A task map: where AI helps, where it hurts
The useful mental model is not "AI writes my paper" but a division of labor. AI is strong at mechanical-but-verbal work: searching and triaging literature, summarizing papers you then read properly, first-pass qualitative coding, cleaning and analyzing data through code, reformatting citations, and language-level editing. It is weak — and dangerous — exactly where research earns its authority: deciding what is true, judging what matters, and generating references from memory.
A defensible rule: AI proposes, the researcher disposes. Every AI output that enters your study should pass through a human judgment you could defend to a reviewer — and for anything load-bearing (a quote, a number, a citation), verification against the primary source.
The failure mode that ends careers: fabricated citations
General chatbots generate plausible-looking references that do not exist — realistic authors, journals, page numbers, DOIs. Papers containing fabricated AI citations have been retracted, and checking for them is now routine at many journals. The fix is architectural, not behavioral: never accept a reference an AI produced from its memory. Citations must come from a real retrieval step — a database search, your reference library — where the paper demonstrably exists and you can open it.
The same discipline applies to quotes and numbers: an AI summary of a paper you have not opened is a hypothesis about the paper, not a fact about it. Verify before it enters your text.
What journals and institutions actually allow
The policy landscape has converged on a few stable rules. Nearly universal: AI cannot be an author (ICMJE, COPE, Nature, Science and most major publishers agree, because authorship implies accountability no tool can carry). Widely required: disclosure of substantive AI use — in the methods, acknowledgments, or a dedicated statement, per venue. Commonly prohibited: uploading confidential manuscripts to public chatbots during peer review, and presenting AI-generated images as data.
Universities vary more, especially for student work — some course policies prohibit what a journal would allow. Two habits keep you safe everywhere: check the specific policy of the venue or course before you rely on a tool, and when in doubt, disclose. Nobody has been sanctioned for over-disclosing.
- Keep the venue’s author guidelines and your institution’s AI policy in the project folder from day one.
- A disclosure sentence costs nothing: "AI tools were used for language editing and literature search; all analyses, interpretations and citations were verified by the authors."
- For sensitive data (participant interviews, unpublished results), check where the tool processes and stores data — GDPR and ethics approvals often require EU processing and no training on your content.
Patterns that work in practice
Literature: use AI search over real indexes to cast a wide net, then read the papers that matter yourself — AI triage plus human reading beats either alone. Qualitative analysis: let AI propose first-pass codes, then do the interpretive work — merging, naming themes, hunting disconfirming cases — yourself, and say so in the methods. Statistics: AI-generated analysis code is genuinely good, but you own the assumption checks and the interpretation. Writing: drafting from your outline and your sources works; asking a model to "write the discussion" from nothing produces confident emptiness that reviewers now recognize on sight.
The compounding trick is grounding: the more of your actual material — papers, transcripts, data, drafts — the AI can see, the less it invents. This is why purpose-built research tools constrain models to work from retrieved sources rather than memory.
Where Kahubi fits
Kahubi is built around exactly these guardrails: the agent cites only papers that exist in your library or its live index searches — marking [CITATION NEEDED] where sources cannot support a claim — statistics run as real computations on your CSVs, interview transcription is processed in the EU with no training on your content, and every AI edit is versioned so you can show precisely what the human accepted. The responsible-use checklist above is, in effect, the product’s architecture.
Last updated 2026-08-07.