How AI is changing work: we ask, measure and keep tracking

#Published surveys (questionnaires)

These results come from asking respondents directly in a questionnaire. Each figure is for that round only—a share among respondents—and each card shows the largest result.

• One-time survey How users of AI tools say they would react if the tools raised their prices (September 2026) Survey period: September 21, 2026–September 22, 2026 / Valid responses: 297 people
  Who we asked: Respondents to an open online call for participants who said they used an AI tool in the past 30 days If that tool's monthly price rose to 2 times its current price, what would you do? 29.1% Cancel and switch to a different paid tool 16 of 55 people (base: People who pay for AI tools themselves)
• One-time survey Employer rules on generative AI and how much workers pay out of pocket per month (employees and civil servants) Survey period: September 18, 2026 / Valid responses: 276 people
  Who we asked: Respondents to an open online call for participants who said they are employees (including contract and temporary staff) or civil servants What is your company's policy on using AI for work? 44.2% No policy 122 of 276 people (base: respondents)

#Every 5 daysAI Practical Work Benchmark (measured results)

We give 4 AIs (ChatGPT, Gemini, Claude and Copilot) exactly the same work tasks under the same conditions and measure their accuracy and how long they take to answer. All scoring is done by a program; no AI scores the answers. Each task is run 3 times, so the numbers also show how much the answers vary.

Accuracy combined across 13 runs and 856 trials in total. As of November 2, 2026.

• #1 ChatGPT Accuracy: 90.7% (204 of 225 trials)
• #2 Claude Accuracy: 78.8% (164 of 208 trials)
• #3 Copilot Accuracy: 65.7% (130 of 198 trials)
• #4 Gemini Accuracy: 59.6% (134 of 225 trials)

#How we run the surveys

All tallying is done by a program; we do not let AI do the tallying. We also do not change the questions after collection starts.

• STEP 1 Set the questions We first fix in writing who we will ask and what the questions are. They do not change after collection starts. If we revise the wording, we raise the version number, and rounds with different versions are not joined by a trend line. Questions finalized down to the option wording
• STEP 2 Recruit respondents We recruit respondents through an open online call. It is first come, first served, one response per person, and limited to people with a good record of past responses. Participation is voluntary, and AInformation pays the reward. Reward paid by AInformation
• STEP 3 Check quality A program flags duplicate, blank and internally inconsistent answers using several rules. The checks are set for each survey, and a person decides case by case whether to exclude each response. Checks for inconsistencies
• STEP 4 Tally and publish We publish the results for each question together with the survey method. They are also available as JSON and CSV. JSON and CSV released at the same time

#How we report numbers

These rules are written into the program. We do not change them from round to round to suit ourselves.

#Data available now

Separate from the surveys, these are datasets AInformation's editors collect and publish themselves. All of them are available as JSON and CSV.

• AI Model Benchmark 8 models Measured results from giving the same questions

See /en/data/index.json for the list and About us for how you may use it. You can get the list of published surveys from /en/data/research/index.json.

#FAQ

Common questions about using the surveys and how we handle personal information.

• QWho runs these surveys?

  AInformation (run by Shunta Fujii) plans the surveys and pays for them itself. No one commissions or sponsors them. We never create numbers to serve anyone's advertising.

• QHow many people do you survey?

  We set a target number of respondents for each survey before recruiting. Surveys published so far: 3, with 602 valid responses in total. Each survey's page shows how many responses we collected and how many we removed in quality checks. The base for each question is on the same page.

• QCan I use these figures as “X% of people in Japan”?

  No. The survey asked people who answered an open online call for respondents, and the figures are shares of those respondents. The sample cannot be used to estimate figures for Japan as a whole, so we do not give a margin of error (±X%) either.

• QDo you use AI to tally the results?

  All tallying is done by a program, not by AI. We use AI only for groundwork when reading open-ended answers and picking candidates for articles, and a person makes the final decision.

• QCan I use the data in my own materials or articles?

  Yes. Please cite the AInformation survey name and URL as the source. The tallied figures are available as JSON and CSV. You may modify and redistribute them.

• QCould anyone tell which answers are mine?

  No. We never publish anything in a form that could identify individuals. We do not publish the original open-ended answers either. Anything used in articles is chosen by a person after removing details that could identify an individual or a company.

• QCan I withdraw my answers?

  Yes. Contact us through the contact form. If we have not tallied the results yet, we will remove your response. We keep responses for 2 years and delete them after that.

• QWhat will you survey next, and when?

  Next, we will start a recurring survey that asks the same questions every time. Asking the same questions over time lets us show changes from the previous round in percentage points. We will post the topics and timing on this page when recruitment starts.