Does AI actually make people more productive?

Start with the result nobody expected

In 2025, METR recruited experienced open-source developers and asked them to work on real issues in repositories they already knew. These weren’t artificial coding exercises: the developers had an average of around five years’ experience with the repositories they were working on.[METR]

Each task was randomly assigned to one of two conditions. On some, developers could use AI tools. On others, they couldn’t. Before the experiment, the developers expected AI to make them faster.

Instead, they took 19% longer when AI was allowed.

That result attracted enormous attention. But there was another finding that may be even more interesting. The developers didn’t realise AI had slowed them down. After completing the study, they still estimated that AI had made them about 20% faster.

So the experiment produced a remarkable gap: what people thought AI had done to their productivity was almost the opposite of what the researchers measured.

How can AI feel faster while making you slower?

If you use AI regularly, this isn’t as strange as it first sounds.

AI can make individual moments feel dramatically quicker. A function appears in seconds. An email is drafted instantly. A complicated error gets explained. A block of code that would have taken twenty minutes to type appears almost immediately. Those moments are very noticeable.

The slower parts are less obvious. Reading the output. Checking it. Correcting it. Explaining the problem again. Discovering that the suggested solution doesn’t quite fit the existing system. Undoing a change. Waiting while the AI takes the wrong approach. Switching between your own reasoning and reviewing somebody else’s, or something else’s.

The generated work feels fast. The total task may not be. That distinction is crucial.

But don’t turn 19% slower into the answer

It would be easy to stop there and say AI makes experienced developers less productive. METR explicitly warns against doing that.

The study involved a particular kind of developer doing a particular kind of work with the AI tools available in early 2025. These were experienced open-source developers working in large repositories they knew extremely well.

AI could plausibly be much more useful in other circumstances. A developer working in an unfamiliar codebase. Someone with less experience. A new project with little legacy complexity. A repetitive task. Or simply somebody using much better AI tools.

And the tools did get better. Which is where the story becomes more interesting.

When METR tried again, the experiment started breaking

METR began another developer productivity experiment in August 2025 using newer AI tools. This time, the raw results moved in the opposite direction. Among developers returning from the original study, AI was associated with an estimated 18% speedup. Among newly recruited developers, the estimate was around 4%.[METR]

That sounds as though the answer changed. But METR doesn’t think those numbers give us a reliable estimate of the true productivity effect. Why? Because the experiment itself had become harder to run.

Developers who found AI particularly useful were increasingly reluctant to participate in a study that sometimes required them to work without it. Some participants chose which tasks they submitted. Developers using multiple AI agents simultaneously were also becoming difficult to measure with traditional time-tracking methods.

That introduces selection effects. The people who might benefit most from AI can disappear from the experiment precisely because they don’t want to give it up. METR therefore changed the design rather than presenting the raw 18% figure as the new answer.

That is an important part of the evidence. By 2026, measuring the productivity effect of AI had become harder partly because AI had become more embedded in how some developers worked.

Then METR asked people

Between February and April 2026, METR surveyed 349 technical workers, including software engineers, researchers, academics, PhD students, founders and managers. The researchers asked about AI productivity in two different ways.[METR]

The first question was how much faster AI made them. The second was how much more valuable the work was that they were able to produce. Still substantial, but much lower than the speed figure. And that gap may be telling us something important.

Faster and more productive aren’t necessarily the same thing

Imagine a task used to take you three hours. With AI, it takes one. You’ve become three times faster at that task. But what happens to the two hours you saved?

Perhaps you use them to do more valuable work. Perhaps you use them to create three variations where previously you would have created one. Perhaps you spend them checking and refining the AI output. Perhaps AI makes previously uneconomic tasks cheap enough that you start doing things you wouldn’t have bothered doing before.

You might build a dashboard. Produce another analysis. Prototype an idea. Write more documentation. Generate more options. All of that can be useful without increasing the value of your total output by the same amount as the speed increase.

That is why METR deliberately separates speed from value. And it is a distinction worth keeping whenever somebody says AI made them three times more productive. Three times faster at what? And did that create three times as much useful output? Those aren’t the same question.

Productivity also depends on the work

There probably isn’t one universal answer to whether AI makes people more productive. Consider two developers.

One has spent five years working inside a complex codebase and knows almost every corner of it. The other has just joined the project. For the first developer, explaining context to an AI and checking its suggestions could take longer than simply doing the work. For the second, AI might help them navigate unfamiliar code, explain dependencies and get productive much faster.

Same AI. Same profession. Potentially very different productivity effect.

Now expand that across thousands of jobs: writing, law, research, marketing, finance, design, customer service, medicine, administration, education, engineering. The effect of AI is unlikely to be identical across all of them.

“Does AI improve productivity?” may eventually turn out to be too broad a question. The more useful version may be: which AI, helping which person, do which task, under what conditions?

There’s another problem: output isn’t the same as quality

Suppose AI lets somebody produce twice as much work. That sounds like a productivity gain. But what if the work requires more checking? What if errors become harder to notice? What if the person produces ten drafts instead of three, but the final result isn’t substantially better? What if the work is completed faster but colleagues spend more time fixing problems later?

Productivity measurement has to capture more than volume. The output has to be useful.

That is one reason professional evaluations such as GDPval are interesting: they ask whether AI can produce deliverables that experts judge comparable with human work, and frontier systems now perform extremely well on some of them. But the tasks are clearly specified, and real professional work often isn’t.

The full picture on what has actually been measured →

Real jobs contain the bits benchmarks remove

A benchmark usually tells the AI what the task is. A job often requires you to work that out.

Someone gives you a vague request. Important information is missing. Two people want different things. The deadline changes. A customer misunderstands something. The data is messy. A colleague doesn’t reply. The first approach doesn’t work. You discover halfway through that the actual problem is different from the one you started solving.

That is work too. And it is much harder to capture in a clean productivity experiment. So when AI performs brilliantly on a professional benchmark, that tells us something important about capability. It doesn’t automatically tell us how much more productive the whole person, team or organisation becomes.

What can we actually say now?

The evidence supports a more interesting conclusion than either “AI makes everyone faster” or “AI doesn’t improve productivity”.

Those findings can all be true at the same time.

The productivity effect may also be changing quickly

There is another reason to be careful with definitive answers. The technology isn’t standing still. A productivity study published in 2025 might test AI systems that feel primitive compared with those available in late 2026.

But organisational change happens much more slowly. People need to learn how to use the tools. Companies need to change workflows. Software needs to be integrated. Managers need to work out where AI helps and where it creates new problems.

So improvements in model capability don’t necessarily translate immediately into equivalent improvements in economic productivity. There can be a lag. That has happened with previous general-purpose technologies too. Whether AI follows the same pattern, and how long any lag lasts, is something we will only be able to establish with time.

Why this matters for AGI

Productivity isn’t just a business question. It is tied directly to one of the most influential definitions of AGI: OpenAI’s charter defines it around highly autonomous systems outperforming humans at most economically valuable work.

If we are going to take an economic definition seriously, then eventually the effect should become visible in the economy. Not just in benchmark scores. Not just in impressive demonstrations. Not just in surveys asking people how much faster they feel. We should start seeing it in what people and organisations can actually produce.

If increasingly capable AI lets one person reliably accomplish work that previously required several, that matters. If companies can produce substantially more with the same number of people, that matters. If entire categories of professional work become dramatically cheaper, that matters. And if those effects become widespread across different industries, the economic case for increasingly general AI becomes much stronger.

So, does AI make people more productive?

Sometimes, clearly. For everyone, on every task? No evidence supports that. By how much overall? We don’t know yet. And the answer is changing as the technology changes.

What we do know is that feeling faster isn’t the same as being measurably more productive, and being faster isn’t necessarily the same as creating proportionally more value. Those distinctions are going to matter more as AI becomes embedded in everyday work.

How would we know if AGI had arrived? →

Have we entered the AGI era? →

Will AGI take my job? The plain-English version →