SciSpace 发布七步流程，核查 AI 生成引用与论断

AI can save hours during a literature search, but one unverified output can undo that advantage.

A fabricated citation can enter a proposal. A genuine paper can be used to support a claim it never made. Mixed findings can be rewritten as a confident but false unanimity.

The practical question is whether you can trace every citation and claim back to the original evidence before the error reaches your work.

Drawing on Prof. Joanne Donoghue’s SciSpace session, this guide provides a 7-step workflow for checking AI-generated research. It moves from identifying the type of error to verifying the source, reading the full paper, locating the supporting evidence and auditing the final reference list.

The AI hallucination verification process at a glance

1. Recognise how AI hallucinations appear in research

Start by identifying the type of failure. Different hallucinations require different checks.

An AI hallucination is not limited to an invented paper. The source may exist while its DOI, authors, findings or interpretation are wrong.

A citation may refer to a real article while containing the wrong DOI or author list. Every bibliographic detail may be correct while the paper is still being used to support something it never reported.

Common forms of AI hallucination in research

Dr. Donoghue demonstrates several of these failures. In one example, a genuine study was summarised as showing that muscle was preserved, even though the study did not establish that conclusion or measure some of the outcomes implied by the summary.

Two checks are required

• Citation verification: Is this the correct source?
• Claim verification: Does the source support this statement?

2. Define the evidence criteria before searching

Before checking individual papers, make the search itself inspectable.

A vague request can return plausible papers without giving you a clear basis for deciding whether they address the same question.

Consider this request: Find studies on GLP-1 medications and muscle loss. Give me ten papers.

It does not define the population, study design, outcomes, publication type or exclusions.

A more useful request would be:

Find peer-reviewed studies examining the effect of GLP-1 medications on lean mass, muscle quality, strength or physical function in adults. Prioritise randomised controlled trials, longitudinal cohorts and meta-analyses. Exclude animal studies, editorials and studies that do not report a muscle-related outcome. Show the source and DOI for every study.

This gives the search clear boundaries and gives you criteria against which the results can be checked.

What to include in your research request

Dr. Donoghue compared the results of a short, general request across several platforms and then repeated the search with explicit inclusion and exclusion criteria. The more detailed request produced a more focused set of papers and made the output easier to inspect.

A detailed request can improve the relevance and transparency of the search.

3. Verify the identity of every citation

Once the results arrive, verify each reference as one complete bibliographic record.

Check:

1. Title: Does the exact title exist?
2. Authors: Are all authors listed correctly and in the right order?
3. Journal or publisher: Was the work actually published there?
4. Year and volume: Do the publication details agree?
5. DOI: Does it resolve to the same title and authors?
6. Version: Are you citing the preprint, accepted manuscript or published article you actually read?

Why checking only the DOI is not enough

A DOI can resolve successfully and still belong to the wrong paper.

In Dr. Donoghue’s demonstration, the generated reference list contained:

• A DOI that led to an entirely different article
• A real paper paired with the wrong DOI
• A paper where only the first author was correct
• The same title listed with two different author groups

Each citation looked credible until its components were checked against the publisher record.

Citation verification table

When a DOI fails, search the exact title. Take the correct DOI and publication details from the journal, publisher or a trusted bibliographic record.

A working DOI is not sufficient evidence that the citation is correct. The complete record must match.

4. Retrieve the full paper before relying on it

A verified citation is not yet usable evidence. You still need the material behind it.

An answer based on an abstract is not equivalent to one based on the full paper.

What different access levels can support

Use a clear source status

Dr. Donoghue’s practical rule is to obtain the PDF before relying on a paper. Full-text access allows the researcher to inspect the methods, results and limitations rather than relying on the most compressed version of the study. She also stresses that every citation, statistic and claim remains the author’s responsibility.

Use precise wording when full text is unavailable

5. Trace every important claim to the original evidence

Once the source record matches, test the sentence you plan to write.

Do not verify only the paper’s existence. Identify the exact passage, table, figure or result that supports the claim.

Check:

1. Population: Who was actually studied?
2. Design: Was the study experimental, observational, qualitative or a review?
3. Comparison: What was the intervention or exposure compared with?
4. Outcome: What did the researchers actually measure?
5. Result: Is the number or finding reported accurately?
6. Uncertainty: Are confidence intervals, limitations or inconsistent findings missing?
7. Interpretation: Did the authors claim association, prediction or causation?
8. Generalisability: Is the result being applied beyond the study’s population or setting?

Worked example

Suppose an AI answer states: 'AI-assisted feedback improves academic writing performance'

The paper may exist, but the claim could still be too broad.

Ask:

• Did the study measure writing performance or student satisfaction?
• Was there a comparison group?
• Was the outcome independently assessed?
• Did the study examine one assignment or sustained improvement?
• Was the sample drawn from one course or several institutions?
• Did the authors claim that the feedback caused the difference?

A defensible version may be: 'In one course, students receiving AI-assisted feedback achieved higher scores on the assessed assignment than the comparison group.'

The revised sentence preserves the study’s population, context and measured outcome.

Build a claim-evidence map

The purpose is not to make every sentence longer. It is to prevent the published sentence from becoming stronger than the evidence behind it.

6. Check whether the evidence actually supports a consistent conclusion

Several accurate source summaries can still produce a false overall conclusion.

Studies may use similar terminology while examining different populations, medications, outcomes, designs or follow-up periods.

Dr. Donoghue demonstrates this with research on GLP-1 medications and muscle loss. The AI suggested that the findings were consistent across studies, but the underlying studies reported mixed findings, examined different medications and did not measure the same outcomes in the same way.

Before drawing an overall conclusion, ask:

• Are the studies measuring the same outcome?
• Are their populations and settings comparable?
• Do stronger studies agree with weaker ones?
• Are null findings being overlooked?
• Does one frequently cited study dominate the summary?
• Are different interventions being treated as equivalent?
• Would "mixed evidence" be more accurate than "consistent findings"?

Example:

• Weak summary : Several studies show that the intervention is effective.
• Stronger synthesis : Three studies reported improvement, but they used different outcome measures and follow-up periods. Two larger studies found no clear difference. The available evidence therefore suggests a possible benefit, but does not establish a consistent effect.

The stronger version explains whether the studies can support a shared conclusion.

Separate evidence strength from evidence quantity

Ten studies pointing in the same direction do not automatically outweigh three stronger studies. Compare the methods and evidence quality before deciding which findings can reasonably be combined.

7. Build and audit a verified evidence set

Do not let the final bibliography become a second, unverified AI output.

Separate paper discovery from the reference list that enters the manuscript.

A practical evidence workflow

1. Search for relevant papers.
2. Apply your evidence criteria.
3. Verify the bibliographic record.
4. Retrieve and read the full paper.
5. Save it to a project folder.
6. Add the record to Zotero or Mendeley (or reference manager you use)
7. Record the claim it supports.
8. Note the supporting passage, table or figure.
9. Recheck every in-text citations before submission.

Dr. Donoghue recommends downloading the papers you intend to use, keeping them in a structured collection and using a reference manager for the final bibliography. She highlights importing verified PDFs into a controlled library and working from that collection rather than allowing the system to search freely beyond the papers already selected.

Suggested evidence table

A restricted PDF collection reduces risk, but does not remove it

Working only with papers you have already verified can reduce the chance that unrelated references enter the response. It does not guarantee that every extracted value or interpretation will be correct.

Dr. Donoghue reports that she has still seen AI systems introduce additional citations after the user supplied a closed collection of PDFs. Her recommendation is to write the final work yourself, use a reference manager for the bibliography and check every source that enters the manuscript.

Final reference audit

Before submission, check that:

• Every in-text citation appears in the bibliography
• Every bibliography entry is cited in the manuscript
• Every DOI leads to the correct paper
• Author names and publication details match
• Each source supports the sentence beside it
• Every statistic can be traced to the study
• Guideline recommendations appear in the original guideline
• Mixed findings have not been rewritten as consensus
• Abstract-only evidence is labelled accurately
• AI-assisted work has been reviewed and disclosed where required

Use SciSpace to shorten the path from claim to evidence

Use SciSpace to move from an AI-generated claim to the original evidence behind it. Search the literature, inspect source records, read full papers, locate supporting passages, compare findings and check the references in your manuscript.

The expert guidance behind this guide

Dr. Joanne Donoghue

Joanne Donoghue is a Professor and Director of Clinical Research at the New York Institute of Technology College of Osteopathic Medicine. Her research expertise includes exercise physiology and nutrition, and her work also includes esports medicine.

Her SciSpace session provides the verification workflow behind this guide, including examples of fabricated citations, incorrect DOIs, misrepresented study findings, invented recommendations and false consensus.

Her central principle is straightforward: every citation, statistic and claim remains the author’s responsibility, even when AI assisted with the search or summary.

Frequently asked questions

注意转载授权 · 阅读原文