Responsible AI decision desk
A policy-grounded guide and four-page record for deciding what AI may receive, which judgment remains human, how output is verified, and what must be disclosed.
Manusights · Evidence checked September 15, 2026 · Four-page working record · CC BY 4.0
Prepared by the Manusights editorial team. Publisher and institutional instructions take precedence.
Use this in the work
Download the Research AI use record
Print it, complete it with collaborators, or adapt the original framework with attribution. The PDF is generated from the same reviewed evidence record as this page.
The first decision is not the vendor
A useful product can still be the wrong environment for a particular input or decision. Record what would leave your approved boundary, what scientific judgment could be affected, and who is accountable before comparing features.
Public or non-sensitive input, reversible assistance, no delegated scientific judgment, and a defined human check.
Examples: Formatting a public reference list, brainstorming search terms, or explaining public documentation.
Analysis, code, synthesis, hypotheses, translation, or writing that can materially affect the research record.
Examples: Require approved inputs, reproducible records, domain validation, claim checking, and any required disclosure.
Confidential manuscripts or reviews, grant-review material, participant data, protected health information, credentials, secrets, or delegated final judgment.
Examples: Proceed only if a controlling policy and approved protected environment specifically permit the use.
Original Manusights permission map
What information would leave your approved environment?
Evidence: Data class, confidentiality owner, consent or agreement, retention terms, and approved system.
What decision is the system being asked to make, and who remains accountable?
Evidence: Human owner, prohibited delegation, decision impact, review step, and escalation path.
How will every material claim, citation, transformation, calculation, and image be checked?
Evidence: Verification method, primary evidence, code or prompt record where appropriate, and reviewer.
Could another qualified person understand and repeat the material workflow?
Evidence: Tool and version, settings, input description, change log, retained code, and output provenance.
What do the journal, funder, institution, collaborators, or participants require you to report?
Evidence: Controlling policy, disclosure text, placement, observation date, and approver.
Hard boundaries from current guidance
AI cannot be an author. Humans remain accountable for accuracy, integrity, originality, attribution, and any required disclosure. Confidential submissions require protection. Source A.
Generative AI is prohibited in NIH peer-review responsibilities. Treat review material as confidential and follow the current NIH process. Source C.
Wellcome requires applicants to declare substantive generative AI use beyond language assistance and keeps applicants responsible for the application. Source E.
Minimum validation record
Disclosure skeleton
"We used [tool and version] for [bounded task]. [Named human role] verified [material outputs] against [evidence]. No [protected input category] was provided. The authors remain responsible for the work."
Fictional structure only. Use the exact journal, funder, institution, and collaborator requirements.
Risk management
The NIST Generative AI Profile provides a broader risk-management frame. This worksheet applies that posture to a single research use: establish governance, map the context, measure material risk and performance, then manage the approved use and its record. Source D.
Evidence record
We separated tool choice from permission. The original map starts with information boundaries and accountable judgment, then requires validation, reproducibility, and disclosure appropriate to the controlling policy.
International Committee of Medical Journal Editors. Locator: Authorship, accountability, disclosure, and confidentiality.
Supports: AI cannot be an author, humans remain accountable, use should be disclosed, and confidential material requires protection.
Open official sourceNIH Office of Extramural Research. Locator: Research integrity and fabricated material.
Supports: Research-integrity responsibilities when using AI and the risk posed by fabricated references or other research records.
Open official sourceNIH Office of Extramural Research. Locator: Confidentiality and generative AI in peer review.
Supports: The prohibition on using generative AI for NIH peer-review responsibilities.
Open official sourceNational Institute of Standards and Technology. Locator: Govern, map, measure, and manage functions.
Supports: A risk-management structure for governing, mapping, measuring, and managing generative AI risk.
Open official sourceWellcome. Locator: Declaration and applicant responsibility.
Supports: Declaration of substantive generative AI use in applications and continued applicant responsibility for the submission.
Open official sourceCite this resource
Free to use and cite, with attribution. See the reuse policy for what you may republish.
Use AI with verifiable evidence
Helpful research tools still need a source-level verification step before their output reaches a manuscript.
