How to Humanize a Systematic Review Draft After ChatGPT

A systematic review is not a literature review with a longer bibliography. It is a protocol-bound map of studies that met inclusion criteria you defined in advance. ChatGPT will happily “include” (Chen et al., 2018) for a trial that never existed and smooth every result into “significant improvements across outcomes.”

Humanizing a systematic review means polishing synthesis and discussion voice after the screened set and extraction table are locked. It does not mean rewriting Methods until PRISMA numbers drift.

Hubs: ethical AI use checklist, literature review humanize, citations, research papers.

What reviewers punish

Element	Weak AI draft	Strong systematic review
Protocol	Vague “we searched widely”	Databases, dates, registers named
Inclusion	Flexible after results	Locked criteria; deviations reported
PRISMA counts	Round numbers that “sound right”	Match screening log exactly
Extraction	Narrative only	Table from papers you opened
Synthesis	Every study “important”	Heterogeneity and disagreement named
Citations	Plausible ghosts	Every row openable

Detectors may flag smooth Discussion prose. Examiners flag fake included studies first.

Lock inclusion criteria before any model

Write the criteria by hand (or paste from your registered protocol):

Domain	Your locked rule (example)
Population	Adults with X diagnosis
Intervention / exposure	Y as defined
Comparator	Usual care or Z
Outcomes	Primary outcome named
Study designs	RCTs only (example)
Languages / dates	As in protocol
Exclusions	Case reports, conference abstracts only–if protocol says so

Do not invent studies to fill a thin yield. Empty cells and honest “no eligible studies for subgroup Q” beat hallucinated rows.

Build the extraction table only from PDFs you screened in:

| Study ID | Design | n | Intervention | Primary result | Risk of bias notes | PDF open? |

Prompt with the table and: “Draft Results synthesis from these rows only. Do not add studies, years, or effect sizes. Name where results conflict.”

PRISMA and Methods: what not to humanize away

Section	Humanize?	Why
Database names, search dates	No	Audit trail
n screened / included / excluded	No	Must match log
Inclusion / exclusion lists	No (wording light only)	Protocol fidelity
Extraction numbers	No	From software / hand extract
Narrative synthesis paragraphs	Yes, one pass	Where AI flattens debate
Discussion limitations	Yes, after true limits listed	Cut “more research needed” mush
Reference list	No	Accuracy

Paste stiff synthesis into Human Writes once. Put back study IDs, years, and statistics exactly. Re-open every cited PDF.

Before and after

AI Results sludge:

Numerous high-quality studies have demonstrated significant benefits of the intervention. Smith (2019) and hypothetical cohorts alike highlight promising outcomes, underscoring the importance of the approach across settings.

After extraction table + one pass:

Three RCTs met inclusion criteria (n = 412 total). Ahmad (2020) and Li (2021) found modest gains on the primary score at 12 weeks; Ortiz (2019) found no difference once baseline severity was adjusted. Between-study heterogeneity was high (I² reported from your software)–pooling was descriptive only.

Same assignment type. Second version survives “show me the included list.”

Workflow

1. Freeze protocol / registration details before drafting prose.
2. Complete screening log; export PRISMA counts from your tool.
3. Fill extraction table from open PDFs only.
4. Draft Results from the table–ban new citations in the prompt.
5. Write Discussion limits you actually observed (search gaps, bias, sparse n).
6. One Human Writes pass on synthesis and Discussion framing.
7. Diff-check every number and study ID against the table.
8. Supervisor review; disclose AI if policy requires (ethical checklist).

Related methods writing: lab report, research abstract, annotated bibliography.

What not to do

• Let ChatGPT “find additional key papers” for the included set.
• Humanize until effect sizes or confidence intervals change.
• Hide empty subgroups behind vague “the literature shows.”
• Copy PRISMA flowchart numbers from a template blog post.
• Treat systematic review voice like a persuasive essay (argumentative essay is a different genre).

Discussion without inventing implications

AI Discussion sections love “these findings have important implications for policy, practice, and future research” with no tether to your included set. Anchor Discussion to:

1. What the included studies actually measured
2. Where they disagree (population, dose, follow-up)
3. Limits from your search and bias assessment
4. One cautious implication you can defend in viva / defense

Discussion move	Allowed	Not allowed
Compare two included trials	Yes	Cite a third trial you excluded without saying so
Note sparse evidence	Yes	Invent a “growing consensus”
Suggest research gap	Yes	Promise clinical guidelines your review did not support

Humanize the sentences that sound like every other review’s closing page–not the claims underneath. Soft polish: Human Writes after the gap list is handwritten.

For narrative lit reviews without a protocol, use literature review–different genre, same ban on ghost citations. Broader ethics: ethical AI checklist.

Supervisor red flags to self-check

• Included n does not match the table
• Search date in prose ≠ Methods
• Risk-of-bias tool named but never applied in text
• Forest plot numbers that do not appear in extraction
• Perfect parallel “Author (Year) found that…” stacks with no debate

Fix facts before any second voice pass. A smoother wrong review is still wrong.

Bottom line

Systematic reviews win on protocol fidelity and a true included set. Human Writes is the voice pass after inclusion criteria and the extraction table are locked–never a license to invent studies.

Paste stiff synthesis on Human Writes when every cited paper is one you screened and can open.