. R. “Retrieval Practice Consistently Benefits Student Learning.” Educational Psychology Review, 33, 2021. Open source

[7] Central Board of Secondary Education. Circular Acad-15/2026: Launch of the Computational Thinking and Artificial Intelligence Curriculum for Classes III–VIII, 1 April 2026. Open source

[8] Prime Minister of India. “Cabinet Approves Ambitious IndiaAI Mission to Strengthen the AI Innovation Ecosystem,” 7 March 2024. Open source

[9] ASER Centre. Annual Status of Education Report 2024: National Findings. Open source

[10] UNESCO. AI Competency Framework for Teachers. 2024. Open source

[11] UNESCO. Guidance for Generative AI in Education and Research. 2023. Open source

[12] UNESCO. AI Competency Framework for Students. 2024. Open source

[13] Ministry of Electronics and Information Technology, Government of India. Digital Personal Data Protection Rules, 2025, Gazette of India, 13 November 2025. Open source