Nibeditaa SpeaksHow India can separate credible AI risk from alarm—and build trusted communication and lasting learning
By Prof. Ujjwal K. Chowdhury
Artificial intelligence has reached a point where the central question is no longer whether machines can perform useful tasks. It is whether humans will retain meaningful control over the decisions, knowledge and systems built around them.
The recent emphasis by Chinese President Xi Jinping on human-in-the-loop artificial intelligence—models and operations that remain subject to human control—captures one side of this debate. AI can widen access to knowledge and make work faster. It can also amplify fraud, error, surveillance and intellectual dependence.
India therefore needs neither panic nor complacency. It needs risk-based safeguards, trusted communication and an education system that continues to make a learner’s own thinking visible.
Two ordinary mornings
Consider a Class VIII classroom in a small town in Maharashtra. A student uses an AI assistant to produce a polished explanation of a geometry problem. The teacher changes one measurement and asks the class to predict what happens to the proof. The student who submitted the answer cannot begin.
The machine completed the assignment. The learner missed the lesson.
Now consider 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, encounters a question the software cannot answer and brings in a nurse.
Here, the technology has helped communication. The professional remains responsible for the advice.
These are illustrative situations, not reported cases. But they point towards the central question of the AI debate: what happens to human judgment, trust, knowledge and power when a system can produce a convincing answer at almost no cost?
AI can extend a person’s reach. It can also conceal whether that person knows enough to judge its output. The consequences depend largely on how institutions design the work around the technology.
Why the warnings keep multiplying
Warnings about AI have multiplied because capability, availability and consequence are increasingly moving together.
General-purpose systems can now translate, write code, generate images, summarise documents and assist with specialised research. The 2026 International AI Safety Report records significant gains in mathematics and coding while also stressing that performance remains uneven. A system can succeed on a difficult benchmark and still fail at a task that appears simple.
At least 700 million people use leading AI systems each week, although adoption and access remain uneven across regions.
The technology is also moving beyond conventional chat interfaces towards agents capable of browsing, calling tools, editing files and carrying out multi-step tasks.
The difference matters.
An incorrect answer in a private study session can be corrected by a teacher. An incorrect answer automatically sent to thousands of customers—or used to flag someone for investigation—can have consequences on an entirely different scale.
Speed and scale leave less time to detect mistakes and increase the number of people affected by them.
The warnings now come from multiple directions: researchers observing capability changes inside laboratories; workers watching routine tasks being reorganised; teachers seeing students submit work they cannot explain; and journalists trying to distinguish authentic events from synthetic media.
The risk agenda extends beyond chatbots. It includes fraud, privacy, labour conditions, public information, concentration of infrastructure and the possibility that future autonomous systems could become difficult to supervise.
These are legitimate reasons for scrutiny. A warning can help society prepare before a failure becomes normal.
But warnings themselves can also become powerful narratives, amplified by news organisations, investors, politicians and technology companies with different interests in the outcome.
When fear becomes part of the AI economy
In The Business of Fear, Bhabani Shankar Nayak argues that dramatic warnings from AI leaders can help sell the inevitability of AI, sustain investor expectations and legitimise security infrastructure serving corporate and state power.
The argument raises an important political question: who gains influence, market share or public authority when people are frightened?
That question is more useful than making quick judgments about an executive’s sincerity.
A company may have genuine safety concerns while also having a commercial interest in presenting itself as the responsible steward of powerful technology. A warning may encourage safeguards, but it may also contribute to a regulatory structure that only the largest companies can afford to meet.
Likewise, a call for security may protect users while potentially normalising surveillance beyond what the risk requires.
Motives cannot be established from headlines. They have to be examined through proposals and actions.
Fear also has an obvious advantage in the attention economy. A forecast of imminent extinction travels farther than a careful explanation of uncertainty. “AI will end humanity” turns a complicated 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 expansion of security systems. His sharper claims about coordinated corporate campaigns or the motives of particular leaders should, however, be treated as arguments rather than established findings.
The same discipline should apply 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 does that remedy give the public greater control—or simply give powerful institutions a larger role?
Separate present harm from future catastrophe
An honest assessment of AI risk should distinguish between the strength of current evidence and the severity of possible future harm.
Present-day misuse is the clearest category.
AI can make phishing messages more persuasive, generate fraudulent material, impersonate voices and assist some cyber operations. The 2026 International AI Safety Report documents cases of AI-generated influence operations, while noting that evidence of such content already manipulating people at scale remains limited.
It also reports increasing use of AI tools in real cyber operations, although the overall effect on the frequency of attacks remains unclear.
A second category is malfunction.
AI systems can invent sources, misunderstand instructions, reproduce biases or confidently provide incorrect answers. Such failures may be inconvenient in low-stakes work but dangerous when they affect diagnosis, lending, school safeguarding or government eligibility.
The problem becomes particularly serious when AI is used in settings where the affected person has limited ability to challenge the outcome.
Systemic effects raise another set of questions: job quality, the loss of entry-level work, dependence on a small number of platforms and the transfer of public knowledge into privately controlled systems.
Employment evidence remains mixed. Some studies have found no relationship between AI exposure and overall employment levels, while others have identified declines in early-career employment in certain exposed occupations.
This is reason to protect the apprenticeship and practical experience through which young workers become professionals. It is not evidence that a fixed percentage of jobs will inevitably disappear on a predetermined schedule.
Then comes the question of human control.
Future systems could potentially evade oversight, pursue long-term objectives or resist attempts to stop them. Researchers and technology companies disagree widely about the likelihood of such scenarios. Current systems show some early warning signs in controlled tests, but they do not presently possess the relevant capabilities at levels that would amount to a demonstrated loss of control.
That uncertainty should not lead either to dismissal or certainty.
Catastrophic scenarios deserve serious study because their potential harm is enormous and safeguards take time to build. But a scenario does not become an established forecast simply because a prominent person assigns it a probability or date.
A practical response is to test systems for dangerous capabilities, stage access as risks increase, retain the ability to pause deployment and require independent review before high-consequence systems enter service.
The human is still part of the system
The word “AI” can conceal the human decisions that make a system consequential.
A model does not need human feelings or motives to cause harm. People determine its training data, permissions, business model, safety checks and applications.
At the same time, saying that “a human made the decision” is insufficient if automation has already narrowed the available choices, concealed the reasoning behind a recommendation or left the human with only seconds to object.
Accountability must extend across the entire chain—from design and procurement to deployment and redress.
India needs more than an app
India has both an opportunity and a responsibility.
AI could make expertise more accessible across languages and distances. It could assist teachers preparing examples, small businesses reaching customers beyond their districts, public hospitals organising records and researchers mapping local evidence.
The IndiaAI Mission, approved with an outlay of ₹10,371.92 crore, includes public compute, indigenous models, datasets, future skills and a Safe and Trusted AI pillar. Its stated ambition includes building national capability and using AI for public good.
India is also beginning to introduce computational thinking and AI across school subjects. The CBSE curriculum framework for Classes III to VIII for 2026–27 includes problem-solving, logical reasoning, digital literacy and ethical use among its objectives.
The opportunity is to teach children not merely which buttons to press, but how AI works and how to question it.
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 device. Ownership was 36.2 per cent among boys compared with 26.9 per cent among girls.
A phone in the household does not automatically mean private study time, reliable connectivity, a quiet room or equal permission to use the device.
Language presents another challenge.
An AI model may handle standard Hindi or English reasonably well while struggling with a regional idiom, code-switched sentence, tribal language or the social meaning of a name.
Translation can open a door only when the person using it can correct the translation and be heard.
India’s AI education and communication strategies must therefore account for local users, multilingual realities, low-bandwidth environments and offline access. A paid subscription should never become the hidden price of participating in education.
Communication is an ecology
In Building an Ecology of Communication in the Age of Artificial Intelligence, Cedric Prakash calls on journalists and communicators to engage with the places where the future is being shaped: laboratories, technology companies, schools, media organisations, institutions and communities.
Communication is more than content production.
It is a relationship between speaker and listener, source and audience, evidence and interpretation, message and consequence.
AI can create a false photograph, voice or quotation quickly and cheaply. It can also transcribe a field interview, translate a public notice, describe an image for someone with low vision or help a community organisation 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 a crucial condition. A generated citation can make an unsupported claim appear researched.
For a newsroom, AI-assisted transcription can save time. But reporters still need to retain the original recording with the interviewee’s permission, check the transcript against the audio and verify names, numbers and quotations.
The machine can assist with clerical work. It cannot determine whether a speaker meant what the transcript says or whether publishing a quotation could endanger a source.
Trustworthy communication therefore requires more than an “AI-generated” label.
People need to know what was generated or materially changed, how claims can be traced to evidence, whether consent was obtained before someone’s face or voice was reproduced, and who is responsible for correcting an error.
Watermarks and automated detectors can help, but neither is sufficient on its own. Provenance must be combined with editorial judgment, source verification and accessible correction mechanisms.
Authenticity does not require banning machine assistance.
It requires honesty about how a message was produced and accountability for its consequences.
The learner must take the first step
Education may be the most important arena in which this distinction becomes visible.
A polished submission is not necessarily evidence that learning has occurred.
The OECD Digital Education Outlook 2026 distinguishes between general-purpose AI, which can improve the quality of student work without necessarily producing lasting learning, and tools designed around teaching, questioning and feedback.
A randomised field experiment involving nearly 1,000 high-school mathematics students in Türkiye found that unrestricted GPT-4 access improved practice performance but could reduce performance on an unaided test after access was removed. A version designed with tutoring guardrails reduced that learning penalty.
The study concerns a specific subject, country and tool. It cannot settle every classroom question.
Its narrower lesson is nevertheless important: when AI performs the cognitive work, a polished answer may conceal an unpractised skill.
Learning requires retrieval, effort, feedback and the opportunity to transfer an idea to a new problem.
Students need opportunities to attempt a problem, explain their reasoning, identify errors and later retrieve the concept without the screen.
A simple principle can preserve that sequence:
Human first. AI in the middle. Human last.
The learner begins by stating the question, making an attempt or developing a plan. AI can then perform a defined task—offer a 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 machine may assist throughout the process. It should not silently perform the entire process.
Assessment must reveal what students know
Institutions often respond to AI by trying to detect it.
But detection software can misclassify human writing and can sometimes be defeated by small changes. A detector result is therefore weak evidence for a high-stakes decision about cheating.
A more durable response is to redesign assessment.
Students can be permitted to use AI openly for research, translation, brainstorming, design or iteration, while being required to demonstrate independent capability at purposeful points.
An open assessment lane could require students to disclose AI use, preserve a short record of important decisions, verify claims and reflect on what the system got wrong.
A secure assessment lane could involve supervised examinations, practicals, studio reviews, oral defences or live debugging where independent capability needs to be established.
The two approaches can coexist.
For major assignments, institutions can collect an evidence stack: the student’s initial attempt, a record of meaningful AI assistance, the finished work, a short demonstration and a reflection explaining what the learner accepted or rejected.
The central question is not whether a student touched AI.
It is whether the student still understands the work.
Keep human authority real
“Human in the loop” can easily become a reassuring label placed 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 happens after that disagreement.
Where an AI output can affect health, safety, education, employment, credit, liberty or access to a public benefit, there should be an explicit and accessible human appeal mechanism.
Institutions should classify AI uses according to consequence.
Low-risk activities, such as generating practice examples or improving the layout of a brochure, can be encouraged. Work involving factual claims, personal data or a person’s opportunities requires 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 not be permitted.
The practical principle is simple: automate assistance that can be reversed; impose stronger oversight when a decision is difficult to reverse.
Privacy requires equal clarity.
Schools and universities should keep student names, marks, health information, counselling notes, family circumstances and unpublished work out of unapproved public AI tools.
Procurement agreements should specify what data are collected, how long they are retained, whether prompts are used to train models, how breaches are reported and how institutions can delete records.
Fairness must also be tested in the languages and circumstances in which AI will actually operate.
Institutions should examine whether systems understand code-switching, local names and dialects, and whether error rates differ among girls, students with disabilities, first-generation learners and children from tribal communities.
Being “available in a language” does not automatically make a system accurate, respectful or useful in that language.
A 100-day roadmap
The first month should begin not with buying a platform but with mapping real tasks.
What problem is the technology solving? Which human capability must remain? What could go wrong? Who could be affected? What data would be used? Who can reverse an error?
Institutions can establish a baseline for quality, time, learning or access and create a simple green-amber-red policy distinguishing low-risk assistance, verified work and tasks where AI cannot make the final decision.
The second month should involve a small pilot designed with teachers, learners, staff and the people who will use the service.
Tools should be selected according to the task, with attention to multilingual and low-bandwidth access. Offline alternatives should be available, and adults responsible for supervision should be trained.
Participants should know what the system does, what it does not do, what information it stores and how an error can be reported.
In the third month, institutions should compare the pilot with existing practice.
Did students remember more a week later? Could they transfer the idea to a new problem? Did a local-language message retain its meaning? Did teachers spend more time with learners? Could users challenge the system? Did error rates vary across groups? What did the AI cost in money, time, computing and energy?
If the output looks better while unaided understanding, trust or fairness worsens, the system should be paused and redesigned.
After 100 days, scaling should occur only where there is evidence of human benefit.
Government and education boards can establish procurement standards, data protections, incident reporting and public evaluation. Universities can independently test AI tools and train students to assess generated claims. Newsrooms can make provenance, correction and consent routine. Employers can create transition pathways while protecting early-career learning.
No single safety pledge can replace these responsibilities.
What survives when the screen goes dark?
India can pursue useful AI with confidence while taking its risks seriously.
The evidence supports immediate safeguards against known harms, careful testing of higher-risk capabilities and an honest public conversation about uncertain futures.
It also supports a communication culture in which people understand how a message was created and who stands behind it.
And it supports an education system in which AI can widen practice, feedback and access while teachers continue to see the learner—and learners remain capable of explaining their work.
The final test is simple enough to apply in a classroom, newsroom, hospital or ministry:
After the system has spoken, can a person still question it, verify it, explain the decision, correct the mistake and care for the people affected?
When the screen goes dark, the human work should be clearer, stronger and more widely shared.
Add comment