Most people are trying to understand artificial intelligence from the top down. They begin with jobs, industries, salaries, layoffs, workflows, and headlines.
That is the wrong direction.
For this class, we begin at the bottom. We begin with the primitive operation.
Information technology, at its most basic level, is the management of stored information. In database language, this is CRUD: create, read, update, delete. CRUD is the ordinary shorthand for the four basic operations performed on persistent data. You create a record. You read a record. You update a record. You delete a record. That is the backbone of ordinary software.
This sounds simple because it is simple. But it changed the world.
Information technology made it cheap to manage information that already existed.
That is the key phrase: already existed.
A bank record exists. A payroll record exists. A customer address exists. An inventory count exists. A photograph exists. A song exists. A transaction exists. A reservation exists. A tax record exists. A school transcript exists. Software lets us create it, store it, retrieve it, edit it, transmit it, copy it, sort it, search it, delete it, and protect it.
That is the world of information technology.
It is CRUD at planetary scale.
The computer, in this older sense, is an arithmetic machine. It manipulates symbols according to procedures. An algorithm is a step-by-step procedure for solving a problem or accomplishing an end. That procedure can be simple or complex. It can calculate payroll, compress a song, render a video, sort a database, route a package, or produce a spreadsheet forecast.
But the deep operation is still procedural computation over information.
This is why the original computer story matters so much.
Before computers were machines, “computer” was a human job title. Human computers worked in banks, observatories, laboratories, government agencies, and aerospace programs. They performed calculations. NASA describes the human computers at the Jet Propulsion Laboratory as people, often women, who performed mathematical work before and alongside early machine computers. The Smithsonian notes that around 1970, the job title “Computer” at JPL was replaced with “Engineer” as the work evolved into programming and engineering.
So yes, the machine computer eliminated the human computer.
But that is not the lesson.
The lesson is that cheap arithmetic did not reduce the amount of arithmetic in the world. Cheap arithmetic caused arithmetic to explode.
Once arithmetic became cheap enough, we put it everywhere. We put it into banks, schools, hospitals, logistics networks, factories, phones, cameras, cars, music, maps, calendars, stores, reservations, publishing, games, and eventually every office desk.
The job title disappeared. The operation multiplied.
That is the pattern students need to understand.
When a fundamental operation becomes cheap, the old job title can disappear while the total demand for that operation explodes.
The same pattern is now happening again.
But the operation is different.
AI is not fundamentally about arithmetic. It uses arithmetic, of course. Every neural network runs on computation. Every model depends on chips, memory, electricity, data centers, and mathematical operations. But that is implementation. That is not the economic primitive.
The economic primitive of AI is prediction.
This is the clean break.
Information technology manages information.
Generative AI predicts information.
That is why this is not merely another software upgrade. It is a new economic layer.
A traditional computer stores and manipulates information that has already been captured. A digital camera captures light from the world and converts it into stored information. Photoshop manipulates that captured information. A spreadsheet stores numbers and runs formulas over those numbers. A database stores records and lets us create, read, update, and delete them.
Generative AI operates differently.
A generative image model does not need a camera. It does not need a lens. It does not need a captured photograph. It has learned a statistical structure from prior examples, and from that structure it predicts a new image. The output is not a retrieved photograph. It is not merely an edit of an existing photograph. It is a generated prediction of what an image matching the condition should look like.
This is why the photography example is so powerful.
Digital photography was arithmetic applied to captured light. It made photographs cheaper to take, store, copy, edit, and distribute. That was enormous. It put a camera in every smartphone and made the global image supply effectively infinite compared with the film era.
But generative AI goes one level deeper.
It does not just make photographs easier to manage. It makes image formation itself predictive.
That changes the center of gravity.
In the arithmetic world, Photoshop is king because the photograph already exists. The task is manipulation. Remove this. Brighten that. Crop this. Correct that. Composite these. Sharpen those.
In the prediction world, the first question is no longer, “How do I manipulate the photograph?”
The first question is, “What should the image be?”
That is a completely different question.
The Vogue cover example makes the point. In the old world, the cover begins with a photo shoot. There is a model, a photographer, a studio, lighting, makeup, wardrobe, direction, cameras, files, editors, and then manipulation. The photograph is captured first. Software comes afterward.
In the prediction world, there does not have to be a photo shoot. A prediction machine can generate the cover concept directly. It can predict the model, the pose, the lighting, the background, the wardrobe, the typography, the mood, and the finished composition. The artifact arrives without the original photographic event.
That does not mean photographers disappear. It does not mean models disappear. It does not mean magazines disappear. It means the work moves.
Some old tasks shrink. New tasks appear around taste, direction, authenticity, brand judgment, prompt architecture, model selection, rights, provenance, cultural meaning, final approval, and the relationship between synthetic imagery and reality.
The mistake is to look only at the old task and declare the job market smaller.
That is exactly the mistake people made with computers.
They saw the human computer disappear. They did not yet see the software engineer, database administrator, systems analyst, IT manager, cybersecurity analyst, digital designer, web developer, cloud architect, data scientist, product manager, network engineer, ERP consultant, CRM administrator, and smartphone app developer.
They saw arithmetic replace a job.
They did not yet see arithmetic create a civilization.
AI is now doing the same thing with prediction.
Prediction has always been expensive because the only general-purpose prediction machine we had was the human brain. A doctor predicts disease. A teacher predicts confusion. A salesperson predicts buyer intent. A manager predicts what will go wrong. A lawyer predicts risk. A parent predicts a child’s needs. A designer predicts what will feel beautiful. A leader predicts what the room is ready to hear.
Prediction is everywhere, but until now it was locked inside scarce human attention.
That is what AI changes.
Ajay Agrawal, Joshua Gans, and Avi Goldfarb made this point in the economic literature: artificial intelligence lowers the cost of prediction, and prediction is an input into decision-making under uncertainty.
That sentence is the article.
AI lowers the cost of prediction.
Once prediction becomes cheap, prediction will be used everywhere.
That is the job-creation argument.
Not because AI is nice. Not because technology always creates jobs automatically. Not because every displaced worker is instantly rescued by a new title. That is lazy optimism, and students should reject it.
The stronger argument is mathematical and economic.
When the price of a useful input falls by orders of magnitude, demand for that input expands into places where it was previously too expensive to use.
Arithmetic became cheap, so arithmetic entered everything.
Prediction becomes cheap, so prediction enters everything.
This is why the radiology example matters.
For years, radiology was used as the obvious example of AI job elimination. If AI can read scans, people said, then we will need fewer radiologists. But that is not what has happened so far. The American College of Radiology was still describing a radiologist workforce shortage in 2026, and a 2025 JACR workforce projection estimated 37,482 radiologists providing care to Medicare patients in 2023, with the number projected to grow substantially by 2055 under both modeled scenarios.
Why?
Because the demand for medical interpretation is not fixed.
If image interpretation becomes cheaper, faster, and more available, we do not simply reduce the number of people interpreting images. We increase the number of images. We screen more people. We catch more cases earlier. We monitor more conditions. We move diagnostics closer to the patient. We create more follow-up, more explanation, more treatment planning, more patient communication, and more clinical accountability.
The prediction layer expands the service layer.
This is the part that most AI job-elimination arguments miss. They assume a fixed amount of work.
There are 1,000 scans. AI reads 500. Therefore, humans read 500 fewer.
That is not how economies usually behave when a valuable operation becomes cheaper.
The better model is this:
There are 1,000 scans because interpretation is expensive.
Make interpretation cheaper, and soon there are 10,000 scans.
Now the bottleneck moves. The new bottleneck is not raw image reading. The new bottleneck is clinical judgment, patient trust, workflow integration, regulation, liability, explanation, prioritization, and treatment.
That is where the new work appears.
The same logic applies far beyond healthcare.
If prediction becomes cheap, schools can predict student confusion earlier. Companies can predict customer churn sooner. Farms can predict disease and yield more precisely. Cities can predict maintenance failures before breakdowns. Small businesses can predict cash flow, staffing needs, marketing response, and inventory problems. Individuals can predict their own health, finances, schedules, and learning paths with far more precision.
Each case creates new work because prediction is not the final product.
Prediction is an input.
A prediction must be interpreted. It must be trusted or rejected. It must be translated into action. It must be explained. It must be delivered in the right form to the right person at the right time. It must be embedded into a system. It must be governed.
That is where humans remain essential.
The valuable future worker is not the person who competes with AI at raw prediction. That is the wrong contest. The valuable future worker knows what predictions are worth making, which predictions are dangerous, which predictions are actionable, which predictions require human confirmation, and how to turn prediction into finished work.
This is why the job categories will look strange at first.
The new jobs will not all be called “AI jobs.” Most of them will be ordinary jobs rebuilt around a prediction layer.
Teacher becomes learning diagnostician.
Nurse becomes continuous care interpreter.
Manager becomes exception architect.
Designer becomes synthetic creative director.
Lawyer becomes risk translator.
Salesperson becomes intent interpreter.
Analyst becomes model-and-market explainer.
Local business employee becomes AI operations coordinator.
The title comes later. The work comes first.
That is exactly what happened with information technology. At first, computers were machines in special rooms. Then they became departmental tools. Then they became office tools. Then they became personal tools. Then they became phones. Then they disappeared into everything.
AI will follow the same absorption path.
At first, AI is a product. Then it is a tool. Then it is a layer. Then it disappears into the work.
The World Economic Forum’s Future of Jobs Report 2025 projects 170 million new roles created by 2030 and 92 million displaced, for a net increase of 78 million jobs. That is not an AI-only forecast; the report includes technological change, demographics, geoeconomic tensions, economic pressure, and the green transition. But AI and information-processing technologies are clearly part of the transformation.
The precise number is less important than the principle.
The principle is this:
AI creates jobs because prediction was expensive.
When prediction becomes cheap, society discovers demand it could never previously afford to satisfy.
This is the argument students should learn.
Do not ask only, “What task can AI do?”
Ask, “What happens when prediction becomes cheap enough to put everywhere?”
That question opens the future.
Information technology made it cheap to manage existing information.
Artificial intelligence makes it cheap to predict new information.
CRUD gave us the digital office, the database, the spreadsheet, the website, the smartphone, the cloud, and the modern white-collar economy.
Prediction will give us the next layer: synthetic diagnosis, synthetic design, synthetic tutoring, synthetic administration, synthetic coordination, synthetic forecasting, synthetic research, synthetic planning, and synthetic decision support.
Some jobs will disappear. Some titles will become historical artifacts, just as “human computer” became a historical artifact.
But the operation will not disappear.
The operation will multiply.
The future of work is not that humans stop working because machines can predict.
The future of work is that prediction becomes so cheap, so abundant, and so embedded that human work moves upward toward judgment, trust, meaning, responsibility, relationship, and final form.
Arithmetic created the information economy.
Prediction creates the intelligence economy.