03 · What You Need to Know
What Counts as AI in the Research Process?
Start With What an AI System Does
There is no single universally accepted definition of artificial intelligence. A useful contemporary reference is the OECD definition of an AI system: a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions. AI systems can also differ in their levels of autonomy and their capacity to adapt after deployment.
For researchers, the important word here is outputs. An AI system may predict whether an image belongs to a particular category, recommend potentially relevant information, identify patterns in a dataset, produce computer code, transform existing text, or generate new content. Those functions are different, but they can all fall within the broad territory of AI.
This is why AI in research should not be treated as synonymous with ChatGPT, chatbots, or even generative AI. Generative AI is one particularly visible family of AI technologies, but machine learning used for prediction or classification has been part of scientific work long before conversational generative AI became widely accessible.
AI Can Be the Object of Research or a Tool Used to Conduct Research
The phrase “AI in research” can describe two fundamentally different relationships between AI and a study.
Research about AI
AI itself is an object of investigation. A study might evaluate an algorithm, examine bias in an AI system, develop a new model, or study how people interact with AI.
Research using AI
AI is part of the research workflow. A researcher might use it to classify data, assist with coding, summarize material, generate code, or support another research task.
The two can overlap, but they should not be confused. A historian using an AI system to transcribe archival material is using AI in research without necessarily conducting AI research. Conversely, a computer scientist evaluating the performance of a new model is studying AI and may also use other AI systems during the project.
AI Can Enter Research at More Than One Stage
AI use is not confined to data analysis. Depending on the discipline, research design, available systems, and applicable policies, AI may enter the workflow before data are collected, while evidence is being processed, during analysis, or while findings are communicated.
| Research activity |
Possible AI role |
What still requires researcher judgment |
| Exploring a topic |
Suggesting concepts, terminology, questions, or alternative formulations |
Determining whether the problem is worthwhile, grounded in scholarship, feasible, and appropriately framed |
| Working with literature |
Supporting discovery, classification, extraction, summarization, or comparison of documents |
Locating and reading authoritative sources, checking representations of the literature, and judging evidential relevance |
| Designing research |
Generating possibilities for instruments, procedures, variables, code, or analytical approaches |
Justifying methodological choices and ensuring that the design actually addresses the research question |
| Processing data |
Classifying, detecting patterns, transcribing, extracting features, predicting, or assisting with coding |
Assessing data quality, model suitability, assumptions, errors, bias, and the meaning of outputs |
| Analyzing evidence |
Supporting computation, programming, pattern detection, comparison, or preliminary interpretation |
Determining whether the analysis is valid and whether the conclusions are warranted by the evidence |
| Writing and communication |
Drafting, revising, translating, summarizing, formatting, or generating other forms of content |
Checking accuracy, preserving the intended meaning, verifying citations, and taking responsibility for the final work |
These are examples, not permissions. Whether a particular use is appropriate depends on the research context, the capabilities and limitations of the system, confidentiality and data-protection requirements, intellectual-property considerations, institutional rules, funder requirements, and publication policies. The European Commission's current living guidelines on generative AI in research, for example, frame its use around continuing researcher accountability, transparency, responsibility, and research integrity.
Using AI Does Not Necessarily Mean Generating Research Content
One reason discussions about AI become muddled is that very different activities are placed under the same label. A model that detects tumors in medical images, software that predicts molecular properties, and a chatbot that drafts a paragraph all involve AI, but the AI is doing something different in each case.
Even within generative AI, asking a system to suggest alternative keywords for a database search is not equivalent to asking it to write an interpretation of your results. The relevant distinction is not merely whether AI appeared somewhere in the workflow. You need to examine what cognitive or technical task was transferred to the system and what happened to its output afterward.
This is also why it helps to distinguish AI-assisted work from AI-generated material. The boundary can sometimes be fuzzy, but the distinction draws attention to the actual contribution made by the system rather than treating every use of AI as equivalent.
AI Assistance Exists on a Spectrum
Researchers often talk about AI use as a binary choice: either you used AI or you did not. In practice, the degree of reliance can vary considerably.
Limited assistance
AI performs a bounded task, such as suggesting search terms, correcting syntax in code, or proposing alternative wording.
Substantive assistance
AI produces material that contributes meaningfully to a research task, such as preliminary code, classifications, summaries, or analytical suggestions that the researcher subsequently evaluates.
High reliance
AI performs substantial parts of a task or produces outputs that materially influence methods, analysis, interpretation, or reporting.
These are useful conceptual categories rather than universal regulatory classifications. What matters is that the consequences of an AI error generally become more significant as researchers rely more heavily on the output for consequential parts of the study.
That makes the more useful question, “What did the AI do?” rather than simply, “Was AI used?” A researcher who uses AI to rephrase one sentence and a researcher who uses an AI system to classify the study's primary data have both “used AI,” but those uses should not be evaluated as though they were methodologically identical.
AI Is Not a Researcher
Conversational AI can make this distinction easy to forget. It can respond in complete sentences, explain apparent reasoning, propose hypotheses, criticize an argument, and imitate scholarly prose. That interface can make interaction feel less like operating software and more like consulting another person.
But linguistic fluency should not be mistaken for scientific authority. A system can produce an answer that sounds methodologically sophisticated while containing unsupported assumptions, factual errors, nonexistent references, inappropriate analytical choices, or conclusions that exceed the evidence. Understanding how large language models produce and handle information is therefore particularly important when their outputs enter scholarly work.
Watch Out
An AI output can be plausible without being correct. Treat fluency, detail, confidence, and polished academic language as properties of the output, not as evidence that the underlying claim has been verified.
AI Use Does Not Automatically Make a Study AI Research
If you use an AI assistant to edit a paragraph in a survey study, the study does not suddenly become an “AI study.” Similarly, using AI-assisted transcription does not necessarily change the substantive research design into an artificial-intelligence methodology.
AI becomes methodologically central when the system itself performs a consequential analytical or inferential function that is part of how the evidence is produced or interpreted. In those cases, details about the model, data, validation, performance, parameters, reproducibility, or other technical characteristics may become part of what readers need in order to evaluate the study.
The exact reporting requirements depend on the field, research design, venue, and role of the system. There is therefore no single disclosure sentence that adequately describes every form of AI use.
Traditional Research Software and AI Systems Should Not Be Treated as Interchangeable
Researchers have always relied on software. Statistical packages calculate estimates, reference managers organize citations, spreadsheets transform data, and qualitative analysis platforms help researchers manage evidence. The arrival of AI does not suddenly make computational assistance a new feature of research.
The difference is that some AI systems infer outputs rather than merely executing a transparent, predetermined transformation specified by the researcher. Generative systems add another complication because they can produce new text, images, code, audio, and other content in response to instructions.
That is why the comparison between generative AI and traditional research software requires more than asking whether both are “tools.” The relevant issues include how outputs are produced, how predictable they are, how they can be validated, and what risks arise when researchers rely on them.
“AI in Research” Says Nothing by Itself About Whether the Use Was Appropriate
Calling a workflow “AI-assisted” does not tell you whether it was rigorous, ethical, permitted, transparent, or useful. Those are separate questions.
An AI system can support careful research when it is suited to the task and its outputs are appropriately evaluated. The same system can weaken a study when researchers use it for a task it cannot reliably perform, supply information that should remain confidential, accept fabricated or distorted information, conceal consequential use, or substitute generated material for scholarly judgment.
Conversely, the mere presence of AI does not establish that a study is less rigorous. The methodological question is whether the research process remains defensible. This is why debates over whether generative AI affects research rigor need to examine actual use rather than the label “AI” alone.
06 · What This Means for You
Describe AI by Its Role, Not Merely by Its Presence
If you use AI in research, begin by abandoning the question “Did I use AI?” as your only diagnostic. It is too coarse to tell you what matters.
Instead, map the AI system to the specific research task. Ask what information entered the system, what the system returned, what decisions depended on that output, how you verified it, and what would happen if the output were wrong.
A simple decision framework
If AI only supports a low-consequence peripheral task
Check its output and determine whether any applicable policy requires disclosure or imposes restrictions.
If AI processes research data or contributes to analysis
Evaluate methodological suitability, validation, data governance, reproducibility, bias, and reporting requirements as part of the research method.
If AI generates claims, interpretations, references, or other scholarly content
Verify the underlying evidence independently. Do not treat generated content as an authoritative source merely because it appears plausible.
If AI would perform a consequential task you cannot independently evaluate
Do not assume that prompting the system transfers expertise to you. Reconsider the task, obtain appropriate expertise, or use a method whose validity you can defend.
As the role of AI becomes more consequential, the standard of scrutiny should generally rise with it. This is particularly important when an AI output can influence the evidence, analysis, interpretation, or conclusions rather than merely the efficiency of a peripheral task.
And regardless of how sophisticated the system becomes, responsibility does not simply migrate to the software. If an AI-generated or AI-influenced error enters your research, the question of who remains responsible for that error cannot be answered by saying that “the AI did it.”
