Lessons0 of 20 done

Module 5: Judgment. Lesson 19 of 20, 8 minutes.

How to read AI news and claims.

How to tell a benchmark from a real task, a demo from a product, and a capability from something you can rely on, with six questions for any claim.

In one week you might read that a new model has passed a professional exam, that AI will replace a profession by the end of the decade, that the whole thing is a bubble, and that a chatbot gave someone absurd cooking advice. All four stories can be true at the same time, and none of them tells you whether the tool on your screen will do your work well tomorrow.

This lesson is a set of habits for reading those stories. It takes no side. The point is to come away knowing what was actually shown.

A test score is not your job

A benchmark is a fixed set of questions used to compare models: math problems, multiple-choice exams, coding puzzles, reading comprehension. When a headline says a model scored higher than most lawyers or beat human experts, a benchmark is almost always behind it.

Benchmarks are useful. They are the same for everyone, so researchers can see whether a change helped. But they measure something narrow. A model that scores well on a bar exam has shown it can answer bar exam questions. It has not shown it can handle your client, your deadline or the document with the typo in clause 14. Real work is messier, longer and less clearly graded than any test.

Three more cautions. Benchmarks get saturated: once every model scores near the top, the test stops telling you anything. They can leak: if the questions were on the web, they may have been in the training data, so a high score partly measures memory (this is called contamination). And they are chosen by the people announcing the result, who naturally pick the ones that look best. None of this makes a score meaningless. It makes it a starting point rather than a verdict.

A demo is not a product

A demo is a recording, or a live run, of a model doing something impressive once. It is usually real. It is also the best take.

What a demo cannot show you is how often the thing works, what happened on the attempts that were cut, how carefully the prompt was tuned beforehand, and whether the feature exists yet for ordinary users. A product is the same capability wrapped in reliability, limits, support and a button you can press today. The gap between the two can be weeks or years.

So the question to ask of a demo is not whether it is fake. It is: can I do this myself, today, with my own material, and what happens when I try five times?

What AGI and superhuman usually mean

Two words appear in a lot of headlines and carry less meaning than they seem to.

AGI stands for artificial general intelligence: a system that can do roughly any intellectual task a person can. It has no agreed definition and no agreed test. One person means a model that passes a bundle of exams; another means something that can hold down a job; another means something that learns new skills the way a human does. Arguments about whether AGI is near are often arguments about the definition. When you read the word, ask which one is meant.

Superhuman usually means better than people at one specific, measurable task. A chess program beat the world champion in 1997. A Go program beat one of the best players in the world in 2016. Both were real and remarkable, and neither generalized: the chess program could not play Go, and neither could file your taxes. When a claim says superhuman, find the task. It is almost always narrow. Narrow is not nothing, but it is not everything.

Can do once versus does every time

This is the distinction that matters most for anyone using these tools at work, and the one headlines blur most.

Capability is what a model can do at its best: solve this proof, write this function, read this scan. Reliability is how often it does it: on the tenth try, with slightly different wording, on your data instead of the demo's.

A model can be capable of a task and wrong a third of the time. For a brainstorm that is fine. For anything you act on without checking, a third is a disaster. Most of the frustration people feel with AI tools comes from meeting a capability in a demo and a reliability in practice. When you read a claim, ask which of the two it is about. Most are about capability.

Six questions for any claim

Keep these somewhere you will see them.

  1. Who measured it? The company that built it, an independent group, a journalist trying it out, or a stranger online? Each is useful, and each has reasons to see what they saw.
  2. On what task, exactly? A benchmark, a staged demo, or real work by real users? How close is it to what you would do?
  3. Can I try it myself? If yes, do, before forming a view. Five minutes with your own material beats an hour of reading.
  4. How often does it fail, and how? A claim with no failure rate is a claim about capability only. Ask for the error rate the way you would ask a surgeon.
  5. What does it cost, in money, time and checking? A result that takes a specialist an hour to verify may save less than it seems.
  6. What would change my mind? If you cannot name evidence that would move you, you are reading for confirmation rather than information.

The two cycles

New technologies tend to follow a familiar arc: a surge of excitement, inflated promises, a dip when reality falls short, then a slower, steadier climb as people learn what the thing is actually good for. AI has been through this before; the AI winters of earlier decades were the dips.

What is newer is the counter-cycle. For every claim that AI will do everything, there is a claim that it does nothing, that it is all hype, that the bubble is about to pop. These stories have the same shape as the hype they answer: confident, total, and more interested in a verdict than in a measurement. Both attract attention, and attention pays.

The even-handed position is less exciting. These tools are genuinely useful at a wide, jagged set of tasks, genuinely unreliable at others, improving at a pace that is hard to call, and surrounded by people with strong incentives to round up or down. You do not have to join either side. Read the claim, find the task, ask the six questions and try it yourself. That is the whole method, and it works on every headline you will see.

Try it

Find one AI claim you saw this week: a headline, a post, a colleague's remark. Write down which of the six questions you can answer from the claim itself. Usually it is one or two.

Then test reliability directly. Copy any paragraph from a news article and type: Summarize the following in exactly 25 words. followed by the paragraph. Send it five times in five fresh chats and count the words each time. Whatever the result, you now know something no headline told you: how often this particular task works for you.

Key terms

Benchmark
A fixed set of test questions used to compare models, useful for researchers but narrower than real work.
Contamination
When benchmark questions were in the training data, so a high score partly measures memory rather than skill.
Demo
A single impressive run of a capability, which shows what is possible but not how often it works.
AGI
Artificial general intelligence, a system able to do roughly any intellectual task a person can; the term has no agreed definition or test.
Capability
What a model can do at its best, as distinct from reliability, which is how often it does it.
Reliability
How consistently a model succeeds across repeated attempts, wordings and real data.

Check yourself

1.A headline says a model scored higher than most doctors on a medical licensing exam. What has been shown?
2.What is the difference between capability and reliability?
3.Which of the six questions usually gives you the most information for the least effort?

Find the task behind the claim, separate what a model can do once from what it does every time, and try it yourself before you believe anyone, including the skeptics.