In September 2017, the people building the James Webb Space Telescope ran out of spare time.
They had planned for this. Six years earlier, after a string of missed dates, NASA had rebuilt the schedule for an October 2018 launch and put 13 months of schedule reserve into it: extra time set aside, as government auditors put it, “in the event of delays or unforeseen risks.” By September 2017, problems in testing and assembly had used all of it.
So NASA asked for a later launch window, March to June 2019, which put up to 4 months of fresh reserve back into the plan.
Shortly after, the prime contractor reported that it would need up to 3 of them.
The rule that counts itself
It takes longer than you expect, even after you’ve allowed for that.
Douglas Hofstadter coined the law in his 1979 book Gödel, Escher, Bach, in a passage about predictions that a computer would soon be world chess champion, a day that, ten years on, still seemed more than ten years away.
The joke is in the second half. The obvious fix for a late project is padding: guess, then add a cushion. Hofstadter’s version says the cushion is part of the guess and can fail the same way.
On Webb, the cushion was meant for more than known trouble; the auditors’ definition covers unforeseen risks too. What fell short was how much was set aside and how the risks were judged. When an independent review board looked at the slip in 2018, its chair, A. Thomas Young, named five causes. Three of them were lack of experience in areas such as the sunshield, systems complexity and “Excessive optimism.” A reserve gets sized by the same people, with the same view of the risks, as the estimate it protects. If that view is too hopeful, both numbers are.
Berlin’s new airport ran into a similar wall at a different scale. Ground was broken in 2006; it opened on 31 October 2020, after several delays and, by the German government’s count, more than 20,000 defects found in the fire detection systems and emergency lighting alone.
Watch the cushion, not the plan
None of this proves every project overruns. Hofstadter offered his law as a wry aside, not as a measured finding.
What Webb’s auditors did do was count. In January 2014 they reported 14 months of reserve. In December 2014, 11. In December 2016, 6, still more than the space center required. By September 2017, none. The level looked acceptable until late; the steady drop had been visible for years.
They had also asked for a fresh forecast. Back in 2012 they recommended that NASA update the project’s joint cost and schedule confidence level, an estimate rebuilt from current risks. NASA agreed and didn’t do it. “An updated JCL may have portended the current schedule delays,” the auditors wrote in 2018.
Half the padding gone by week two
Imagine you’re moving your team’s monthly reports to a new data system. Your honest guess is three weeks, and because you’ve been burned before, you promise five.
By the end of week two you’ve hit an old spreadsheet nobody documented and a permissions mess in the new system. The visible work is about half done. At this pace the job takes four weeks. That still fits inside five, but half your padding is already spoken for.
That moment is the signal. If you read how fast the padding is going instead of how far off the deadline is, you can tell your manager now, with three weeks of room to cut scope or add help. Wait until it’s nearly due to report, and there’s little room left to adjust.
The machine in the estimate
The law was born from a forecast about computers. AI tools have given it a fresh test.
In early 2025 the research group METR ran a randomized trial with 16 experienced open-source developers working on 246 real tasks in projects they knew well. Before starting, they expected AI tools to speed them up by 24 percent. With the tools, they took 19 percent longer. Afterward, they still believed AI had sped them up by 20 percent.
That was early-2025 tools in one setting. In February 2026 METR said it believed developers were likely more sped up by then, but that a follow-up had been skewed by developers who would no longer work without AI, so its new data was only very weak evidence. The same note warned that self-reported speedups “can be quite unreliable.”
You can check your own sense of it without a lab. Before you hand a task to an AI assistant or agent, write down how long you expect it to take, start to finish, including your time reading and fixing what comes back. Afterward, write down what the clock says. Ten such records won’t tell you how good the tools are. They are a place to start calibrating your own estimates, and the gap between the two columns is the first thing they show you.