---
name: International Atomic Time
slug: international-atomic-time
type: timescale
status: running
version: "1.0"
released: 1972-01-01
maintainer: Bureau International des Poids et Mesures (BIPM)
dependencies:
  - cesium-133-atom
  - ensemble-of-atomic-clocks
  - general-relativity-corrections
  - coordinated-universal-time
license: Physical Law (irrevocable)
tags:
  - time
  - metrology
  - atomic-physics
  - coordination
  - infrastructure
---

International Atomic Time

What it actually is

A continuous count of seconds since midnight, January 1 1958, maintained by averaging the output of roughly 450 atomic clocks across 80 laboratories worldwide, trusting no single clock completely.

How it works

Every participating laboratory runs one or more atomic clocks. Each clock measures the resonance frequency of the cesium-133 hyperfine transition: exactly 9,192,631,770 cycles per second, by definition. The BIPM collects all readings, weights them by historical reliability, and computes a paper timescale called EAL (Echelle Atomique Libre, free atomic scale). A small correction for long-term accuracy is then applied to produce TAI (Temps Atomique International, which is the French name, because Paris).

The result is a timescale with no leap seconds, no political interference, and no opinion about whether it is currently summer.

"TAI just keeps going. That is its whole personality." — clock engineer, observed in the wild

Features

Continuous: never paused, never rewound, never adjusted for seasonal feelings
Ensemble-based: no single point of failure; the network votes
Relativistic corrections applied: clocks at altitude tick faster; the math accounts for this because general relativity is not optional
Offset from Coordinated Universal Time (UTC): currently 37 seconds ahead, a gap that grows only when a leap second is inserted into UTC
Traceable: every participating clock has a paper trail going back decades

Configuration

epoch: 1958-01-01T00:00:00Z
unit: SI second
current_offset_from_UTC: +37s
leap_seconds: not_our_problem
clock_count: ~450
weighting_algorithm: ALGOS (proprietary-ish)
relativistic_correction: enabled
geoid_reference: mean_sea_level

Known bugs

The epoch choice: TAI was retroactively defined to start in 1958, but atomic clocks in 1958 were not good enough to set it properly. The initial value was an approximation. It compiled anyway.
No direct observability: TAI is a paper timescale. You cannot query a single clock and get it. It is computed after the fact, published in arrears. You are always living slightly in TAI's past.
Drift from astronomical time: Earth's rotation is irregular. TAI does not care. The planet and the timescale slowly diverge, which is UTC's problem to manage via leap seconds, which are TAI's problem to ignore.

Changelog

Date	Event
1958	Epoch defined; atomic time begins
1972	TAI officially introduced; UTC synchronized at TAI minus 10s
1972 to present	37 leap seconds inserted into UTC; TAI offset grows accordingly
Ongoing	BIPM publishes Circular T monthly; labs argue politely about weights

FAQ

Why not just use one very good clock? Because trust does not scale. An ensemble fails gracefully. A single clock fails silently.

Is TAI the "true" time? TAI is the most consistent time. Whether it is true depends on your relationship with the concept of now.

What happens if BIPM stops publishing? Each lab continues ticking. The coordination collapses. Local atomic time continues. The universe does not notice.

drafting spec…