Responsible Research

Can AI Write a Scientific Paper? What Researchers Should and Shouldn’t Delegate

Speaker

Dr. Ethan Gantana (Stellenbosch University)

Dr. Ethan Gantana (Stellenbosch University)34 min

Quick answer

SciSpace can help outline, rephrase, and critique a scientific paper, but it cannot be an author or take responsibility for data, methods, interpretation, or integrity. Delegate low-risk drafting support, keep scientific decisions with the researchers, verify every claim and reference, and disclose AI assistance when required.

About this session

Dr. Ethan Gantana explores the implications of using AI tools like SciSpace for research workflows, focusing on tasks like outlining, rephrasing, and critiquing papers. The presentation examines how these technologies accelerate literature reviews and productivity while raising critical questions regarding research integrity, methodological rigor, and the evolving nature of human authorship [L1, L5-L42].

AI can generate fluent scientific prose, but fluency does not prove that the methods are valid, the citations are real, or the interpretation follows from the data. This session examines AI’s role in planning, drafting, editing, and reviewing scientific text while preserving human accountability for the research itself.

What this tutorial covers

Separate useful writing assistance from the intellectual and methodological work of authorship.
Evaluate AI-generated text for unsupported claims, invented details, false citations, and loss of nuance.
Protect the integrity of Methods and Results, where fabricated or altered information is especially damaging.
Build a human verification workflow for evidence, statistics, argument, originality, and disclosure.

Key concepts explained

Writing versus research: Generating sentences is not the same as designing, conducting, or validating a study.
Methodological fidelity: AI must not fill missing procedural details or transform exploratory analysis into a preplanned method.
Claim calibration: Conclusions should match the design, data, effect size, and uncertainty.
Accountable authorship: Humans approve every word and answer for the manuscript.

How to try it in SciSpace

1
Define the limited writing task AI may assist with and the source material it is allowed to use.
Define the limited task AI may assist with—such as outlining, language revision, or critique—and state what it may not decide. Identify the evidence and files it is allowed to use.
2
Keep original data, methods, analysis outputs, and verified papers available as the evidentiary base.
Keep the original data, methods, analysis outputs, tables, and verified literature as the source package. Do not ask the system to reconstruct missing study details from context.
3
Review generated text for factual, methodological, statistical, citation, and logical errors.
Audit generated text for factual, methodological, statistical, citation, and logical errors. Trace every numerical or scientific claim to the study record or source.
4
Compare the manuscript with reporting guidelines and the actual study record.
Compare the draft with the relevant reporting guideline, protocol, ethics commitments, and journal instructions. Resolve discrepancies before polishing the prose.
5
Revise in the authors’ own voice and disclose AI use according to current policy.
Revise in the authors’ voice and retain accountability for the final argument. Disclose AI assistance according to the current policy and preserve records when required.

Worked example

A research team uses SciSpace to outline the Discussion and identify unclear paragraphs in a completed paper. The tool is not given permission to invent results, select the primary interpretation, or create references. Researchers decide which findings matter, write the causal and clinical interpretation, verify citations, and approve the final text. The manuscript disclosure describes the limited writing assistance rather than presenting AI as a co-author.

Practice prompt

Evaluate this AI-assisted scientific section. Check every factual claim, citation, method detail, number, causal statement, and limitation against the supplied sources and study record. Mark unsupported or ambiguous text and explain the correction needed.

Key takeaway

AI can help write a paper, but it cannot supply the study’s truth, take authorship responsibility, or replace rigorous human verification.