The twenty minutes ChatGPT doesn't count

The sentence sat in an essay I almost scrolled past. I did not. I sat with it the way I sit with a ChatGPT draft: longer than I meant to, slightly annoyed at myself for caring this much about a stopwatch.

Five minutes of generation. Fifteen of checking. I have done that split. I did not feel faster. I felt like I had been given a head start and then asked to run the same distance.

I was not angry at the tool. That would have been simpler. What sat with me was smaller and slightly embarrassing: the five minutes were not the work. The work was the checking. And nobody had mentioned the checking when they told me this would give me my evenings back.

I went looking for the numbers behind the sentence. Glean's Work AI Institute, with researchers at Stanford and a handful of other universities, had asked 6,000 office workers in the United States, the United Kingdom, and Australia a question everyone likes. How much time had AI given back? Eleven hours a week, they said. The survey ran across December 2025 and January 2026 and was published in June as the Work AI Index.

Then the researchers asked the question almost nobody asks. How much time do you spend feeding these systems context, watching what comes out, fixing the mistakes, and cleaning up afterwards? 6.4 hours.

They gave the residue a name. Botsitting. Rebecca Hinds, who heads the institute, called it the largely unrecognized, unbudgeted, untracked labor of making AI usable.

I laughed, which is what I do when something is a little too accurate.

There is no productivity dashboard on my phone. There is just the draft, and the extra minutes, and the quiet decision not to send the thing at all. The survey was about offices. The arithmetic was not.

The essay I started with had the measurement trick in plain sight. Adoption metrics count seats, sessions, prompts. They do not count the second attempt, or the fifteen minutes after the answer appears. On the stopwatch, nothing has changed. On the dashboard, everything has. The tool has not saved those minutes. It has reclassified them.

The simple version would be: stop using it, then you don't have to check. I don't believe that. The checking is also how I stay inside the sentence. Without it, the words go out wearing my name and none of my judgement.

The same survey had another number I did not enjoy. 69% of people using AI at work admitted to what Glean calls botshitting: shipping output they had not reviewed, did not fully understand, or could not defend if asked. That is one combined category, not a single confession. The piece that is measured on its own is ugly enough. 41% had sent work they could not explain if asked.

I have not done that with anything that mattered. I have come close with smaller things. A message I almost sent because it was "fine." Fine is how you disappear from your own messages.

Here is the part that complicates the sermon I was preparing. The people Glean calls high AI achievers, the ones who report both speed and quality, spend more of their time botsitting, not less. Forty percent of their AI time, against 33% for the people getting less out of it. They also refuse the tool on some jobs on purpose. The checkers are not the ones falling behind. The checkers are doing the part the dashboard cannot see.

In 2025 a research group called METR ran a trial with experienced software developers. Before they started, the developers guessed AI tools would make them about 24% faster. They finished 19% slower. Afterward they still estimated they had been 20% faster. METR now says those speed results are out of date; the gap between what the developers felt and what the clock recorded is the part that remains useful. I am not a developer. I am a person who felt quicker while sitting still. That gap, between the feeling of speed and the clock, is the part I cannot shake.

I thought I was going to land somewhere cleaner than this. Something like a rule. I don't have one. I still want the five-minute draft. I also refuse to pretend the fifteen minutes are leisure, or a personal failing, or proof I "don't get it." The getting-it is the checking. The tool is very good at producing something that looks finished. Looking finished is not the same as being willing to sign it.

I still use it. I still check. Naming botsitting did not make the minutes shorter. It made them honest. I am not behind because I check. I am doing the part that makes the words mine. The dashboard, if there had been one, would have counted the five. I am keeping the other number for myself.

I don't know how long you sit with the draft after it arrives. I suspect you have a version of those twenty minutes. I'd be curious what you do with them.

Kevin note: Next time a draft arrives in seconds, write down the time you actually send. The gap is the job. If you cannot say what you changed, do not send it.

A few terms came up in this one — chatbot, prompt. Both in the glossary if either is new to you.

If someone asks: "I read something about botsitting, the unpaid checking after ChatGPT drafts an email. The time the tool 'saves' is often the time you spend making sure it still sounds like you. That checking is the actual work."