Input–output–contribution matrix
When does an early idea become material?
Brainstorming is often described as casual, yet it frames what will be considered. A proposed topic may exclude alternative perspectives; a list of categories may become the analytical framework; a suggested hypothesis may quietly direct data interpretation. Literal copying is therefore a poor threshold. Ask whether the exchange helped determine a choice that remains visible in the finished work.
A one-off question that produced no usable idea need not be inflated into a major contribution unless local rules demand a complete record. Priority belongs to suggestions that were saved, developed, inserted into searches or used as a decision frame. Capture them while working: date, purpose, output category and decision usually fit in a short log. A retrospective account written at submission is more likely to confuse what the system proposed with what later emerged from reading.
Make the scholarly decision visible
A language model offers possibilities. Academic authorship lies in defining criteria, finding evidence, recognising counterpositions and defending the final choice. A plausible topic is not necessarily novel, feasible or ethical. Before retaining it, check the literature, available data, key definitions, timescale and assessment scope. In the record, avoid writing “AI developed my research question” if you actually rejected most of the proposal. State the useful element and how you transformed it.
- Origin: Which perspective or term entered through the exchange?
- Criterion: What made it relevant enough to test?
- External check: Which database, original source or academic discussion informed the choice?
- Transformation: How does the final version differ?
- Responsibility: Who can explain and defend the decision?
This separation also prevents AI suggestions from being cited as evidence. If a system claims a “research gap”, locate and assess the literature yourself. The absence of a quick search result is not proof of novelty.
Worked case: narrowing a topic
A student plans to study trust in news and requests ten narrower questions. She rejects “Does AI make news untrustworthy?” because its concepts are broad and the wording implies causation. Two other proposals introduce the terms “source cue”, “synthetic media disclosure” and “perceived credibility”. Database searches lead her to focus on labels attached to synthetic images in a defined setting.
Her log stores the prompt, displayed model label, date, complete suggestion list and brief appraisal. The submission statement does not reproduce every abandoned sentence. It explains the consequential role: search vocabulary and initial topic alternatives. The methods section then reports the actual literature search. The tool is neither concealed nor treated as a scholarly authority. A continuing AI use log makes this easier than reconstructing the trail.
Common disclosure failures
- Minimising: “Only ideation” conceals that the adopted research question began as a generated option.
- Over-reporting: Every predictive search completion is called generative research although it changed nothing.
- Mis-citing: The response is used to support a claim about the field.
- Archiving prompts alone: No record explains which idea survived or why.
Sensitive projects need a further boundary. Do not place personal or confidential details in a public appendix. A redacted log can describe the input category and decision without exposing protected material. The general AI declaration template provides wording for controlled evidence.
Final questions for a concise statement
Could another person identify which idea came from the system? Have you explained the academic criteria used to test and revise it? Do tool, date and visible version match the working log? Are statements about the research field supported by real publications rather than generated summaries? Is the disclosure located where the assessment instructions specify? Clear answers allow a brief statement to remain informative.
The AI transparency hub covers adjacent records for prompts, models and verification. For the wider research process, consult the academic writing overview. A text classifier cannot recover this decision history from the final prose.