Between July and September 2025, I interviewed 42 high school students as part of ongoing research into career aspirations and problem-solving. One pattern kept recurring, and it unsettled me more each time I saw it.

Students who regularly used AI tools gave curiously superficial answers to everyday problems. They would address the obvious aspect of a question, then stop, almost as if waiting for a “next suggestion” that never came. But when I asked simple follow-up questions such as “What other possibilities exist?” or “Is there an alternative perspective?”, something shifted. They began weighing trade-offs, generating options, thinking out loud in ways they hadn’t a moment before. The capacity for deeper analysis was clearly there. They just weren’t reaching for it on their own.

AI and Critical Thinking: What Studies Reveal

This pattern aligns with what researchers call cognitive offloading, one of the more quietly consequential effects of living alongside AI. The term describes our long-standing habit of delegating mental tasks to external tools like notebooks, calculators, and calendars to lighten cognitive load. But AI changes the nature of the offload. A calculator stores a number. AI generates the finished thought. When a tool not only holds information but produces the answer itself, students can begin expecting ready-made solutions rather than exercising their own problem-solving capacity. What is meant to scaffold learning risks becoming a crutch that quietly atrophies the very skills it was meant to build.

This is not a fringe worry. A 2025 mixed-methods study of 666 participants across age groups found a significant negative relationship between frequent AI tool use and critical thinking performance, with cognitive offloading identified as the mediating mechanism, and the effect most pronounced among younger, heavier users. My 42 interviews cannot establish that kind of statistical relationship on their own, but they offered a close-up view of the same mechanism in action, most visibly in that pause where a student seemed to be waiting for a “next suggestion” that never came.

A 2026 study on generative AI and learner agency distinguishes between two very different modes of AI use, dependent offloading, where the tool substitutes for a student’s own thinking, and autonomous offloading, where AI scaffolds thinking without replacing it. The dependent form threatens autonomy by making choices on the student’s behalf, even when it may feel helpful. A parallel systematic review frames the same divide as amplification versus substitution, where guided, metacognitively aware AI use extends a student’s cognitive reach, while passive, unreflective use displaces the mental effort deep learning actually requires. My follow-up questions seemed to pull a few students from dependent mode into autonomous mode. The capacity for deeper analysis was clearly still there. They just were not reaching for it on their own.

From McDonald’s to the Algorithm

Thirty years ago, sociologist George Ritzer described what he called McDonaldization, the spread of fast-food logic, efficiency, calculability, predictability, and control, into hospitals, universities, even relationships. It was a troubling process, but at least it was visible. We could see the golden arches multiplying, see the assembly line replacing the artisan.

What we are living through now is harder to see, because it inverts the logic. Call it AI-zation. Where McDonaldization imposed visible uniformity, AI-zation offers something that feels like the opposite — personalisation. Our search results are tailored to us. Our feed reflects our interests. Our recommendations are, supposedly, uniquely ours.

This is the seduction of AI-zation. It feels personal while quietly manufacturing a new kind of conformity, not the visible sameness of a fast-food counter, but an invisible narrowing of thought, preference, and possibility. Recommendation systems are trained on existing patterns, so they inevitably steer people toward what is already popular, already “successful,” already proven. The menu looks infinite. The actual range of what people consume keeps narrowing.

I think of this as programmed spontaneity, the feeling of free choice operating inside an algorithmically constrained space. The pattern is easiest to notice somewhere trivial. Watch one dance reel, and the next fifty look almost identical to it. The platform has correctly identified what will keep you watching and, in doing so, has quietly narrowed the world to a single, endlessly repeated genre. Nobody decided this for us, exactly, and yet our options shrank the moment we engaged.

The same logic operates in domains that matter far more than reels, and it does not require any AI to appear at all. As a professor who has placed students in internships across dozens of organisations, I have watched a version of this play out for years. A student’s developmental need is exposure to something unfamiliar; if they have already worked in one area, the internship that would serve them best is often in a different one entirely, so they build range rather than repetition. But host organisations want the opposite. If a student has prior exposure to, say, employee engagement work, an organisation wants exactly that student, because they arrive already useful and need less onboarding. The organisation optimises for efficiency, minimum ramp-up, and maximum immediate output. The student needs breadth; the system rewards depth in a single, already-proven groove. Multiply that logic across a career, and a person can end up highly efficient at one narrow thing and never discover the other things they might have been.

This is the same tension recommendation algorithms formalise and accelerate, not invent. Efficiency, for any system, means doing more of what has already worked. Development, for a person, means doing something not tried yet. AI-zation is what happens when that older institutional logic gets encoded into infrastructure that operates continuously, at population scale, and largely out of sight.

Consider a career platform’s job suggestions, or a course recommendation engine, operating on the same principle as the internship market. It is optimising for patterns it has already seen. If your interests do not fit an existing category, the system is unlikely to help you find your way there, and the more you see what “people like you” are doing, the more that pattern starts to feel natural rather than constructed.

Why This Should Worry India Specifically

India has long been a civilisation organised around multiplicity, languages, philosophical schools, and ways of solving problems coexisting, often in productive tension. That diversity was never just decoration. It was, and is, epistemological, made of different ways of knowing and different definitions of a life well lived.

AI-zation threatens this precisely because it operates through standardisation disguised as personalisation. Search algorithms are globally standardised. Ed-tech platforms structure content around what optimises engagement metrics, not around pedagogical diversity. Jugaad, the distinctly Indian capacity for improvisation, depends on encountering a problem the “official” system has not already solved, and then building a workaround. But if an algorithm is always ready with the “right” answer before a student has fully sat with the question, where does that improvisational instinct come from?

This is not a case for rejecting technology. Digital tools have brought genuine benefit, access to information, connection, and efficiency gains that matter enormously in a country still building out infrastructure

This is not a case for rejecting technology. Digital tools have brought genuine benefit, access to information, connection, and efficiency gains that matter enormously in a country still building out infrastructure. The distinction that matters is between technology that augments human capability and technology that quietly programs human behaviour. Increasingly, we are getting more of the latter than the former.

The Illusion of Control

Defenders of algorithmic systems often point out that users retain control, that you can adjust settings, opt out, switch platforms. This misses something important. When algorithms mediate access to jobs, credit, education, and healthcare, individual opt-out becomes practically impossible for most people. And the “preferences” we express are themselves shaped by prior exposure. You keep choosing certain content partly because the algorithm keeps showing it to you. The system trains you as much as you train it.

The same illusion holds for the reasons AI systems offer when they do explain themselves. Research on explainable AI has repeatedly found that these explanations are often generated after the decision, plausible stories rather than faithful accounts of how the system actually arrived at its answer, and that even explanations carrying no real information can produce as much user trust as genuine ones. Demanding explanations from algorithms is necessary, but it is not sufficient. A system can learn to produce a persuasive explanation without that explanation being true.

Scale up the pattern from my student interviews and ask what happens when algorithms do our remembering (search), our navigating (maps), our reading choices (feeds), and increasingly our writing (generative AI). Each instance looks helpful in isolation. Together, they add up to something closer to the outsourcing of cognition itself.

What Can Actually Be Done

Four responses seem worth taking seriously, none of them rejecting technology.

First, algorithmic literacy needs to become collective, not just individual. When people understand that their “personalised” experience is shaped by hidden, profit-driven choices, they can begin to question the pattern rather than simply live inside it. For educators specifically, this means treating AI tools as objects of critical scrutiny in the classroom, not just productivity aids, and explicitly teaching students to pause and probe past the first answer, the way my follow-up questions did in those interviews.

Second, we need to protect spaces of deliberate non-optimisation. Not everything should be made efficient. Deep learning requires struggle. Creativity requires wandering. We need to consciously build and defend spaces, in classrooms, in workplaces, where algorithms do not intrude by default.

Third, India already has a structural alternative worth naming directly. The Open Network for Digital Commerce (ONDC) is public infrastructure built on the same logic that made UPI transform payments, an open protocol that keeps any single company from owning the whole stack of app, algorithm, and data. It has scaled fast, past 500 million transactions by mid-2026. Adoption is still uneven outside metro cities, but this is what building alternative infrastructure looks like in practice.

Cooperative ownership is a related, distinct model, workers or citizens owning the platform itself. Europe’s Smart cooperative serves over 100,000 freelancers; Switzerland’s MIDATA lets citizens govern their own health data. New York’s Drivers Cooperative is the cautionary case. Launched as a driver-owned Uber alternative, it now struggles because collective ownership has to compete with venture capital willing to lose money for years to win the market. Whether India’s cooperative tradition can extend into education technology or data governance, alongside infrastructure like ONDC, remains an open question.

Fourth, algorithmic systems that affect access to opportunity need to be explainable and challengeable. This requires regulation, but it also requires organised public demand, because voluntary transparency from platforms whose business model depends on opacity is not something to wait for.

The Stakes

Mahatma Gandhi’s idea of swadeshi, self-reliance rather than dependence on external systems, has an obvious digital-age analogue, something like cognitive sovereignty, the capacity to think, choose, and imagine outside the boundaries an algorithm has already drawn. In education specifically, this means treating AI as scaffolding to be gradually withdrawn, not a permanent support to lean on.

We are not choosing between technology and tradition. We are choosing who controls the cognitive infrastructure young people grow up inside, and whether that infrastructure preserves the messy, effortful, sometimes inefficient work of actually thinking something through. My conversations with those 42 students suggested the capacity for that kind of thinking has not gone anywhere. It just needs to be asked for.

EDITOR’S FACT-CHECK

Key claims independently verified against primary sources: Gerlich (2025), Societies 15(1):6, on cognitive offloading and critical thinking (n=666); Zhu et al. (2026), Frontiers in Psychology, on dependent vs. autonomous cognitive offloading; ONDC’s 500-million-transaction milestone (reported July 2026); Smart cooperative’s membership figures; and the trajectory of New York’s Drivers Cooperative.

India’s renewable-energy transition will require more than adding solar panels and wind turbines. As variable renewable power expands, the electricity system will also need ways to store power when generation is high and release it when demand rises or renewable output falls. That makes the cost and availability of batteries increasingly important.

The global lithium-ion battery market exceeded $150 billion in 2025, growing by more than 20% from the previous year, according to the International Energy Agency (IEA). Batteries are now important not only for electric vehicles but also for electricity-grid flexibility and backup power.

But the dominance of lithium-ion technology also creates supply-chain concerns. Different lithium-ion chemistries use different combinations of materials, including lithium, graphite, nickel, manganese, iron and cobalt. Processing of several battery materials is concentrated geographically, creating vulnerabilities as battery demand grows.

This is driving interest in alternative chemistries, including sodium-ion batteries.

Why Sodium-ion Batteries Matter

Sodium-ion batteries work on the same basic principle as lithium-ion batteries but use sodium ions rather than lithium ions to store and release energy.

Their attraction is partly linked to materials. Sodium-ion batteries do not require lithium or graphite, although some sodium-ion chemistries still rely on materials such as nickel or manganese. The IEA therefore cautions that sodium-ion technology can reduce some mineral dependencies but does not eliminate critical-mineral or supply-chain risks altogether.

The technology also has a significant limitation: energy density. The latest sodium-ion cells can reach around 175 Wh/kg, compared with up to about 205 Wh/kg for the latest LFP lithium-ion cells and 265 Wh/kg for NMC cells, according to the IEA. Lower energy density makes sodium-ion less attractive for applications where battery weight and size are critical.

That does not rule it out for stationary storage. A battery installed alongside a solar farm or connected to the electricity grid does not have to be lightweight. Cost, reliability, durability and the availability of materials can become more important considerations.

The IEA therefore identifies battery stationary storage, along with smaller electric cars, urban light commercial vehicles and two- and three-wheelers, as potential applications for sodium-ion technology.

From Research to Commercialisation

Sodium-ion batteries are not a new idea. The technology has been studied since the early 1980s, but its commercial development has lagged far behind lithium-ion. The first sodium-ion-powered electric car was introduced in China in late 2023, while global sodium-ion production in 2025 was still less than 1% of lithium-ion production.

That is beginning to change. The IEA says 2026 could be a pivotal year for sodium-ion’s scale-up. CATL has announced commercial-scale deployment of its second-generation sodium-ion batteries across multiple sectors from 2026, while other manufacturers are also investing in the technology.

Yet lithium-ion remains difficult to displace. The IEA says highly optimised lithium-ion batteries, particularly LFP chemistry, continue to have advantages in energy density, supply-chain maturity and cost.

What MIT Researchers are Working On

The challenge of making sodium-ion technology more practical is also being explored at the Massachusetts Institute of Technology (MIT). MIT News highlighted Hugh Smith, a fifth-year PhD candidate in the Department of Materials Science and Engineering who studies sodium-ion batteries. His work focuses on balancing cost, performance, sustainability and reliability rather than trying to maximise a single battery characteristic. Smith’s research reflects an important distinction in battery design.

A smartphone battery needs to be compact, lightweight and energy-dense. A grid-storage battery does not face the same constraints. For stationary applications, Smith says, cost and reliability can be more important than weight and size.

The sodium-ion batteries he studies use materials including sodium, iron and manganese. MIT says the technology could eventually provide lower-cost options for electrical grids and some electric vehicles. It also notes that sodium-ion batteries can largely be manufactured using infrastructure developed for lithium-ion batteries, potentially easing the path towards commercialisation.

The MIT article is a profile of Smith’s ongoing doctoral research rather than an announcement of a new sodium-ion battery breakthrough. Its significance lies in illustrating the broader research effort to develop battery chemistries suited to different applications.

Why India Has a Stake

For India, the issue is particularly relevant because the country needs more storage while remaining dependent on imports for lithium. In March 2026, the Ministry of Heavy Industries said India’s entire current demand for lithium is met through imports, making the sector sensitive to external shocks. India is simultaneously building domestic battery manufacturing capacity.

The government’s Production Linked Incentive scheme for Advanced Chemistry Cells has an outlay of ₹18,100 crore and aims to establish 50 GWh of domestic manufacturing capacity. As of March 2026, 40 GWh had been awarded to four companies, while 1 GWh had been installed. The government said domestic demand for advanced chemistry cells continued to be met largely through imports.

Storage requirements are also expected to grow. The Central Electricity Authority’s National Electricity Plan estimates that India will need 8.68 GW/34.72 GWh of battery energy storage systems by 2026-27. For 2031-32, the requirement rises to 47.24 GW/236.22 GWh. When pumped-storage hydropower is included, total projected storage requirements reach 82.37 GWh in 2026-27 and 411.4 GWh in 2031-32.

The government is now also targeting stationary storage specifically. In July 2026, the Ministry of Heavy Industries opened a process to select manufacturers for 10 GWh of Advanced Chemistry Cell manufacturing capacity for grid-scale stationary storage under the PLI scheme.

India is Already Exploring Sodium-ion

Sodium-ion technology is not merely an international research topic for India. The Ministry of New and Renewable Energy published an assessment of the global sodium-ion battery landscape and its potential in India under the India-UK strategic partnership’s ASPIRE programme in December 2024.

Indian research and industry are also exploring the technology. In December 2025, MNRE sanctioned a project at IIT Roorkee for the development of sodium-ion battery technology as a cost-effective alternative to lithium-based systems.

In April 2026, the Technology Development Board also announced financial assistance for an Indigenous Energy Storage Technologies project in Roorkee aimed at commercialising hard carbon derived from bio-waste and agricultural waste for sodium-ion batteries. Hard carbon is used as an anode material in sodium-ion cells.

A Second Option, Not a Replacement

Sodium-ion batteries are unlikely to replace lithium-ion technology across the board. Their lower energy density remains a disadvantage, while their manufacturing and supply chains are far less developed. The IEA says current sodium-ion manufacturing capacity is only a little over 1% of lithium-ion cell capacity, and announced sodium-ion projects for 2030 amount to about 7% of committed lithium-ion manufacturing capacity for that year.

There is also an important supply-chain caveat. Nearly all existing sodium-ion manufacturing capacity is currently located in China, and the IEA estimates that China could account for more than 95% of global sodium-ion manufacturing capacity in 2030 when announced projects are included.

For India, therefore, sodium-ion is not a simple route to eliminating import dependence. Its potential lies in diversification. A long-range electric vehicle may require the energy density of lithium-ion technology, while a grid battery can place greater emphasis on cost, reliability and material availability. Developing several battery chemistries could allow India to match different technologies to different needs.

As the country’s renewable-energy system expands, the question will not only be how much clean electricity India can generate, but how affordably and reliably it can store that electricity.

Sodium-ion batteries are still an emerging technology. But their development shows why the future of energy storage may not be about finding one perfect battery—it may be about finding the right battery for each job.

Artificial intelligence is becoming part of the cyberattack toolkit, but its most immediate impact may be less dramatic than the idea of fully autonomous hackers suggests. Instead, AI is helping attackers improve existing techniques — from reconnaissance and phishing to analysing stolen information and exploiting software vulnerabilities — while cybersecurity agencies warn that these capabilities could become more powerful as AI systems advance. A recent case involving the North Korean-linked cyber group Kimsuky illustrates the shift.

From Phishing Lures to AI Infrastructure

In August, South Korean cybersecurity company Genians reported finding evidence that Kimsuky had assembled infrastructure for running and managing AI models locally.

According to Genians, the infrastructure included tools such as Ollama, GPT4All and Msty, as well as retrieval-augmented generation (RAG) technology, AI-agent development frameworks, speech-to-text software and Cursor, an AI-assisted coding tool. The significance is not that Kimsuky had developed its own large language model. Rather, the reported infrastructure suggests an attempt to bring existing AI capabilities into a cyber-operation workflow.

Genians said local AI systems could allow operators to analyse documents without sending sensitive information to external AI services. It also assessed that the tools could support malware development, analysis of stolen material, attack automation and more convincing phishing campaigns.

The company said it also identified finance- and cryptocurrency-themed documents that appeared to have been generated with AI and designed to resemble legitimate investment reports and workplace documents. Reuters reported that Genians’ evidence could not be independently confirmed. The claims should therefore be understood as a cybersecurity firm’s assessment rather than independently established evidence of Kimsuky’s capabilities.

Still, the reported activity fits a wider trend identified by cybersecurity authorities.

AI Strengthening Existing Cyber Capabilities

The UK’s National Cyber Security Centre (NCSC) has repeatedly warned that AI is already affecting the cyber threat landscape.

Its 2025 assessment of the impact of AI on cyber threats through 2027 says AI is likely to make elements of cyber intrusion more effective and efficient. The agency expects AI-enabled tools to improve threat actors’ ability to exploit known vulnerabilities and warns that the period between vulnerability disclosure and exploitation could become even shorter.

The important point is that AI does not need to independently carry out an entire cyberattack to be useful to an attacker. A cyberattacker can use AI to perform individual tasks more quickly: understand technical information, research a target, process large quantities of text, generate or modify code, or produce convincing communications.

This can reduce the amount of human time needed for different stages of an operation. The NCSC’s assessment is therefore more measured than the idea of an imminent era of completely autonomous hacking. It describes AI primarily as a technology that can enhance existing cyber capabilities, while acknowledging that the technology is developing rapidly and that technical surprises are possible.

Phishing: One of the Clearest Risks

Social engineering is particularly suited to generative AI. Phishing traditionally depends on persuading a victim to trust a message, link or document. Poor grammar, awkward phrasing or obvious inconsistencies can expose fraudulent communications.

Generative AI can reduce some of those weaknesses by producing coherent text and adapting content to a particular context. The risk is not simply better-written emails. AI can help attackers work with information about potential targets and produce different versions of messages at much greater scale.

This is why cybersecurity agencies have focused on AI-enabled social engineering alongside other forms of cyberattacks. But there is an important distinction between AI-assisted phishing and autonomous phishing operations. The evidence available today supports the former much more strongly than the latter.

Stolen Data Could Become More Useful

The Kimsuky case also highlights another potential application: analysing information after it has been stolen. Large language models are designed to work with large volumes of text. When combined with retrieval systems, they can make it easier to search and retrieve relevant information from a collection of documents.

RAG, or retrieval-augmented generation, is not itself a cyberattack technology. It is a general AI architecture that allows a model to retrieve information from an external knowledge base and use it when generating an answer. Its presence in Kimsuky’s reported infrastructure is therefore significant because of how the technology was allegedly being incorporated into the broader operation, rather than because RAG itself is malicious.

The same principle applies to the other tools identified by Genians. Ollama, GPT4All, Msty and Cursor are legitimate AI or software-development tools. Their appearance in a suspected cyber-operation does not make the tools themselves malicious. The concern is how legitimate AI capabilities can be repurposed.

Cyberattacks: Why Local AI Matters

The reported use of locally operated AI models introduces another dimension. When an AI system runs locally, information can be processed on infrastructure controlled by the operator rather than necessarily being sent to an external AI service.

For legitimate users, local processing can provide privacy, control and offline functionality. For a cyberattack, the same characteristic could make it possible to process sensitive or stolen material without relying on an external AI provider.

That does not make local AI inherently unsafe. It illustrates a broader cybersecurity principle: capabilities designed for privacy and control can have both legitimate and malicious uses.

The UK AI security: Distinct Cyber Issue

Governments are now responding not only to the use of AI by attackers but also to vulnerabilities within AI systems themselves. The UK published its Code of Practice for the Cyber Security of AI in January 2025. The voluntary framework applies to AI systems, including generative AI, and sets out security principles across the AI lifecycle. The UK’s approach has subsequently moved towards international standardisation.

The European Telecommunications Standards Institute (ETSI) developed EN 304 223, a standard for the cybersecurity of AI. The UK government says the standard draws from the UK’s AI Cyber Security Code of Practice. In July 2026, the UK Department for Science, Innovation and Technology also published a mapping of global AI security standards, regulations and guidance against ETSI EN 304 223.

This reflects an important change in policy thinking: AI security is increasingly being treated as part of cybersecurity rather than as a separate question of AI ethics or safety. The UK is also strengthening wider cyber resilience

AI-specific policy sits alongside broader UK cybersecurity measures. The Cyber Security and Resilience (Network and Information Systems) Bill is currently progressing through Parliament. Its stated purpose is to strengthen the security and resilience of network and information systems used in connection with essential activities. As of August 13, 2026, the Bill is in the House of Lords, with committee stage scheduled to begin on September 1.

The legislation is broader than AI. But that is important because AI-enabled attacks ultimately target the same networks, organisations and digital infrastructure that conventional cyber threats target.

The NCSC has also stressed that cyber resilience cannot be treated solely as an IT concern. In June 2026, it warned that the rapid pace of frontier AI development means assumptions about cyberattack can become outdated within months rather than years.

The Bigger Risk is Convergence

The Kimsuky case points towards a more important question than whether hackers will use AI. They already are. The question is how deeply AI will become integrated into the different stages of a cyber operation. Today, the strongest evidence points towards augmentation: AI helping humans perform existing tasks more quickly or at greater scale. Tomorrow’s risk could lie in the increasing connection between those individual capabilities — reconnaissance, information retrieval, social engineering, coding and vulnerability exploitation.

That does not mean AI will suddenly produce completely autonomous cybercriminals. Current evidence does not justify that conclusion. But it does suggest that cybersecurity is entering a period in which the traditional boundary between a human attacker and a software tool is becoming less clear.

For defenders, that makes speed important. If AI allows cyberattackers to analyse information, identify vulnerabilities or tailor social-engineering attempts faster, defensive systems will need to detect and respond at comparable speed. The UK’s emerging policy framework reflects this shift: secure the AI systems themselves, strengthen the resilience of the infrastructure around them, and prepare for AI to become part of both offensive and defensive cybersecurity.

The Kimsuky case, if Genians’ findings are borne out by further evidence, could be an early example of that transition, not because AI has replaced the hacker, but because the hacker is beginning to use AI as part of the machinery of the cyberattack.