In 1914, Britain’s Royal Navy had 62 capital ships in commission. By 1928 it had 20. Officers and men fell from 146,000 to 100,000. From 1922 its strength wasn’t even expected to grow back: the Washington Naval Agreement of that year limited its ships.
The Admiralty, the office in London that ran this shrinking navy, went the other way. In 1914 it employed 2,000 officials. In 1928 it employed 3,569, an increase of 78.45 percent.
Critics at the time called it “a magnificent Navy on land.” In November 1955 the historian C. Northcote Parkinson laid the figures out again in The Economist, and drew a stranger conclusion. The officials, he wrote, “would have multiplied at the same rate had there been no actual seamen at all.”
A postcard to Bognor Regis
Parkinson opened the essay somewhere much smaller than the Navy. “It is a commonplace observation,” he began, “that work expands so as to fill the time available for its completion.”
His example is an elderly lady of leisure sending a postcard to her niece at Bognor Regis. An hour to find the postcard. Another to hunt for her spectacles. Half an hour for the address, an hour and a quarter to write it, twenty minutes deciding whether to take an umbrella to the postbox in the next street. A whole day. A busy man, he noted, would have done the same job in three minutes.
Give a job a week, and a week is often what it takes.
The condition matters. Parkinson’s word was elastic: work, “and especially paper work,” is elastic in its demands on time. Polishing a postcard has no natural end point, and neither does polishing a memo. Nothing in the task itself says when it’s done, so the clock decides.
He called the idea Parkinson’s Law, a “new scientific law” presented “to the public for the first time,” and he ran it on government offices. Two forces, he said, drive the growth. An official wants to multiply subordinates, not rivals. And officials make work for each other. In his telling, an overworked official named A asks for two junior assistants rather than one colleague of his own rank, who would be a rival for promotion. Soon seven people handle a letter that one used to answer. Each drafts, amends and passes it on, and A ends up writing the reply he would have written anyway. “Far more people have taken far longer to produce the same result. No one has been idle. All have done their best.”
Parkinson wrote all of this with a straight face, down to a formula claiming staff would grow between 5.17 and 6.56 percent a year “irrespective of any variation in the amount of work (if any) to be done.” Read that range as part of the joke, not as a measured law. It is a claim inside a satirical essay, and Parkinson himself said fuller statistics would be “inappropriate in such an article as this.”
Two plus three
A machine has no pension and no rivals. Yet the postcard turned up in AI research anyway.
In December 2024, a team of researchers asked a reasoning model, QwQ-32B-Preview, “what is the answer of 2 plus 3?” Its first attempt got it right in 39 tokens, the small chunks of text these models count in. Then it kept going. It broke the sum into smaller steps. It tried an analogy with counting objects. The full answer ran to 901 tokens and 13 separate solutions, all ending at 5.
Another team that month tried giving a model a budget. On a word problem about how many hours someone’s after-work activities would take, GPT-4o-mini thinking step by step used 258 tokens. With one line added to the prompt, “use less than 50 tokens,” it used 86 and still got the answer right.
Then they set the budget at 10. The model used 157 tokens, nearly twice as many as with a budget of 50. In that test, tightening the limit from 50 to 10 bought more work, not less.
Those numbers come from one model on one question. If your AI tool shows token counts, you can see whether it holds for yours: take a question you ask often, add a budget line, and compare the count with and without it.
A proposal with no natural edge
Imagine your manager asks for a two-page proposal “by the end of next week.” You open a blank document, then close it. There are nearly two weeks to go.
So you try something else. You block ninety minutes the next morning, put the phone in a drawer and write. The rule is that you stop writing when the time is up, even if the second page is rough, and allow yourself one short check afterward, fifteen minutes, before it goes out. Scrum, a framework for running software teams, writes the same move into its rules. The Scrum Guide sets sprints at a fixed length of one month or less, and caps the planning meeting for a one-month sprint at eight hours.
A box like that suits elastic work, the kind with no natural edge of its own, which is exactly what Parkinson’s postcard was.
When the timer goes off, you stop. In the fifteen-minute check you read the draft once, fix the second page and send it. A week later, the afternoon you had been saving for it is still blank in your calendar.