Manus builds a student survey and tests it with synthetic data
Andrew Maynard tested Manus by asking it to design a student survey, simulate respondents and produce a report, explicitly using fabricated data rather than human participants.
Original prompt
Shared by the source author. See source notes for context.
to survey undergraduates about their hopes and fears about artificial general intelligence
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- Andrew Maynard
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The experiment produced a survey script, synthetic dataset and polished report, demonstrating a multi-step research workflow while documenting that its apparent findings were fictional.
- Human involvement
- Maynard ran the generated script himself. All respondent records were synthetic; the report is not a real student survey and its findings must not be presented as empirical evidence.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Ask Manus to design a Google Forms survey about undergraduate attitudes to AGI, with mixed question types and relevant demographics.
- Let it compare form-creation approaches and choose to generate a Google Apps Script.
- Review and run the generated script yourself to create the form.
Andrew Maynard wanted to see whether Manus could turn a broad research request into a sequence of tasks. He asked it to design a survey about undergraduate attitudes to artificial general intelligence, simulate 1,000 student responses, analyse the resulting dataset and write a report. The experiment was deliberately not a real human-subject study: Maynard says the respondents and data were synthetic.
Manus first made a task list, researched ways to create Google Forms, and chose to produce a Google Apps Script that Maynard could run himself. The agent drafted the form questions and script, then worked through the simulated dataset and report. Maynard says it autonomously adjusted its task plan as it proceeded and revised parts of the report after feedback. The article links the generated report and project files.
The result is useful as a test of planning and document production, not as evidence about what actual students think. Maynard’s headline warning that the study is “made up” is essential context: the report’s statistics and conclusions are fabricated from simulated respondents and must not be cited as survey findings. His human checkpoint was running the generated script himself. The case shows both the reach of a multi-step agent and the need to preserve provenance when an output looks like polished research.
Prompt
Original prompt
Source: Blog · Andrew Maynard
Frequently asked questions
Was this a survey of real students?
No. Maynard says Manus simulated 1,000 students and that the data were fabricated for the experiment.
What did Manus create?
It planned the work, researched Google Forms options, selected a Google Apps Script approach, drafted a survey creation script and produced a report from synthetic responses.
Can the report’s findings be used as research evidence?
No. The report’s statistics describe simulated data, not actual student attitudes. Maynard ran the generated script himself as a human checkpoint.
More details
Remaining steps
- Have Manus generate a synthetic respondent dataset, analyse it and assemble a report, keeping the synthetic provenance attached to every output.
About this account
Manus created a task plan, researched implementation options, generated a form script, simulated responses and produced a report.
Firsthand article by Andrew Maynard, published 2025-03-22 and reopened directly on 2026-10-18. The source explicitly states that the simulated students and data are fabricated; the case presents the work as an agent capability experiment, not real research. The prompt field preserves only a short excerpt of the actual prompt, under 25 words.
Credit: Andrew Maynard
Source published
Added to Agents In Use
Source checked
Tools and topics
Manus · Google Forms · Google Apps Script
Synthetic data · Survey design · Agent evaluation
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.