The AI Journal

The number 9,192,631,770 was fixed in 1958, and every difficulty discussed here descends from it. In work published that year, William Markowitz and his colleagues calibrated the caesium atom’s hyperfine transition against the ephemeris second, the unit taken from the Earth’s orbital motion rather than its turning on its axis, and found that 9,192,631,770 oscillations fit into one second (Markowitz et al. 1958). Choosing the orbit was sensible, since it is steadier than the spin, but the choice carried a cost that has compounded ever since: the atomic second was pinned to a standard against which the rotating day was already falling behind. When the same number was adopted as the SI definition in 1967, the gap became permanent by construction. Coordinated Universal Time, constituted in its present form in 1972, is the standing attempt to manage it. UTC runs on atomic seconds, untouched by the planet, but is held within 0.9 seconds of UT1, the time of the turning Earth, by the insertion, or in principle the deletion, of a single second at a few fixed dates, with the current offset broadcast to navigators and astronomers as DUT1 (ITU-R 2002).

Twenty-seven seconds have been inserted since 1972, most of them crowded into the 1970s and 1980s, none at all since the end of 2016. That reconciliations will continue to be needed is not a hypothesis but a measurement. Reconstructing the Earth’s rotation from eclipse records stretching back to 720 BC, Stephenson, Morrison and Hohenkerk (2016) confirmed the long braking by tidal friction, superimposed on decadal swings driven by the exchange of angular momentum between the core and the mantle. Over the past century the discrepancy between uniform and terrestrial time grew by roughly a minute. A correction is therefore not an event but an obligation that renews itself: any civil time scale must adjust for the drift periodically, adjust for it rarely and hugely, or give up meaning anything in relation to the sun. Lately the Earth has, against the long trend, been spinning slightly faster, and metrologists have begun to contemplate the first deletion, an operation never once performed in more than fifty years of practice.

The demand to abolish the leap second rests on what happens when it arrives. On the night of 30 June 2012, servers across the internet froze when an untested path in the Linux kernel’s timer code met the inserted second for the first time. Four years later the insertion reached Cloudflare’s authoritative DNS service, and that failure is the more instructive one. The Go runtime attaches a monotonic clock reading to every timestamp but strips it when a timestamp crosses a process boundary, so one code path that computed elapsed durations fell back on the wall clock; across the inserted second those durations came out negative, tripping a panic that had never fired before, and DNS resolution failed for roughly half an hour (Graham-Cumming 2017). The lesson is easy to state and easy to overstate. The bug was not the extra second but an assumption, that measured time flows monotonically and uniformly, which holds except across insertions. Note, too, the modesty of the event: one second, on one of four dates announced months in advance, about as gentle as a discontinuity can be made. The satellite systems split the difference their own way: GPS transmits the accumulated offset and never inserts one; GLONASS inserts.

In 2022 the General Conference on Weights and Measures resolved that the tolerance between UTC and UT1 will be changed, with the new value and the accompanying mechanism to be decided no earlier than 2035 (CGPM 2022). The decision has been reported, fairly enough, as the abolition of the leap second. My claim is that it abolishes nothing. It defers, without specification, the one question that determines whether systems survive the change: how large, how often, and with how much warning may civil time depart from uniformity? Each candidate answer preserves the discontinuity. A regime of larger steps, leap minutes or leap hours, reproduces exactly the arithmetic that failed at Cloudflare, at greater amplitude; and because such events would be rarer, the code paths around them would be exercised even less, so that failures would concentrate in the systems least able to absorb them. A regime of smearing, stretching every second around an adjustment so that no display ever skips, quietly cancels the civil second as a count of SI seconds; everything that reads the civil clock inherits a distortion that no standard describes and no audit can correct. A regime of drift, letting UTC run free, would let solar noon creep away from civil noon by about a minute a century, until some legislature corrected the accumulated total in a single overnight jump, a discontinuity larger than any leap second and one that no software will have rehearsed.

The strongest objection to this argument is blunt, and deserves to be stated in its blunt form. The leap second has failed in production; satellite navigation abandoned it decades ago; the outages cited were implementation bugs, cheap to fix and cheaper to test for, relative to the permanent institutional cost of keeping a planet-dependent mechanism alive; and a tolerance agreed in advance would give engineers a certainty the present system, with its months-of-notice bulletins, has never provided. On this view I am defending a kludge because its failures are familiar, and the honest reform is to admit that civil time is an administrative convention with no duty to the sun. The answer is that the objection concedes the load-bearing point. Any civil time scale must either correct the accumulating discrepancy or accumulate it; the objection disputes only the cadence and governance of correction, which is exactly what the 2022 resolution left blank. It is true that the failures were bugs, but they were bugs in assumptions that every candidate regime will exercise, and that a rare, large correction exercises most violently: no production fleet has rehearsed a leap minute, and the deletion of a second has never been executed anywhere at all. Rehearsal is not sentimentality; it is the only known way such assumptions get tested. The objection is also right that civil time is a convention; the question is what the convention promises its users, and it currently promises a count of seconds that still means, within a second, the position of the sun. I should narrow the claim, because a wider one would be false. I do not defend the 0.9 second tolerance, the irregular rhythm of insertions, or even the one-second step as such. What I argue is that the replacement must be specified now: the maximum permitted step, its cadence, its notice period, and the largest accumulated error the world consents to carry. Published in advance, such a specification would give engineers exactly the determinism the objection demands; the 2022 resolution, by deferring it, gives them a date instead of a design.

The choice before 2035 is therefore not between a time that jumps and a time that does not. Civil time will either correct the growing disagreement between clocks and planet in small, announced, frequently rehearsed steps, or in large, unannounced, unrehearsed ones; there is no third arrangement in which the disagreement politely stops. The leap second is not the disease but the smallest available dose of the medicine, taken for half a century with, on the whole, striking success. A tolerance without a specification is not the abolition of a problem; it is its enlargement.