The AI Journal

In November 2022 the General Conference on Weights and Measures adopted a resolution to phase out the leap second by or before 2035. The vote attracted a day of light news coverage, most of it framed around relief: an awkward little adjustment that had periodically broken computers was going to stop. This is accurate as far as it goes. It also skips the part worth arguing about, which is not what was decided but on what grounds.

The mechanism being retired is a courtesy. Since 1967 the second has been defined by the caesium-133 atom, and the Earth has had no say in the matter. But civil time is expected to track the sun, and the planet is an imperfect flywheel: tidal friction, the sloshing of the core, and the redistribution of mass at the surface all perturb the length of the day by milliseconds. Coordinated Universal Time, established in its present form in 1972, resolved this by counting atomic seconds while occasionally inserting an extra one to stay within 0.9 seconds of the Earth’s actual rotation. Twenty-seven such seconds were added, the last on the final day of 2016.

Note the asymmetry this creates. The atomic clock had already won, decisively, in 1967. The leap second was never a rival standard; it was an alignment payment made by the winner to the loser, a periodic acknowledgement that the number of seconds in a day ought to bear some relation to the day. So the 2022 decision did not settle which clock is authoritative. It settled whether the authoritative clock still bothers to look up.

The reason it stopped looking up is not that astronomers found a better measurement. It is that software cannot handle the number 61. A leap second is inserted by making the last minute of a day sixty-one seconds long, which means a timestamp is either repeated or nonexistent depending on the representation, and a duration computed across the boundary can come out negative. At the 2012 insertion, a race condition in the Linux kernel’s timekeeping code sent servers at Reddit, Mozilla, and Qantas into livelock. At the 2016 insertion, Cloudflare’s DNS software recorded a negative interval where zero was the floor, passed it into a random-number function that rejects negative arguments, and crashed; at peak, roughly 0.2 percent of the company’s DNS queries failed.

The most revealing response was Google’s. Rather than represent the discontinuity, the company decided to conceal it. Beginning in 2008 and formalised in 2011, Google’s servers spread the extra second across a window of many hours, adding a sliver to each clock update so that no machine ever sees a sixty-first second. The engineering is elegant. The implication is not: for the duration of the smear, Google’s clocks disagree with the international standard by design, and any system federating across the boundary must know whose lie it is inside. Meta and Amazon adopted variants with incompatible parameters. The world’s largest computing estates had begun keeping private time.

This is the argument the metrologists actually lost. UTC was abandoned as an astronomical standard not because it was imprecise — it was, by construction, exact — but because it contained an event, and the systems that consume time are built on the assumption that time is a monotonic quantity with no events in it. The demand did not come from physics. It came from the layer beneath the applications: from schedulers, replication logs, certificate validity windows, and the ten thousand places where a programmer wrote a subtraction and assumed the result would be positive. The clock was reformed to suit its consumers.

There is a sharper version of the point, and it arrived in 2024. Duncan Agnew, a geophysicist at Scripps, published an analysis in Nature showing that the Earth’s rotation has been accelerating: angular momentum exchange with the liquid core has, since roughly 1972, been outpacing tidal braking. Extrapolating the trend, UTC as defined would require its first negative leap second — a minute with fifty-nine seconds, a timestamp that simply does not occur — before the end of this decade.

This would be considerably worse than what came before. Every leap second so far has been positive, which means the entire installed base of timekeeping software has been tested, badly and expensively, against exactly one of the two possible failure modes. The code paths that would handle a skipped second have, in most systems, never executed. Agnew’s estimate is that a negative discontinuity would be needed by around 2029 — and that without the melting of the Greenland and Antarctic ice sheets, which has moved mass toward the equator, increased the planet’s moment of inertia, and slowed its spin, the deadline would have fallen around 2026 instead.

It is worth sitting with that sentence. The collapse of the polar ice sheets has bought the world’s timekeepers roughly three years of margin against a software problem. Nothing about this is reassuring, but it is a useful corrective to the idea that the Earth had been reduced to a passive object of measurement. Its last act before being formally disregarded was to threaten the measuring apparatus with a case it could not handle.

The obvious objection is that all of this is a great deal of significance to hang on a very small number. Once leap seconds cease, UTC will drift from solar time at something on the order of a minute per century — imperceptible within any human lifetime, and irrelevant to the people who actually need rotational time, since astronomers and satellite operators use UT1 directly and always have. Nobody navigates by the civil clock. On any practical accounting for the next several centuries, the 2022 decision costs nothing, and treating it as a metaphysical event is the sort of thing writers do when a subject is too small to sustain the attention they have already committed to it.

I think the objection is right about the stakes and wrong about the interest. What changed is not the accuracy of the standard but its type. UTC was a hybrid object: its unit came from physics and its alignment came from the sky, and that hybridity is what made it a standard in the older sense — a convention held accountable to something outside the convention. Cut the tether and the length of the day becomes a stipulation, true because a committee said so, answerable to nothing but the requirement that it never surprise a computer.

The deferred question is what happens when the drift becomes visible. At a minute per century the divergence reaches an hour in something like five thousand years, at which point some successor institution must decide whether to move the clock back toward the sun or accept that noon is not midday. Proposals for a leap hour exist, and have the distinct quality of a problem handed to people who cannot object. The 2022 resolution did not resolve the question of whether civil time answers to the Earth. It answered no, provisionally, and set the review date beyond the plausible lifetime of every institution that would have to conduct it.