Existing abstract · same-language editing
Editing an abstract with AI without changing numbers, negation or claims
Ask for an edit you can compare sentence by sentence, not a more confident conclusion. Lock the design, counts, effect measure and uncertainty before changing the wording, and leave unresolved facts for the authors.
This method edits an existing abstract that you are entitled to use, in its original language. It does not invent a study from a title, translate the text or run statistics. The handoff should contain an edited draft, original–revised comparisons, reasons and questions.
The passages below are self-written synthetic teaching material and an editorial demonstration, not real study findings, a customer manuscript or an AI run. ICMJE guidance supports accurate reporting and author responsibility, not automatic software certification or permission to submit.
The deliverable first: synthetic original and edit
Original v1: “We retrospectively studied 240 adults at one center. Within 30 days, 36 of these adults were readmitted. The adjusted OR for each 1-point increase in the baseline score was 1.30 (95% CI 0.95–1.78); this study cannot establish causality.”
Editorial example v2: “This single-center retrospective study included 240 adults, 36 of whom were readmitted within 30 days. For each 1-point increase in the baseline score, the adjusted OR was 1.30 (95% CI 0.95–1.78). This study cannot establish causality.”
Only the first two sentences were combined and the wording adjusted. No covariates, missing-data methods, P values or study dates were added. This is not a complete journal-ready structured abstract. Ask about missing background, outcome definitions or limitations rather than completing them through editing.
Four locked-fact rows: what an edit must preserve
Freeze factual content and strength of wording before editing. These are facts stipulated in a synthetic passage, not estimates calculated from real data.
| Check | Preserved in both versions | Do not rewrite as | Human verification |
|---|---|---|---|
| Design and population | Single-center, retrospective, adults | Multicenter, prospective or randomized trial | Check the Methods: the original abstract itself may need correction. |
| Counts and time | 240 adults; 36 readmissions; 30 days | A different denominator, omitted event count or follow-up period | Check participant flow, outcome definition and final outputs; the example does not establish complete follow-up. |
| Measure and unit | Per 1-point increase; adjusted OR 1.30 | Risk ratio, a 30% risk reduction or a between-group difference | Check score direction, reference, model and covariates; if absent, ask. |
| Uncertainty and causality | 95% CI 0.95–1.78; cannot establish causality | A significant increase, no effect or proven causation | An interval crossing 1 proves neither no effect nor equivalence; do not invent a P value. |
Unchanged numbers do not ensure unchanged meaning. Turning “cannot exclude” into “excluded”, or “associated with” into “caused”, changes the claim even when every number stays the same.
Materials to prepare before editing
Authorized original and scope
Provide abstract v1, paragraph or sentence IDs, original language and permitted scope. With an abstract alone, label the review abstract-only; do not claim the full paper was checked. Do not upload identifiable patient information or unpublished text you may not share.
Facts that authors can verify
Prepare relevant manuscript passages, confirmed aggregate tables/figures and analysis notes for checks of counts, units, measures, negation and limitations. Missing materials create questions; do not reverse-engineer results with a model.
Editing goal and journal rules
Specify clarity, compression or restructuring and the actual word limit, headings and current author instructions. Translation is a separate task. Verify counting rules in the journal or submission system.
A five-step, reviewable editing process
- Number every v1 sentence. List design, population, sample and event counts, time, units, measures, intervals, negation, uncertainty, causal language and citations before polishing.
- Choose one goal, such as removing repetition. Add no study content and do not automatically strengthen the tone. Preserve ambiguities or conflicts for questions first.
- Deliver v2 with original sentence, revised sentence, reason and protected facts. Explain deletions as well as rewrites; a clean paragraph alone is insufficient.
- Reread v1, v2 and the supplied authoritative material side by side. Check numbers, units, reference, direction, tense, negation, modality and limitations, including consistency with the manuscript.
- Create the next version only after author confirmation and retain unresolved issues. Follow current journal rules on truthful AI disclosure and human responsibility. AI is neither an author nor a primary evidence source.
A complete suggested AI request
Edit only my authorized existing abstract [file/version v1, original language], in the same language. Goal: [reduce repetition/improve clarity/compress to a verified limit]; scope: [sentences/paragraphs]. First list protected facts: design, population, counts and denominators, events, time, units, effect-measure type and adjustment, intervals, negation, uncertainty, causal limits and citations. Then deliver draft v2, a sentence-by-sentence original–revision–reason–preserved facts table, deletion notes and questions for authors. Use only the abstract and [available manuscript/aggregate tables and figures/journal rules]; if there is no manuscript, say abstract-only. Add no methods, covariates, data, P values, references, findings or ethics facts. Do not turn association into causation, an interval crossing the null into no effect/equivalence, or remove essential limitations. Flag conflicts, ambiguities or important content that cannot fit the word limit; do not guess or silently correct. Separate text checks from author-confirmed semantics. Do not claim certification, human expert review or submission readiness. Materials: [authorization, numbered original, goal and available sources].
Author acceptance: more than a word count
- Are counts, denominators, events, time, units and intervals identical item by item?
- Are OR versus other measures, per-unit/reference meanings and adjustment status preserved?
- Have negation, may/suggests, tense or association/causation strength changed?
- Did deletion or merging omit key findings or limitations, or turn not reported into did not occur?
- Have authors checked the manuscript and final outputs, resolved questions and verified current format and AI disclosure rules?
Questions researchers ask
Does passing number checks establish semantic fidelity?
No. Negation, references or causality can change while numbers do not. Deterministic checks are preliminary; authors must review all meaning. Fluent prose is not permission to submit.
Can a title alone be expanded into a complete abstract?
Not by this method. Editing cannot generate missing study facts. Supply an authorized draft and actual methods/outputs, or discuss an explicitly labeled writing structure only.
May nonsignificant findings or limitations be deleted for length?
Do not selectively conceal important results or inflate conclusions to meet a limit. Remove redundancy first; authors must resolve genuine conflicts with journal rules and retain change notes.
Verified sources
The official sources below were checked on 2026-10-19. Access conditions and update dates are rechecked before publication.
- ICMJE: preparing the abstract
Abstract section checked: accurate reflection of the paper, important limitations, no overinterpretation and journal-specific format.
- ICMJE: AI use by authors
Full author-AI guidance checked: disclosure, human review and responsibility; AI is not an author or primary source.
Start with your text and a review goal
Hand off an existing-abstract editing task in RigorMed
Describe a same-language editing task in the browser workspace, provide authorized materials and adapt the full request above. Check actual support, materials and credit information before execution; review the draft, changes and open questions afterward. Software prechecks are not semantic approval, human expert opinion or permission to submit.