03 · What You Need to Know
What Researchers Must Be Able to Explain About AI-Assisted Research
Explain Why AI Was Used in the First Place
The first explanation concerns the scientific or practical purpose of introducing AI into the workflow.
A researcher should be able to identify the task, the expected contribution, and why the selected system was considered suitable.
For example, a generative AI system might be used to propose preliminary qualitative codes, assist with developing statistical code, or improve the clarity of manuscript language. These activities have different methodological implications.
The explanation should establish whether AI provided a capability needed for the research, reduced routine workload, or supported exploratory thinking.
It need not demonstrate that AI was the only possible method. However, where AI materially affected the evidence, the researcher should be able to justify its suitability compared with realistic alternatives.
Identify the Specific AI System and Its Relevant Configuration
Researchers should be able to identify the system used to the extent necessary for evaluating its contribution.
Relevant information may include the provider, product, model identifier, available version or snapshot, dates of use, and consequential configuration settings.
For some applications, prompts, retrieval settings, generation parameters, or connected tools may influence the output.
A statement such as "ChatGPT was used" may be adequate for some limited disclosures but insufficient to reconstruct a consequential analytical procedure.
Where a provider does not expose the exact model version, researchers should record the available information and acknowledge the uncertainty.
They should not invent model identifiers or claim to have controlled settings that were inaccessible.
Explain What Information the AI Received
The inputs supplied to an AI system can determine the content and reliability of its outputs.
Researchers should be able to describe whether the system received public literature, selected data extracts, de-identified records, synthetic examples, statistical summaries, or other research material.
For analytical tasks, the explanation may also need to cover preprocessing, variable selection, data formatting, and relevant exclusions.
For example, a qualitative classifier receiving short excerpts may produce different interpretations from one receiving complete interview responses with contextual information.
Researchers should therefore explain the information available to the system and any important limitations introduced by that selection.
Descriptions of inputs must respect confidentiality. Methodological transparency does not automatically authorize publication of participant data or restricted prompts.
Explain What the AI Was Asked to Do
Knowing which model was used does not establish how it was instructed.
A system asked to generate inductive codes may perform a different task from one instructed to apply a predefined deductive coding framework.
Similarly, an AI system asked to identify possible methodological alternatives is performing a different function from one instructed to choose a final statistical model.
Researchers should be able to describe the instructions that materially shaped the output.
Where prompts were revised, it may be important to explain why the revisions occurred and whether they changed the analytical procedure.
For consequential workflows, preserving relevant prompts and instructions can support later inspection.
Explain Which Outputs Were Accepted, Modified, or Rejected
AI-assisted research often involves several stages of human selection and revision.
A researcher may accept some generated codes, reject others, combine categories, revise analytical scripts, or discard unsupported interpretations.
These decisions can be scientifically important because they determine how AI output enters the final research.
For example, if an AI system proposes ten qualitative categories but researchers retain only four, the final analytical framework cannot be explained solely by identifying the model.
The researchers should describe the criteria used to evaluate the suggestions and the reasoning behind consequential changes.
Where these decisions influence findings, a documented audit trail may be necessary.
Explain How AI-Generated Results Were Verified
Verification is central to the defensibility of consequential AI contributions.
Researchers should be able to identify which outputs were checked, what procedures were used, and what evidence supported acceptance.
For literature synthesis, this may involve comparing generated claims with original publications.
For statistical code, verification may include code inspection, independent calculations, tests using known inputs, and methodological review.
For qualitative analysis, researchers may examine classifications against original source material and document interpretive disagreements.
Merely stating that the output was reviewed is less informative than explaining what the review established.
Verification should be appropriate to the research task rather than represented as a universal procedure.
Explain the Difference Between Technical Execution and Scientific Validity
A computational procedure can execute successfully without being appropriate for the research question.
For example, AI-generated regression code may produce coefficients and significance tests while ignoring dependence among observations or applying an unsuitable model specification.
Researchers should be able to explain both how the procedure was implemented and why it was methodologically appropriate.
Technical explanation
Describes what the software or AI-assisted procedure did, including relevant inputs, operations, and outputs.
Scientific explanation
Justifies why the procedure was suitable, what assumptions it required, and what conclusions its results can legitimately support.
Both forms of explanation may be necessary for consequential research activities.
Technical correctness cannot compensate for an unsuitable design, and methodological appropriateness does not excuse an incorrect implementation.
Explain the Methodological Assumptions Behind AI-Assisted Decisions
When AI influences method selection or analysis, researchers should understand and explain the assumptions that matter.
These may concern sampling, measurement, independence, missing data, model identification, confounding, or the relationship between evidence and inference.
For example, an AI system may recommend structural equation modeling for survey data. The researcher should be able to explain the model's purpose, relevant identification and measurement considerations, and limitations of interpretation.
Researchers do not necessarily need to reproduce every mathematical derivation. They should, however, understand enough to identify conditions under which the chosen procedure may produce misleading conclusions.
Where specialist expertise is required, qualified consultation can support the explanation, but the research team remains accountable for the final work.
Explain How AI Influenced Scientific Judgment
AI may contribute more than text or code. It can influence which questions researchers ask, which methods they consider, and how they interpret evidence.
These influences may be productive when they encourage researchers to examine overlooked alternatives.
However, researchers should be able to explain why a consequential recommendation was accepted rather than relying on the model's apparent authority.
Suppose AI suggests excluding observations because they appear anomalous. The researcher must establish whether the exclusions follow defensible criteria and whether the decision changes the findings.
The scientific explanation should identify the evidence and methodological reasoning supporting the final choice.
Where AI materially changed a research decision, documenting the change can help establish accountability.
Explain What the AI System Could Not Reliably Establish
A complete explanation includes limitations.
Generative AI may produce unsupported citations, inaccurate calculations, contextually inappropriate classifications, or interpretations that exceed the evidence.
A system may also change between uses, making exact reproduction difficult.
Researchers should explain limitations relevant to the task rather than reciting a generic statement that AI can make mistakes.
For example, if a model was evaluated only on English-language survey responses, the researcher should not assume equivalent performance for responses in other languages without appropriate evidence.
Limitations should be connected to the validity and interpretation of the reported findings.
Explain the Difference Between AI Model Explainability and Researcher Accountability
AI explainability concerns the extent to which a system's operations or outputs can be understood and interpreted.
Researcher accountability concerns the responsibility to justify the methods, evidence, and decisions incorporated into scientific work.
These concepts overlap but are not identical.
A researcher may use a proprietary model whose internal weights are unavailable while still documenting the workflow, evaluating performance, and explaining why the output was accepted.
Conversely, access to detailed model architecture does not establish that the researcher's analytical choices were appropriate.
The NIST AI Risk Management Framework treats explainability, interpretability, transparency, validity, and accountability as related characteristics requiring contextual assessment.
Researchers should not confuse a technically elaborate explanation of a model with a defensible explanation of the research conducted using it.
What Level of Explanation Is Appropriate for Different AI Tasks?
| AI activity |
What the researcher should explain |
Level of methodological concern |
| Language editing |
Nature of assistance, author review, and whether scientific meaning changed |
Usually limited unless substantive claims were altered |
| Literature synthesis |
Source selection, AI's summarization role, and verification against original publications |
Potentially consequential for scholarly claims |
| Statistical code generation |
Code function, model selection, validation, and analytical interpretation |
High when primary results depend on the code |
| Qualitative coding |
Coding approach, prompts, human decisions, source-data review, and limitations |
Depends on the analytical role and methodology |
| AI-based measurement |
Model identification, construct definition, validation, output stability, and uncertainty |
High when outputs constitute primary measurements |
| Participant-facing decisions |
Authorization, validation, safeguards, oversight, and decision responsibility |
Potentially high because participant welfare may be affected |
The table illustrates why a uniform disclosure statement cannot adequately explain every AI-assisted research activity.
The required detail should follow the scientific significance of the contribution and the expectations of the relevant discipline, journal, or institution.
Can Researchers Explain Work They Did Not Personally Execute?
Yes. Scientific collaboration routinely involves specialized contributions.
A statistician may develop analytical code, a laboratory specialist may operate an instrument, or a qualified collaborator may review a complex computational procedure.
The principal researcher need not personally perform every technical operation, but should understand its scientific purpose and implications sufficiently to evaluate the resulting work.
Where responsibility is distributed across a team, the relevant expertise and verification responsibilities should be clearly assigned.
AI does not change this principle. It provides another source of assistance, but it does not become an accountable member of the research team.
What if the Researcher Cannot Explain a Consequential AI Contribution?
Inability to explain an important analytical decision is a reason to pause acceptance of the output.
The researcher may need to examine the code, consult methodological references, obtain qualified assistance, or repeat the procedure using a better-understood method.
If the model's internal information is unavailable, the researcher should determine whether independent validation and workflow documentation are sufficient for the intended claim.
Where essential details cannot be established, the limitation should not be concealed through a generic AI disclosure.
A defensible response may involve changing the workflow rather than attempting to justify an analysis that cannot be adequately evaluated.
