AI (artificial intelligence)

The word does sales work far more often than descriptive work. It appears on products doing something genuinely new and on products doing what a spreadsheet formula did a decade ago, and the label reads the same in both cases. Translating a claim about AI into a claim about behaviour, what goes in, what comes out, how often it is wrong and what happens when it is, is most of what buying well in this market consists of. It also guards against the opposite mistake of dismissing the whole category because so much of the marketing is thin.

The boundary of the word has always moved, and it moves in one direction. Chess programs, spam filtering, handwriting recognition and route-finding were all called artificial intelligence when they were new, and all of them are now just called software. Nothing about them changed when they crossed over; they simply became reliable and familiar. The label tracks unfamiliarity rather than any technical property, which is worth remembering when a product's newness is being sold to you as its substance.

Nearly everything sold under the word today rests on the same underlying change of method. Instead of a person writing rules for the software to follow, a system is shown a great many examples and derives its own patterns from them. That single shift explains both the strength, which is coping with messy input nobody could write rules for, and the weakness, which is that it cannot tell you why it produced an answer and behaves unpredictably on material unlike anything it was shown.

Underneath the umbrella sit families doing quite different jobs, and they fail in quite different ways. Systems handling language, systems handling images, systems handling speech and systems predicting a number from a table of past records are separate lines of work. A product excellent at one is not thereby competent at another, and the shared label is what makes that easy to forget when a vendor moves between them in one sentence.

What all of them share is that they produce likely answers rather than certain ones. The output is a best guess, and the confidence behind it is not shown to you, which is why every deployment that matters has something checking the result. Expecting one of these systems to know a thing, in the way a database knows a thing, is the mental model behind most of the disappointment in the field.

• Ask a general assistant to summarise a document and then to judge whether the summary is fair to it, and notice the second request is much harder although both arrive under one word.

  ChatGPT

• Meet the same label as a feature inside a product you bought for something else entirely, where it names a few additions to a notes tool rather than the tool itself.

  Notion AI

• Search a company workspace and get results ranked by what you meant, which is the same two letters covering a job with nothing conversational about it at all.

  Glean

Is AI the same as machine learning?

Not quite, though in current commercial use they are close to interchangeable. AI is the older, broader label, and covers rule-based systems that involve no learning at all. Machine learning names the specific approach of deriving patterns from examples, which is what nearly everything now sold under the wider word actually does.

Does AI-powered on a product tell me anything?

Very little on its own. No threshold, standard or certification sits behind the phrase, so it appears on genuinely capable products and on ordinary ones. Treat it as a signal about the marketing department and go straight to what the feature does, on what input, and how you would notice when it is wrong.

Is AI actually intelligent?

That depends entirely on what you take the word to mean, which is why the question generates more heat than light. What these systems demonstrably do is produce useful output on tasks that used to need a person. Whether that constitutes intelligence is a philosophical argument, and it has no bearing on whether a given tool helps you.

Where should a business start?

With a task you already do, that is frequent, dull and easy to check, rather than with a technology you would like to have used. Run it through a general assistant for a fortnight and measure. Starting from a capability rather than a task is a well-trodden route to an expensive pilot that quietly ends.

• AI labels a category rather than describing a product, so a vendor claiming to use it has told you almost nothing.
• The boundary moves in one direction: work stops being called AI once it becomes reliable and familiar.
• Nearly everything under the word today learns patterns from examples rather than following written rules.
• The families beneath it fail differently, so competence at one job implies nothing about another.
• All of them produce likely answers rather than certain ones, which is why a checking step is not optional.

• ChatGPT

  The general assistant most people meet the word through.

• Notion AI

  The same label as a feature inside a product bought for other reasons.

• Glean

  Workspace search, where the label covers a job with nothing conversational about it.