Large language models can’t read analog clocks. That’s not a punchline — it’s a benchmark result. When ClockBench launched in 2025, humans averaged 89% accuracy reading analog clock faces. The best-performing AI model managed 13.3%. The leaderboard has improved since, but the gap remains vast. This is frontier AI — systems that can pass medical licensing exams, write legal briefs, and generate working code — stumped by something most of us learned in primary school.

I discovered this by accident over a weekend, spending more than an hour testing Claude Opus 4.6 with images of analog clocks. What I expected to find was a reasoning failure. What I actually found was something more interesting: the reasoning was fine. When I described where the hands were pointing, Claude could calculate the time without difficulty. The failure was upstream — in the vision encoder, the component that converts an image into the vector representations the model reasons over. The evidence points to the encoder not preserving angular-spatial information with enough fidelity for the reasoning layer to know where the hands are.

So I kept testing. Grok. Gemini. ChatGPT. Claude. DeepSeek. Every major frontier model, the same result — wrong answers, confused hand identification, confabulated explanations. I showed Grok 4.3 and Claude Opus 4.7 the same 6:00 clock — hands in a simple vertical line. Both identified the hands correctly, then swapped which was which, and calculated 12:30. Grok spent 85 seconds reasoning, oscillating between 6:00 and 12:30 before landing on the wrong answer. Claude decided the clock must be stopped. The reasoning was meticulous in both cases. But no amount of reasoning depth can correct an error that happens before reasoning begins. The vision encoder feeds the wrong spatial data in, and the intelligence faithfully processes it to a wrong conclusion. Days later, Anthropic released Claude Opus 4.7 with triple the image resolution of its predecessor — up to 2,576 pixels, roughly 3.75 megapixels — and a 98.5% visual acuity score on computer-use benchmarks. It still can’t reliably read an analog clock. More pixels don’t help when the encoder doesn’t represent angular-spatial relationships in the way the task requires. This isn’t a resolution problem. It’s a representational one.

Then I tried Qwen 3.6, Alibaba’s open-source model family. And something unexpected happened: it got them right. Not just the easy ones — I gave it a clock lying on its side and it identified the rotation, mentally corrected for it, and read the time accurately. When I flipped a clock upside down, it recognised what I’d done but couldn’t work out the time from that orientation — still, it was the only model that even identified the spatial transformation. Across Grok, Gemini, Claude, and ChatGPT, none could reliably handle even standard clock orientations. Qwen was the only model that came close.

The real kicker? Qwen 3.6-35B-A3B — a Mixture of Experts model with 35 billion total parameters but only 3 billion active at inference — outperformed Claude Opus 4.7, ChatGPT 5.4, Gemini, Grok, and DeepSeek on straightforward clock reading in my tests. This isn’t a formal benchmark result — it’s a pattern from repeated informal testing. But it’s a useful clue. Whatever Qwen’s vision encoder is doing differently, it appears to retain more of the angular-spatial structure that other architectures flatten or lose in translation.

And that discovery led me somewhere I wasn’t expecting. Because when you spend an hour watching an AI fail at clock reading, you start to notice just how much invisible work you’re doing when you glance at a clock. The long hand is on the 8 — so that’s 8×5=40 minutes. You just do that. Instantly. Without thinking. I caught myself mid-calculation and thought: oh. I didn’t even realise I do that.

But here’s the question that actually matters: could you still do it?

The Quiet Disappearing Act

A YouGov poll found that 95% of Americans over 65 can read an analog clock instantly. Among adults under 30, only 43% can. A 2017 Oklahoma study found only one in five children aged 6–12 could read one. Some UK schools reportedly began replacing analog clocks in exam halls as early as 2018 because students couldn’t use them to manage their time. In New York, a viral moment followed the city’s phone ban in schools — students could finally focus in class, but they couldn’t tell what time it was because they couldn’t read the wall clocks.

A longitudinal study across grades 3–12 found students didn’t reach 83% proficiency on numbered clocks until 5th grade, struggled with numberless clocks until 8th grade, and roughly 10% of high school students never reached proficiency at all.

Schools still teach it. The skill just doesn’t stick because the environment no longer demands its practice. One teacher put it plainly: “We still teach this. They just don’t practice using it anymore.”

It’s Not About the Clock

Here’s what most people miss when this topic comes up: reading an analog clock is not the same cognitive task as reading a digital one. A digital display gives you a number. An analog clock asks you to interpret a spatial relationship — the angle of two hands against a circular scale, one overlaid on the other, each operating at a different rate.

That’s visuospatial reasoning. It involves fractions, angular estimation, and a continuous circular mapping of a linear quantity. And it’s fuzzy — beautifully, usefully fuzzy. Nobody reads an analog clock to the minute. You say “just gone one” or “nearly half past” or “quarter to three.” You’re doing spatial estimation and then translating it into natural language that itself has flexible social conventions. Digital time gives you 13:47. Analog time gives you a felt sense of where you are in the hour.

That fuzziness matters more than people realise.

The Tyranny of False Precision

I was on a call recently where someone joined at one minute past the hour and apologised for being late. One minute. On an analog clock, one minute past the hour is indistinguishable from the hour itself — the minute hand hasn’t visibly moved. You’re “on time” because the representation doesn’t have the resolution to say otherwise.

But 13:01 on a screen? That’s a fact. You’re late. It’s logged. The Teams notification reads “started 1 minute ago.”

Digital time didn’t just change how we read time — it changed how we experience it. It manufactured a granularity of accountability that analog never imposed. The fuzziness of analog clocks wasn’t a flaw. It was a social grace period built into the technology itself. We replaced a humane imprecision with a punishing exactness, and then we wonder why everyone is anxious about punctuality.

The Bigger Pattern: When We Stop Interpreting the World

We’ve already watched this happen once, in real time.

When TomTom and Google Maps arrived, paper map reading disappeared within a generation. Map reading required you to orient yourself physically — rotate the map, triangulate landmarks, hold a mental model of where you are relative to where you’re going. Satellite navigation reduced that to “turn left in 200 metres.” The cognitive load dropped to almost zero, which felt like progress. But we know now what it cost.

Research published in Scientific Reports found that habitual GPS use is associated with measurably worse spatial memory — and critically, the impairment transfers to situations where GPS isn’t available. The GPS use appeared to be causing the decline, not compensating for it. London taxi drivers who completed The Knowledge — learning every street by navigating it — had measurably larger hippocampi. Their brains physically grew in response to spatial navigation demands. GPS users don’t get that growth. They’re using a different brain system entirely — the caudate nucleus, which handles stimulus-response learning. Efficient, but it doesn’t build maps.

We replaced something cognitively hard with something easy, and a measurable cognitive capacity shrank as a result. That’s not a theory — it’s a documented outcome. And it’s the same pattern now playing out with analog clocks, compasses, manual gearboxes, cooking by instinct rather than following an app timer, estimating distances by eye. All of these require you to take an abstract quantity and ground it in physical, spatial awareness. You’re not reading a number — you’re interpreting a relationship between your body, a tool, and the world. And one by one, digital convenience is removing the need.

I’ve never held a compass. Not because I couldn’t learn — because the environment never once demanded it. That’s how these skills disappear. Not through active rejection, but through quiet irrelevance.

What Else Are We Losing?

The children who can’t read analog clocks and the AI systems that can’t read analog clocks are failing for structurally similar reasons. Not because the knowledge is missing — both have been exposed to clocks. But because neither has had the spatial reasoning practice that builds the underlying representational capacity. The knowledge is there. The navigation of that knowledge isn’t.

And nobody noticed, because the surface task is handled. Your phone tells you the time. The digital display works. The skill disappeared and the world kept turning — if you’ll excuse the expression.

That’s what makes this unsettling. Analog clock reading is visible enough that someone eventually thought to test it. But how many other spatial-interpretive skills have already gone without anyone building a benchmark? Nobody is measuring whether people can still estimate distances by eye, or read a topographic map, or judge how full a container is without a measuring line. These aren’t skills anyone tests for. They’re the kind of thing you only notice you’ve lost when you need it and it’s not there.

Kris Perry, executive director of Children and Screens, framed the question well: is this a cognitive downgrade or just a replacement? I’d push on that framing. It’s not an either/or. The surface task — knowing what time it is — has been replaced perfectly well by digital displays. But the cognitive exercise that analog clocks provided — spatial estimation, circular mapping, fuzzy interpretation, translating between representations — that hasn’t been replaced. It’s just gone.

The Language Death Parallel

A language dies roughly every 40 days. Of the approximately 7,000 languages currently spoken, researchers predict that 1,500 could disappear by the end of this century. The 20 most common languages, each with over 50 million speakers, are spoken by half the world’s population. The consolidation is relentless.

Languages don’t die because they’re wrong. They die because the environment stops requiring them. And with each one goes not just vocabulary but an entire way of encoding reality — concepts, relationships, and patterns of thought that have no equivalent in the dominant tongue. When a language disappears, a cognitive framework for understanding the world goes with it.

Spatial-interpretive skills are undergoing the same process. Analog clock reading is one dialect. Map reading is another. Compass navigation another. Each one is a way of understanding quantity, direction, and position through physical interpretation rather than digital readout. Each one is quietly going extinct — not because it was rejected, but because the environment no longer selects for it.

The parallel with language death is not exact, but it rhymes uncomfortably well. We’re converging on a smaller and smaller set of cognitive tools — predominantly digital, predominantly precise, predominantly disembodied — and losing the diversity of ways humans have historically interfaced with the physical world.

Does It Actually Matter?

The honest answer has layers.

For clock-reading specifically? Probably not. It’s a cultural artefact. Nobody mourns the sundial.

For what clock-reading indexes? That’s where it gets uncomfortable. Analog clocks are one of the first places children encounter a continuous circular mapping of a linear quantity. That’s the same representational family as compass bearings, phase angles, cyclical patterns, seasonal reasoning. When you strip the training environment of that exercise, you’re not just losing “tell me what time it is” — you’re losing practice on a whole class of spatial-quantitative thinking.

For the convergence between human and AI cognition? That’s the real question. We’re shedding embodied reasoning capabilities from above while AI sits below, never having developed them. And neither side is alarmed because the digital layer handles the surface task perfectly well.

But here’s the thing about redundancy: you don’t miss it until you need it. We’re building an increasingly AI-augmented world in which both the humans and the AI share the same blind spot — a reduced capacity for spatial, embodied, physically-grounded reasoning. The backup system is degrading, and the primary system never had it.

What happens when both partners in a collaboration can’t see the same thing?

Maybe it doesn’t matter. Maybe digital precision is simply better, and the fuzzy spatial reasoning of analog tools was always just a workaround for not having screens. Maybe my inability to use a compass is no more tragic than my inability to shoe a horse.

Or maybe — just maybe — we’re optimising for convenience in a way that silently removes the cognitive exercise that built capabilities we haven’t yet learned to value.