Humanly research and responsible study design
Humanly Cloud can provide a configured environment for studying writing and in-platform AI assistance. Researchers can inspect task submissions, activity records, source composition and available replay to investigate what happened inside that environment. A research protocol still needs its own participant permissions, sampling design and interpretation rules.
What does the published paper establish?
Zhu and colleagues' Humanly paper, arXiv:2607.21758v1, submitted on July 23, 2026, describes a configurable writing system and evaluates a research version through user and red-teaming studies. The paper is a preprint, not a service-level guarantee for every Cloud deployment.
The paper reports 30 valid responses for each of three participant roles: Owner, Writer and Verifier. Mean usability and perception ratings were above the neutral point on the study's five-point scale. These are participant ratings under the study conditions; they do not demonstrate improved writing quality, learning outcomes or the fairness of a disciplinary process.
In the reported red-teaming experiment, the evaluated set contained 17 human-operated and 20 agent-operated submissions. The typing detector achieved 85% detection of agent-operated submissions, 0% observed human false positives and 0.994 AUROC. Those results concern that small sample, operating threshold and browser-agent setup. An observed zero false-positive rate is not a guarantee of zero future errors. The experiment does not test every route for external AI assistance. See the full methods and limitations when citing these findings.
Which product records can inform a study?
The Cloud writing workflow connects a task's settings to its submissions and certificates. Final-text source composition describes the recorded origins of remaining text. Process input volume captures a different quantity: recorded input over the writing process. Activity and replay add temporal context; enabled detectors provide additional signals whose availability must be reported.
Treat these as different observations. A typed-character count is not a measure of original thought. Editing time is not a measurement of every minute a participant spent thinking about the task. An AI message is evidence of an in-platform interaction, not evidence that its advice was accepted or that no other assistance occurred. The evidence guide defines the practical boundaries.
A minimal protocol
- State the research question and decide which recorded behavior can actually answer it.
- Create the assigned task, instructions, resources and AI permissions. Keep a record of the configuration used by each study condition.
- Explain recording and access to participants, and arrange consent and any required institutional review before collecting data.
- Inspect completed submissions and their available evidence. Distinguish absent, disabled and inconclusive detector results from negative findings.
- Report the product version or study date, task settings, inclusion decisions, measures and limitations with the analysis.
For example, a study of permitted AI polishing could compare revision patterns under a predefined instruction. It should not silently substitute “percentage typed” for writing quality or participant independence. Specify the intended analysis and verify that the workflow exposes the required records before recruitment; these pages do not promise a general research-data export API.
Data access and research claims
Participant documents and certificates are not public research datasets merely because the product has public reference pages. Sharing a certificate and giving research consent are different actions. Review the privacy policy and the study's own data-handling arrangements.
The Cloud research page links the paper and product discussion. Use education or peer review for the corresponding task workflows, and attribute study findings to the paper rather than presenting them as guaranteed Cloud outcomes.
Reviewed: 2026-09-24.
Applies to: Humanly Cloud release baseline 75eba744, checked against source revision d5818c71 on 2026-09-24. Product descriptions cover capabilities shared by that release and source, not a new deployment. Available controls and evidence depend on the writing setup. Research findings refer to arXiv:2607.21758v1.
Canonical: https://cloud.writehumanly.net/reference/research