What artificial intelligence actually is
Artificial intelligence is not the name of a technology, it is the name of a goal: programs that do work which until yesterday needed human judgment. What sits behind the name today is almost always a model that built its behaviour from data, rather than a rule somebody wrote in advance.
- Lesson 1 of 2
- Beginner
- Free, no signup
Six jobs collected under one name
- Language
- Image
- Video
- Voice
- Code
- Agents
These six are not separate and their boundary keeps fading: one model today reads text and images alike. The split is here to be understood, not as an official taxonomy.
Last checked: Facts and tool names in this lesson are re-checked against their sources on this date.
What exactly gets called artificial intelligence
Artificial intelligence is the name of a goal, not the name of a method: building programs that do work which until yesterday needed human judgment. Under that single name sit several decades of completely different approaches, and that is where most of the confusion starts.
What gets sold under the word today is specific though: a model that built its behaviour from data. The concrete example is the spam filter. The old version was a hand written list; if a certain word shows up, drop the message in the spam folder. The current version has seen a large volume of email that people themselves marked as spam or not, and derived its patterns from that. Nobody wrote in the second version that the word means spam; the program pulled that out of the data itself, and that is what gets called learning.
A simple test that stays useful to the end of this path: ask which line can be pointed to as the one that made the decision. If the answer is a particular rule, you are dealing with a rule based program. If the answer is only that it was trained on such and such data, you are dealing with a model. Both boxes may say AI, and both may be telling the truth.
The practical difference sits elsewhere and gets expensive at exactly the wrong moment: you fix a rule based program by changing a line, and you cannot fix a model that way. Fixing a model means changing its data or its training or its instructions and trying again, without knowing what else that same change moved.
How machine learning, deep learning and generative models differ
These three words keep getting used for one another, and they are in fact nested. Artificial intelligence is the widest circle and takes in the old rule based methods too. Machine learning is a subset of it: any method that builds its behaviour from data. Deep learning is a subset of machine learning and works with many layered neural networks. And a generative model is the narrowest ring: a model that produces new output, text and images and sound, rather than only labelling something or predicting a number.
What today tools are, the chatbots and the image models, is that inner ring. Which means that when somebody says AI, they almost always mean the smallest of these four subsets rather than the largest. Once you know that, a good deal of the news reads more precisely.
Why the distinction earns its keep: a vendor saying their product has AI in it may be telling the exact truth and mean a system that could have been written thirty years ago. The right question is not whether it is AI, because the answer is almost always yes. The right question is which ring, and trained on what.
Four nested rings, from the centre outward
-
1
Generative model
Produces new output: text, images, sound. The chatbots live here.
-
2
Deep learning
Many layered neural networks. Being generative is not required; image recognition sits here too.
-
3
Machine learning
Any method that builds its behaviour from data, including simple old ones.
-
4
Artificial intelligence
The outermost ring, including rule based systems that have seen no data at all.
The numbers are an order of nesting, not of importance or of time. Each outer ring takes in more methods, including ones with no learning in them at all.
Where the name came from, and why it collapsed twice
The name artificial intelligence was made in the summer of 1956, at a summer workshop at Dartmouth College, so around seventy years ago. From the start it promised more than the day could deliver, and twice those promises drove the funding away; the stretches that were later called the AI winters. Anyone who imagines this wave is the first has not seen the two before it, and may read this one badly too.
What was different this time was three things arriving together: the volume of text and images the web produced, graphics hardware that made training on that volume possible, and one architectural idea from 2017 called the transformer, published in a paper titled Attention Is All You Need, which nearly every language model in use today descends from. None of those three is about thinking. What changed was the scale, not the nature of the work.
Four things AI is not
It is not a database of truth. There is no table of facts sitting inside a model that can be queried. What it has are weights that produce the most probable continuation, and probable is not the same as correct. That is why a source can look entirely credible and not exist at all.
It is not conscious. When we say a model understands or thinks, we are using a metaphor, exactly as when we say a clock knows what time it is. Output that resembles understanding proves nothing about having experience, and the philosophical argument about it has stayed open for seventy years.
It is not a search engine. Unless it has been given search tools; and then what you are looking at is two pieces, a search and a model summarising its results. The simple test: if it did not hand you a link you can open, no search happened and what you read came out of training memory.
It is not deterministic. The same question asked twice can get two different answers, and that is not a defect but part of the design. Anywhere you need repeatability, such as something meant to run inside a business process, that has to be accounted for from the start.
Four sentences you hear a lot, and what is actually true
What actually happens
- It builds the most probable continuation; probable is not correct
- Its output resembles understanding, which proves nothing
- Only if it has search tools, and then it hands you links
- Two different answers is normal and part of the design
What gets said
- It fetches the answer from a trusted database
- It understands what you are saying
- It searches the internet
- The same question gets the same answer
The right column does not reject the claim, it makes it precise. None of these four means the tools are useless.
Where the word turns into marketing
The word AI on a product tells you almost nothing, because the definition is loose enough that few programs fall outside it. Three questions sort out very quickly, in a sales meeting, what the other side actually has: which model, trained on what data, and what happens when it answers wrongly and who carries that. The third question usually goes unanswered, and going unanswered is itself the answer.
Our position is blunt: asking whether something is AI is a useless question, because the answer is almost always yes and it changes no decision. The question worth asking is who signs the output. If the answer is you, then the model is a tool like any other. If the answer is nobody, what you bought is not a tool, it is a risk.
Which work to hand it, and which not to
The tasks you can hand over safely share one property: checking the output is faster than doing the job. A first draft, a translation you proofread yourself, a summary of something you have also read, converting one data format into another, explaining an error message. In all of them, if the model wanders off, you notice within seconds.
The tasks not to hand over share a property too: being almost right is worse there than being plainly wrong. Numbers, dates and proper names, anything carrying legal or medical weight, and any text somebody will later decide on when you cannot check it line by line. Add to those anything holding another person private data, for a reason that is not accuracy and comes in the next block on this page.
The rule we use ourselves is one sentence: if checking the output takes longer than doing the work, the model has not helped, it has only moved the work. That test depends on no particular tool, so it still works on next year one.
The fast path, with AI
The fastest way to find out whether AI is any use to you is not reading another article; it is a ten minute audit of your own last week. What separates this from generic advice is that you make the model argue against itself. A frontier model suits this better than a fast cheap one, because it is judgment work rather than mechanical work; our current pick is written up in the AI section of this site.
- Write down the five tasks that took the most of your time last week. Not the five most important ones, the five slowest; those two are usually not the same list.
- Beside each one mark two things: can the output be checked faster than the job can be done, and does it hold anybody else data.
- Run the recipe below with that same list and keep the answer. The text deliberately requires the model to rule at least one item out.
- Trial only the item that ranked highest, for one week, and at the end compare the time it took against the time it used to take. Without this step the rest is guesswork.
Copy-ready recipe
Role: an adviser who is not paid to sell tools and whose job is to prevent a pointless purchase.
My five tasks from last week:
1. {task, in one sentence} - checking the output: {faster / slower than doing it} - other people data: {yes / no}
2. {same shape}
3. {same shape}
4. {same shape}
5. {same shape}
What I want:
1. Rank these five from most to least worth doing with AI, one sentence of reasoning each.
2. Rule at least one of them out and say why it should not go near AI. If you ruled none out, look again.
3. For the top item, write exactly what I have to check after every output.
4. Write what breaks if I automate this task and the model gets it wrong, and how long it would take me to notice.
5. Name no product, and write no number you have no source for.Before you trust the output: Take one thing in the answer seriously and believe one thing not at all. Take seriously the item it ruled out, because that is usually the soundest part of the reply. Believe no number it gives you about time saved; only your trial week produces that number. And if the model ruled nothing out, rule something out yourself: the pull toward agreeing is a known weakness of these tools, named by a vendor in its own published document on how its model should behave.
AI in this kind of work
For this particular topic, understanding what AI is, the best use of these tools is one thing: explaining a specific document, not answering from memory. Hand the model a vendor own page and ask what the text actually claims and what it deliberately does not say. Our position: asking a model a general question about itself is the lowest value thing you can do with it.
Tools that actually help
- NotebookLM You supply the sources yourself and the answer is built only from them, with a reference to the passage it came from. For this exact job, reading several official documents on one subject, it is the best shape available. Google own help says it runs in the same regions the Gemini app runs in, and Iran is not on the supported-countries list.
- Gemini It carries a separate study mode called Guided Learning which asks questions instead of handing over answers and recaps at the end of a topic. That suits a concept like this lesson. Google own page says the Gemini web app runs in over 230 countries and territories, and Iran is not on that list.
- Claude Give it a text and ask what the document does not say, and the answers are more useful than an open ended question gets. Iran is on neither of Anthropic two supported-countries lists; we read that on Anthropic own page rather than measuring it.
Where it backfires
Two specific risks live in this topic and the vendors wrote both of them down themselves. First, the pull toward agreement: models tend to affirm what you said, and OpenAI put a separate clause about avoiding exactly that sycophancy in its published document on how its model should behave. So if you ask from a wrong premise, the odds are good the answer confirms the premise. Second, where what you type ends up: depending on the plan and the settings, the text you send may be used for training, which Anthropic explains on its own privacy page. The practical consequence for a business is simple: contracts, customer lists and any data that is not yours do not go into this box, not even for a quick summary. For how access and payment for each of these tools stands from Iran, see the buying guide.
Sources: OpenAI Model Spec: avoid sycophancy Anthropic: is my data used for model training Anthropic: supported countries Google: where the Gemini web app is available
Where this advice stops
This lesson draws boundaries and teaches you no tool; for that, read the lessons that follow in this path. Second, no standards body owns the phrase artificial intelligence, so any definition including this one is a working definition rather than a ruling; in legal texts and in each country regulation the boundaries are drawn differently and sometimes more widely. Third, anything tied to a specific model or a specific price goes stale within weeks, which is why none of it is on this page.
From our own work
On this very site an AI classifier decides whether a new listing in the marketplace, freelance and ads sections goes live or gets rejected. What we learned while writing it fits in two numbers, both of them in wp-content/mu-plugins/rgb-modai.php: publishing happens at 60 percent confidence, and rejecting needs 78. The two are deliberately not equal, because the two mistakes do not cost the same: a wrong publish means we pull a listing down by hand, while a wrong rejection means a real seller was turned away for nothing. The model did not take the judgment away from us; it moved it into two numbers somebody had to choose, and setting them equal would have been the wrong answer.
Real follow-up questions
Are AI and ChatGPT the same thing?
No. AI is the name of a field, and ChatGPT is one specific product from one specific company, built on one family of models. In the same way that a search engine is a concept and Google is a product. The distinction starts to matter when you want to switch tools: the habits you learned transfer, the details of any one product do not.
Will AI take my job?
Anyone giving you a number and a date for this is selling something. What we have seen in our own work is a shift at the level of tasks rather than jobs: the mechanical parts got faster and the parts needing judgment got heavier, because now you also have to check a machine output. Where that balance lands in your profession depends on how much of your work is mechanical.
What does artificial general intelligence or AGI mean?
It is a goal, not a product: a system that could learn any new intellectual task the way a person can, not only the tasks it was trained for. There is no agreed definition of it, which is exactly what makes a claim of having reached it impossible to check. Anyone giving a date is making a prediction, not a report.