Ethical Tech CoLab · Fall 2025 Cohort

AI-Powered Assistance in
Formulating Research Questions

A survey of how AI helps researchers move from a broad area of interest to a well-formed, answerable research question — and the guardrails that responsible use demands.

Yorke E. Rhodes III, Nathaniel Fossella, Alexa Shamie, Kirsten Co, Taylor Badt, Amanda Lindsey, Grace Driscoll, Mohagani Townsend, Pegi Bracaj, Vedant Jain

Five ways AI assists question formulation

The paper surveys where AI genuinely helps across a researcher's workflow — from scouting the literature to drafting a candidate question.

01

Identifying research gaps

NLP gap finders (GapFinder), keyword co-occurrence graphs, and knowledge-graph analysis surface topics that remain under-investigated.

02

Formulating new questions

Systems like SCIMUSE (a 58M-paper graph plus GPT-4) and AGATHA generate candidate questions — about 25% of SCIMUSE's ideas were rated highly interesting by expert reviewers.

03

Summarizing the state of the art

Tools like Elicit, Scispace, and Consensus condense the literature so a new question is grounded in what is already known.

04

Comparing & contrasting findings

Scite and NLP contradiction detectors expose disagreements in the literature — often the most productive place to pose a question.

05

Integrated research assistants

End-to-end platforms combine search, summarization, and suggestion into one interactive workflow.

Guardrails

The same generative power that suggests high-value questions can hallucinate citations, fabricate conceptual links, or overstate consensus. Responsible use asks researchers to:

A red-team agent. The paper proposes pairing an ideation agent with a second AI acting as a quality-assurance layer — one that audits citations against real sources, flags unsupported claims, and surfaces missing perspectives. AI for ideation, AI for critical interrogation.

Read & reuse

The paper, plus the practical materials it ships with.