A paper can pass peer review, collect citations, and still come down years later.

An analysis of 2,047 PubMed-indexed papers retracted as of 2012 found that retractions arrived an average of about 33 months after publication, and about 24 months for papers published after 2002. 

By then the work has been read, cited, and built on, and the retraction carries every name on the byline with it.

Careful labs treat research integrity as a set of checks they run while the work is still theirs to fix. They verify data against the record, confirm approvals, settle authorship, and test whether each claim stays within its evidence before anyone outside the lab reads the manuscript.

That discipline extends across the full research process, from day-to-day lab practices and pre-submission checks to clear standards for misconduct, authorship, approvals, and evidence.

‍

What Research Integrity Requires at the Level of Practice

Research integrity becomes useful when it translates into behavior a lab can verify. The European Code of Conduct, revised in 2023, names four principles, and each one maps to something a lab does or fails to do.

• Reliability: Covers the quality of design, methodology, analysis, and use of resources. In practice, that means recording analysis choices as you make them and keeping the raw data behind each figure. 
• Honesty: Asks that research be developed, reviewed, reported, and communicated in a transparent, fair, full, and unbiased way. That rules out quietly dropping the replicate that didn't fit.
• Respect: Extends to colleagues, research participants, and research subjects. It shows up as approvals obtained before work begins and protocols followed as written.
• Accountability: Runs from idea to publication and includes supervision and mentoring. A PI who signs off on a figure without knowing who can trace it to raw data has a gap here, even if every number is correct.

‍

‍

The Integrity Checks That Should Be Completed Before Submission

Each check below is easier to settle before submission than after an editor or reader raises it. Work through the list with every author, including those who contributed a single experiment.

• Data match the record: Every figure and table traces to dated raw data. Every image reflects the original acquisition, and any adjustment or splice follows the target journal's image policy and is disclosed where that policy requires.
• Methods are reported in full: Follow the reporting guideline for your study design from the EQUATOR Network, such as CONSORT for trials or ARRIVE for animal studies. Report every condition, exclusion, and analysis you ran.
• Ethics approvals are on file: Research ethics review has to come before the work itself. Federally supported human subjects research cannot begin before IRB approval, unless it is exempt or waived, and PHS-supported animal work requires an approved Assurance and IACUC review. Confirm that the approval numbers match the protocol you actually ran.
• Authorship is settled: Every listed author should meet the ICMJE authorship criteria, including agreement to be accountable for all aspects of the work. Resolve order and contributions now, while everyone still agrees on who did what.
• Text and citations are accurate: Check that overlapping text is quoted and credited, and that each citation supports the specific claim it sits beside.
• AI use is disclosed: COPE's position is that AI tools cannot be authors and that authors must disclose how they used them. The practical questions this raises for lab work came up in a recent discussion of how AI will affect science and scientists.
• Claims stay within the evidence: Read the abstract and discussion against the results. A claim that generalizes past the cell line, cohort, or model the data come from needs narrowing or more data.

‍

Research Misconduct in Practice: Fabrication, Falsification, and Plagiarism

US federal policy defines research misconduct as fabrication, falsification, or plagiarism in proposing, performing, or reviewing research, or in reporting results. The same definition excludes honest error and differences of opinion.

Fabrication

Fabrication means making up data or results and recording or reporting them. Across seven surveys, a pooled average of about 2% of scientists admitted fabricating, falsifying, or modifying data at least once. In 12 surveys that asked about colleagues, about 14% reported knowing of falsification.

Falsification

Falsification refers to manipulating materials, equipment, or processes, or changing or omitting data, such that the record misrepresents the research. Image data are one place it shows. 

In a screen of 20,621 papers containing Western blot images, from 40 journals published between 1995 and 2014, 3.8% contained problematic image duplication, and at least half of those showed features suggesting deliberate manipulation. Less visible forms include dropping outliers or unfavorable conditions without disclosure, which misrepresents what the experiment showed.

Plagiarism

Plagiarism is using another person's ideas, processes, results, or words without appropriate credit. Reusing your own published text falls outside the federal definition, though journals can still retract duplicate publication on their own grounds. 

Journals that use Crossref Similarity Check can compare submissions against a database of more than 78 million full-text items.

‍

How Research Misconduct Is Detected and What Follows

Concerns reach the formal process through several routes. Journal screening catches some problems before publication. Readers who reanalyze images and statistics flag others after publication, and people inside the lab raise others still.

When an allegation involves PHS-supported research at a US institution, a finding of research misconduct requires a significant departure from accepted practice, committed intentionally, knowingly, or recklessly, and proven by a preponderance of the evidence. The updated rules apply from January 1, 2026. The institutional process moves through defined stages, and journals act on a separate track.

‍

‍

The journal step runs on its own logic. COPE's retraction guidelines advise retraction when editors no longer have confidence in a paper's results, and an expression of concern when they are uncertain or a decision is delayed. The guidelines describe retraction as a correction to the literature.

Retractions also say something about cause. In an analysis of 2,047 retracted biomedical and life-science articles indexed as of 2012, 67.4% were attributable to misconduct as the authors defined it, a category broader than the federal one that includes fraud or suspected fraud, duplicate publication, and plagiarism. Error accounted for 21.3%.

‍

How to Build Research Integrity Into Your Workflow From the Start

Survey data put outright misconduct in the single digits and questionable practices far higher. 

A survey of 6,813 researchers across all fields in the Netherlands estimated that 51.3% had frequently engaged in at least one questionable research practice. Most fall short of the federal definition of misconduct and still distort the record. The habits below reduce the room for them.

• Write the analysis plan before the data arrive: Preregistration or a dated internal plan fixes outcome measures, exclusions, and tests in advance. Deviations are fine when they're reported.
• Keep raw data separate and dated: Store original files read-only, work on copies, and log every step between the instrument and the figure.
• Record contributions as the project runs: Note who designed, performed, and analyzed each experiment, and revisit authorship at each draft.
• Train the whole lab: NIH requires trainees and fellows on its training and career-development awards to complete instruction in the responsible conduct of research at least once per career stage and at least every four years. Extending the same training to technicians and visiting students covers people those rules leave out.
• Get a critical read before submission: Ask someone outside the project to read for overreach and check whether each conclusion follows from the data shown.

‍

Where Claim-Level Review Fits Before Submission

The checks above protect a paper on two fronts. Records, approvals, and authorship show that the work was done properly. The claims still have to match what the work shows, and that test is hard to run on writing you know well.

QED Science built Paper Review for that step. It breaks a manuscript into its core claims, evaluates the evidence behind each one against the existing literature, and returns structured findings with specific recommended fixes. 

The same claim-level scoring ran blind across 57,455 bioRxiv preprints to identify the most original and valid work of the past year. Paper Review reports stay private in your account, and authors should still check their target journal's policy on AI tools.

‍

Upload your own manuscript and see where its claims hold up before it goes to reviewers. Paper Review is free for academic researchers.

‍

FAQs