Everyone Says AI. Almost Nobody Means the Same Thing

In 1997 a computer beat the world chess champion, and it was called artificial intelligence. Today your phone plays chess better than that computer did, and nobody calls it anything. It is just a chess app.

This happens every time. Handwriting recognition was AI until banks started reading cheques with it. Spam filtering was AI until it became a checkbox in your email settings. Voice transcription was AI until it appeared free in every phone.

The computer scientist Larry Tesler summarised this decades ago. AI is whatever has not been done yet. The moment something works reliably, we stop calling it intelligence and start calling it software.

Which tells you something important. AI is not a technology. It is a label we attach to whatever machines currently do that still surprises us, and it moves every few years.

What the word is covering right now

When someone says AI today, they usually mean one of four different things, and confusing them is the source of most bad arguments about the subject.

Systems that classify. Given an input, produce a category. Is this transaction fraudulent. Is this leaf diseased. Is this scan abnormal. Unglamorous, enormously deployed, and the bulk of what actually runs in production. Most of my own research sits here.

Systems that predict. Given what happened, estimate what comes next. Demand forecasting, traffic estimation, equipment failure warnings.

Systems that generate. Produce text, images, code or audio that did not exist before. This is what caused the recent explosion of public attention, and it is the reason people who paid no attention to the previous two categories suddenly have strong opinions.

Systems that act. Take actions in an environment and adjust based on results. Robotics, self-driving vehicles, automated trading.

These share underlying methods and almost nothing else. A model that flags diseased potato leaves and a model that writes an essay raise entirely different questions about accuracy, about responsibility and about what happens when they fail. Discussing them as one thing produces the confused public conversation we currently have.

Is any of it actually intelligent

The honest answer is that the question is worse than it looks, because we cannot define intelligence precisely enough for the answer to mean anything.

What can be said clearly is what these systems do and do not have.

They have pattern recognition at a scale no person can match. They have consistency, since they do not get tired at four in the afternoon or let a bad morning affect a judgement. They have speed measured in milliseconds.

They do not have understanding of cause. A model can learn that two things occur together without any notion that one produces the other, which is why it will confidently use a correlation that a child would recognise as coincidence.

They do not have a model of the world. A system that describes a photograph accurately does not know that the objects in it continue to exist when the photograph ends.

They do not know what they do not know. This is the one that causes real damage. Ask a human expert something outside their field and you usually get hesitation. Ask most AI systems and you get an answer delivered in exactly the same confident register as everything else. The confidence carries no information about reliability, and people naturally read confidence as competence.

So calling it intelligence is a stretch. Calling it useless because it is not intelligence is a bigger one. A calculator does not understand mathematics either, and no one argues it is not worth having.

Three questions that deserve better than a sentence

Work

The debate is usually framed as whether AI will take jobs, which is the wrong question because it invites a yes or no about something that is neither.

The better question is which specific tasks within a job become cheap, since jobs are bundles of tasks and automation removes tasks rather than whole roles. A radiologist does not disappear because software reads scans. The reading task gets cheaper, and the job shifts towards the parts software cannot do, which are the ambiguous cases and the conversation with the patient.

The pattern that has held so far is that routine work automates first, whether it is physical or cognitive. Predictable, high-volume, clearly-specified tasks go. Work involving judgement under ambiguity, physical dexterity in unpredictable settings, or persuading another human being stays much longer.

For Bangladesh this is not abstract. Our economy leans heavily on exactly the kind of routine, high-volume work that automates well, in garments and in outsourced services. The country’s exposure is real and the timeline is uncertain, and treating it as a distant problem because the technology is being built elsewhere is a mistake. Where it is built has never determined where it lands.

Decisions

The more immediate issue is not machines replacing people. It is machines making decisions about people.

Loan approvals. Job applications sorted before a human sees them. Medical triage. Insurance pricing. Each of these is already partly automated somewhere.

Two things go wrong reliably. The system learns whatever bias was in its training data and applies it with perfect consistency, which is worse than a biased human because there is no variation and no bad day that lets an exception through. And the system usually cannot explain its reasoning, so a person told no has nothing to appeal against.

My position on this is not neutral, since explainable AI is one of my research areas. A model that cannot justify its output should not be making consequential decisions about people, whatever its accuracy. Accuracy is not the same as legitimacy, and a system that is right 94 percent of the time is still simply wrong about the person standing in front of it.

Dependence

This one gets the least attention and I think about it most, because I watch it every week.

My students have access to AI tools that will write an essay, solve a problem set or explain any concept on demand. Used well, this is the best tutor most of them will ever have, available at midnight, infinitely patient, and free.

Used as a substitute for thinking, it produces students who can obtain answers and cannot construct them. The output looks identical. A completed assignment is a completed assignment. The difference only surfaces later, when they need to reason through something no tool has seen.

I do not think banning these tools is realistic or even desirable. I do think we are running an uncontrolled experiment on a generation, and that assessment which cannot distinguish understanding from retrieval will stop measuring anything at all.

The question I care about most

Almost every significant AI system in use today was built somewhere else, trained on data from somewhere else, for problems defined somewhere else.

That has consequences that are easy to miss. A model trained mostly on English fails on Bangla, which is why my group has worked on Bangla smishing detection, since fraudulent SMS in Bangla is a solved problem in English and a barely-addressed one here. A crop disease model trained on American agriculture does not know our crops. A speech system trained on standard accents does not understand a Chattogram one.

The default outcome is that we import tools built for other people’s problems and adapt as best we can. The alternative is doing research here, on problems that matter here, with data collected here. That work is less prestigious than whatever is being announced in California this month. It is also the only version where the benefits arrive in the places that need them.

That is the reason I do research from Chattogram rather than treating it as somewhere to leave. The interesting problems are not all in the places with the largest budgets. They are in the gap between what the technology can do and where it has actually been pointed.

---

I research machine learning, quantum computing and AI-driven cybersecurity, and teach Computer Science at Sunshine Grammar School and College in Chattogram. My publications are on the research page.