Is generative AI the same thing as AI?

No, it is one branch of it. The broader field includes sorting, scoring and prediction, most of which has been in ordinary business software for years without anybody calling it AI. What changed recently, and what nearly every current product is built on, is the ability to produce new material on request.

How do we judge whether the output is good enough?

Not by an accuracy figure, because there is no single correct answer to compare against. Judge it the way you would judge a junior colleague's draft: against the standard the job actually needs, on a sample of real cases rather than demonstrations, and with somebody named as the person who checks it.

Why does it sound certain when it is wrong?

Because certainty is a property of the writing rather than of the knowledge behind it. The system produces the most plausible continuation, and plausible prose is fluent prose. Nothing in the process marks the difference between an answer it has grounds for and one it does not, so both arrive in the same voice.

Does it learn from what we type into it?

That depends entirely on the product and the plan, not on the technology. Business tiers commonly commit not to train on customer content and consumer tiers often reserve the right to. It is a contractual question rather than a technical one, and the answer belongs in writing before real material goes in.

Will it replace the work or assist it?

In practice it changes where the effort sits rather than removing it. The drafting gets faster and the checking gets more important, so roles tend to shift towards judging output rather than producing it. Teams that plan for the checking get value sooner than teams that plan only for the speed.

Where should an organisation start with it?

With a job that is repetitive, low-stakes and currently done by hand, where somebody can tell quickly whether the output is acceptable. Starting with the highest-value process is a common instinct and a poor one: the stakes make the checking expensive before anybody has learnt what good output looks like.

Glossary

Generative AI

Generative AI is software that produces new material on request, such as text, images, audio or code, rather than only classifying, ranking or predicting from things that already exist.

In plain terms

Older business software answered questions about things it already had. Which of these emails is a complaint, which customer is likely to leave, what is the total for this quarter. This newer kind writes the reply, drafts the image, produces the code. The difference sounds small and is not: sorting has a right answer you can check against records, while producing has a range of acceptable answers and no ledger to check them against.

Why it matters

It changes what you are buying and how you judge it. A classifier can be measured against a set of known-correct answers, so a procurement conversation about it is largely a conversation about accuracy figures. A generator cannot be judged that way, because there is no single correct output to compare against, so the questions that matter become different ones: is it good enough for this job, who checks it, and what happens when it is confidently wrong. Most of the difficulty organisations meet with AI comes from applying the first set of questions to the second kind of system.

How it works

The model predicts what should come next, over and over. Given the material so far, it produces the most plausible continuation, adds that to the material, and repeats. Whether the output is a paragraph, a picture or a function, the shape of the process is that same loop running until the thing is finished.

Plausibility is not truth, and the gap between them is the whole risk. The system is optimising for output that fits the pattern of its training, which usually coincides with being correct and sometimes does not. Nothing inside the process distinguishes the two cases, which is why a wrong answer arrives with the same confidence as a right one.

The same machinery serves every medium. Text, images, audio and code are different formats over the same idea of continuing a pattern, which is why one vendor ships all four and why capabilities tend to arrive across the board rather than in one medium at a time.

What you supply steers it more than most people expect. The instruction, the examples and the documents attached to the request all shape the output, which is why two people using the same product get results of very different quality and why the skill of asking well is a real one rather than a fashion.

It produces something every time, including when it should not. There is no state in which the system declines because it does not know, unless it has been deliberately built to recognise that case, so an absence of information becomes an invented answer rather than an empty one.

Nothing is retained between requests unless a product adds that. Each request is answered from what it was given, which is why the same question can get different answers on different days, and why continuity across a conversation is a feature the product builds rather than something the model does.

The ability and the use of it are separated in time, and that separation explains most of the odd behaviour. The capability was fixed during a long training process that ended on a particular date, while every request happens afterwards against a world that has moved on. So the system can be fluent about a subject and wrong about its current state, can be unaware of your organisation entirely, and can be excellent at the shape of a task while ignorant of the facts it needs. Products close that gap by supplying material at request time rather than by the model knowing more, which is why attaching the right document so often matters more than choosing the cleverest model.

Two kinds of question

The two columns are not two products, they are two different questions to ask software, and almost every difficulty organisations meet with AI comes from asking the second kind while still using the habits of the first. On the left the work is closed: the email either was a complaint or it was not, the customer either left or did not, and a figure can be produced afterwards saying how often the system was right. On the right the work is open. Three good drafts of the same brief will differ from each other, all three can be acceptable, and no percentage describes that. What replaces the accuracy figure is not a weaker version of it but a different apparatus: a sample of real cases rather than demonstrations, a named person who checks, and a standard set by the job rather than by the tool.

Seen in the wild

Ask an assistant to draft the same short brief three times in a fresh conversation each time, and read how far the three differ while all three remain acceptable.
ChatGPT
Give a model a document and ask for a summary, then ask for a fact the document does not contain, and watch which of the two it declines.
Claude
Ask for the same answer with sources attached, where the product retrieves material before producing anything, and compare how checkable the result becomes.
Perplexity
Put the same drafting step inside an automation so it runs on every incoming record, which is where output nobody reads becomes a governance question rather than a productivity one.
n8n

Common misconceptions

People assume

This is the same technology we already had, with a better interface.

In fact

The previous generation sorted, scored and predicted from existing records. This one produces material that did not exist before. Both are useful and they fail in opposite ways: a classifier is wrong in ways you can count, and a generator is wrong in ways you have to read.

People assume

If it produces something, it has understood the request.

In fact

Producing an answer is what it always does, whether or not the request was understood or the information was available. Output is not evidence of comprehension, which is why an instruction that was ambiguous to a person is usually answered rather than queried.

People assume

A better model would fix the mistakes we are seeing.

In fact

Most of the mistakes teams meet early are about what the system was given rather than how capable it is. A model with no access to your policies cannot follow them, and a stronger model with the same gap is confidently wrong faster. Supplying the right material usually beats upgrading the model, and costs less.

People assume

It only matters for creative work.

In fact

Most of the value organisations get from it is unglamorous: drafting routine correspondence, restructuring documents, writing the first version of something a person then corrects. Creative use is the visible part and rarely the largest.

Telling them apart

Generative AI vs Machine learning generally

Generative AI

Produces new material on request, judged by whether the result is good enough for the job.

Machine learning generally

Learns patterns from examples to sort, score or predict, judged against known-correct answers.

Ask what the output is. A new artefact means the first; a label, a score or a number means the second.

Generative AI vs An AI agent

Generative AI

Produces something and hands it back. One request, one result, and a person decides what happens next.

An AI agent

Uses that same ability repeatedly to pursue a goal, taking steps and reacting to what each one returns.

Count the decisions. If the software decides what to do next rather than only what to write, it is the second.

Questions

Is generative AI the same thing as AI?: No, it is one branch of it. The broader field includes sorting, scoring and prediction, most of which has been in ordinary business software for years without anybody calling it AI. What changed recently, and what nearly every current product is built on, is the ability to produce new material on request.
How do we judge whether the output is good enough?: Not by an accuracy figure, because there is no single correct answer to compare against. Judge it the way you would judge a junior colleague's draft: against the standard the job actually needs, on a sample of real cases rather than demonstrations, and with somebody named as the person who checks it.
Why does it sound certain when it is wrong?: Because certainty is a property of the writing rather than of the knowledge behind it. The system produces the most plausible continuation, and plausible prose is fluent prose. Nothing in the process marks the difference between an answer it has grounds for and one it does not, so both arrive in the same voice.
Does it learn from what we type into it?: That depends entirely on the product and the plan, not on the technology. Business tiers commonly commit not to train on customer content and consumer tiers often reserve the right to. It is a contractual question rather than a technical one, and the answer belongs in writing before real material goes in.
Will it replace the work or assist it?: In practice it changes where the effort sits rather than removing it. The drafting gets faster and the checking gets more important, so roles tend to shift towards judging output rather than producing it. Teams that plan for the checking get value sooner than teams that plan only for the speed.
Where should an organisation start with it?: With a job that is repetitive, low-stakes and currently done by hand, where somebody can tell quickly whether the output is acceptable. Starting with the highest-value process is a common instinct and a poor one: the stakes make the checking expensive before anybody has learnt what good output looks like.

Key takeaways

Generative AI produces new material; older AI sorted and scored material that already existed.
It predicts a plausible continuation, and plausible is not the same as correct.
One approach serves text, images, audio and code, which is why capabilities arrive together.
It always produces something, including when the honest answer is that it does not know.
Judge it against the standard the job needs, not against an accuracy figure.

Tools that use this

ChatGPT
The general case: one assistant producing text, images and code from the same box.
Claude
Where the difference between summarising a document and inventing beyond it is visible.
Perplexity
Production with retrieval attached, so the output can be checked against sources.