Ask an image generator for 3:15 and it hands back a clock at 10:10

A watch drawn to order still reads 10:10

Ask an image generator for a wristwatch showing 3:15. Almost always, it hands back a watch reading close to 10:10, both hands raised toward the top of the dial, one near the 10 and one near the 2. Ask for 7:20 and you get the same face. The pull is strongest for wristwatches and watch-like faces. Wall clocks come out at more varied times.¹

A reasonable guess is that the model read the words 3:15, worked out where the hour and minute hands belong, and drew them roughly there. That is not what happens. Nothing inside the generator ever converts 3:15 into the position of two hands.

The generator renders the most likely clock

A diffusion image generator does not draw a clock the way a person would, by choosing a time and placing each hand to show it. It begins with a field of random static and takes the noise away a little at a time until a picture remains, a process explained in full in An image generator turns a field of static into your picture. What matters here is the single thing the model was trained to do. Its training rewards one outcome: producing images whose statistics match the images it was shown. The paper that introduced these models describes the goal as a chain trained to produce samples matching the data after a finite number of steps.² Nothing in that objective rewards a rule that applies to a single image, such as make the hands show the requested time. It rewards looking like the data in aggregate.

That turns the question of what the model draws into a question about what it saw. A clock it renders will look like the clocks in its training images, in whatever way those clocks tended to look.

Ten past ten is a sales pose

Clocks and watches in photographs are not posed at random times. Since the 1950s, watch advertisements have set the hands at 10:10, where the decades before favored 8:20.³ Two forces hold the convention in place. The raised hands frame the maker's logo, which usually sits just below the 12, so 10:10 keeps the brand name clear.⁴ And the two lifted hands read as a smiling face. In a controlled study, viewers shown a watch at 10:10 reported more pleasure than the same watch at 11:30 and rated the 10:10 face as resembling a smile. The pose also raised their stated intention to buy, and it did so without their noticing.³

No one has counted the times shown on clocks across a training set the size these models use, so this is an inference rather than a tally. What is documented is the convention and the fact that product and stock photography, which make up the bulk of clock and watch imagery online, follow it.¹ The well-supported inference is that the faces the model saw lean heavily toward 10:10, binding the idea of a watch to hands near the 10 and the 2. The binding is tightest for watches. In one documented test, a generator asked for watches returned 10:10 almost every time, while the wall clocks it drew came out at scattered times.¹

Nothing turns 3:15 into two hand angles

A text prompt does change what the model draws. The words 3:15 enter the same way the word watch does, as a description the denoising is steered toward. But the model can only make the image look like a watch, and the watches it learned from look like 10:10. It has no separate step that reads 3:15, works out that the minute hand should point straight down and the hour hand a quarter of the way past the 3, and places them there. That computation was never part of what training rewarded, so the model never built it.

This is a sharper failure than a model simply being bad at detail. When a generator draws the wrong number of apples, it is missing one count. The same kind of gap turns up in an image generator turning "six apples" into "some apples" and in an image generator never counting the fingers it draws. The clock combines two specific gaps at once: a training set that leans hard toward one arrangement, and the absence of any rule connecting a requested time to an angle. It is the single-mode version of a pattern already seen in text, where a chatbot asked for a random number tends to answer 7, drawing from a lopsided distribution it learned rather than from any source of chance. The clock has one dominant answer where the random number had a few.

The pull toward 10:10 is strong enough to surface even when a model is asked to read a clock rather than draw one. On a benchmark of instrument reading, one open model answered 10:10 for 72.88% of the real clock photographs it was asked to read, regardless of the time the hands actually showed.⁵

The behavior is not fixed, and it is coming apart model by model. A documented test in August 2026 ran one prompt, a request for wall clocks reading 3:27, through ten image generators.⁶ Several still produced 10:10. One, GPT Image 2, placed every minute hand at :27 and the hour hands within about six minutes of correct. That is a demonstration rather than a measured study, and it cannot say what changed: more clock-specific training images, an added planning or tool step, or a different design would all look the same from the outside. What it shows is that the missing rule is becoming a solved problem in some systems and not others.

Reading a clock face hits the same missing step

The reading failure is the same absence seen from the other side. A vision-language model handed a photograph of a clock has to turn two hand angles back into a time, the task laid out in a chatbot slicing your photo into squares and turning each into a word. It was no more trained to compute angle-to-time than the generator was to compute time-to-angle, so it stumbles on a task most people find trivial. On ClockBench, a set of 720 questions about analog faces, people score 90.7% and the best model shown scores 66.7%.⁷ At the benchmark's launch in September 2025, the best model managed 13.3%, missing by a median of about an hour where people missed by about three minutes.⁸

That climb from 13.3% to 66.7% in a year is the part still in motion, and some of the movement is deliberate. On a set of 12,483 real-world clock photos, a base model read the full time correctly 1.41% of the time. After training aimed squarely at the habit of swapping the hour and minute hands, the same model reached 46.22%.⁹

That marks the line between what is settled and what is open. Settled: a diffusion model reproduces the statistics of its training images, and clock imagery leans toward one pose. Open: how far targeted data or an added tool closes the gap, and whether models that plan their reading or call a clock-reading tool clear it entirely. A watch coming back at 10:10 is the settled part on display. Whether a model can be made to draw or read the time you actually asked for is the part still being decided, system by system.