Training
Feeling Fast Isn't Being Fast
The feeling of speed is the most convincing thing AI gives you, and it's often wrong.
In 2025, a research group called METR ran a careful experiment. They took experienced software developers, people who knew their own projects inside and out, and gave them real work from those projects. Some tasks were done with AI tools and some without, decided at random.
Before they started, the developers expected AI to make them about 24% faster. When they finished, they believed it had made them about 20% faster.
It had made them 19% slower.
That result has stayed with me more than almost anything I've read about AI. It reaches well past software, into how easily any of us mistakes the feeling of speed for the real thing.
The Study
The details close off the easy excuses. The developers were experienced, working in code they knew well, on tasks they would have done anyway, and the comparison was randomized, so nobody picked the easy tasks for the AI.
The time went to places that don't feel like work: writing out what you want, waiting for the answer, reading it, checking it, and fixing the parts that were almost right. Each step feels small, but added up, they cost more than the help saved.
And the people doing it couldn't tell, careful and capable as they were, because they were measuring with the one instrument everybody uses: how the work felt.
Why It Feels Fast
AI is fastest at the part of the work you can see. You ask, and a full draft appears in seconds, whether it's an email, a plan, a page of code or a summary of a document you haven't read. Watching that happen feels like a leap forward, and the first draft really is faster.
The cost shows up later, and it's spread out where you don't notice it: the reread, the second prompt, the fact you have to look up, the paragraph you rewrite because it doesn't sound like you, the mistake someone else catches a week later.
Feeling measures effort, so when the hard-looking part is done for you, the job feels mostly done, when often it has barely started.
Almost Right Is the Expensive Part
In Stack Overflow's 2025 survey of developers, the most common frustration with AI, named by 66% of them, was answers that are "almost right, but not quite."
Almost right is more expensive than wrong, because wrong is obvious and you throw it out, while almost right looks finished, so you trust it, build on it, send it, and pay for it later, usually at a worse time and in front of more people.
This is true far beyond code: a contract summary that drops one clause, a number in a proposal that's close to the real one, a client's name spelled the common way, which happens to be wrong for them. Each one reads perfectly well, and each one costs more than it saved.
What It Looked Like for Me
I've lived this. When I started building with AI, I was producing at a pace that felt unreal. Systems, apps and sites that used to take a team months were coming together around me in weeks.
Then I took an honest look at what I'd built. Over time I deleted about 60% of more than a million and a half lines of code. In one place, a single person's name had ended up spelled 29 different ways across the systems, because I had never decided, once, what that thing was called.
While it was happening, it felt like the fastest I had ever moved. The speed was real, and so was the rework it created, and the feeling only ever told me about the first one.
What changed my results was deciding things once, building on a clean foundation, and checking my work against reality every step of the way. That takes skill, which I wrote about in AI Is a Skill.
Measure the Whole Job
To find out whether AI is actually helping you, you don't need a research lab, only a measure of the right thing.
Time the whole job. Start the clock when you begin and stop it when the work is truly done: checked, fixed, sent and accepted. The draft is the smallest part of that number.
Pick something you do often. A weekly report, a proposal, a type of email. Do it a few times with AI and a few times without, and compare the whole-job times honestly.
Count what comes back. Corrections, follow-up questions, things you had to redo. Rework is the cost the feeling never reports.
Judge it the way a stranger would. A result is real when someone who wasn't there could check it and agree. If the only evidence is that it felt productive, you don't know yet.
Sometimes the honest answer is that AI saves you enormous amounts of time, and sometimes a task is faster without it. Both are worth knowing, and you only find out by measuring.
Field Work
The Check
Before you trust or send anything AI helped you make, run a short check, the same kind of tool I wrote about in The Checklist. It takes a minute, and it's where the real speed comes from:
- What would make this wrong? Ask it of the work, and ask the AI too. Then look there first.
- Did I verify the facts? Names, numbers, dates, quotes and sources. These are where "almost right" hides.
- Does it sound like me? If you wouldn't say it that way, the person reading it will notice.
- Would I sign it without rereading? If not, it isn't done.
- What did I fix, and what's the rule? Every correction is a rule in disguise. Write it down, and the next draft starts closer to right.
The goal is to reach work you'd stand behind sooner than you could before. Measure that, and you'll know.
Sources: METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (July 2025); Stack Overflow, 2025 Developer Survey.