Generative AI and research ethics
A verifiable and accountable AI workflow
Soorathep Kheawhom
Tasks and verification burden
Language editing: check meaning and quantitative claims.
Literature support: open and verify the original sources.
Code assistance: test units, assumptions and outputs.
Research interpretation: retain human scientific judgment.
Human responsibility remains central
AI tools do not qualify as authors under COPE’s position.
Researchers remain responsible for manuscript content.
Journal and institutional requirements both apply.
A claim can change during language editing
Input: “Conversion was 75% for one feed at 350 °C.”
Edited output: “The catalyst achieved superior activity under all operating conditions.”
Audit: scope expanded, comparator missing, activity undefined.
Verification before reuse
Trace each claim to a source or measurement.
Check numerical values, units and denominators.
Read the cited source and confirm it supports the claim.
Save what you accepted, changed or rejected.
An example of publisher-specific disclosure
Elsevier requires a declaration for substantive AI help in manuscript preparation.
Basic grammar, spelling and punctuation checks are exempt.
AI use within the research method needs a reproducible Methods description.
Policy checked 6 October 2026. Recheck the target journal before submission.
Image categories have different rules
Elsevier allows explanatory AI images with disclosure.
Data plots must derive faithfully from data and reproducible methods.
AI must not fabricate primary research images.
General-purpose image generators cannot create Elsevier graphical abstracts.
Publisher-specific example, checked 6 October 2026.
Confidentiality before a prompt
Does the material belong to you, collaborators or a reviewer assignment?
Do you have authority to share it with this service?
What do the service terms say about retention and training?
Use approved tools and the minimum necessary information.
A numerical audit catches a changed claim
64% to 75%
11 percentage points
17.2% relative increase
“An 11% increase” changes the denominator and is ambiguous.
A bounded editing prompt
“Edit this paragraph for clarity. Preserve every numerical value, unit, comparison and limitation. Add no factual claims or references. List changes that may alter meaning.”
The author still checks every change.
Workshop: audit an AI-assisted paragraph
Pairs, 10 minutes. Use Worksheet 4A.
Find unsupported scope, numerical and citation claims.
Rewrite the paragraph using only the supplied evidence.
Record an accept, revise or reject decision for each change.
Code needs scientific tests
Check dimensions and units before trusting an output.
Test a limiting case with a known answer.
Compare against a hand calculation or independent method.
Record versions, dependencies and data provenance.
Three ethics scenarios
An AI tool improves the wording but invents a reference.
A collaborator uploads an unpublished industrial dataset.
An image generator creates a plausible SEM micrograph.
A useful AI use record
Tool and version, date and purpose
Material shared and the authority to share it
Output retained and verification performed
Required disclosure and responsible author
An AI use log records the check
Task: improve the clarity of a nonconfidential paragraph.
Verification: checked every value, unit, comparison, limitation and cited source against the originals.
Decision: restored the tested scope, rejected an invented reference and retained only verified edits.
Record the actual tool, date and applicable version. Disclose use as the target journal requires.
A disclosure must describe actual use
Name the tool and the task it performed.
Describe human review and responsibility.
Use the location and wording required by the journal.
Exit ticket & anonymous feedback
3 minutes
One concept check
One evidence-based reflection
One point that is still unclear
seminar3-cheme-feedback.soorathep-k.chatgpt.site/session/4
Exit ticket: student ID optional. Feedback: no name or student ID.
Reading and policy sources
Elsevier: Generative AI policies for journals
COPE: Authorship and AI tools