Seeing a generated image

The next screen shows two pairs. First, banks of brass post office boxes on a plain wall in a federal building in Arkansas. Then two shelves of a general store — tinned food above, glass canning jars below — in a museum in Texas. One picture in each pair came out of a camera. The other did not exist until we asked for it.

Each of the made-up ones took one sentence, about a minute, and no skill at all. That is the part worth sitting with — not that a machine can produce a convincing picture, which you knew already, but that producing these particular pictures cost about as much effort as sending a text message.

So the demo asks you the same fair question twice, and both are worth answering before you read on. Look at each pair. Decide which one came out of a camera. Then find out what your decisions were actually worth.

Image B and Image C were each made in about a minute from one sentence, with the photograph they stand beside handed to the model as a reference so that the subject, the framing and the light would match. Nothing below depends on seeing any of the four — it is all written out.

Neither difference is a glitch. No extra fingers, no melted lettering, nothing obviously broken. And the two are not the same difference twice, which is the reason there are two pairs here rather than one.

Image D's jars carry their names in the glass itself: raised letters, no ink, moulded in when the jar was made. So a word is only ever a bright edge and a dark one, and whether you can read it depends on what happens to be behind the glass and where the light fell. On that one shelf two jars say the same three words, and on one of them the words have almost gone. Under several of the jars is a smaller line that cannot be read at all.

In Image C every jar can be read. The strokes hold one even weight, close all the way round each letter, and keep the same contrast at the top of the jar and the bottom, over a pale background and a dark one. That is lettering laid on a surface rather than displaced into it. The model has been shown a great many pictures in which words are made of ink, so it made these of ink.

Which is a different fault from the first pair's. There the numbers looked like numbers and did not count: what failed was meaning. Here what fails is material. The two are not sealed off from each other — one small jar in Image C is lettered with a phrase that is not a phrase — but they are different enough that one pair could not have shown both, and there is no reason to think there are only two.

Everything on the last screen is true of these four pictures and of nothing else. Ask for a thermostat on the wall and the next one has a thermostat. Ask for lettering half lost in the glass and it will be half lost. None of it is a defence, because every item is one sentence away from being fixed, and the version that fixes it will not mention that it has.

That is why this ends here rather than with a checklist, and why there are two pairs rather than five. Two was worth doing because the two go wrong in different ways, which one pair could not show. Five would be practice at a skill this site argues against having. Two right answers is what a coin manages a quarter of the time; getting one wrong, on pictures as ordinary as these, is a useful thing to have learned here rather than somewhere that matters.