W1. AI and Future of Work

Author

Munir Makhmutov

Published

September 7, 2026

1. Theory

Week 1 frames the whole course: where artificial intelligence sits in economic history, why this wave differs from previous automation, how it reshapes jobs task by task, and what societies and individuals should do about it.

1.1 From Mechanisation to Intelligence

Four industrial revolutions structure the story. The First (1760–1840) brought steam power and mechanisation; the Second (1870–1914) electricity, assembly lines and mass production; the Third (1970–2000) computers, electronics and digital technology; the Fourth (2010–present) adds artificial intelligence and autonomous systems. The Fourth Industrial Revolution means intelligent digital technologies — AI, robotics, IoT, big data and cloud, cyber-physical systems, advanced connectivity — woven into the physical, economic and social world. * What makes this one different: earlier revolutions multiplied physical productivity. This one reaches cognitive work — language, analysis, decision support, creativity, communication, knowledge production. Machines no longer only automate muscle; they participate in thinking.

1.2 Potential Benefits and Risks

Benefits: higher productivity, less dangerous and repetitive work, better products and services, faster science, new industries and professions, expertise made accessible, help with ever more complex problems. Risks: deeper inequality, destroyed or transformed jobs, skill gaps, power concentrated in big tech firms, surveillance and privacy erosion, bias baked into automated decisions, wider gaps between countries and groups. * Key Pitfall: technology does not affect everyone equally — averages hide the distribution, and the distribution is the policy problem.

1.3 Will AI Eliminate Human Work?

History counsels against panic and against complacency. Every wave — mechanised textiles, farm machinery, factory automation, office computers, industrial robots, internet services — triggered fears of the end of work; English textile workers (the Luddite movement (English textile workers smashing machines in the early 1800s). A 1960s US commission concluded technology kills particular jobs, not work itself: jobs are eliminated, transformed, and created. But adaptation was never automatic or painless — communities endured long disruptions. * Should we worry? Yes, but about the right things. The question is not “will this job disappear?” but “how will the tasks inside it change?” The realistic risks are rapid content change, unequal access to opportunities, skill decay, wage polarisation, botched transitions and unfairly shared gains — not overnight disappearance of all jobs.

1.4 Jobs Are Collections of Tasks

A job bundles many tasks: a teacher explains, prepares, grades, motivates. The famous 47% figure (Frey and Osborne, 2017) estimated the share of US employment at high risk of computerisation — at risk, not doomed; technical possibility is not adoption. Automation hits tasks, and the routine–cognitive map predicts where: traditional machines took routine manual work (assembly) and routine cognitive work (data entry); modern AI reaches into non-routine cognitive tasks, while non-routine manual work (repair, care) stays hardest. * Two modes: substitution (the robot does the task instead of the person) versus complementarity (technology makes the person faster, better, safer). Exposure is broadened by predictable, repetitive, standardised, measurable work plus education, income, location, age and retraining access — but exposure is not replacement: AI-assisted teachers may serve more students and grow the profession.

1.5 Generative AI, Adoption and Adaptation

Generative AI pushes automation into professional and creative occupations once thought safe — yet it confidently produces falsehoods, inherits training bias, lacks accountability and real-world grounding, needs verification, and cannot supply trust, empathy or responsibility. Human oversight stays mandatory for high-impact decisions. And capability is not deployment, adoption takes time: institutions and societies adapt slower than technology, and wealthier countries adopt first. * Successful adaptation playbook: governments — mobility, safety nets, infrastructure, affordable retraining, modernised education; schools — foundations, AI literacy, critical thinking, creativity, communication; companies — gradual responsible rollout with workers involved, retrain before roles die; individuals — keep technical knowledge fresh, build AI-complementary skills, verify AI output, deepen domain expertise. Routine skills depreciate; critical thinking, complex problem-solving, creativity, emotional intelligence, leadership, ethical judgment, domain mastery, adaptability and AI literacy appreciate. The strongest professionals will fuse human judgment with machine capability.

1.6 What AI Means: Four Definitions

There is no single definition — the classic Russell & Norvig map crosses the goal (think vs act) with the standard (humanly vs rationally). Thinking humanly models actual human cognition (reaction-time fits, fMRI-backed architectures); thinking rationally follows laws of thought (logic, probability, optimal inference); acting humanly passes as human from outside (the Turing-test chatbot); acting rationally picks the best action given goals and evidence (braking assistants, game-playing search). The course lives mostly in the right column — rational agents — while remembering the left one exists. * Six areas, two directions: the textbook map names natural language processing among six concern areas; the lecture spotlights two live directions — computational creativity and biomimicry, learning from evolved solutions rather than only from human engineering.

1.7 Intelligence Is Situational, Then Offloaded

Intelligence has no culture-free definition: what counts as smart depends on who judges — Indigenous navigation, animal and even plant problem-solving all embarrass a narrow IQ-style view. Technology meanwhile offloads intelligence out of people: nobody hand-computes steering torque anymore (power steering), clutches assist, and driving as a whole is sliding toward the car. Offloading is gradual — each step looks like assistance until the human is out of the loop. * Key Pitfall: defining AI once and for all. Situational means the definition moves with the judge, the culture and the decade.

1.8 Bioinspired and Embodied AI

Engineers copy more than humans: swarms coordinate without a leader, flocking birds and foraging ant colonies solve allocation problems, evolution plus breeding plus genetics is a working optimisation lab. And intelligence wants a body: a robot without a program is soulless, an AI without a body is a ghost — modern systems arrive embodied, physically (robots) or digitally (Siri, Alexa, game characters as agents). Humans eagerly anthropomorphise both kinds, which is exactly why oversight in 1.5 matters. * Old dream: Pygmalion’s statue and Daedalus’s animated figures — the wish for made minds is ancient; only the engineering is new.

1.9 Strong Versus Weak AI

Weak AI solves specific tasks and may or may not look human doing it — intelligence simulation, and the overwhelming majority of real work. Strong AI aims at human-based, general design. Everything deployed around you today, from spellcheckers to recommenders, is weak; treat any claim otherwise as marketing until shown a general agent. * Key Pitfall: confusing fluent weak systems (chatbots) with general understanding — fluency is performance on a task, not generality across tasks.

1.10 Bottlenecks and What Comes Next

Frey and Osborne’s own analysis names three bottlenecks slowing computerisation: perception and manipulation (unstructured physical world), creative intelligence (genuine novelty), social intelligence (trust, persuasion, care). Their 2013 predictions have only partly come true — bottlenecks bite. The course toolbox for what follows is already on the agenda: decision trees, game trees, expert systems and neural networks. * Key Pitfall: reading any single forecast as destiny — bottlenecks, adoption lag (1.5) and economics all stand between “can” and “does”.

2. Definitions

  • Fourth Industrial Revolution: Integration of intelligent digital technologies (AI, robotics, IoT, big data, cyber-physical systems) into physical, economic and social life, from ~2010.
  • Cognitive work: Tasks involving language, analysis, decisions, creativity and knowledge — newly exposed to automation.
  • Substitution: Technology performing a task previously done by a person.
  • Complementarity: Technology helping a person perform a task faster, better or more safely.
  • Thinking humanly: AI modelling actual human cognition (reaction times, neural fit).
  • Thinking rationally: AI following laws of thought — logic, probability, optimal inference.
  • Acting humanly: AI behaving indistinguishably from a person (Turing-test behaviour).
  • Acting rationally: AI choosing the best action for its goals and evidence.
  • Strong AI: General, human-based intelligent design (aspiration, not deployment).
  • Weak AI: Specific-task problem solving, possibly with no human likeness; nearly all deployed AI.
  • Situational intelligence: Intelligence judged relative to culture and context, not absolutely.
  • Offloading: Moving intelligence out of the person into technology (power steering → self-driving).
  • Embodied AI: Intelligence with a physical (robot) or digital (assistant, agent) body.
  • Bottlenecks of computerisation: Perception and manipulation, creative intelligence, social intelligence.
  • Routine task: Predictable, repetitive, standardised activity — the easiest to automate.
  • Task exposure: Share of a job’s tasks technically automatable; high exposure does not imply replacement.
  • Generative AI: Systems producing novel text, images and code; powerful but unverified, biased and unaccountable by default.
  • Adoption lag: Gap between technical feasibility and real organisational uptake.