In November 2022, a chatbot called ChatGPT reached one million users in five days. Instagram took two and a half months to hit the same number; Netflix took three and a half years. Four years later, the technology behind that chatbot has moved from a novelty search-bar experiment into something that touches nearly every part of daily life – how people search for information, how doctors read scans, how farmers plan a harvest, and how governments think about the next decade of economic policy. Understanding what artificial intelligence actually is, and is not, has stopped being optional for anyone who wants to make sense of the world it is quietly reshaping.
Artificial intelligence, in its broadest and least controversial definition, is the field of computer science concerned with building systems that perform tasks which would normally require human intelligence – recognizing a face, translating a sentence, predicting which product a shopper is likely to buy next. That definition is broader than most people expect. It covers the spell-checker in a word processor and the algorithm that flags a fraudulent bank transaction, not only the chatbots and image generators that dominate headlines.
A more useful distinction separates AI into three tiers. Narrow AI, also called weak AI, is built to do one thing well – play chess, filter spam, recommend a video – and cannot transfer that skill to an unrelated task. Every AI system in commercial use today, including ChatGPT, falls into this category, however impressive its range of abilities looks. General AI, sometimes called strong AI, refers to a hypothetical system with human-level reasoning across any domain; no such system exists yet, and serious researchers disagree sharply on how close it actually is. Superintelligence, a step beyond that, describes a system that would exceed human intelligence across essentially every field – a concept that remains theoretical and is debated more in philosophy departments than in engineering labs.
The field’s origin point is usually traced to 1950, when the British mathematician Alan Turing published a paper proposing what became known as the Turing Test – a simple criterion in which a machine could be considered intelligent if a human conversing with it, in writing, could not reliably tell it apart from another human. The term “artificial intelligence” itself was coined six years later, in 1956, at a summer workshop held at Dartmouth College, organized by John McCarthy, Marvin Minsky, and a handful of other researchers who believed the problem could be substantially solved within a single generation.
That optimism collided with reality more than once. The 1970s and again the late 1980s produced what historians of the field call “AI winters” – periods when funding collapsed after early techniques failed to scale, and the phrase “artificial intelligence” briefly became something researchers avoided using in grant applications. The current boom traces to a specific, datable turning point: a 2012 competition called ImageNet, in which a neural network named AlexNet, built by Geoffrey Hinton’s research group, cut the error rate on an image-recognition benchmark by roughly ten percentage points over the next-best entry – a margin so large it convinced much of the field that a particular technique, deep learning, deserved serious investment. A second turning point followed in 2017, when Google researchers published a paper titled “Attention Is All You Need,” introducing the transformer architecture that now underlies nearly every major language model, ChatGPT included.
Most of what gets called AI today is a specific technique within it called machine learning, and machine learning itself rests on a fairly simple idea once the mathematics is set aside: instead of a programmer writing explicit rules for every situation, the system is shown a large number of examples and adjusts itself to recognize the pattern connecting them. Show a system ten thousand photographs labeled “cat” or “not cat,” and it gradually tunes millions of internal numerical values until it can classify a photo it has never seen before with reasonable accuracy.
Deep learning is the subset of machine learning built on neural networks – layered mathematical structures loosely inspired by, though far simpler than, the neurons in a biological brain. Each layer transforms the data it receives and passes a new representation to the next; a network recognizing handwritten digits might have early layers detecting simple edges and curves, with later layers combining those into recognizable shapes of numbers. Modern large language models extend this idea to enormous scale – GPT-4, for instance, is estimated to contain somewhere in the range of a trillion parameters, though OpenAI has not published an exact figure, and training a model of that size can cost tens of millions of dollars in computing time alone.
The AI systems most people now interact with daily belong to a category called generative AI – models that produce new content (text, images, audio, code) rather than simply classifying or predicting a single value. Large language models, the text-generating branch of this category, work by predicting the next word in a sequence, one word at a time, based on probabilities learned from enormous quantities of text scraped from books, websites, and other written material during training. What looks like reasoning to a user is, at the mechanical level, an extraordinarily sophisticated form of pattern completion – which is also why these systems can produce fluent, confident-sounding answers that are factually wrong, a failure mode researchers call hallucination.
This distinction is worth keeping in mind every time these tools get used for something that matters: a large language model does not “look things up” the way a search engine does unless it has been explicitly connected to a retrieval system or live internet access. Its knowledge is fixed at the point its training data was collected, subject to a training cutoff date, and its fluency should never be mistaken for verified accuracy.
In medicine, AI systems trained on retinal scans can now flag diabetic retinopathy with accuracy that rivals trained ophthalmologists, a capability already deployed in parts of India and increasingly piloted in Pakistani hospitals facing a shortage of specialist doctors relative to population size. In agriculture, satellite imagery combined with machine learning is used to predict crop yields and detect pest outbreaks weeks before they would be visible to a farmer walking the field – a genuinely significant application in a country where agriculture still employs close to forty percent of the labor force. In finance, the algorithms that approve or decline a credit card transaction in under a second, and that flag suspicious patterns for the State Bank of Pakistan’s anti-money-laundering systems, are machine learning models refined continuously on new transaction data.
Education is another area worth a closer look: adaptive learning platforms now adjust the difficulty of practice questions in real time based on a student’s answers, a technique already used in several international tutoring platforms and increasingly by ed-tech startups worldwide building tools of their own.
Pakistan approved its first National Artificial Intelligence Policy in 2023, developed under the Ministry of Information Technology and Telecommunication, setting out goals that included establishing AI centers of excellence at public universities, building a national AI fund, and integrating AI literacy into higher education curricula. The policy explicitly frames AI adoption as an economic necessity rather than an optional pursuit, projecting that AI-related industries could add billions of dollars to the country’s GDP over the following decade if adoption keeps pace with regional competitors such as India and Bangladesh.
Several public-sector pilots have followed. The Punjab Information Technology Board has experimented with AI-based chatbots for citizen services, and NADRA has explored facial-recognition and document-verification systems to speed up identity processing. None of this is without friction – Pakistan’s internet penetration, electricity reliability, and the sheer cost of the specialized computing hardware AI training requires remain real constraints on how quickly these ambitions can be realized, a tension worth keeping in mind rather than treating the policy as a settled success story.
Bias is the most extensively documented concern. Because machine learning systems learn from historical data, they can absorb and even amplify the biases present in that data – a hiring algorithm trained on a company’s past hiring decisions will reproduce whatever demographic skew existed in those decisions, unless specifically corrected for. A widely cited 2018 study by researchers Joy Buolamwini and Timnit Gebru found that commercial facial-recognition systems from several major technology companies had error rates below one percent for light-skinned men but above thirty percent for dark-skinned women, a gap that prompted several of those same companies to pause or restrict police use of their facial-recognition products.
Job displacement is the concern most frequently raised in public debate, and the honest answer is that economists are genuinely divided on its scale. A frequently cited 2023 analysis from Goldman Sachs estimated that generative AI could affect roughly three hundred million full-time jobs globally in some way, though “affect” in that study covers a wide range from full automation to partial task assistance, and history offers a mixed precedent – automation has repeatedly eliminated specific tasks while creating new categories of work that did not previously exist. Misinformation is a newer and arguably more urgent problem: the same generative models that can draft an essay can also produce a convincing fake photograph or a synthetic voice recording, collapsing the cost of producing disinformation to nearly zero and complicating efforts by election commissions and news organizations worldwide, including in Pakistan’s own recent election cycles, to verify what is genuine.
The European Union’s AI Act, which entered into force in phases beginning in 2024, is the first comprehensive attempt by a major regulatory bloc to categorize AI systems by risk level and impose corresponding obligations, banning a small number of “unacceptable risk” applications outright, such as social scoring by governments, while imposing lighter transparency requirements on lower-risk uses. The United States has taken a more fragmented approach, relying on a mix of executive orders, sector-specific agency guidance, and state-level legislation rather than a single federal framework. Pakistan, along with most developing economies, does not yet have dedicated AI legislation, though the National AI Policy signals an intention to develop a regulatory framework, and questions of data protection connect directly to the country’s 2023 Personal Data Protection Bill, still moving through the legislative process at the time of writing.
Most people will never train a model or read the EU AI Act in full, and none of that is necessary to get real value from understanding AI at the level this article covers. What is useful is the ability to tell the difference between a tool that genuinely reasons and one that is very convincingly guessing – a distinction that matters the next time an AI-written summary shows up in a news feed, a job application gets filtered by an automated system before a human ever sees it, or a voice on the phone turns out not to be a person at all. It also matters for anyone weighing how AI might change their own work: knowing that today’s systems are narrow rather than general is the difference between reasonable caution and either dismissing the technology outright or overestimating what it can be trusted to do unsupervised.
The technology is not going to stay still long enough for any single article to remain complete. What is worth carrying forward is the underlying framework: narrow AI versus the still-hypothetical general AI, the mechanical reality of how a language model actually generates its answers, and the gap – a gap regulators and employers are all still working out how to close – between what these systems can convincingly appear to do and what they can be reliably trusted to do.
No Comments