Ask people whether a tool they chose, use daily and enjoy makes them faster, and you will get a confident yes. That is not cynicism. It is how surveys work.
Keep that in mind when you read the newest AI productivity number.
What OpenAI measured
OpenAI's Economic Research team published Science at the Frontier this week. The team analysed 580,186 scientific messages from 1,819 scientists in its ChatGPT for Academic Researchers pilot between August 20 and September 22, 2026, then surveyed 251 of them.
The usage findings are genuinely interesting. Scientists used ChatGPT across the whole research process: literature reviews, analysis, code and writing. 68% used it to develop research ideas. When they did, 35% of the time they brought a specific question, 36% of the time they brought an idea to refine together, and 23% of the time they asked ChatGPT to suggest the research question.
The headline productivity figure is different in kind. Surveyed participants estimated that the same work, at the same quality, would have taken 1.9 times as long without AI. And 98% of those who answered expected AI to speed up discovery in their field.
Usage data versus perception data
These are two very different types of evidence, and they deserve to be treated differently.
The message analysis is behavioural. It shows what 1,819 people actually did, classified by automated tools without researchers reading the content. That is cold data, and I would trust its broad shape.
The 1.9x figure is a counterfactual estimate. Each scientist imagined how long the work would have taken without a tool they had been using for weeks.
Nobody ran the same project twice. Nobody timed a control group. And the sample was users of a free pilot who chose to answer a survey from the company that provides the tool.
OpenAI is reasonably candid about this. The post says understanding whether expectations translate into faster discovery will take more research. The trouble is that "1.9x" is the number that travels, and the caveat stays behind.
The same mistake inside your company
I see this pattern in businesses every month. A team adopts AI tools, someone runs an internal survey, and people report saving six hours a week. The number goes into a board deck. Headcount plans get built around it.
Then the quarter closes and nothing in the fundamental metrics moved. Same output, same pipeline, same cycle times. The hours were real in the sense that tasks felt faster. They were not real in the sense of showing up anywhere the business could count.
There are a few reasons. Saved time often gets absorbed by more revision, more meetings or simply more polish. Work that felt faster sometimes created rework downstream, which I wrote about in the AI efficiency gain nobody audited. And people overestimate gains from tools they like, the same way they underestimate time spent on tasks they hate.
One signal worth taking seriously
Buried in the study is a detail I find more useful than the headline. OpenAI measured how often scientists followed up to clarify, check or revise ChatGPT's responses, and reports that scientists stay involved even after delegating tasks.
That is the healthy pattern. Experts using AI as a collaborator, then checking the work, is exactly how productivity gains become durable. The risk is the opposite pattern, where checking stops because the output sounds right.
If you measure anything internally, measure that. How often does a human verify AI output before it ships? Where does that rate drop to zero? That is where your hidden risk sits.
How to measure AI productivity properly
None of this requires a research lab. It requires the discipline to set a baseline before the tool goes in.
Pick two or three workflows with clear outputs: proposals produced, campaigns launched, tickets resolved, reports delivered. Record how long they take and how many come back for revision over a normal month. Then introduce the AI tooling and measure the same things for the next month.
Add one quality measure that someone outside the team owns. Client approval rates, error rates found in review, conversion rates on what was produced. Speed without quality is just faster rework.
Finally, translate time into money or output only when it actually moves. If a team saves 20% of its hours and produces 20% more, that is a real gain. If it saves 20% and produces the same, you have a capacity opportunity, not a result.
In your token bill is not an output metric, I made the same argument about AI spend. Inputs feel like progress. Only outputs pay.
The scientists may well be right that AI makes them faster. Many of them probably are. But "people believe it" and "the numbers show it" are different claims, and the gap between them is where budgets get wasted.
Measure before you celebrate. The baseline is the whole point.