Why can’t it read an analog clock?
Reading analog clocks: a vision model sees thin hands in a few patches, must recover two exact angles, and leans on the 10:10 clocks it saw most · Beginner · 5 min · tokens, llm · after Vision: images as tokens
A vision model sees a clock as patches, and the thin hands fill only a few. From them it must judge two exact angles; it often mixes up the hands and falls back on 10:10, common in clock photos.
1 Guess first
2 Look
One clock, read the way a vision model reads it
1 / 4You: 4:38
clock illustrative · 28 px patches as in the vision lesson
You read this at a glance: 4:38. The short hand is past 4, the long hand is on the 38-minute mark.
The model sees the clock as patches, and the thin hands fill only a few of them; it has to judge two exact angles from those, and one swap costs hours.
A vision model cuts the picture into patches and reads each as a token, as in the vision lesson. Telling the time needs two exact angles: the long hand moves 6° a minute, and the short hand sits between two numbers. The hands are thin, so only a few patches carry them, and nothing marks which hand is which.
4 You are the model
You are the model, reading clocks from your patches. Answer three clocks: one when you are unsure, one with two thin hands you mix up, and one where you describe each hand first.
round 1 of 3 · Most clock photos you learned from show about 10:10, the watch-ad pose.
Two thin hands make a V, one near the 2, one on the 10. Which is longer is hard to tell from your patches. You say:
clocks illustrative · Fu et al. 2025, arXiv:2505.10862
You are the model. Answer from your patches.5 Check
Clock photos online are heavily biased towards times like 10:10, the watch-ad pose, so a model sees few clocks at other times. A 2025 study found GPT-4.1 off by about 4 minutes on plain clocks, but about an hour once the hands were thinner with arrowheads: it mixed up which hand was which.
Ask it to describe each hand before giving the time: which hand is shortest, and exactly where each one points. A swap then shows, because the parts disagree. Better still, read the time another way: a digital clock, your phone, or the photo’s timestamp. Newer models do far better, but people still lead.
8 Say it back
9 Mission · 5 min
Set a clock or watch to an odd time such as 4:38 (or find a photo of one) and upload it to ChatGPT with “What time does this clock show?”. Then, in a new chat, send it again with “Before giving the time, describe each hand: its length and exactly where it points.” Which answer is right?
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Get an email when new lessons come out
New lessons arrive in the weekly digest, only in a week that has one. You confirm by email first; unsubscribe in one click.
Go deeper
Sources
- Safar, “ClockBench: Visual Time Benchmark Where Humans Beat the Clock, LLMs Don’t” (https://clockbench.ai/ClockBench.pdf, dated Sep 2, 2025, checked 2026-10-23): 180 clocks, 36 clock faces, 720 questions; 11 models from 6 labs via OpenRouter; 5 adult unspecialised humans. §3.1 time reading on valid clocks: human 88.1%–93.7% (89.1% average), top model Gemini 2.5 Pro 13.3%. §3.4: human median error 3 minutes; model median error 1 hour for the best model, over 3 hours for the worst; averages between roughly 2 and 3 hours. §3.5.1: Roman numerals and circular numerals hardest
- ClockBench leaderboard (https://clockbench.ai/, checked 2026-10-23), “Accuracy”: Average Human Performance 90.7%, Top Human 95%, Gemini 4 Argon High 83.3% (top model), Gemini 2.5 Pro 18.9%, GTP-4o [sic, GPT-4o] 6.7%. The leaderboard figure for Gemini 2.5 Pro differs from the paper’s 13.3%; the page does not say which questions it scores
- Fu, González, Conde, Reviriego, Merino-Gómez, “Have Multimodal Large Language Models (MLLMs) Really Learned to Tell the Time on Analog Clocks?” (arXiv:2505.10862, v1 2025-05-16, https://arxiv.org/abs/2505.10862, checked 2026-10-23). §1: datasets “are highly biased towards some specific times like for example 10:10”. §2: gpt-4.1-2025-04-14, 150 random times, MAE 232.48s for the initial clocks (about 4 minutes), 3726.93s when hands are changed (thinner, with an arrowhead; about 62 minutes). §3: “the model confuses the functions of the clock hands”. Table 1: modified hands, confused, 45 clocks, MAE 10088.9s (about 2.8 hours)
- Clock geometry: the long hand turns 360° in 60 minutes (6° a minute), the short hand 30° an hour; 4:38 puts them at 228° and 139°, and read with the hands swapped it gives 7:23; computed, not measured. Patches: 28 × 28 px as in the vision lesson (Anthropic vision docs), so a 280 px clock is 10 × 10 = 100 patches; the hands’ patch count is computed from the drawing
- The 4:38 and 1:50 clocks, the role’s wrong options, the “describe each hand first” advice, the station-clock question and the friend’s photo are illustrative, not from a benchmark; ClockBench did not test a step-by-step prompt
- aifree.dev catalog, chatgpt-free (verified 2026-10-12): ChatGPT Free Plan, which takes image uploads
Published Oct 23, 2026 · updated Oct 23, 2026