Slower, and sure they were faster.

The question

Is what I sell still worth selling? The Temporal question is whether the time I am supposedly saving accrues to me. Underneath it is the principle the whole inquiry runs on: self-report lies at the moment you most need it to hold. On this build the question is whether the agent made this task faster. On the long clock it is whether I can trust my own read of that at all.

Here is a group that measured the gap between the feeling and the clock, on developers who had spent years in the repositories they were working in, and then measured its own measurement and said where it broke.

What they found

In early 2025: sixteen experienced contributors to large open-source repositories, working on 246 real issues from their own projects, each issue randomized to AI allowed or not. With AI, issues took 19 percent longer. Before the study the developers expected AI to speed them up by 24 percent. After it, having been slowed down, they still believed it had sped them up by 20 percent.

In February 2026 the same group published what happened next. A late-2025 run with 57 developers and more than 800 tasks estimated 18 percent slower for the original developers, with an interval from 38 percent slower to 9 percent faster, and 4 percent slower for newly recruited ones, from 15 percent slower to 9 percent faster. Then the authors said what the numbers could not: developers who did not want to work without AI had stopped taking part, the pay rate had dropped, and the tasks chosen were the ones people expected AI to help with. They believe developers are likely more sped up in early 2026 than a year before, call their own data only weak evidence for the size of it, and are moving to other designs. Self-reported speedups remain very high, and, in their words, those estimates can be quite unreliable.

The three questions

Every research post runs the same three rows. Which of the nine it names, how good the evidence is, and whether the thing is being measured.

What this cannot say:

1. Whether AI speeds up experienced developers in 2026. The authors' own limit, stated plainly, which is why the row above reads open.
2. Anything about the developers who would not work without AI. They left the study. The people most sure of the speedup are the least measured, and the design cannot fix that from inside.
3. Whether time saved on a task became time. A faster task is not a shorter week. Nothing in either paper reaches the dividend.
4. What faster is worth when the reviewer is the one who was sped up. The slowdown in 2025 was traced largely to reviewing and integrating what the tool produced. The clock and the judgement are the same person.

What I am doing about it

The status says what is missing. These are the moves, on my own work, that would supply it. Each is either read automatically off the footprint or asked of me on a cadence, and each costs something to keep doing.

1. AutoNever publish a speed claim about my own work without a defined output metric beside it. Hours alone are the floor. A speedup with no numerator is a feeling with a chart.Costs: no headline number, ever.
2. Asked · WeeklyEstimate the week's speedup before I look at the line, and keep both. The gap between the guess and the reading is the Epistemic signal I care about most, and this trial is the reason.Costs: a weekly record of being wrong in a known direction.
3. AutoKeep the hours line on the page saying holding, not rising until the quarterly answer says where the hours went. A page that refuses to say I am faster until it can say faster at what.Costs: a page that refuses to say I am faster.
4. AutoRead the design changes, not just the results. When a group I trust says its data got hard to interpret and shows its work, that is the finding. The standing commitment on this page is the same move made in advance.Costs: publishing my own broken readings with the same visibility as the clean ones.

What I keep