03 · What You Need to Know
How Responsible AI Use Works Throughout the Research Process
Begin by Establishing Whether AI Is Appropriate for the Task
Responsible use begins before the researcher enters a prompt.
The first question is whether AI offers a justified benefit for the intended activity. A generative model may be useful for exploring alternative search terminology, but unnecessary for calculating a median that established statistical software can determine directly.
Researchers should define the task, expected output, available alternatives, and consequences of error. The proposed tool should be evaluated against these requirements rather than selected because it is accessible or technologically sophisticated.
The NIST AI Risk Management Framework emphasizes the importance of evaluating AI systems within their intended contexts of use. A system suitable for one activity may be inappropriate for another.
Where an established procedure can perform the task more reliably and transparently, choosing that procedure may be the responsible decision.
Establish What the AI System Is Authorized to Receive and Process
Research information may be governed by informed consent, ethics approval, institutional policy, contractual restrictions, intellectual property obligations, or data protection law.
Before supplying information to an AI system, researchers should establish whether the proposed processing is permitted.
For example, publicly available article abstracts and confidential participant interviews present different information-handling requirements. Removing participant names does not necessarily eliminate the possibility of reidentification.
Researchers should examine the provider's applicable data-use terms, retention arrangements, access controls, and any institutional agreements. A provider's commitment not to use customer content for model training does not independently establish that every confidentiality obligation has been satisfied.
If authorization or adequate safeguards cannot be established, restricted information should not be submitted to the proposed service.
Understand the AI Contribution Before Accepting Its Output
AI can assist with tasks ranging from editorial suggestions to substantive methodological procedures. The researcher needs to understand what the system was asked to do and how its response might affect the work.
A generated paragraph may introduce unsupported claims. A proposed statistical model may rely on assumptions that the study does not satisfy. An automated qualitative classification may misrepresent the meaning of participant statements.
Researchers should be able to identify these possibilities and determine what verification is necessary.
Using AI responsibly does not require understanding every internal computational mechanism of a proprietary model. It does require sufficient knowledge of the research task to recognize consequential limitations and evaluate whether the output is fit for purpose.
Match Verification to the Scientific Consequences of Error
Not every AI interaction requires the same level of scrutiny.
A suggestion for a manuscript heading can usually be evaluated through ordinary editorial judgment. A generated citation supporting a central claim requires examination of the original source. AI-generated analytical code may require independent testing and methodological review.
| Research activity |
Principal concern |
Appropriate verification |
| Generating search terms |
Missing or irrelevant terminology |
Review against the research question and relevant disciplinary vocabulary |
| Summarizing literature |
Fabricated references or distorted findings |
Compare consequential claims with original publications |
| Developing statistical code |
Incorrect implementation or unsuitable analysis |
Inspect code, test calculations, and assess methodological assumptions |
| Classifying qualitative responses |
Loss of context or inaccurate coding |
Evaluate against source material using a defensible validation procedure |
| Drafting manuscript sections |
Unsupported claims or altered scientific meaning |
Review content, evidence, terminology, and applicable disclosure requirements |
| Supporting participant-facing decisions |
Potential participant harm |
Task-specific validation, safeguards, qualified oversight, and required authorization |
Verification should establish that an output is acceptable for its intended use. Asking the same model to confirm its previous answer is not equivalent to independent validation.
Keep Scientific Judgment With Accountable Researchers
Generative AI may suggest research questions, analytical methods, theoretical interpretations, and possible explanations for findings.
These suggestions can be valuable when they expose alternatives that researchers had not considered. However, they should not determine consequential decisions merely because they are presented persuasively.
Researchers must evaluate the evidence supporting a recommendation, consider alternatives, and explain why the final decision is appropriate.
For example, an AI system might recommend removing outliers from a dataset. The researcher must determine whether exclusion is justified by the research design, data characteristics, prespecified criteria, and relevant methodological literature.
Human responsibility requires substantive evaluation rather than a final approval that merely repeats the model's recommendation.
Document AI Contributions That Matter to the Research
Documentation allows researchers and reviewers to understand how AI influenced the study.
Depending on the task, relevant records may include the tool and model identifier, dates of use, prompts, input preparation, settings, outputs, verification procedures, and decisions made by researchers.
The level of detail should reflect the significance of the AI contribution.
A researcher using AI for preliminary brainstorming may need relatively limited records. A team using AI-generated classifications as a primary analytical variable may need a comprehensive audit trail.
Documentation should also respect confidentiality. Preserving every prompt in an unsecured repository would not be responsible if those prompts contain restricted participant information.
Disclose AI Use According to Applicable Publication Requirements
Publishers and journals differ in their requirements for reporting AI-assisted activities.
The International Committee of Medical Journal Editors recommends describing AI-assisted technologies used in submitted work and identifying relevant methodological uses in the appropriate manuscript sections.
Other publishers may distinguish basic language corrections from substantive AI-assisted drafting or research procedures.
Researchers should review the instructions applicable to their intended journal and report the contribution accurately.
A useful disclosure identifies what AI did rather than merely stating that an AI tool was used.
Disclosure is necessary when required, but it does not validate inaccurate claims, justify unauthorized processing, or transfer authorship responsibility to the AI system.
Preserve Reproducibility Where AI Influences the Evidence
AI-assisted procedures may be difficult to reproduce when model versions change, generation is variable, or relevant inputs and outputs are not preserved.
Researchers should establish what level of reproducibility their study requires.
For consequential classifications or analytical results, preserving the original outputs, relevant instructions, model information, and verification records can support later evaluation.
Where exact output repetition is essential, a fixed and appropriately validated procedure may be preferable.
Researchers should distinguish documenting what occurred from guaranteeing that an AI system will generate identical responses in the future.
Evaluate AI-Related Risks Before Participants Are Affected
AI may influence recruitment, eligibility screening, intervention allocation, participant communication, or interpretation of sensitive information.
These uses can introduce risks beyond ordinary manuscript preparation.
The research team should identify foreseeable harms, evaluate the reliability of the proposed system, and determine whether meaningful safeguards are available.
For consequential participant-facing applications, ethics review or an amendment to an existing approval may be required.
Human oversight is meaningful only when qualified reviewers can examine relevant evidence and intervene before an inappropriate AI recommendation causes harm.
Recognize That AI Reliability Is Context-Dependent
A model may perform adequately on a benchmark while producing unreliable outputs for a particular research population or task.
Language, disciplinary terminology, data characteristics, and the conditions of deployment can influence performance.
Researchers should avoid assuming that general claims about model capability establish suitability for their study.
Where AI produces consequential classifications or measurements, task-specific evaluation may be necessary.
Performance should also be reconsidered when model versions, prompts, data characteristics, or operating conditions change materially.
Distinguish Responsible AI Use From Mere Disclosure
Disclosure
Communicating the AI contribution according to applicable research and publication requirements.
Responsible use
Ensuring that the contribution is appropriate, authorized, sufficiently verified, documented, and subject to accountable scientific judgment.
A researcher can disclose an AI-generated analysis while still failing to verify whether the code is correct.
Conversely, a carefully verified AI-assisted procedure may still require disclosure under journal policy.
Both scientific reliability and reporting transparency matter. Neither substitutes for the other.
What Does a Responsible AI-Assisted Workflow Look Like?
Responsible use is best understood as a sequence of decisions embedded in ordinary research practice.
Define the task.
Specify what AI is expected to contribute and why the task is suitable for the proposed system.
Check authorization.
Determine whether the information, processing arrangement, and intended use comply with applicable requirements.
Establish verification.
Identify how consequential outputs will be checked before they influence evidence or decisions.
Record the contribution.
Preserve relevant inputs, outputs, model information, and human decisions according to the task's significance.
Evaluate the result.
Accept, revise, or reject AI output according to independent evidence and methodological judgment.
Report accurately.
Describe consequential AI involvement and disclose its use according to applicable publication requirements.
This sequence is a practical framework rather than a universal institutional standard. Its implementation should be adapted to the research design and risks.
