03 · What You Need to Know
The Essential Components of a Responsible AI-Assisted Research Workflow
Begin With the Research Design Rather Than the AI Tool
The research question should determine the methodology, and the methodology should determine where AI assistance may be useful.
Researchers should avoid designing a study around the capabilities of a particular AI service without first establishing the scientific objective.
For example, a qualitative study intended to understand participants' experiences should not automatically become a predictive classification study simply because an AI platform offers convenient categorization.
Similarly, a quantitative study should not adopt an advanced analytical model solely because AI can generate the necessary code.
Begin by identifying the research question, intended evidence, methodological assumptions, and expected analytical outputs. AI should then be evaluated as a possible component of that design.
Create an Inventory of Proposed AI Tasks
A responsible workflow begins with identifying where AI will be used.
The inventory should describe each task, the proposed tool, the information required, the expected output, and how that output may influence the research.
This helps distinguish low-consequence assistance from activities that materially affect scientific findings or participants.
| Research stage |
Possible AI contribution |
Principal control |
| Research planning |
Generate possible questions and identify methodological alternatives |
Independent scholarly evaluation |
| Literature review |
Suggest search terms or summarize retrieved publications |
Source verification and documented selection criteria |
| Instrument development |
Propose candidate items or interview questions |
Content evaluation and appropriate validity evidence |
| Data preparation |
Generate cleaning or transformation code |
Verified rules and preserved transformation records |
| Data analysis |
Assist with statistical code or qualitative classification |
Task-specific methodological validation |
| Interpretation |
Suggest alternative explanations or identify possible limitations |
Independent evaluation against original evidence |
| Manuscript preparation |
Assist with structure, language, or preliminary drafting |
Author verification and applicable disclosure |
The inventory should reflect the actual planned activities. A task described as language editing may require stronger controls if the system is permitted to change substantive scientific claims.
Evaluate Whether AI Is Appropriate for Each Task
Not every research activity benefits from generative AI.
Researchers should compare the proposed AI procedure with realistic alternatives, including established statistical software, reference-management applications, rule-based processing, and qualified human analysis.
For exact calculations, a deterministic and validated procedure may provide stronger reproducibility and transparency.
For exploratory language-based tasks, AI may offer useful flexibility.
The decision should consider methodological suitability, expected benefits, verification effort, information protection, and potential consequences of error.
Choosing not to use AI for a particular task can be an appropriate outcome of responsible workflow planning.
Classify AI Tasks According to Their Risks
Different AI tasks create different scientific and ethical risks.
An incorrect heading suggestion may have little consequence. An incorrect data transformation may alter a study's findings. An inaccurate participant safety recommendation could create more serious harm.
Risk assessment should consider the likelihood of error, severity of potential consequences, detectability, affected individuals, and available safeguards.
The NIST AI Risk Management Framework provides principles for identifying, measuring, and managing AI-related risks in context.
Researchers may use a task-specific classification scheme to organize decisions, but should distinguish an internally developed framework from a formally validated risk instrument.
The purpose of classification is to determine appropriate controls, not merely to assign a descriptive label.
Establish Authorization Before Sharing Research Information
AI use may involve information governed by consent agreements, ethics approval, institutional policies, contractual restrictions, or data protection requirements.
Researchers should establish whether the proposed system is permitted to process the information required for the task.
Confidential interview transcripts, identifiable student records, unpublished manuscripts, and proprietary datasets may require different protections.
Removing names does not necessarily make information anonymous, particularly when contextual details permit reidentification.
Researchers should examine the relevant provider terms, data retention arrangements, access controls, and institutional agreements.
If the processing cannot be authorized or adequately protected, the proposed AI workflow should be modified or rejected.
Determine Whether Ethics or Security Review Is Required
Some AI applications may change the procedures or risks described in an approved research protocol.
For example, introducing AI-assisted participant screening or sending interview transcripts to an external service may require additional institutional assessment.
The applicable requirements depend on the jurisdiction, institution, research design, and nature of the proposed activity.
Researchers should consult the responsible ethics or information-security authority when the proposed AI use changes participant risks or data-processing arrangements.
Approval of a general-purpose AI tool does not automatically authorize every research application.
Where required, authorization should be obtained before the relevant processing begins.
Define What AI Is Allowed and Not Allowed to Do
A workflow should establish the boundaries of AI assistance.
For example, AI may be permitted to suggest preliminary qualitative codes but prohibited from determining final themes without researcher review.
It may generate statistical code but not independently select the final analytical model.
It may assist with organizing literature while being prohibited from introducing unverified citations into the manuscript.
These boundaries help prevent a supporting task from expanding into an unexamined decision-making role.
Researchers should specify which operations are automated, which require human approval, and which remain outside the authorized use of AI.
Assign Responsibility for Each Consequential AI Contribution
Research teams should establish who is responsible for evaluating AI-generated outputs.
A statistician may review analytical code, a qualitative researcher may examine generated codes, and a principal investigator may oversee compliance with the approved protocol.
These responsibilities can be distributed according to expertise.
However, assigning a reviewer does not establish that review was effective. The person must have sufficient competence, relevant evidence, and authority to correct or reject outputs.
Researchers should avoid relying on nominal human approval as a substitute for meaningful verification.
Specify Verification Before Generating Consequential Outputs
Verification should be planned before AI-generated results enter the analysis.
For factual claims, researchers may require checking original publications.
For numerical calculations, independent computation or reviewed code may be appropriate.
For AI-generated classifications, the team may need a reference dataset, relevant performance measures, and procedures for resolving uncertain cases.
For qualitative interpretation, verification should be consistent with the chosen methodology and grounded in original research material.
Planning these procedures early reduces the risk of accepting outputs that cannot subsequently be substantiated.
Preserve a Traceable Record of AI-Assisted Work
Documentation allows researchers to explain what occurred during consequential AI-assisted procedures.
Relevant records may include the provider, model identifier, version information, dates of use, prompts, inputs, outputs, configuration settings, validation results, and human decisions.
Not every interaction requires identical documentation.
A preliminary brainstorming session may warrant limited records, while AI-generated variables used in the primary analysis may require a detailed audit trail.
Records should be stored securely and according to the project's retention requirements.
Preserving sensitive prompts in an unrestricted repository would not constitute responsible transparency.
Plan for AI Model Updates and Output Variability
Generative AI systems may change during a long research project.
Model updates, altered system instructions, or differences in generation settings can affect outputs.
This is particularly consequential when AI-generated classifications or scores are compared across different periods.
Researchers should determine whether a fixed model version is available and whether changes require renewed validation.
Where exact reproduction is essential, an unstable model may be unsuitable.
The workflow should include a procedure for identifying material changes and evaluating their effects on the research.
Establish a Procedure for Handling AI Errors
Even a carefully designed workflow may encounter incorrect AI outputs.
Researchers should determine how errors will be detected, documented, corrected, and escalated.
For example, an incorrect generated citation may require replacing the reference and checking whether the supported claim remains valid.
An analytical coding error may require rerunning affected analyses and reviewing dependent conclusions.
An AI error affecting participant welfare may require immediate protective action and reporting under applicable institutional procedures.
The response should reflect the consequences of the error rather than treating every mistake as equally serious.
Separate AI-Generated Suggestions From Accepted Research Evidence
Exploratory suggestions should not automatically become empirical findings.
An AI-generated hypothesis may be useful for planning but does not establish that the proposed relationship exists.
An AI-generated thematic label may suggest a possible interpretation but does not demonstrate that the theme is supported by participant material.
A generated statistical explanation may identify a plausible mechanism without establishing causation.
Researchers should preserve a clear distinction between provisional AI assistance and evidence accepted through appropriate methodological procedures.
This distinction should remain visible during analysis and manuscript preparation.
Integrate AI Disclosure Into Research Reporting
Publication policies differ in how they regulate and require disclosure of AI-assisted activities.
Researchers should review the applicable journal requirements and accurately describe material AI contributions.
Where AI affects the methodology, the methods section should explain the relevant procedure and validation.
Additional declarations may be required for AI-assisted drafting or other uses.
The International Committee of Medical Journal Editors emphasizes human responsibility for AI-assisted scholarly work and the need to describe relevant use.
Disclosure supports transparency but does not substitute for scientific validation or ethical authorization.
Audit the Workflow Before Manuscript Submission
A final audit should examine whether the AI-assisted procedures were performed as planned and whether the reported findings are supported.
The team should verify consequential references, code, calculations, classifications, interpretations, and disclosures.
It should also examine whether relevant records are preserved and whether model changes or unexpected errors affected the analysis.
Where deviations occurred, the researchers should determine whether they require additional validation, protocol amendments, or changes to the manuscript.
The audit should identify unresolved limitations rather than create an appearance of completeness through documentation alone.
What Does a Complete AI-Assisted Workflow Look Like?
A practical workflow can be organized into successive research decisions.
Define the research task.
Specify the scientific objective and the contribution expected from AI.
Evaluate suitability and risk.
Compare AI with alternatives and identify consequential failure modes.
Confirm authorization.
Establish appropriate ethics, privacy, security, and institutional permissions.
Assign human responsibility.
Identify who will evaluate outputs and make consequential decisions.
Implement and verify.
Use AI within defined boundaries and apply task-specific validation.
Preserve the record.
Document consequential inputs, outputs, configurations, and human decisions.
Audit and report.
Evaluate the complete workflow and accurately describe the AI contribution and limitations.
This sequence is a practical organizational approach, not a universally mandated research protocol. Its implementation should be proportionate to the study's scientific and ethical requirements.
