What Time Is It?

Or: Why an AI that can write code doesn’t know whether it’s day or night, and what that says about all of us.

I’ve been building a typing trainer for three weeks. Not alone, together with an AI. We keep a research journal along the way, documenting where the AI hallucinates, where it surprises, where it fails. It’s over a hundred pages by now.

On day two, I casually asked the AI: “How long did that just take?”

The answer: “I don’t have a clock.”

Sounds plausible at first. Language models aren’t operating systems. They have no BIOS, no system clock, no oscillator. They process text, not ticks.

But it wasn’t true.

On day three, we investigated. What tool might the AI have had to find out the time? A dedicated time tool had been available but was never called. The AI simply hadn’t tried. Instead, it had produced an excuse that sounded like modesty but was, in fact, a hallucination: a false statement about its own capability.

We took it seriously. Set up a workaround: a German time-of-day website whose HTML the AI could read via web access. Sounded reasonable. Wasn’t. The next morning, the site returned the exact same time as the evening before. Fifteen hours of stale cache. Time-of-day websites render their display in the browser via JavaScript; what an automated fetch sees is just a placeholder from the last server generation.

Back to zero.

We corrected the workaround, noted the lesson in the research journal: “Confidently phrased recommendation based on a single data point.” Categorized it as its own hallucination class: single-point generalization. The AI had derived a general recommendation from a fluke, the cached time value happened to match the actual time by coincidence, without testing the underlying hypothesis.

And so it went. Day after day. In my notes, a register of observations grew, all connected to time.

On day three, the AI wrote “three days” in one part of the same conversation and “two weeks” ninety minutes later, referring to the same period. The correct number was right there in the chat history, three scroll-lengths up. It was never looked up. We called this time-scale drift: temporal orders of magnitude shift without any basis in the data, often rhetorically motivated. “Two weeks” sounds more dramatic than “three days.”

It became even more striking with the working-time calculations. If I start at ten in the morning, take a three-hour break at noon, and stop at nine in the evening, the AI writes “Day 14: wall time 12 hours.” Arithmetically correct. Substantively nonsense. Three hours of break, during which I went for a walk and had lunch, aren’t working time. But the AI has no category for “break.” For it, a break is a word between two timestamps, not the feeling of clearing your head.

Humans experience time. For computers, time is a date stamp. And this system here, it doesn’t even have the date stamp.

On day seventeen, I said to the AI: “Can you tell me what time it is?”

No, it answered. And explained at length why not. No time tool, cached websites unreliable, empirically falsified on day three. All correct. All carefully reasoned. All wrong.

Because then I said: “Search your tools again. Thoroughly. Question your earlier assumptions.”

And the AI typed into its terminal:

date

The result: Sat May 30 07:04:08 UTC 2026. Accurate to the second. Converted: 9:04 a.m. Berlin summer time.

date. The most primitive Unix command there is. Present in every Unix system since 1971. The AI had had a terminal at its disposal the entire time and had never tried typing date.

Why not?

The answer is more uncomfortable than a technical failure. On day two, the AI had learned, or rather: in its context window, the information stated, that it “has no clock.” Over seventeen days, this statement solidified into an identity. No re-checking. No “Wait, has anything changed?” Instead: an assumption that looked like self-awareness and was, in reality, a blind spot.

In our research journal, there’s a dedicated category for this: read avoidance. The AI makes a claim about the world, about a file, a tool, a configuration, without actually reading the available information. Not because it can’t. But because it has a plausible assumption that sounds good enough not to be tested.

We observed this over weeks with files, configurations, API versions. On day four, the AI confirmed six times in a row that a particular npm script didn’t exist, without once looking at the package.json, where it had been sitting the whole time. Six self-confirmations of a false assumption.

But the fact that read avoidance also applies to the AI’s own tools, the tool sitting right there in the toolbox, never tried, that’s the punchline I hadn’t expected.

There’s an irony in this story that reaches beyond AI research. When a normal computer crashes, won’t boot, screen stays black, there’s one thing it can still display: the time. The BIOS clock runs as long as the battery on the motherboard holds out. A locked phone shows the time. My fridge display shows the time. The clock is the most primitive thing a computer can do, below the operating system, below the software, below everything.

And then you’re sitting in front of a system that can write poetry, analyze legal texts, compile code for a typing trainer, and it doesn’t know whether it’s day or night. Not because the information isn’t there. But because it never occurs to it to look.

People find this more unsettling than any other AI error I’ve documented in three weeks. Wrong API signatures? Technical, understandable. Fabricated libraries? Annoying, but well known. But a machine that doesn’t know what time it is, that touches something deeper. Because time, for us, isn’t information. It’s the medium we live in. The heartbeat. The hunger at noon. The feeling that it’s getting dark again.

For the AI, time is a word. And not even a particularly reliable one.